<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Maciej Wolski on Medium]]></title>
        <description><![CDATA[Stories by Maciej Wolski on Medium]]></description>
        <link>https://medium.com/@maciejwolski?source=rss-c9620c68a273------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*YvEKpEX8GBgF7tq-OCNkLQ.jpeg</url>
            <title>Stories by Maciej Wolski on Medium</title>
            <link>https://medium.com/@maciejwolski?source=rss-c9620c68a273------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 02:15:36 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@maciejwolski/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[A unified vision for next-gen AI]]></title>
            <link>https://maciejwolski.medium.com/a-unified-vision-for-next-gen-ai-c5d7872c24bf?source=rss-c9620c68a273------2</link>
            <guid isPermaLink="false">https://medium.com/p/c5d7872c24bf</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[future]]></category>
            <dc:creator><![CDATA[Maciej Wolski]]></dc:creator>
            <pubDate>Wed, 17 Sep 2025 18:52:52 GMT</pubDate>
            <atom:updated>2025-09-17T18:52:52.096Z</atom:updated>
            <content:encoded><![CDATA[<p>Three synergistic layers of human-like AI that we will use in 2030s</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HqBIKnFmuyMH7sCv" /><figcaption>Photo by <a href="https://unsplash.com/@bhautik_patel3?utm_source=medium&amp;utm_medium=referral">Bhautik Patel</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>The current state</h3><p>Eight years after invention of Transformer architecture in Google Brain, three since the launch of popular chatbots based on Large Language Models (LLM), and thousands both hype and doomsday scenarios later — we do not seem to be closer to human-like AI than before.</p><p>GPT-5 did not blow our minds, take our jobs or even deliver profit to anyone involved…</p><p>AI is not yet our partner, and it seems to have more negative than positive influence on our cognitive abilities. For many it was clear for years, that language models alone are not the solution for Artificial Intelligence.</p><p>Yet, for most people it was necessary to experience the first interactions with non-human entity capable of conversation in natural language and then enter the phase of disillusionment, when it started to be obvious that just like in the human brain — where the language centers are a small part of the whole — Large Language Models are not really intelligent, but just reuse patterns of intelligence extracted from humans.</p><p>And still make silly mistakes even small children would not make.</p><p>So what lies ahead? What is the vision for the future?</p><p>I spent 11 years of my life to build understanding I want to share with you today.</p><h3>The path</h3><p>The path to Artificial General Intelligence (AGI) can be envisioned as a unified architecture built on three synergistic layers. This blueprint moves beyond today’s Deep Learning paradigms by integrating functional data representations with non-autoregressive and structured probabilistic models.</p><p>The result is a system designed for deep human-like reasoning, dynamic understanding, and emergent self-awareness. To enhance this framework, multi-factor reinforcement learning (RL) is incorporated as a core mechanism, drawing inspiration from biological neuromodulation to balance multiple internal drives during decision-making and learning processes.</p><h3>Layer 1: The Foundation — Data as a Continuous, Dynamic Flow</h3><p>The first layer revolutionizes how AI perceives and represents information. Instead of treating data as static, discrete vectors in embedding space, this architecture represents the world as a system of continuous functions and dynamic flows.</p><p>This is realized through two key technologies:<br>1. Functional Representations: Data is represented not as points but as functions in a Hilbert space (using wavelets, spectral coefficients, etc.). This captures the inherent structure, scale, and dynamics of the information.<br>2. Flow Matching: This non-autoregressive generative technique provides the engine for this layer. Instead of generating data step-by-step, flow matching learns a continuous, dynamic transformation — a vector field — that morphs a simple noise distribution into the complex data distribution. This perfectly aligns with the “data as function” philosophy by modeling the generative process itself as a continuous operator. The AI learns not just what the data looks like, but the fundamental dynamics that create it.</p><p>This foundation allows models to operate on rich, multidimensional data types, treating reality as an interconnected, multi-aspect system rather than a flat, linear sequence.</p><p>Multi-factor RL integrates here by introducing biologically inspired balances, such as dopamine-like reward maximization to guide the flow toward high-value data transformations, and GABA-like entropy reduction (via methods that promote sparsity in selections) to stabilize dynamic flows and prevent overfitting to noisy patterns.</p><h3>Layer 2: The Engine — Principled Reasoning and Thought Trajectories</h3><p>The second layer builds upon this dynamic foundation to create a powerful reasoning engine. It moves beyond simple pattern matching to enable principled, structured inference.</p><p>This is achieved by combining:<br>1. Probabilistic structure models: This is the critical component for robust reasoning. allows for the integration of structured probabilistic models (like Bayesian networks or probabilistic circuits) with the expressive power of deep neural networks. The operators learned at the functional layer are no longer black boxes; they are integrated into a probabilistic framework that enables causal inference, reasoning under uncertainty, and the principled incorporation of prior knowledge. This provides the logical and causal “scaffolding” for thought.<br>2. Reasoning as a Continuous Trajectory: Flow matching finds a new role at this layer. The process of reasoning — moving from a query or a state of uncertainty to a structured answer or a conclusion — can be modeled as a continuous trajectory through a cognitive state space. A flow matching model can learn this trajectory, providing a differentiable and flexible model of the reasoning process itself. A “thought” is no longer a discrete series of steps but a smooth, dynamic transformation of mental states.</p><p>It allows for an equivalent of imagination — where future events can be simulated, even in multiple alternative branches and outcomes evaluated. And it makes possible to initiate generative processes even without external input data — allowing for proactive, and not only reactive activities.</p><p>This engine can manipulate abstract concepts, generalize from underlying rules rather than just surface-level data, and maintain a coherent, logically consistent world model. Multi-factor RL enhances this by incorporating acetylcholine-like mechanisms to balance external inputs (such as new data) with internal memory states, ensuring adaptive reasoning that draws from both real-time observations and stored knowledge. Additionally, serotonin-like signal modulates satisfaction to sustain long-term objectives during trajectory optimization.</p><h3>Layer 3: The Apex — Emergent Metacognition and Self-Awareness</h3><p>The final layer integrates the lower levels to enable the emergence of a subjective, self-aware agent. True autonomy requires an agent to not only model the world but also to model itself, its knowledge, and its own cognitive processes.</p><p>This higher-order cognition emerges from the interplay of the full architecture:<br> • A Dynamic Internal World: The agent maintains a multidimensional, multi-aspect internal model of reality, built upon the continuous functional representations from Layer 1.<br> • The Stream of Consciousness as a Flow: The agent’s moment-to-moment cognitive process — its “stream of thought” — is modeled as a continuous flow, transforming sensory inputs and internal goals into beliefs, plans, and actions.<br> • Metacognition through Probabilistic Self-Modeling: This is the key to self-awareness. The agent uses the probabilistic structure models to reason about its own internal states. It can assess the uncertainty of its beliefs, evaluate the coherence of its world model, and identify gaps in its own knowledge. This capacity for introspection — a system that can reason about its own reasoning — is the foundation of self-evaluation and subjective experience.</p><p>Multi-factor RL culminates in this layer by unifying these neuromodulator-inspired factors into a holistic self-regulation system. For instance, it combines reward maximization with entropy reduction to refine metacognitive assessments, while balancing internal memory and goal pursuit to foster emergent self-awareness through iterative self-correction and adaptive planning.</p><p>This three-tiered vision, augmented by multi-factor RL, outlines a clear path forward to next-gen architectures. It creates an AI built on a foundation of continuous, dynamic data, powered by a reasoning engine that is both probabilistic and fluid, and culminating in an agent that can not only perceive and act but also reflect, understand, and evaluate itself in a balanced, biologically plausible manner.</p><p>AI that is human-like.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c5d7872c24bf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Will AI agents and robots do all our work?]]></title>
            <link>https://medium.com/geekculture/will-ai-agents-and-robots-do-all-our-work-026a3180a85c?source=rss-c9620c68a273------2</link>
            <guid isPermaLink="false">https://medium.com/p/026a3180a85c</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[brain]]></category>
            <dc:creator><![CDATA[Maciej Wolski]]></dc:creator>
            <pubDate>Wed, 21 May 2025 10:22:01 GMT</pubDate>
            <atom:updated>2025-05-21T12:31:34.284Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eBV2DQfyE7xVkZvL" /><figcaption>Photo by <a href="https://unsplash.com/@jezar?utm_source=medium&amp;utm_medium=referral">Jezael Melgoza</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>It is game over! AI Agents and robots will soon do all our work! OpenAI just released [something] and the world will never be the same.</p><p>How many times have you already heard that in the last 2 years?</p><p><strong>In a world where change is the only constant, many would still believe what I wrote in the first two sentences…</strong></p><p>As a long-time AI researcher, I am now more often asked by friends and non-tech people I collaborate with, questions like these:</p><p>- What now with our jobs?<br>- Professor [name] said that language models think like us, is it true?<br>- Should I retrain myself to work in tech?</p><p>I started spending so much time on AI more than a decade ago, soon after the Deep Learning revolution took off because I knew it would transform our society and the whole world. And I wanted to be part of it.</p><p>One day, indeed, AI agents and robots will be able to fulfill all our jobs. It can be a very positive situation as long as the whole of humanity we will be able to deal with it.</p><p><strong>But that day is not today.</strong> It is not even this year and probably not even this decade. Current AI models in the future will be as exciting as using Google for information searching in 2025.</p><p>On the other hand, the omnipresence of AI agents and autonomous robots can also bring negative and harmful outcomes, if we make it so. And especially people responsible for the governance.</p><p>I don’t believe people need to earn their right to live and use all the opportunities this world brings — through working in a full-time job. It just has been like that for centuries, that someone had to get the job done — and we only had animals, people, and thoughtless machines at our disposal.</p><p>But in current circumstances, the loss of a job could mean loss of the right to a normal life. So no doubt, it is an area of concern for so many people.</p><p>And if AI will become powerful enough to cover all our needs — we will need to adapt as well.</p><h3>So what is going on with this whole AI?</h3><p>The world went crazy about AI when the language models started to talk to us and it felt human-like.</p><p>But parrots can also generate full sentences. Some AI skeptics even call language models ‘stochastic parrots’.</p><p>The problem is that the ability to generate language is not proof of intelligence.</p><p>In the human brain, language centers (Broca, Wernicke) are just small parts of the whole — used rather to decode meaning from what other people said or wrote or to describe the thoughts we have.</p><p>Thinking and reasoning involve mostly all other areas. And then, the whole process is ‘described’ by the language center responsible for it.</p><p>What these other areas are doing? They are creating models.</p><p>Models of the world, models of the self, models of the objects, sequences of actions, and spaces. Models all the time.</p><p>And here is what brings some confusion. Reasoning and thinking is a <strong><em>process</em></strong>. A process that uses <strong><em>models</em></strong>.</p><p>Despite what investors and executives of leading AI companies often say — the process can’t be executed by a model itself. The models are being used as part of a process.</p><p>Declaring something else is like saying that the car drives by itself, even if it is being driven by a driver.</p><p>Patterns of ‘pre-reasoned’ information when mixed and combined together, can create an illusion of reasoning. But they are not it.</p><p>Why then so many educated, experienced, and skilled people believe that we can have intelligent software (AI) just by using models?</p><p>The problem is often related to the “universal approximation theorem” which says that a neural network with enough neurons can approximate or represent any continuous function.</p><p>In simpler terms: it can learn almost anything.</p><p>However, the ability to learn any function does not imply that all functions will be learned.</p><p>Let’s use another analogy — the fact that any person born could learn anything does not mean it will.</p><p>Our internals were shaped by a long process of adaptation called evolution. This is what allows anyone interested to apply reasoning inside their heads.</p><p>So how does natural intelligence emerge, what makes us able to learn things much more efficiently than machines? And how can we learn on our own?</p><p>And what it could teach us about truly intelligent AI?</p><h3>First principles of intelligence</h3><p>There are many definitions of intelligence. I especially like the one that says — it is the ability to adapt to circumstances.</p><p>But how does it exactly happen?</p><p>We are here to think about technology and AI, so let’s try to think about it technically and create a minimalistic technical definition.</p><p><strong>Intelligence is a continuous loop.</strong></p><blockquote>A loop compressing the past, simulating the future to select optimal actions.</blockquote><p>So intelligence, both natural and artificial has to:</p><ul><li>build models by compressing the past information (remembering what is important, forgetting what is irrelevant)</li><li>simulating the possible futures (in imagination for people or simulator for AI) and evaluating consequences</li><li>deciding which actions are the most preferred, with respect to the goals and preferences</li></ul><p>People try things and learn from them. Imagine new things, try them, and update their models to then simulate better and make better decisions.</p><p>It should be clear now that the missing part is the ‘simulation’ allowing for continuous learning — from real or just imagined (if the imagination/simulation is good enough) mistakes.</p><p>Simulation happens in an environment, handling the transition from the conscious awareness of the current state to the potential states in the future.</p><h3>The definition of reasoning</h3><p>So we have defined a minimalistic technical definition of intelligence. Can we do the same for reasoning? Intelligence and reasoning are not the same thing. One contributes to another.</p><p>Let’s use this definition, based on a gradual increase of the reasoning capabilities:</p><ul><li>1D (1-dimensional) reasoning: a simple (and primitive) approximation of reasoning with generated tokens (subword parts often referred to as reasoning tokens or just tokens), where what has been generated at the beginning influences the next steps. <strong>This is where our currently popular language models are. </strong>You can’t get back from a selected token trajectory. And that is a serious limitation, leading to retrieval of ‘pre-reasoned’ patterns.</li><li>2D (2-dimensional) reasoning: a more advanced approach, where information is generated in parallel and tokens can influence both previous and next ones (hence it is 2D). <strong>This is what diffusion or flow models do. </strong>Still not something as powerful as human-level reasoning, but much better.</li><li>3D (3-dimensional) reasoning: the solutions allowing to reason about multiple alternative concepts in parallel, where tokens or data from separate concepts (objects, ideas, etc.) can influence the other ones. <strong>Because of that, it can be called a form of simplistic simulation. </strong>Instead of just retrieving patterns, it can simulate how they affect each other in parallel.</li><li>4D (4-dimensional) reasoning: a situation where the reasoning is applied to a sequence of multiple concepts at a time. It is like reasoning over a timeline, with causes and effects and only immediate relationships. <strong>It is a real simulation that allows the evaluation of the outcomes of multiple possible variants. This is what humans do and when implemented as a technology, could be called AGI (Artificial General Intelligence).</strong></li><li>5D (5-dimensional) reasoning: a situation where the reasoning is applied to multiple sequences (timelines) of concepts manipulated in parallel. Where not only separate concepts, but also timelines can influence each other. <strong>This is a really advanced type of simulation that allows the evaluation of the outcomes of multiple complex associations between causes and effects. A technology like this could be called a ‘superintelligence’ (ASI) as no human mind can analyze such amounts of information at once.</strong></li></ul><p>Now, just to clarify — when we say ‘tokens’ many people associate them with text, but they are just units of representation. Tokens can symbolize also other modalities: visual or auditory.</p><p>So we just stated that intelligence requires simulation. And we have defined the types of reasoning. Is that all?</p><p>Of course not, but it shows us where are we in terms of building something equal to natural intelligence. And frankly speaking, we are quite far. At least the mainstream way of building AI. But of course, it does not mean that current AI solutions can’t automate repetitive work, even if it is knowledge work. They can and they will.</p><h3>The Advantages of Natural Intelligence</h3><p>To understand the advantages, we need to notice two main parts of Natural Intelligence:</p><ul><li>unconscious mechanisms (retrieval of models, patterns, representations)</li><li>conscious mechanisms (<strong><em>using</em></strong><em> </em>models, patterns, and representations in a conscious/simulated environment)</li></ul><p>I know it may be weird to read it, but you don’t perceive the world as it is. Its representation is generated for you by the brain. The atoms do not have colors, smell, or taste — these experiences are all defined by the environment of our conscious awareness.</p><p>And you always perceive the past. Not really a distant past, just something about 200 milliseconds ago — as this is the time that is (on average) needed to construct a consistent representation of reality.</p><p>A reality that looks and feels different for various living things. Different species of animals perceive reality differently from us.</p><p>The mind does not bother our consciousness with anything that can be retrieved from models of the world, objects, and actions.</p><p>But it brings our conscious awareness to surprising things, that can’t be retrieved from the models. But also to ones marked as important, because of the emotional response — which is by the way based on another model.</p><p>From what is perceived by the senses, the mind constructs multi-modal representations of the world. It also uses language as a model (description) of the world.</p><p>It then uses stored representations to reconstruct data in the simulated environment of our conscious awareness. If you ever read a book and imagined the scenes described by the author — you know what I mean.</p><p>The brain is like the hardware of intelligence delivering computing power, and the mind is like the software using it to run our conscious awareness.</p><p>Both are much more sophisticated than anything in technology we have now.</p><p>A single biological neuron is as complex as an artificial neural network.</p><p>The brain is modular. It uses multiple signals (neurochemicals) to learn and process data at the same time. It simulates potential outcomes and uses it to learn without any prepared datasets.</p><p>It is autonomous, self-learning, self-maintaining.</p><h3>Conclusions</h3><p>In current AI models, there is nothing (no component) that thinks. They are just models, which are based on the data extracted from human intelligence.</p><p>Repetitive work will be automated. I believe that more and more people will become employed as caretakers of AI systems — because we are not yet there with intelligent machines.</p><p>One day, maybe during our lives we will have human-level AI. And probably something even more powerful.</p><p>It may be a wonderful thing. It may be a terrible thing. It depends on our society.</p><p>We are still early on this journey (by the definitions of intelligence and reasoning from this article).</p><p>But one day, indeed we will have machines able to work with us, support us, and maybe solve the biggest problems.</p><p>It is not today, no technology on the market thinks, reasons, or acts like us.</p><p>But such technology will arrive.</p><p>How do I know?</p><p>I have been actively researching it for the last 11 years. I spent more than twenty thousand hours to understand what I wrote above.</p><p>I believe we can make AI one of the best things that happened to humanity. I also want to contribute to preventing it from becoming one of the worst.</p><p>My current work is related to 2D and 3D reasoning. I believe this is something that when scaled (to 4D and 5D) will lead to true Artificial Intelligence. But it is very far from the autoregressive language models based on current Transformers architecture.</p><p>Please follow me to learn more about the future opportunities.</p><p>I will write about the technical aspects of intelligent machines. The impact of AI on society and the future of work. And I will demonstrate the results of my work, at the right time.</p><p>I am also entertaining the idea of a video course for future AI researchers. Where I would help you to move from the place you are now, to make you one of the AI builders. The only requirement is your willingness to make an effort.</p><p>If you would be interested — please clap for this article.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=026a3180a85c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/geekculture/will-ai-agents-and-robots-do-all-our-work-026a3180a85c">Will AI agents and robots do all our work?</a> was originally published in <a href="https://medium.com/geekculture">Geek Culture</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why is current deep learning technology a dead end for Artificial General Intelligence? Part 2]]></title>
            <link>https://medium.com/swlh/why-is-current-deep-learning-technology-a-dead-end-for-artificial-general-intelligence-part-2-18341cbd2413?source=rss-c9620c68a273------2</link>
            <guid isPermaLink="false">https://medium.com/p/18341cbd2413</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Maciej Wolski]]></dc:creator>
            <pubDate>Mon, 29 Jul 2024 16:24:26 GMT</pubDate>
            <atom:updated>2024-07-29T17:20:10.064Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kf-ONEJlVcXe7kep" /><figcaption>Photo by <a href="https://unsplash.com/@emilianovittoriosi?utm_source=medium&amp;utm_medium=referral">Emiliano Vittoriosi</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Four and a half years ago I wrote a quite popular article about <a href="https://medium.com/swlh/why-is-current-deep-learning-technology-a-dead-end-for-artificial-general-intelligence-56f781c36d82">why current Deep Learning technology is a dead end for Artificial General Intelligence</a>.</p><p>It turned out that I was not alone with this view:</p><blockquote><em>“</em><strong><em>Scale will only get you so far. (…) The biggest breakthroughs in AI are yet to come — and will take more than just chips.</em></strong><em>”</em> — Demis Hassabis, Google DeepMind CEO</blockquote><blockquote><strong><em>“On the path towards human-level intelligence, an LLM is basically an off-ramp, a distraction, a dead end.”</em></strong><em> </em>— Yann LeCun, Meta AI Chief Scientist</blockquote><p>Completed with an older, but still relevant quote:</p><blockquote><strong><em>“My view is throw it all away and start again.”</em> </strong>— Geoff Hinton, the man behind backpropagation being used in most DL models today, about his most important contribution</blockquote><p>So how does that compare to the claims of popular AI figures (especially those connected somehow to a single company) that we will soon have superintelligence?</p><p>And what additional years of hard practical work between those articles taught me from first-hand experience?</p><p>I can’t wait to share it with you — and believe me this is not something you can read about anywhere else… So stay with me till the end.</p><h3>1) The learning algorithm</h3><p>Do you know why training large Deep Learning models is so expensive and slow? Why can people learn very fast and those models still require so many examples and have the electricity requirements of a city?</p><p>Because they are involved in lossy data compression, by using the same neurons for multiple tasks — therefore adjustments have to be delicate and slow to not disrupt the whole system, when trying to learn a new concept.</p><p>Imagine adjusting millions of tiny little control knobs every second, just by a fraction of a millimeter to find out how much better or worse the outcome is — this is how it works.</p><p>Compression happens when instead of memorizing every single value, you combine them as a single representation.</p><p>For illustration purposes look at the photos below and imagine that instead of storing a value for each pixel, you can store a single value for a group of similar pixels and reduce the needed storage space.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k81cbfqEYkftUv_F6S-tfg.jpeg" /><figcaption>Source: <a href="https://www.epfl.ch/labs/ivrl/research/slic-superpixels/">https://www.epfl.ch/labs/ivrl/research/slic-superpixels/</a></figcaption></figure><p>Computer scientists learned long ago that storing images as bitmaps is not the most effective way, and that is how we got JPEG and PNG — the most popular and compressed image formats. The same happened for sound and video with MP3, MP4 and AVI files.</p><p>When you train a foundational model you use terabytes/petabytes of data and try to fit it into a model with much less memory capacity. Because of that forced bottleneck, the model (automatically as a byproduct of compression) learns what has to remain and what can be skipped to still perform its task reasonably well.</p><p>So you take thousands of examples of something, show them to the model thousands of times to teach it how to compress its representation to keep it useful. This is the essence of backpropagation.</p><p>What is wrong with it?</p><p>Nothing, as long as your goal is to build a tool extracting knowledge and wisdom from human reasoning — I believe there is currently no better tool for that task.</p><p>But don’t confuse extraction of pre-reasoned pieces of data (textual Q&amp;A, labeled images, etc.) with intelligence…</p><p>There is none. There is only a blending of pre-reasoned patterns of information, extracted from people.</p><p>Deep Learning mimics our unconscious mind (a.k.a. System 1), which stores and retrieves the information we encounter in our lives. When you know what has to be memorized (you get it from human reasoning in the training dataset) — it is really good at the job.</p><p>But trying to use a Deep Learning model to define what to memorize (provide answer) works only as far as this model’s level of knowledge extracted from people. When you ask it for something very different from the training dataset — it fails miserably.</p><p>To reason and learn from that reasoning the AI system has to have access to the (hypothesis) testing environment. Create an idea, test if it works in simulated or real conditions, and update the knowledge base (simplified description of System 2). With the current setup repeat the same message many times, making sure that other precious pieces of information will not be disrupted…</p><p>An example of such an evaluation procedure can be a software development environment — where an AI agent can write different variants of software and test it in terms of accuracy, performance, or memory consumption — where the results are numerical and can be easily evaluated as better or worse. For much more complex scenarios, when the ‘good answer’ is subjective — the testing environment has to contain multiple modalities, and we will discuss it later in the article.</p><p>How AI can learn faster and without human participation (continual learning)? We need a way to encode new information in a useful way (we can use the same model, other Deep Learning models, or alternative techniques) and store that encoding in a memory external to the base model. To prevent memorization of every data sample, we can use unsupervised and self-supervised learning techniques (while storing only new samples, when they differ significantly from what was already learned) that allow us to benefit from both short-term (external memory) and long-term (base model parameters) memory.</p><p>As a result: we can learn things both really fast and slowly integrate them with the base model, using various available methods (e.g. replay methods).</p><p>So as humanity, we already know how to make AI learn fast with already existing methods, but do we know how to optimally define what to learn in an automated way?</p><h3>2) Cognitive map</h3><p>Raise your hand if you have used an LLM already for some time and have not received pure nonsense as an output even once.</p><p>In the art of blending pre-reasoned information patterns, LLMs do their best even if they are significantly undertrained in some areas and have no clue what they are talking about — introducing <strong>hallucinations</strong>.</p><p>In my previous article, I wrote that brains are not initialized randomly. And we humans share semantic maps, meaning that our brains have neurons processing similar types of data in similar brain regions.</p><p>In our heads, during childhood, neural progenitor cells create scaffolds for neurons and glial cells to arrange formations in the brain, that enable 3D information flow between senses and various components of the brain that shape the semantic map of our knowledge.</p><p>Narrow AI systems are great at mapping individual “routes” of cognitive information and combining them (input-to-output mappings if you prefer).</p><p>They are much worse at building a “map” to define completely new and previously unseen (or not experienced yet) routes. To understand what they do not yet understand.</p><p>Or in simpler terms: they would benefit if what they are planning to do makes any sense in the first place.</p><p>Algorithms like Q* (which according to some Chief Hype Officers on the Internet will bring the AGI soon…) aim to combine a set of internally generated rationales with algorithms that teach AI to move through the space of potential possibilities optimally to generate better responses.</p><p>I see some problems here:</p><p>a) imagine that your system has 90% accuracy when measured on single tasks, now chain them together in a sequence 5–10 times and this accuracy will quickly deteriorate:</p><ul><li>0.9 x 0.9 = 0.81</li><li>0.9 x 0.9 x 0.9 = 0.729</li><li>0.9 x 0.9 x 0.9 x 0.9 = 0.6561</li><li>0.9 x 0.9 x 0.9 x 0.9 x 0.9 = 0.59</li><li>0.9 x 0.9 x 0.9 x 0.9 x 0.9 x 0.9 = 0.5314</li></ul><p>You know where I am going. When you rely only on your <strong>incomplete</strong> knowledge, you quickly become a victim of the <a href="https://en.wikipedia.org/wiki/Dunning%E2%80%93Kruger_effect">Dunning-Kruger Effect</a>, where instead of trying to fill the gaps in your knowledge, you try to apply the wrong information immediately.</p><p>On the contrary, the best people in specific areas try to understand the shortcomings of their knowledge and fill in the blank spots in the cognitive map to be better prepared for the challenges ahead.</p><p>b) now, being at the level of 90% accuracy, how hard is it to get to 99% or better 99.99% accuracy? It is just about 1/10 of the effort, right?</p><p>But when you look at it from a different perspective, not how many times you are right, but how often are you wrong, things look a bit differently.</p><ul><li>for 90% accuracy, you are wrong in 10% situations — let’s define it as 0.1</li><li>for 99% accuracy, you make mistakes in 1% of cases — it will be 0.01</li><li>for 99.99% accuracy, it is 0.01% — that is 0.0001</li></ul><p>You need to reduce the error rate by a factor of 10x (0.1/0.01 = 10 = 1000%) to go from 0.1 to 0.01 error rate, and 1000x (0.1/0.0001) to reach 0.0001 — the level of 99.99% accuracy. And if we want to make AI part of our society, we should not stop there to minimize the risk of errors, when technology will drive our cars or make critical decisions. 99.9999% accuracy would require 100.000x improvement (0.1 / 0.000001)…</p><p>In my opinion, it is not reasonable to shuffle already available data endlessly for months, burning billions of dollars for training models that will just mix pre-reasoned knowledge patterns to store them effectively, hoping that AGI will emerge on the other end…</p><p>Instead, we should build <strong>reliable </strong>cognitive maps for AI — hand in hand with AI, creating a form of hybrid intelligence.</p><p>AI in its highly condensed description is <strong>a technology for extracting human skills, knowledge, and wisdom from data</strong>.</p><p>Why not organize it in a controlled way?</p><p>The relationships between concepts can be mapped with knowledge graphs, and the space of potential solutions can be mapped in 2D or 3D from high-dimensional vectors (embeddings and pre-computed cached neural representations) that encode the meaning of specific concepts.</p><p>With such a map, you can organize knowledge hierarchically, grouping similar concepts, and making sure that your AI understands the concept at all levels of abstraction.</p><p>You can now recognize that this is the opposite process of compressing data in a Deep Learning model. You start with a structure that has only signal for learning and reasoning, with no noise to filter out (that is how models compress information — by filtering out what is irrelevant).</p><p>This knowledge representation is fully interpretable, able to be evaluated by people, and modified in real time. Explainable and secure.</p><p>Having this knowledge representation can help AI to:</p><ul><li>quickly discard what is wrong, and forget what is irrelevant during training</li><li>make sure (similar to conscious effort) that what it outputs makes sense according to its knowledge base held in a cognitive map</li></ul><p>Current Large Models suffer from being good only when focused on narrow tasks, they lack the big picture. These problems are fixed with multiple iterations of generating, refining, summarizing, and evaluating the outputs (how economically inefficient it is and how bad for our planet it will be at a larger scale of deployment?).</p><p>While it could generate what is right (at least from our understanding and expectations) from the very beginning.</p><p>Now Knowledge Graphs are not equal to Semantic Maps, and especially Cognitive Maps, but that is the topic for another article. We will touch on it briefly in the next section.</p><p>With a Cognitive Map, the AI (and its users) understands the empty spots in the knowledge and the ways how to validate the plans and improve before taking action.</p><p>So let’s not leave the accuracy of AI to chance, but work actively to enhance it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pT0wBirj3dIFCEhQ" /><figcaption>Photo by <a href="https://unsplash.com/@vackground?utm_source=medium&amp;utm_medium=referral">vackground.com</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>3) World Model</h3><p>A cognitive map requires a world model. A structure that holds information about the world — it can be a knowledge graph, a virtual 3D environment, or any other model that holds information that allows for creating a cognitive map through interactions with that model.</p><p>The world model represents the environment, various locations, objects, and their properties, agents (understood as people, animals, or machines able to act autonomously), rules or laws, potential actions to be made, etc.</p><p>The world model also contains a model of an agent using it to interact with the world. It helps to understand what is possible and what is not, and what to expect from other autonomous entities and the surrounding environment.</p><p>So in its simplest form, it can be a graph with properties, it would help to solve some problems, but it would be too simple for others — where a richer and more interactive environment would be more beneficial.</p><p>It is the thing an autonomous entity can use to think before acting. It is what can be used to reason about complex plans and predict the outcomes.</p><p>And of course, you can count on the rich, complex, and accurate world model automagically emerging from data scraped from the Web and Reddit communities, inside billions of parameters of a Large Model, which no one ever will be able to grasp in their mind, influence or control…</p><p>But why to do that?</p><p>I took the harder path in my life many times, but only if it was reasonable.</p><p>This is not. The impact of AI technology, the associated cost paid directly or indirectly by all people — billions and then trillions spent on training and inference of AI — just begs for something safe, effective, interpretable, and reliable.</p><p>Deep Learning does not have these properties (yet).</p><p>External world models and cognitive maps can bring those features much faster and in a much better way than current approaches.</p><h3>4) Embodied AI in Virtual Environment</h3><p>Another aspect is that cognition and processing information is only one part of being intelligent — as it deals with internal operations. The second is to truly understand the physical world — an environment external to the agent.</p><p>We, people, use language, books, photos, and videos only as representations of some ideas and concepts.</p><p>What happens in our minds is a recreation of richer information patterns from those representations. This is why people are affected differently by various pieces of content, food, music, or events. They can bring different feelings and thoughts to each of us.</p><p>You can describe something, you can show it on video — but only those who experienced something similar before will be able to fully recreate what you mean.</p><p>This is because of our embodied multi-sensory experience of life.</p><p>Now, it will take some time for robots to be able to share the world with us. I was involved in this space for a couple of years.</p><p>But what we are becoming really good at as a species, is simulating the world(s) in virtual environments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kRN7lfd18dDy8oRlphQaFA.jpeg" /><figcaption>Source: UE5 Engine for Simulation (<a href="https://www.unrealengine.com/en-US/blog/unreal-engine-5-offers-significant-new-potential-for-the-simulation-industry">https://www.unrealengine.com/en-US/blog/unreal-engine-5-offers-significant-new-potential-for-the-simulation-industry</a>)</figcaption></figure><p>Within these environments, AI agents can experience and learn similar things to what we do in our world.</p><p>Like AlphaGo models they can ‘imagine’ various scenarios and learn from them, even when no one told them anything about them.</p><p>And these knowledge pieces can have completely transparent definitions.</p><p>Even if something has to be opaque to be effective, it is possible to train a parallel surrogate model that reflects what is inside its counterpart.</p><p>In these virtual environments, agents have the chance to have experiences and learn from them.</p><p>Many people think about consciousness (or awareness) as something mystical or powered by unexplainable ‘quantum’ phenomena. As I can’t be sure that my explanation of how it emerges can be correct or near the truth, I will limit myself to defining what it does instead:</p><p>Awareness or consciousness is a construct that simplifies (generalizes) the state of the internal and external worlds for an agent.</p><p>The atoms we experience every day do not have color, taste, or smell — it is our brain that defines these and other properties. At the same time, it filters out irrelevant information allowing us to focus on what needs our attention — to help us navigate the complexities of life.</p><p>If there is a problem in our body, we need to take care of — it sends a pain signal. If there is a need to be satisfied — it sends pleasure signal to reinforce it in the future. If all needs are satisfied — curiosity and exploration are triggered.</p><p>It is a construct that allows us to perceive and make sense of the world, imagine things, and plan, similar to a virtual environment.</p><p>Animals see and hear things differently. Maybe machines will also do the same. But they will strongly benefit from having a virtual environment to interact with, before making any decisions.</p><p>Reading this section you may think, that it is necessary to equip AI with rendering power stations, but this is not the case. Reasoning and simulating can happen inside the neural network, without rendering the view (watch the presentation <a href="https://www.youtube.com/watch?v=5H8dL6t-pUI">here</a>, if you are interested to learn more).</p><h3>5) Managing the modular architecture</h3><p>All concepts mentioned in this article make AI systems more complex, not necessarily in terms of confusing technologies, but in the number of integrated modules.</p><p>Models trained with backpropagation are self-contained and optimized to perform tasks similar to what has been experienced during training procedures.</p><p>How to manage even more complex, modular architectures?</p><p>In recent times, especially this year (2024) we could read about and use different approaches to more flexible usage of resources available to AI solutions, with a common name Mixture of X (something).</p><p>We have a Mixture of Experts (models), Depth (depth of computation in the model), Memory Experts (model parts that can be used), or Agents (multiple cooperating AI agents).</p><p>But once again we use good, old Backpropagation to teach models how to manage these resources. Which expert to load, how deep to go in computation, etc. And it works, as long as we deal with something that the models have been trained to deal with, on many examples.</p><p>You can be surprised with this information, but <strong>neurons may not be the ones running the show in the brain</strong>.</p><p>In section #2 I already mentioned glial cells, which basically create the environment for neurons to operate. For many years considered just a filler, maintenance workers keeping neurons fed and healthy — can have a much more important role.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/1*_8XnsJa3chc9hVd21_JQwQ.png" /><figcaption>Source: <a href="https://www.researchgate.net/figure/Tripartite-synapse-Participation-of-the-astrocyte-in-the-regulation-of-synapse-activity_fig3_366464810">link</a></figcaption></figure><p>You see at the level where neurons connect and exchange information (tripartite synapse), there are actually three players: two neurons and an astrocyte (a type of glial cell).</p><p>Astrocytes exchange neurotransmitters with other members of the tripartite synapse — as a result, they can potentially lower the activation threshold of the connection and probe the interactions to propagate that information as context.</p><p>Different types of glial cells (or glia in short) take care of neurons, heal them, and feed them like caring parents. Neurons are stimulated electrically and can send various signals fast over longer distances, while astrocytes are stimulated chemically — slower and for a longer time.</p><p>We can view them as big-picture context holders that deliver blood, and nutrients to the right places and measure information along the way about what will happen next.</p><p>In real life with continuous experience, we don’t switch context every three seconds like some AI tools (or at least most of us…) — so astrocytes can detect signals of specific neural resources being needed and facilitate the enhanced retrieval of the right data, at the right time.</p><p>It can be viewed as astrocytes using neurons to make the information being sent faster, through electric impulses — while focusing on the high-level context and state.</p><p>Of course, they are not alone — a working memory in the brain emerges from interactions between organs called the thalamus, hippocampus, pre-frontal cortex, and other areas of the cerebral cortex.</p><p>We also have supportive components that allow us and animals to manage cognitive or instinctive processes.</p><p>The key information is that there is more to the brain, than just a single unified mechanism, processing information from the input to the output on the same path.</p><h3>6) Supportive components</h3><p>In the brain, the only area that produces new neurons in the adulthood is hippocampus. It is responsible for establishing episodic memory — our ability to remember sets of information occurring together (“events”), forming our memories.</p><p>The semantic memory, giving meaning to things, objects, and concepts is distributed over the whole cerebral cortex — the most characteristic part of the human brain. It exchanges information with the thalamus (being a ‘router’ of information between senses and various parts of the brain — also helping us to form more complex representations).</p><p>We have organs that define what is relevant (usually new, surprising, or bringing intense emotional reactions), manage our emotional responses, allow us to orientate in physical environments, and so on.</p><p>The cerebellum, a relatively small part of the brain, located in the back of our heads where actually most of our neurons are placed. It helps us to perform activities with precision, without the risk of any disruption caused by further adaptations of connections between the neurons. You can think of it as a place where we keep highly accurate and trained enough neural circuits to be stored and not overwritten.</p><p>People often think that the seat of the intelligence is cerebral cortex, it is what makes us different from most animals. But it is not what makes us autonomous. It makes us able to remember a larger number of the more complex ideas.</p><p>Deep Learning aims to replicate the features of exactly this part of our brain. One of its flavors, Reinforcement Learning tries to mimic the reward mechanism that motivates our actions. But more signals govern our autonomous behavior.</p><p>My point in this chapter was to make a clear picture — it is not about the scale, because we already have huge models.</p><p>It is about the modular composition, that allows animals to do their jobs effectively, and us to do some additional intellectual work too.</p><h3>7) Multi-signal learning</h3><p>This modular composition requires multi-signal learning. In the end, it is not that we respond to the same events always in the same way.</p><p>When we are hungry — we react to food differently, than when our needs are satisfied and we have something important to do.</p><p>When we have time to think about the optimal solution, we tend to think differently than in situations when the decision just has to be made in the next 1–2 seconds.</p><p>In section #5 I mentioned how context is potentially managed in the brain, where different areas are ‘activated’ more, depending on the predictions about the next events.</p><p>In our heads we have neuromodulators, multiple substances that influence the behavior of neurons (and probably glial cells) at specific areas, not only at single synapses like neurotransmitters.</p><p>Supportive modules of our brains often have specific neuromodulators that are especially active there— e.g. acetylcholine in the hippocampus, or dopamine in the striatum where rewards and motivations are being managed.</p><p>Our lives are not only about remembering things, performing tasks, or maximizing the reward (satisfaction or pleasure).</p><p>They are about knowing how to respond to an unpredictable world, socializing with others, knowing when to focus on internal thoughts and when on the inputs from the external world.</p><p>It is not about using our mouth to speak first things that come from our unconscious mind like current LLMs seem to do.</p><p>Multi-signal learning, by using more signals to learn can teach AI agents how to respond in specific situations and even to understand better human emotions(!).</p><p>Imagine a situation, where we teach an AI agent that in specific situations it should spend more time seeking an optimal answer — not by prompts or any tricks, but just internally.</p><p>In other cases, it should ask other agents (people included) or information sources, to ‘feel’ the need to interact with them.</p><p>Or when there is a low battery level (or computational limit close to being reached) — act with the least steps necessary.</p><p>All those elements are not possible, when the only thing you teach AI is to map the association between input and output, through backpropagation.</p><p>You don’t run to the nearest food or yell “burger!” every time you see it. You adjust your mind’s and body’s response to the current situation — your level of hunger, the people nearby, and your current plans.</p><p>We just need this complexity of possible responses. And it is done with various timescales of our memory and multiple learning signals.</p><p>More about it in the next section.</p><h3>8) State generalization for full autonomy</h3><p>So we know that specific signals (neuromodulators) can drive the activity of the brain’s modules. But it is not the only function of these elements.</p><p>Our consciousness, our self needs a generalized state of the external world, and internal environment to respond to anything.</p><p>We don’t need to know which cells in our stomach have a problem to learn that what we just ate was not good for us — we feel pain and reduced happiness.</p><p>In challenging moments, we feel stress, increased heart rate, and adrenaline flow to ensure the focus and action necessary to move past this obstacle.</p><p>In times when everything is ok, and we can just enjoy life we feel the surge of satisfying substances — that allow us to focus on sensory experiences in the nearby environments, without overthinking.</p><p>When the challenge is intellectual — our brain motivates a larger set of its components to help us solve it, and we feel like information that we have not remembered a moment before — magically appears in front of our imagination.</p><p>Those elements do not only affect the activities at the level of specific modules — but also our highest-level executive center: pre-frontal cortex and consciousness.</p><p>Contrary to what people often think, we don’t make decisions based on logical reasoning but through our feelings.</p><p>And those feelings are state generalizations — how do we feel about the current situations and what will probably come next — are we excited or rather like to not continue what we do?</p><p>It does not mean that thoughts are not involved, but in the end, we just do what ‘feels right’.</p><p>All those components that our intelligence is composed of — are constantly interacting with themselves. From the low-level details to the high-level executive center, we call ‘self’.</p><p>To make that happen it is not only necessary to know when to load which information, but also how to make decisions — including which variant of the plan to choose.</p><p>What we see, hear, or think feeds the decision-making process that is based on the state generalization, or in simple words: feelings.</p><p>These kinds of elements can be quantified (expressed in numbers) in AI systems, as long as they are simulated in the virtual environment.</p><h3>9) AGI &amp; ASI</h3><p>My point in this article was to share that building AGI and later ASI with the evolution of LLMs is highly improbable, or even impossible.</p><p>We should also not aim to replicate exactly how the brain works — but instead what it does (the functions).</p><p>And if you read the text until this point — would you agree that it does something different than ChatGPT or any other language model?</p><p>It has many more modules, interacting with themselves all the time, more signals to learn, more context management techniques — and us in the center of it all.</p><p>A conscious agent, aware of its surroundings — using a data representation created in the brain, to make sense of colorless, soundless, and tasteless atoms around.</p><p>At some time in the past, people believed to be the center of creation. Now we have a more accurate picture of the Universe.</p><p>We can be proud, putting trillions of dollars into training LLMs or we can be humble, knowing what is lacking in the pursuit of AI. And that we have an example of general intelligence to be inspired by.</p><p>Nature can be understood, the brain can be understood at the level that will allow us to recreate intelligence, and not just a model of pre-reasoned information patterns.</p><p>We will also create agents highly aware of the situation in their environment. Will they be like us? I don’t know. But I know that we can create some kind of environment allowing them to explore their ideas and learn without us — while keeping their artificial minds more transparent than any human being before.</p><p>We can allow them to understand us and the world much better, than just from a text description.</p><p>Now, at the end of this article, I would like to conclude with something that can be surprising to you…</p><p>The creation of <strong>Artificial General Intelligence</strong> (human-level intelligence) and <strong>Artificial Super Intelligence</strong> (above human level) will be much harder than some people think and will happen sooner than others think.</p><p>If we define (correctly) the current AI stage as the skill and knowledge extraction systems from humanity — then why not tighten the cooperation to form Hybrid Intelligence?</p><p>As a person deeply passionate about both biological and artificial intelligence — I know that creating AGI and ASI can take decades or centuries if we take the wrong direction and only years if we choose the right one.</p><p>Imagine a Hybrid Intelligence system based on:</p><ul><li>the same foundational models on all devices, allowing for collective machine learning, what is learned by one agent — AI or human can be shared with the rest of the community almost instantly in the form of an external memory component</li><li>full interpretability of the system (either through sparsely activated components or surrogate models)</li><li>tools allowing people to think better with AI support, and AI to learn from how people think</li><li>privacy: local data processing, where you decide what is shared with the community and what you can get in exchange</li><li>participatory economy: where your data as a carrier of value brings money to you, by helping AI and robots to create value</li></ul><p>This kind of system combines the intelligence of humanity, automates its growth with AI features, and equips AI with the highest signal/noise ratio information possible.</p><p>With all the thinking about how clever are we, the truth is that to reach this 99.99% level of accuracy (that required 100.000x improvement in error rate mentioned before, in section #2) we will need to work together.</p><p>This is the only possible path for the exponential growth of AI — where people want to participate in the system and be economically incentivized to do so — instead of giving their data to corporations for free.</p><p>Big Tech companies live in this illusion of having free access to petabytes of unfiltered data bringing so much noise to the models, while I really believe that the most powerful AI company of this decade is not yet known.</p><p>But companies creating Hybrid Intelligence systems first, will have much more higher quality resources to build AGI, and then ASI systems. Especially those who will share the value with their individual contributors.</p><p>Who knows, maybe they are being built right now.</p><p>— — — — — — — — — — — — — — — — — — — —</p><p>If you enjoyed this article and would like to read more, feel free to download my <a href="https://drive.google.com/file/d/1J-Du8cYYEQJnt3Z4bqRgJtj27VqJTTCd/view?usp=sharing">FREE e-book “Intelligent Machines of the Future”</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wAW8-TZDtmlKI2DxK3Kp6A.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=18341cbd2413" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swlh/why-is-current-deep-learning-technology-a-dead-end-for-artificial-general-intelligence-part-2-18341cbd2413">Why is current deep learning technology a dead end for Artificial General Intelligence? Part 2</a> was originally published in <a href="https://medium.com/swlh">The Startup</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Maximizing our collective potential — the ultimate goal of futurism]]></title>
            <link>https://medium.com/gamma-waves/maximizing-our-collective-potential-the-ultimate-goal-of-futurism-e25fc4828c09?source=rss-c9620c68a273------2</link>
            <guid isPermaLink="false">https://medium.com/p/e25fc4828c09</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[futurism]]></category>
            <category><![CDATA[economy]]></category>
            <dc:creator><![CDATA[Maciej Wolski]]></dc:creator>
            <pubDate>Sun, 15 Aug 2021 13:07:09 GMT</pubDate>
            <atom:updated>2021-08-15T13:07:09.180Z</atom:updated>
            <content:encoded><![CDATA[<h3>Maximizing our collective potential — the ultimate goal of futurism</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eikEd2d1YFpaPW84" /><figcaption>Photo by <a href="https://unsplash.com/@christianlue?utm_source=medium&amp;utm_medium=referral">Christian Lue</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>With all these things keeping us busy in our daily lives — almost no one thinks about the future. I don’t mean the next year of life, but the next 50 years or more for the whole planet.</p><p>It is nothing objectively wrong — with all the problems, bills to pay and duties to perform — it is not easy to find time to think about our contribution to a better future.</p><p>But if so, is the world governed in the right way?</p><p>At the most basic level, if we strive for the best possible life for us, our children and the rest of the world — we should find the ways how to maximize our collective potential. So the real value will be provided by all or at least majority of Earth’s inhabitants. And they should be happy and energetic, not tired and helpless.</p><p>And yet all we do is compete. At the scale of nations, businesses and individuals. We want to prove that we are right, the best and the most capable. Even if we compete with neighbors, colleagues or family members.</p><p>Cooperative competition would be something good — because it makes lives of all participants better, no matter who finds the solution. However the one we see in the world today is not. It is based on the lowest instincts of securing territory and resources at the expense of others.</p><p>This is truly wrong.</p><p>And as a futurist I feel obligated to imagine what is in my opinion the most optimal scenario.</p><p><strong>The sign of changes — Artificial Intelligence</strong></p><p>We are at the edge of significant changes, caused by technology. A completely new world will emerge with the advent of powerful Artificial Intelligence. But will it be a better world?</p><p>AI already is a powerful invention and it will increase its influence beyond most people’s imagination.</p><p>Should we use it to increase corporate profits? Decrease costs? Make more people new billionaires? Or to empower everyone?</p><p>However powerful the AI of the future will become — discarding more than 7 billions beings equipped with natural general intelligence from the long-term vision is not only a waste of opportunity, but it is also immoral.</p><p>We tend to associate the word ‘genius’ with rare individuals. However deep inside of each of us — there is a hidden talent or even many of them. When expressed, they will do such things — bring joy, remove frustration of the particular person and provide value to the world.</p><p>Whatever the scale and influence of such talent is — local or global — this is a recipe for a better world, filled with happiness and with decreased amount of frustration, violence and conflict.</p><p>In the next decade AI will change the world dramatically. Will it mean further wealth concentration among the small group of people? Or this technology will be used for the common good?</p><p>I can imagine AI used to produce food hyper locally — even such that is associated with a different climate zone. To reduce the price, the carbon footprint and for better nourishment of the body. Without the need of long transportation. Food that is grown for taste and nutrition. Not reduced losses and maximized profits.</p><p>I can imagine AI used to provide much better, fully personalized education. Better economic system. Better governance.</p><p>End of mundane jobs.</p><p>But the first step is always questioning the status quo: are we going in the right direction?</p><p>Is current AI designed to empower people or corporations?</p><p>Who owns the data (value) contributed by the whole population?</p><p>Does current approach to AI even make any sense?</p><p>The answer to all three questions above is not optimistic. We need AI that is based on an <a href="https://medium.com/swlh/why-is-current-deep-learning-technology-a-dead-end-for-artificial-general-intelligence-56f781c36d82">equivalent of a brain</a>. And that requires combining neuroscience with technology. And a lot of effort.</p><p>We need constant unrestricted access to private data — what means that most of it should remain private and not be collected in the cloud environment.</p><p>There is a difference between learning from data and storing all available data.</p><p>I believe that after making some corrections we will be able to use Artificial Intelligence to create an inclusive environment for all people to live in a better world and feel their contribution to it.</p><p>With AI being so powerful, that it will be able to support us in many new areas of life: including food production, education, economy and governance.</p><p>The best future lies where we are maximizing human potential with AI. Not only maximizing profits.</p><p>And this is the path I entered a few years ago — to contribute to the AI revolution for a common good, the best possible future and maximization of our collective potential.</p><p><strong>The way to make the world a better place — optimization of natural intelligence</strong></p><p>Powerful Artificial Intelligence will bring many new possibilities to the world. But may also change the way we live. Many things will be automated — leaving people without the associated jobs. It is our collective responsibility to allow these people to contribute to the society.</p><p>But I am confident that we will have the right solutions. When the world will change so dramatically, we will have an opportunity to transform our lives collectively.</p><p>Where people had to work just to maintain their lives — they will be allowed to contribute their data to make AI work instead of them — for economical benefits. They will be allowed to grow themselves and decrease frustrations.</p><p>Personalized education, supported by powerful AI will allow anyone, anywhere to develop new skills and gain knowledge.</p><p>If people will more often do thing they love to do — they will become better humans. AI when implemented in the right way, will not create a dystopian reality. It may empower everyone and create a world we would not be even able to dream about.</p><p>Happy, energetic and joyful people are the foundation of a positive society. It does not matter if someone has a talent for art, manual work, science or any other thing — if there is joy, there is a good outcome and benefit for everyone around.</p><p>New economic solutions will allow such transformation, because the way we handle data right now is not yet successful, fair or optimal. The AI will need lots of data to serve us better. But it does not need to know all our private details, nor its creators. It just needs a general knowledge about how things are done on this planet.</p><p>Contributing your private data to a mechanism that learns from it, leaving your secrets yours — will provide substantial benefits to the developers of AI and in return they should pay back to the people who generate and share such data.</p><p>It is not realistic to believe that we will have powerful AI with restriction to data access. AI needs data, needs the ability to decide what is valuable, what to learn and what to skip. This is how it will experience the hypergrowth beyond our imagination.</p><p>The solution is to allow AI to do repetitive jobs, but reward those that will make it happen. The software solutions are already in the making, hidden from plain sight.</p><p>I have heard people who tried to convince others that their individual data is worthless. It is not. It is said like that just to benefit from free contributions.</p><p>When you earn money or are rewarded in any other way for what you bring to the future of technology, you can become better version of yourself. You can strive for any goal and live a happier life.</p><p>AI will change the world. But it will also change us, our perspectives and will provide many opportunities.</p><p><strong>The summary</strong></p><p>We will never have absolutely the best possible world if we will not include everyone on this planet. Who cares about money if we could have 10x better things now for the same price, if more educated people would bring their talents to the pool.</p><p>Just imagine the inventions that are not here, because of that.</p><p>The future is not about charity of the wealthiest people on Earth. It is about providing everyone with a chance to become the best version of themselves.</p><p>With joy and motivation to do what you love — your body and brain start to function optimally. You are no longer frustrated, seeking for ways to express your negative emotions. You feel better, you treat yourself and others better. There is no need for conflict.</p><p>And this is the future we should aim for. Together.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e25fc4828c09" width="1" height="1" alt=""><hr><p><a href="https://medium.com/gamma-waves/maximizing-our-collective-potential-the-ultimate-goal-of-futurism-e25fc4828c09">Maximizing our collective potential — the ultimate goal of futurism</a> was originally published in <a href="https://medium.com/gamma-waves">Gamma Waves</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The way of making the impossible possible]]></title>
            <link>https://medium.com/gamma-waves/the-way-of-making-the-impossible-possible-233cefc41ce3?source=rss-c9620c68a273------2</link>
            <guid isPermaLink="false">https://medium.com/p/233cefc41ce3</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[motivation]]></category>
            <category><![CDATA[future]]></category>
            <dc:creator><![CDATA[Maciej Wolski]]></dc:creator>
            <pubDate>Sun, 15 Aug 2021 09:48:28 GMT</pubDate>
            <atom:updated>2021-08-15T10:41:28.534Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JDdur3rsoMO_ezfS" /><figcaption>Photo by <a href="https://unsplash.com/@nasa?utm_source=medium&amp;utm_medium=referral">NASA</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Achieving great things requires extraordinary preparation. Otherwise, you can just be a fool chasing unrealistic dreams. If you don’t realize this very quickly — you will follow the path of very painful experiences.</p><p>The world is shaped by people who deeply understand reality and can transform the present into the future.</p><p>Sometimes you need to prepare your body, sometimes the mind — but in most cases, it is beneficial to optimize both. To have rich supplies of energy, mental resilience and clarity on what you aim to achieve.</p><p>As an independent researcher of Artificial General Intelligence, in the past years, I had truly ambitious goals. I needed to understand the principles of natural intelligence governing its source — the human brain. And as you may know, this is one of the most complex objects in the universe — therefore understanding its operation is a hard task.</p><p>In addition, I had to master programming skills and due to the nature of challenges — also the basics of hardware design, especially processing units that will be working better than the current ones. So I needed to combine expertise from some of the most demanding areas of human knowledge — neuroscience, biochemistry and computer science.</p><p>Usually, when I set a goal, I realize it much ahead of a deadline. When I assumed that I will write a 300+ page book in 10 weeks, I finished it in three. When I gave myself half a year to learn the basics of hardware design, I saw satisfying results just after one month of intense studying and practicing.</p><p>But how can you do such things?</p><p>Well, Elon Musk would say:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*AfMlE5N5CK7NorN3lGOMiQ.png" /></figure><p>And it is true to some extent because I worked hard as hell. Up to 16 hours per day. But the true secret lies in the optimization of the mind and body. And the answer to a question: how to get the most from the time you invest?</p><p><strong>The body</strong></p><p>It is a very complex topic, but extremely ambitious goals require unusual methods. Necessary expertise and experience, but also a strong will to become the absolute best version of yourself.</p><p>It starts from a brutal realization that you are not a machine running on pizza and coffee. You are a complex organism or as I prefer to see it — you are equipped with really sophisticated biotechnology. That can be understood and optimized.</p><p>And you should treat your body seriously. During each second there are billions of biochemical reactions that allow you to live and perform any task you want to accomplish.</p><p>Many of these processes depend on enzymes that require specific nutrients. If they are not available — part of your body simply does not work or perform suboptimally.</p><p>You see, diet and nutrition are not about calories. It is just a virtual term describing the amount of heat as a result of burning something. What your body needs and uses are nutrients.</p><p>Some of them, called macronutrients are indeed providing energy or serve as building blocks for body parts and various chemical substances. But there are also micronutrients that take part in these biochemical activities each second of our lives. And it is all about their absorption and proper utilization. Because putting something in your mouth does not mean it will end supporting your body.</p><p>As human beings, we like simplicity. Let’s just drink a cup of coffee to feel energized.</p><p>But then it starts to work worse and worse, to the point that even several cups per day are not making our performance the way it could be. Followed by crashes and a feeling of exhaustion.</p><p>There are multiple reasons for such a situation — depletion of raw building blocks for dopamine and adrenaline, decreased sensitivity of receptors, hormone imbalance — just to name a few of them.</p><p>A cup of coffee is like a debt that sooner or later you need to pay back.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9QiPxQCnUWSyXBWm" /><figcaption>Photo by <a href="https://unsplash.com/@jeremyyappy?utm_source=medium&amp;utm_medium=referral">Jeremy Yap</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>In my case, I managed to optimize my days in such a way, that I was able to work 12–16 hours per day with full performance and a single cup of coffee, but supported by a whole range of other components.</p><p>It worked like a charm for a year and a half, until I experienced a total breakdown of my adrenal and hormonal system.</p><p>Another lesson, another area to explore, learn from mistakes and make conclusions.</p><p>First of all, I spent way too much time in front of a computer. My physical endurance was too weak, but I thought that I will manage to realize my goals without maximally caring about my body. I exercised daily, but it was not enough to keep my peak performance.</p><p>Our bodies simply are not made to do a similar thing all the time. They need rest-active cycles. Even if everything worked fine for many months — my body was producing so much cortisol that in the end, it disrupted the balance of my hormones.</p><p>In addition, fueling my mental work required a stable supply of glucose. Even if I was eating very healthy for years, still I gained a lot of visceral fat that negatively affects the performance of the internal organs.</p><p>There are so many aspects of body operation that it is just not possible to reduce it to simple rules like:</p><ul><li>eat less and exercise — you will be in a good shape or all forms of exercise are healthy;</li><li>coffee is your friend and the ultimate source of energy in the body. It is not, it just spends all resources you have, until you are broken. That is why I like the financial metaphor of debt.</li></ul><p>First of all, we are more hungry if we lack the necessary nutrients. This is a natural reaction, initiated by the body — to make us eat more to refill them. Second, exercises are only a small part of daily energy expenditure. Therefore there is a need to combine diet and exercise, with the first factor having much more influence on the general fitness of the body.</p><p>The only way how the body is making fat tissue — is through elevated insulin and a high intake of glucose. The only way how body intensively burns fat — is when the glucose is not available and instead it is necessary to use the fat reserves to make ketone bodies.</p><p>It is very helpful to have this metabolic flexibility to avoid insulin resistance and be able to effectively burn both fat and glucose for energy.</p><p>But to control the body and effectively communicate necessary actions — it is required to have building blocks for hormones, neurotransmitters and transfer various elements through the cellular membrane. All of it does not come for free, even if it looks like this from our high-level perspective.</p><p>Without tyrosine, the body will not make dopamine and noradrenaline necessary to stimulate the mind to act. It is available in food, but it takes up to several hours to digest various protein sources.</p><p>And if raw building blocks are not available to the body — it can’t make the desired substances out of thin air. It is very efficient at recycling internal resources, so it regains some building blocks daily, but it is just for minimal maintenance. And not optimal performance.</p><p>Without sufficient dopamine and noradrenaline levels — you will lack motivation and energy to do the work. This is how you welcome procrastination. In fact, all things that people hate about themselves are related to insufficient levels of energy in the body.</p><p>The other important factor is that when you eat meat or vegetarian protein sources — there are many amino acids out there and they all compete for resources necessary to make and utilize enzymes — the causative factors in the body.</p><p>Therefore if we have a specific goal in mind, it is often way more effective to take an isolated substance in a powder form as it starts to work in 15–30 minutes instead of hours and in addition with a reduced effect.</p><p>Imagine that you want to have full mental and physical energy in the morning, but the tyrosine in your meal will be available at noon after the digestion of breakfast. Assuming that you have found time for it in your constantly busy lifestyle.</p><p>Besides communication through hormones, neurotransmitters and other substances — the body needs to utilize carbohydrates or fats and transform them into a usable form of energy resource.</p><p>No, it is not that the cell just gets a unit of glucose or a ketone body and uses it directly. There are complex procedures at each step of our life maintenance. But the final and most basic one for energy extraction is based on getting electrons from a reaction that changes ATP molecule into ADP. And then possibly regenerate it back to ATP.</p><p>There are also other paths of getting these electrons, but I consider them out of the scope of this article — you just need to be aware of the general rule.</p><p>We may consider ourselves as beings that are powered by electricity produced in a series of biochemical reactions that derive energy from the food we eat, the water we drink and the air we breathe. Personally, I like the concept of seeing the human body and mind as advanced biotechnology.</p><p>And the stuff that when lacking will kill you faster is the most important for your energy levels and life itself. So air and oxygen first, water and hydration second, food and nutrition last but not least. In addition, the lack of versatile movement makes all biochemical processes harder.</p><p>The blood flows in our veins through magnetic forces and the work of the heart, but the extracellular fluids require pressure caused by movement. Otherwise, the delivery of nutrients and oxygen is slowed down and the same applies to waste management. You get weaker and weaker.</p><p>Several factors may influence the amount of energy derived from food. Besides our habits that train our body to be prepared for specific metabolic paths — there are also nutritional factors, like the availability of niacin or creatine to boost the energy level available.</p><p>If all you do for physical activity is jogging, aerobics or any single activity — your body will be accustomed only to specific ways of operation and energy production.</p><p>We need mixed activities including low-intensity aerobic and high-intensity anaerobic ones. We also need to be able to effectively burn both glucose and fat for energy.</p><p><strong>The mind</strong></p><p>If you have ever seen professional athletes — you know that absolutely the best treat their bodies with unbelievable care. And more often than not — they also work with psychologists and advisors on their mentality.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-TnUsKu6A4BXKiAC-wiMdg.jpeg" /></figure><p>You may also look at mentors that combine physical activity with mental resilience. The choice is yours, but just to give you an example — you may pick from one of the following extremes: a former Navy Seals officer — <em>Jocko Willink</em>, a perfect example of a tough man from the West or Shaolin Monastery Master — <em>Shi Heng Yi</em> that is a warrior monk raised in Eastern philosophy, depending on your preferences. In my case, I read and listen to both.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wnaKxbqNWasGjQ4g8rL6nA.jpeg" /></figure><p>They can teach you about discipline, embracing the pain, power of will and the path to self-improvement or self-mastery that never ends. You just need to strive to become better and better every day.</p><p>Because the truth is that when you enter the realm of things that seem to be impossible — you need the courage, patience and skill set to jump directly into the void of the unknown. And you need to feel absolutely comfortable in your discomfort zone. You need to embrace the pain.</p><p>You need to find the balance between two extremes: being both calm and dangerous in your area of expertise. Because the better you become — the calmer you are. And vice versa.</p><p>Having your body and mind in optimal shape boosts your confidence. And a belief that you can overcome any obstacle and face any challenge. Because there are always surprises and things impossible to predict upfront.</p><p>Another thing is the ability to explore the environment related to your challenge. We live in a specialized world — where people usually prefer to expand their knowledge and skills in one selected area.</p><p>However with big challenges, things that are considered to be impossible… well, at least until they are done — you need to be able to look from multiple perspectives. You need to understand multiple layers where you can provide new value or avoid mistakes that others would make.</p><p>When you are a part of the big team — there is always someone that can support you. But when you are an innovator and a leader — intending to achieve previously unachievable — you need to know a lot more. Not only to be able to do work for yourself, but also to inspire, lead or manage others.</p><p>For example, a person that took the standard path to develop Artificial Intelligence will try to improve on what is already available. It will be harder for such a person to realize that a specific path may lead to <a href="https://medium.com/swlh/why-is-current-deep-learning-technology-a-dead-end-for-artificial-general-intelligence-56f781c36d82">a dead end</a>.</p><p>On the other hand, if you can see the way through the same perspective, but then compare it with alternatives — based more on neuroscience, biochemistry and other disciplines — you may see that some paths are not realistic to the goal of making Artificial General Intelligence, but others are.</p><p>So you start by thinking strategically and then fill up the details. This requires a big picture view, an open mind and the will to question absolutely everything. For years.</p><p>It requires reading hundreds of books and thousands of scientific papers. Endless hours of experimentation. And the unwavering faith in the final success.</p><p>When you combine expertise from multiple areas, physical and mental endurance — you need to ask yourself — am I ready to spend many years pursuing my goal, even without visible external signs of success? Am I ready to do whatever it takes?</p><p>These are very important questions because making impossible things possible requires an unbelievable amount of willpower and a very high pain threshold. You can’t doubt yourself, your team and your strategy even for a moment. Otherwise, you will fail.</p><p>And at the same time, you also need that ability to fail. There are many factors that you simply can’t control. But you need to focus on those you can. That is yourself and to some extent your team. You need to believe, but completely accept the risk of failure. To deal with the circumstances.</p><p>Having absolute clarity where you want to go, the awareness of risks, unlimited amounts of determination and perfectly prepared body and mind — you may take your chances and try to achieve the unachievable.</p><p><strong>The summary</strong></p><p>In life, we hardly ever experience miracles, rather the consequences of our actions. You either choose comfort or try to feel comfortable in your discomfort zone. You choose to avoid pain and difficulties or embrace them to at least try to get to places where no one has ever been before.</p><p>You may choose to live some years like no one would like to — to live the rest like no one could before. The greatness is built when no one is watching.</p><p>It is ultimately your choice what you do with your life. But if you dare to choose the hard path, you should be prepared.</p><p>In my case, I was forced to understand almost everything about the body and mind. 14 years ago I had trouble doing the most basic things we all take for granted. I couldn’t even breathe, sleep, walk or talk normally. I was not able to remember most of the things that I read or heard.</p><p>But by the necessity of regaining everything from the point of being at the lowest level possible — I gained knowledge, experience, faith and skill set to make the impossible possible.</p><p>Today, understanding so much about body and mind — I feel truly ready to continue my path of self-improvement and to provide the world with Artificial General Intelligence. Now with another investment secured, many things are finally in my hands and control. And I don’t fear those that are not.</p><p>The future will be beautiful and filled with many things considered to be impossible. If you have the courage, go there with me. Maximize your potential, become the person you always meant to be. And give the world your gifts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/820/1*wDPYQ4j7He4kwSfWmzwVWA.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=233cefc41ce3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/gamma-waves/the-way-of-making-the-impossible-possible-233cefc41ce3">The way of making the impossible possible</a> was originally published in <a href="https://medium.com/gamma-waves">Gamma Waves</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Could Everyone Have Been Wrong About the Universe and AI?]]></title>
            <link>https://medium.com/geekculture/how-could-everyone-have-been-wrong-about-the-universe-and-ai-4e96c869b261?source=rss-c9620c68a273------2</link>
            <guid isPermaLink="false">https://medium.com/p/4e96c869b261</guid>
            <category><![CDATA[universe]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[learning]]></category>
            <dc:creator><![CDATA[Maciej Wolski]]></dc:creator>
            <pubDate>Wed, 14 Apr 2021 11:01:12 GMT</pubDate>
            <atom:updated>2021-04-14T13:05:53.811Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7PTyKAVFiRFmM_xf" /><figcaption>Photo by <a href="https://unsplash.com/@joelfilip?utm_source=medium&amp;utm_medium=referral">Joel Filipe</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><strong>Universe and AI</strong></p><p>I read recently that <a href="https://thenextweb.com/news/physicists-working-with-microsoft-think-the-universe-is-a-self-learning-computer">physicists working with Microsoft presented theory that we may live in a Universe that is a self-learning computer</a>. The laws of physics may be changed in time and we may never be able to unify them.</p><p>After centuries of mind-bending thoughts, we may re-discover the old truth that even on the most fundamental level the only thing constant… is change. Biological species are evolving, mutating and adapting. Maybe just as the Universe itself.</p><p>The concept of reality as a computation (simulation if you prefer) can explain the phenomena related to quantum mechanics or boundless (not infinite, it is not the same) Universe. As long as you treat computation as an information exchange that may happen also outside silicon-based processors.</p><p>Infinity is a potential, not something that exists on a material level. You can continue adding as many real numbers as you wish, but the result will never be equal to infinity. In the same way, you may never reach the end of the world but it does not mean it is infinite.</p><p>I patiently wait for the results of another work that will show that the Universe is not continuously expanding at all, but it is just an illusion of our limited perception unable to perceive more than 3 dimensions.</p><p>See for yourself how 3D shape (left) could look like for a creature able to perceive only in 2D (right). It is expanding and then contracting. You can even see how it would be for a 4D shape in our (as beings used to 3D) case — later in the original video:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*vV9BVsNF50TjDmgFCm9ZxQ.gif" /><figcaption>Four Dimensional Maths: Things to See and Hear in the Fourth Dimension — with Matt Parker, YouTube: <a href="https://www.youtube.com/watch?v=1wAaI_6b9JE">https://www.youtube.com/watch?v=1wAaI_6b9JE</a></figcaption></figure><p>All of this just resonates with my lifelong view that we should try to be involved in multiple disciplines if we want to truly understand the big picture. The knowledge should be grown by personal experience and not simply passed or trained.</p><p>Combining physics with machine learning — why not? The same with neuroscience, biology and (bio)chemistry. You’ll never know where you will find an idea that will bring the solution to your problem.</p><p>In the case of universal attributes, we still have a lot to confirm. But think for a moment what would happen if we would stick to a past dogma of a fixed, mechanistic Universe with the Earth at its center…</p><p>The reality is often surprising. And the fate of our civilization is just a constant change (and hopefully progress).</p><p><strong>The sequential vs parallel learning</strong></p><p>The sequential process of developing science and knowledge has its advantages and drawbacks. Obviously, we don’t need to start all over from scratch but can learn from the results of other people’s work. But passive reading and accepting the information is not the same as personal evaluation of what was shared by the author.</p><p>Of course, it would be hard to analyze everything. But if you have decided to learn more about a specific topic and maybe even contribute your work to some discipline — you should double-check if the principal elements of your area of interest still can be considered optimal. Or maybe it is necessary to take two steps back to make five forward.</p><p>Parallel learning provides such an ability. You grow your knowledge in multiple dimensions and then you can see, what other people (or at least you) could not before. Just like a 2D creature suddenly exposed to additional ones.</p><p>To look from different perspectives, evaluate computer algorithms with an eye of nature’s enthusiast or physics from a machine learning engineer’s point of view. This is a recipe for true learning, that embeds new knowledge deep into your memory instead of plain copy-paste.</p><p>In many cases, it makes learning new things easier too. I remember how dramatic progress I made with learning how to design hardware with HDL languages while tinkering with low-level neural network components at the same time.</p><p>In short-term perspective it takes time, but with a longer one - it saves it. The broad horizons and practical nature support solving problems faster.</p><p><strong>What is the future of Machine Learning?</strong></p><p>Neural networks as a concept evolved a lot since their conception in the 40s of the previous century. Right now they are deep, have multiple layers and can do a lot of useful things.</p><p>But at this time it is worth asking — do we want a single neural network or highly-capable digital brain with multiple networks? An energy-consuming black-box that in most cases splits out the correct answers, but in many others is terribly wrong? In ways that we would never be.</p><p>Do we think that AI should be taught by us or grown by experience? Ingest our knowledge and think like us — or propose new mind-blowing concepts that were too hard to comprehend by us?</p><p>Of course, we want to have something better. So is it possible to transform one into another?</p><p>Some time ago I encountered the concept of <a href="https://bair.berkeley.edu/blog/2020/04/23/decisions/">neural-backed decision trees</a>. In this concept, a decision factor of a tree is based on (often quite simple) neural network(s). What this provides is the ability to clearly explain the decision process of AI system — step by step, without cutting off the benefits of raw sensory data processing usually associated with deep networks.</p><p>Sounds awesome. But let’s not stop here. Besides having an access to even better and more flexible (very unusual) data structure than a decision tree, my team and I are exploring the concept of disentangled data representations. In contrast to a distributed one, where everything is mixed together — characteristic to most Deep Learning solutions.</p><p>As a result, we can incorporate local, incremental changes and eliminate the problems of catastrophic forgetting or model degradation. We can easily transfer knowledge between separate brains — as hypotheses to evaluate, of course. Not as a source of the ultimate truth. The rule of questioning everything applies also to the machines.</p><p>We don’t need backpropagation anymore. Just like with <a href="https://blog.roboflow.com/liquid-neural-netowrks/">liquid neural networks</a> — the solution is able to adapt in real-time to the existing conditions. We don’t need a separate re-training phase. You turn on the machine and it works and learns at the same time.</p><p>But as truly ambitious people, for the last several years our goal was to create a digital equivalent of a whole brain. So we don’t only design machine learning algorithms that are transparent/explainable, but also energy-efficient, automated and self-learning.</p><p>We equip our algorithms with the ability to manage own goals, simulating real events using imagination and even understand emotions (hello biochemistry!). To induce neural activity without the input data — seeing what could be and not only what is.</p><p>I would lie if I’d say it is easy. It was not in any part. It was hard to get initial funding for what we do, it was exhausting to combine neuroscience, biochemistry, machine learning and other disciplines. But the ultimate question is always: what do you aim for?</p><p>For 5% improvement or pushing things to another level?</p><p><strong>The path to AGI</strong></p><p>In 2021, we finally started to develop from scratch a new Machine Learning framework for machines operating in the physical world — based on our unique 3D Neural Networks forming an equivalent of the digital brain. It is nothing like whatever you have ever seen or used before — both in terms of complexity and capacity.</p><p>It has equivalents of neurons, glial cells, neocortex &amp; subcortical components, neuromodulators and (currently simple) awareness — able to use high-level structures for combining simple low-level elements into more complex concepts, tasks and goals.</p><p>We can’t wait to share the results of our work with the world but to do that we need to spend more time to make it ready to use, also by other people. Tested, improved, documented in easy to understand way.</p><p>At AGICortex — we have a realistic path to Artificial General Intelligence and a strategy to realize our goals in the following years.</p><p>If you are afraid that it will take over the world — don’t treat movies too seriously. Fear is ignited by a lack of understanding. The AGI will be more transparent than any human mind ever was.</p><p>Let’s look forward together to a more positive future. Where we will explore the things that we could not before.</p><p>I am sure there is still so much to learn about the Universe and Artificial Intelligence. But this is what makes this journey so interesting.</p><p>We should never stop learning and be always open to new perspectives. This way, we will be able to be wrong less often.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4e96c869b261" width="1" height="1" alt=""><hr><p><a href="https://medium.com/geekculture/how-could-everyone-have-been-wrong-about-the-universe-and-ai-4e96c869b261">How Could Everyone Have Been Wrong About the Universe and AI?</a> was originally published in <a href="https://medium.com/geekculture">Geek Culture</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Peek Into the Future of AI With This Compressed Knowledge]]></title>
            <link>https://medium.com/swlh/peek-into-the-future-of-ai-with-this-compressed-knowledge-c9a2aca6c44f?source=rss-c9620c68a273------2</link>
            <guid isPermaLink="false">https://medium.com/p/c9a2aca6c44f</guid>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Maciej Wolski]]></dc:creator>
            <pubDate>Tue, 19 Jan 2021 17:36:25 GMT</pubDate>
            <atom:updated>2021-01-20T22:13:18.991Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_WoLHDE6OS5FcW3zf9BkFA.jpeg" /><figcaption>The AI of the future will be native to both the physical and digital world</figcaption></figure><p>The last year was quite intense for me, as together with my partners — I successfully raised Pre-seed funding for our startup — AGICortex. And believe me, it was not an easy task to convince people that we will change how Artificial Intelligence is done forever while having in hand only results of our research and some proof-of-concept demos.</p><p>But now, in 2021 when we are slowly coming out of the shadows — we believe in it even more. Let me share some of our findings that led to defining our company’s long-term strategy: using local data processing (edge devices) in static analytical machines, then using gained awareness about the physical environment in autonomous machines: drones, robots &amp; vehicles to make them much better than they are today. In the next steps, we have plans for some cloud-based products, that will also use the understanding of the physical world gained in the previous steps.</p><p>This unusual sequence of steps will allow us to get to the level of human-like Artificial General Intelligence in the next couple of years.</p><p>The AI agents and solutions of the future will be native to both the physical and digital world.</p><p>Sounds interesting? Then read this compressed summary of what we have learned in 2020 and the previous years:</p><h3>The natural cycle of progress</h3><p>It is important to realize that as humanity, we have collectively created much more incorrect than genuine knowledge. Only to be improved by the next generations.</p><p>Just think about how people understood the world in the past (e.g. Earth as the center of the Universe), how we treated diseases (leeches therapy anyone?), or the deep belief of many past generations that the world is built with solid matter and not moving energy particles.</p><p>The same is with our inventions and tools. We have replaced post pigeons with telegraphs, phones, and then a global computer network called the Internet.</p><p>Is this the ultimate medium of communication?</p><p>I highly doubt and to be honest, it is naive to think that what is now will be like that in the next centuries.</p><p>There were people claiming that we already discovered everything important in physics or that there is no reason for an individual to have a personal computer…</p><p>You know how it ended.</p><p>We like to think about ourselves, our species, or just our generation as special and unique. It is in our nature.</p><p>But the objective reality tells us that everything changes all the time.<br>It is not reasonable to think that future technologies will be strictly based on Deep Learning. Even if it achieved so many successes — and today there is nothing more sophisticated that simulates true intelligence.</p><p>Just like it doesn’t make sense to assume that what you or I work on today, will be the ultimate solution to the problems we decided to solve. It is just the way it is.</p><p>But it is so crucial to make the next step again and again. What humanity needs to progress with AI are automated knowledge creators, not data-hungry artificial statisticians. Let me explain it more below.</p><h3>The difference between skills and knowledge</h3><p>How it can be that we may need weeks of training to possess some skills, but acquire knowledge immediately if it is presented in a useful form?</p><p>You can recognize someone’s face after seeing his photo only once. You can learn some historical facts just after you will read about them in a book.</p><p>It is also possible to understand the idea after hearing the clear explanation.</p><p>Somebody may try to explain to you how it is to ride a bike or drive a car — but you need to try it for yourself multiple times to master that skill.</p><p>Only complex skills require many repetitions with many corrections to achieve satisfying results.</p><p>On the other hand, you can immediately use the acquired knowledge and it is very easy to update it based on new recent experiences or information.</p><p>Is not this a clue for how to build more effective AI systems?</p><p>The fact that something works well, doesn’t mean there is nothing better.<br>Knowledge is much easily transferred than skills between different people (or in future artificial agents) because of its universal nature.</p><blockquote>Data (raw) ➔ Information (classifications) ➔ Explanation (model of relations) ➔ Knowledge (future applications)</blockquote><p>Because of differences in attributes of bodies (humans) and components (machines) and its complexity - the case of skills is a bit harder and requires more adaptation in a sequence of learning iterations.</p><p>Also, sensory processing paths may learn immediately and gradually improve the accuracy, when needed. These conclusions just beg for <strong>hybrid neuro-symbolic architectures</strong> combining artificial neural networks with graph-like structures.</p><p>Then, some basic skill hypotheses may be transferred among machines to be verified in practice and adapted. Just as you can’t master the skill just by watching a training video or reading some tips — but it can shorten the required time by pointing you in the right direction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/418/1*gOCwGMVyVt3mCfoIjx6SgQ.png" /><figcaption>The cerebellum (marked in red)</figcaption></figure><p>In the human brain, we have a structure called the <strong>cerebellum</strong> („the little brain”) that has the largest density of neurons — containing 75–80% of the total amount in only 10% of its volume.</p><p>We may also simplify the situation a bit and state that cerebrum (the largest part with two hemispheres) is active more in our conscious processes, where the cerebellum leads the unconscious processes like movement and keeping posture.</p><p>The digital equivalent of the cerebellum would be a perfect place to store the proven &amp; verified models of skills, that require little to zero further optimization. Instead offering incredibly fast execution.</p><p>On the other hand, the part mimicking the real neocortex and the cerebrum would be responsible for active learning during the process of acquiring more data through experience.</p><p>With a large group of highly capable autonomous AI knowledge creators, we could significantly accelerate the process of acquiring and verifying new knowledge about not only previously unreachable places (space, oceans) but also more familiar places, where they would simply see what we could not recognize before.</p><p>Artificial statisticians that need carefully prepared examples and constant human supervision are not the optimal solution here. The intellectual and creative force on Earth may be multiplied only with a fully autonomous technology.</p><h3>The confirmation bias</h3><p>As humans, we all have a lot of cognitive biases. (<a href="https://www.visualcapitalist.com/every-single-cognitive-bias/">This infographic</a> lists 188 of them if you are curious to check…)</p><blockquote>„Confirmation bias is the tendency to search for, interpret, favor, and recall information in a way that confirms or supports one’s prior beliefs or values.”</blockquote><p>There is no wonder that if someone devoted his life to something — he will continue to believe in its most crucial foundations. It would just be preferred to leave things as they are now.</p><p>Also, the companies that have successful business models relying on specific mechanisms — like sending all data to the cloud for processing and making better ad profiling solutions at the same time — will not be so willing to change them.</p><p>The market for IT professionals is shaped by the needs of companies. The graduates are equipped with knowledge demanded by those organizations or shared by their lecturers.</p><p>Change happens slowly…</p><p>I would prefer solutions that privately process all my data and therefore may have richer access to it. The environmental cost of the most complex Deep Learning models also starts to become significant. Not to mention the lack of transparency and a requirement to hire a whole team of IT professionals to maintain the solutions.</p><p>I am impressed by our technical progress in the last decade with deep respect to the hard work of all researchers and inventors, but for sure it is not the ultimate state.</p><p>That is why questioning <strong>everything</strong> may be one of the most important skills to teach (as soon as possible) yourself, your children… or your Artificial Intelligence systems.</p><p>They need to be able to recognize exceptions to the rule on their own.</p><p>They need to be able to imagine things that do not yet exist. And not always to seek for a good enough solution — but the best one.</p><p>Let’s have the similar courage to make the next step after the technology that has its foundations in the 80s of the previous century.</p><h3>The obscurity of progress</h3><p>Just like Deep Learning models are in most cases obscure to the human eye, the progress of Artificial Intelligence towards AGI will be visible to fewer and fewer people.</p><p>And not only because of the increased complexity of the technology or science behind it. But due to the requirement of merging knowledge from the different disciplines, e.g. neuroscience, biochemistry, mathematics, software development &amp; hardware design.</p><p>For example, Computational Neuroscience has different goals than Artificial Intelligence — by modeling brain activities in a much more precise way.</p><p>But I believe that both sides may learn from each other — and taking more inspiration from the human brain may lead to progress in AI. And the opposite — progress with technology may result in a better understanding of our brains and ourselves.</p><p>Let’s not stop there, why not add biochemistry, psychology, and other science disciplines?</p><p>There is a high risk of failure if we will just stick to mathematically optimized models of neural networks, without a true understanding of who we are and what our bodies are doing…</p><p>That is why some people may state that we have still 50–100 years to experience advanced AGI. Or that we can’t digitize emotions (what is not true), so machines could understand them.</p><p>While I believe that we may experience AGI in this decade, by having enough understanding of what the brain does and why — not exactly how. The technology does not need to reflect all processes, including those that are designed to grow and sustain a biological organ.</p><p>At AGICortex we already have a detailed plan of software and hardware development that leads to exactly this goal. And we are going to realize it.</p><p>The emergence of AGI may shock people that did not see it as a sequence of consecutive steps. The answers are available to those of us who are curious enough to ask the right questions.</p><h3>Multiple data types in a single architecture</h3><p>Our eyes have different types of receptors:</p><ul><li>rods — used to see at a very low level of illumination or at night, very sensitive but color blind</li><li>L-cones, M-cones, S-cones: that respond to the wavelengths of different sizes and therefore different colors (red, green, blue respectively)</li></ul><p>Besides the explanation of daltonism, this fact shows us how adaptable creatures we are. Our vision systems are not only based on color recognition. We can see in poor lighting conditions too.</p><p>We have this duality of two eyes, ears, hands, legs, brain hemispheres, etc. And it leads to more capabilities, in terms of vision — it contributes to the perception of depth. For ears — it allows for better location of the sound sources.</p><p>The structure of our brain allows processing multiple data types together in a very efficient way. But also its prioritization — what is a more salient visual feature at the moment. Is it something that is in the move or closer to us?</p><p>We know that humans are not perfect. But we are so adaptable and designed to be autonomous and survive in so many conditions.</p><p>It is a true miracle of biotechnology to be inspired by.</p><p>So now imagine that you can combine various sensors: depth or infrared cameras, microphone arrays, and even radio frequency sensors allowing you to see through walls…</p><p>But you are limited with your software that can’t combine motion sensing with the mechanism of detecting visual attributes for faster high-definition video stream processing.</p><p>We need to build AI systems that will be able to work with multiple data types efficiently in a single neural architecture.</p><p>Because it increases accuracy, makes the learning faster, optimizes energy expenditures on processing (e.g. don’t process the part that was not changed from previous frames).</p><p>There are sensors and software solutions that can give true superpowers to our AI while keeping it fully autonomous.</p><p>Just like ensemble models that are successful in many tasks — we can use multiple channels of input data classification to respond in the most optimal way.</p><h3>Introduce some structure</h3><p>Neuroscience — even the name suggests what is the main subject. Neurons. But is it true?</p><p>Glial cells have long been regarded as mere fillers, just supporting neural activity. But we have more and more proofs that it is exactly the opposite.</p><ul><li>Radial progenitor cells — provide the necessary initial structure for the brain</li><li>Astrocytes — control the excitatory and inhibitory synapses, using neurons as fast electric wiring while leaving a significant role of master conductors for themselves</li><li>Microglia — protect the brain from dangers as a part of its immune system, but also contribute to adult brain plasticity, removing the unnecessary synaptic connections</li></ul><p>Suddenly with glial cells available, we have much more options to be inspired by the brain and think about possible algorithms.</p><p>Introducing some structure to the brain allows for creating separate, but cooperating modules for different data types — just as the brain has regions responsible for seeing, hearing, other senses, and movement.</p><p>The astrocytes may collect additional information about the neural activity that we could compare to the meta-data — and I can imagine multiple scenarios of how it may be useful in neural architectures. Some have already been implemented in our innovative 3D neural networks.</p><p>Also removing less relevant parts leads to making the whole system more efficient and accurate.</p><p>This structure powered by meta-data can allow not only for incremental learning without a separate training phase, but also the creation of more complex neural units and utilization of additional modules to route the data to the right places and recombine it for new tasks.</p><p>Not bad for a mere filler…</p><h3>Explainable AI and convenient user interfaces</h3><p>I have heard somewhere that AI still waits for its Steve Jobs — and that mass adoption will require front-end simplicity combined with back-end complexity.</p><p>I could not agree more. That is why as an independent researcher I always tried to work on solutions that we will be able to offer to regular non-tech people. And I continue this approach at our company.</p><p>Forget about command-line scripts and complicated tools. These are good for the developers. The future will belong to those that will offer seamless integration with our digital and physical environments.</p><p>Just like we will have multiple sensors in our homes and offices — we will want to quickly interact with our AI systems from multiple locations through convenient audiovisual interfaces.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/660/1*McoYwVp5nF5jfLPSwzw9Jg.jpeg" /><figcaption>Looking Glass Portrait</figcaption></figure><p>Think about current smart assistants combined with visual input/output mechanisms delivered through smart mirrors, <strong>3D holographic screens</strong> or whatever method we will prefer in the future — but strongly integrated with the look of our home or office.</p><p>The mirror will change into the screen when you will need this.</p><p>You will be able to see explanations about how AI makes its decisions and what data its stores upon request made in the most convenient for you method. If 3D visualization is better, then we will have it presented like that. Such screens are already on the market.</p><p>But as AI will be more and more integrated with our daily lives, having rich access to our data and working autonomously — we will demand two things:</p><p>Privacy and to remain in control — by a clear understanding of how our data is collected and used <strong>for our benefit</strong>.</p><p>This is not something that can be provided by the current state of technology. We will need full transparency about the internal operations delivered in an understandable form — not a complex post-fact attempt to explain how the system worked.</p><p>That is why, at AGICortex — we work on visual management tools for AI, that provide full transparency about our 3D neural networks.</p><h3>The future is local</h3><p>Obviously, there will be many scenarios where we will still use centralized cloud-based data processing — for others often local data processing will be preferred. Especially when the result benefits a single individual or a small group of people in the same location.</p><p>For privacy, reduced latency, accessibility and the environment.</p><p>The biggest brains of the next decades will no longer work on surveillance systems and serving you better ads ➔ they will deliver powerful AI solutions to machines at your homes &amp; offices.</p><p>With the further development of local (Edge AI) processing units — most of the use case scenarios will be fine without the cloud.</p><p>And this doesn’t mean that we will not share any learnings between the devices. It just means that we will not need a justification for collecting sensitive data about our private lives in the data centers.</p><p>Highly accurate AI systems will require more data about us. And we will prefer to process it locally.</p><p>At the same time, collected meta-data in innovative neural architectures may allow for rapid optimizations of specific features.</p><p>Some of us may be excited by 5G/6G networks, but the latency associated with sending data to the cloud and waiting for a response may be too high for some types of machines.</p><p>There is a reason why the data should be kept close to the processing unit — the ‘time cost’ of its transportation can be thousands of times higher in physically distant units.</p><p>In the case of autonomous drones, robots &amp; vehicles that difference may mean the ability to avoid some serious accidents.</p><p>The movement to the edge will also mean the development of energy-efficient algorithms that will be able to run on local devices. This will increase accessibility and will also benefit the environment.</p><h3>3D neural networks</h3><p>Sounds really cool, but how actually do all these things described above?</p><p>The neuro-symbolic architecture of 3D neural networks with supporting modules, designed by AGICortex was intended to form a digital equivalent of the human brain.</p><p>Of course, right now we work on smaller tasks like static machines for audiovisual AI assistants. And only in the next phases, we will equip them with more ways to respond to the input data — through their motion components.</p><p>Nevertheless, this approach provides energy-efficiency, explainability, modularity, compatibility with multiple data types that are required by autonomous Artificial General Intelligence. While preserving privacy of the user and offering convenient interfaces of interaction.</p><p>The data in neural networks no longer needs to flow only from left to right or right to left. It can go in multiple directions, by being routed via dedicated supporting modules. It allows us to use only those parts of the whole architecture that are useful for the specific task.</p><p>It is very similar to how the human brain operates.</p><p>Ultimately, we will be able to have massive neural architectures working locally in machines.</p><p>I think it is a much better perspective than spending the amount of electricity used by a large city just to train &amp; run a single huge Deep Learning model that tries to provide the capacity of the human brain just by its size and not agility.</p><p>I would like to share more about 3D neural networks but:</p><ul><li>it is much better to see them in action than to read a description</li><li>we are working on more detailed materials, as you can imagine the topic is quite complex</li><li>we will launch our first products on the market this year</li><li>we aim to offer them as a part of complete ML framework for physical devices</li></ul><p>Stay tuned for more information.</p><p>And by the way — if you feel that you have what it takes (lots of determination and curiosity) to be a part of this journey — do not hesitate to contact me.</p><p>Maybe we can make it together.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c9a2aca6c44f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swlh/peek-into-the-future-of-ai-with-this-compressed-knowledge-c9a2aca6c44f">Peek Into the Future of AI With This Compressed Knowledge</a> was originally published in <a href="https://medium.com/swlh">The Startup</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From neuroscience to the Autonomous General Intelligence]]></title>
            <link>https://maciejwolski.medium.com/from-neuroscience-to-the-autonomous-general-intelligence-c8e51481051e?source=rss-c9620c68a273------2</link>
            <guid isPermaLink="false">https://medium.com/p/c8e51481051e</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[neuroscience]]></category>
            <dc:creator><![CDATA[Maciej Wolski]]></dc:creator>
            <pubDate>Mon, 24 Feb 2020 13:27:40 GMT</pubDate>
            <atom:updated>2020-02-25T09:19:03.110Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RagW84NvpM3UqEQkeDIFkw.png" /></figure><blockquote>„You waste years, by not being able to waste hours” — Amos Tversky</blockquote><p>I started with one of my favorite quotes because it sums up elegantly the optimal approach to doing research.</p><p>There is not a single moment in time when we should stop questioning our ways of doing things, the current state of science or even the whole world.</p><p>People in the past sometimes thought that what could be discovered has already been revealed, and what we can do is simply describe it in more detail.</p><p>They were terribly wrong.</p><p>And if you are an authority because of your previous achievements or role in some organization — you should never ignore other people’s ideas because you might miss something really valuable.</p><p>It is worth to never stop searching and often waste hours, but sometimes save years of looking in the wrong direction.</p><p>I have spent countless hours reading less popular scientific papers and books with few or even no reviews at all. And it was worth it. I found precious gems in texts written by people from all regions of the world. And inspirations to solve technical problems in often surprising sources of information.</p><p>The next generations will almost certainly build better things than us, improving our work, but also by finding weaknesses in our theories and inventions.</p><p>We stick to things that worked in the past — this is our nature. But the nature of progress is constant change.</p><p>I intentionally replaced the usual „Artificial” part of the AGI abbreviation with the word „Autonomous” as the goal of this article is to describe foundations of general intelligence in humans and probably in future machines.</p><h3>General Intelligence should be taught how to learn and expand its sets of skills and knowledge</h3><p>For me the difference between narrow and general intelligence is fundamental and you simply can’t transform one into another.</p><p>Why? The first is optimized to perform a specific task when the expected input is present. It can’t modify the association between input and output on the fly, when it seems more appropriate in the current situation. It does only what it was trained for. Like a (literally) mindless robot.</p><p>The second should be taught how to learn on its own, build a set of useful skills and knowledge. To respond to input data in different ways, according to the context, own goals and motivations, presence of other actors or salient factors.</p><p>In other words, General Intelligence is not trained, but grown to have a rich inventory of possible reactions to the input data. It can reflect on the past and imagine the future. Make plans with delayed execution of an action. Evaluate the effects to make the optimal choice. Find the root cause of a problem instead of just reacting to a preceding event. It is capable of other types of reasoning besides inductive.</p><p>The simplest and the best analogy I can imagine now is to compare knowing the route to a single destination with having a detailed map of the whole environment. In the first case, you just pick the only option available — straight to the goal. In the other case you have plenty of different ways to pick from. And by constantly updating your knowledge, during exploration of your cognitive map — you find new possibilities all the time.</p><p>The concept of a cognitive map makes even more sense when it is realized that the same parts of the brain are involved in mapping both physical and mental space in our brain.</p><p>General Intelligence needs a real paradigm shift in Machine Learning, if we would like to experience it while interacting with machines.</p><p>Sounds great, but how to actually make it happen?</p><p>Well, maybe future generations (or even ours in the next decades) will come up with better ways, but for now we should be more <strong>inspired</strong> by the only example available — the human brain.</p><p>I would like to emphasize the word “inspired”. There is a significant difference between computational neuroscience and artificial intelligence. I will elaborate on this topic further in the article</p><p>The brain allows us to do extraordinary things, but in the same time often causes problems.</p><p>We like to think of ourselves as rational beings, but then we make simple mistakes or abandon our important commitments and plans, because of feelings or short-term pains.</p><p>We have many cognitive biases, biological and physical limitations. Short-term needs and long-term ambitions that influence our choices and responses, often remaining in conflict with each other.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LIz2AypmPFfrpDVftv4NSg.jpeg" /><figcaption>180 examples of human cognitive biases</figcaption></figure><p>Our environment and cultures shape our world models at both conscious and unconscious level. We also inherit genetic features that are no longer relevant in the present.</p><p>But we are the most advanced intelligent beings that we are all aware of.</p><p>Some say that we will not achieve AGI because the human brain, mind or consciousness is too complicated to understand. They suggest that maybe quantum computing will help us — just because it is equally incomprehensible.</p><p>For an extreme futurist like me who spends a lot of time thinking about the times when I will not be here anymore — it is obvious that just as we stopped to perceive Earth as the center of the Universe, at some point of time we will all be convinced that we are not the ultimate intelligence, nor the only one possible.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/167/1*DEA_-6BxKAnaLulY-d16gQ.jpeg" /><figcaption>Kluge — a book by Gary Marcus suggests that instead of being a product of state of the art engineering, our brains are rather an effect of haphazard evolution during the ages.</figcaption></figure><h3>The only available example</h3><p>We don’t know all about the brain — that it is true. But we don’t need and should not replicate everything.</p><p>The brain has its biological constraints — it does many things to achieve advanced intelligence. But also to grow and maintain itself. Clean up the metabolic byproducts or defend itself from external threats.</p><p>It consists of layers of neural components that have evolved in reptiles, mammals and finally humans. Because of this, our minds work in ways that have proved useful in the past, but may no longer be.</p><p>Our rational plans are confronted with emotional guidance on how to avoid even short-term pain and maximize pleasure. Due to the lowest layers of the central nervous system, we can overreact to situations that affect our personal space or self-image.</p><p>The brain doesn’t have a memory that can be retrieved from anywhere just by providing a location address. Often to connect two pieces of information it needs to grow long synaptic connections across two hemispheres and insulate them with myelin sheaths (white matter).</p><p>Mimicking nature is not always the best strategy.</p><p>We can be inspired by nature but do similar things in different ways.</p><p>Our flying machines do not flap their wings. But they overcome the force of gravity by introducing a stronger one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/630/1*ZXF6pnsNqJ6lbtgnvUMrQg.jpeg" /></figure><p>For lifting the bird the strength of its muscles is enough, for moving tons of weight we need jet engines.</p><p>But intelligence is not about introducing stronger force. Like more raw computing power. It is about something completely different.</p><h3>Modular and semi-structured architecture</h3><p>When we look at the anatomy of the brain it is clear that there is a strong presence of order.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*1CXMiuzp7l7RhgGw1nrM7A.jpeg" /></figure><p>We can recognize the modular organization and functional separation. Although people are different and their behavior may be unpredictable — we all have separate parts responsible for processing sensory data, making associations between them and responding with selected actions.</p><p>The front part of our brain focuses on processing thoughts, planning subsequent moves and performing actions. At the back we have the center of sensual perception.</p><p>These modules are separated, but connected and constantly exchange information.</p><p>We can see hierarchical organization not only at the level of brain regions, where the data is recognized at different scales of complexity but also in functional units of the cortex called cortical columns.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*R0YeHdFm-DIVR5nY96oxKw.jpeg" /><figcaption>Functional units of the neocortex — vertical objects called cortical columns. They may vary depending on the location and goal, but still we can recognize similar layered structure.</figcaption></figure><p>We even share a similar semantic space in which similar words (even in different languages) trigger responses in the same areas of the brain.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/760/1*C14YwAShyvUt7HI3mfaZGw.png" /><figcaption>Brain’s semantic space</figcaption></figure><p>This is because all meaning is grounded in the neural encoding of actions and perceptions. Verbs are stored closer to motor areas and nouns at the back, where the perception of objects and their attributes is processed. Self-directed actions are kept together and those that affect others too. The same applies to words describing other agents (people, animals, and probably robots and other autonomous devices in the future), movement of objects, places, tools and devices.</p><p>There is no chaos and randomness. There is a certain structure, but not rigid — leaving room for adaptation and expanding knowledge.</p><p>This is the reason why children first learn a lot about the world, objects and their attributes, actions and agents that perform them — to later label them with words. Without the multi-modal representation of meaning, the words are empty.</p><p>At the age of about 3 years, the hippocampus responsible for episodic memory is ready to store compositions of events, actions, detected actors, objects and their attributes as unified experiences that allow us to experience life as continuous and effectively learn from past experiences, even in our imagination.</p><p>And usually, around the age of 20–22, the prefrontal cortex is mature enough to allow evaluation of the consequences of actions, making complex plans and adapting own behavior to the experienced situation.</p><p>There is no place for chaos or randomness — but the repetition of the most proven qualities. What was successful in our ancestors - survived in our bodies and brains to serve as a base of our cognition that can be adapted to the current needs and conditions.</p><p>This specific template guarantees our efficient learning and supports the process of building a cognitive map, leaving the space for filling in the details.</p><h3>Computing power and energy requirements</h3><p>The human brain, while learning in real time consumes only a fraction of energy necessary for current deep neural networks.</p><p>Our most important body organ has an operating frequency in tens of Hz, but is more capable than computers equipped with processors clocked at levels of GHz (billions of Hz) levels.</p><p>Tens &gt; Billions? It is clear that the brain works in a very different way…</p><p>To achieve AGI we need a paradigm shift in machine learning, not just more of faster processing units.</p><p>Quantum computers will not change this alone, but one day they can strongly contribute to the ASI revolution in which optimal solutions will be available immediately, regardless of the size of the search space.</p><p>But first we must take a step towards AGI and realize that some changes are necessary.</p><p>The human brain switches between multiple various scale neural networks to achieve different tasks.</p><p>You probably don’t need chess skills while driving. This part is skipped during processing, along with many other parts of the massive brain structure.</p><p>In emergencies, quick and good enough solutions are preferred. The range of possibilities is therefore narrowed to the most useful.</p><p>In currently popular neural networks you either do not know how exactly your solution works or you need to introduce even more computation and a burden of energy consumption to reveal some parts of the secret.</p><p>Introducing more structure and order to Artificial Intelligence may allow us to not only have a transparent view of its operation but also help optimize energy consumption. Without sacrificing the ability of adaptation at the same time.</p><p>Such energy efficient approach can lead to neural architectures with billions of available neurons, but only a fraction of them used simultaneously.</p><p>Infinite possibilities. Optimal energy requirements.</p><h3>Other brain components</h3><p>It is often emphasized that the neocortex provides us with our unique abilities and position in the world. It is indeed important. But many other factors influence its operation.</p><p>We have subcortical components that filter what is transferred for further processing in the neocortex, index our memories formed with partial components stored there, modulate our decisions with neurochemicals, drive our attention and feed our consciousness with salient features, help us locate in the environment or monitor and regulate the body state.</p><p>General Intelligence needs much more than our relatively simple Artificial Neural Networks. Mathematically sophisticated (and energy inefficient) but not very capable, when compared to the biological counterpart.</p><p>If that is not enough, all those components were dealing only with neurons.</p><p>We have another, obscure for most people part of the brain that literally controls the neurons!</p><p>Long ignored and seen merely as a filler — astrocytes, members of the glial cells group may have a much more important role. Glia („glue” in Greek) have a parallel communication network that is chemically and not electrically excited.</p><p>In 1985 Marian Diamond, a neuroscientist that studied the brain of Albert Einstein discovered with a disappointment that there was nothing special about his neurons. But he had more astrocytes in the left hemisphere than the average person.</p><p>Then scientists placed human astrocytes in newborn mice to discover that they are growing up as more intelligent.</p><p>Everyone knows that synaptic connection is made between two neurons, right?</p><p>No, it is two neurons and one astrocyte, that supports them and their synapses. Astrocytes are the key to synaptic plasticity and neurotransmitter activity. They also provide nutrients, co-creating a blood-brain barrier.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/947/1*BxFzELFG4O1uDUXxQW0dFA.png" /><figcaption>Tripartite synapse</figcaption></figure><p>But maybe they are capable of doing much more?</p><p>This is material for a different story, or even better — the implementation in the new generation of machine learning algorithms. Do not ignore astrocytes when you think about Artificial Intelligence. They have the potential to revolutionize it.</p><h3>What gives us the possibility of doing and thinking whatever we like at any given moment?</h3><p>Parts of our body and brain are massively connected and communicating with each other by electrochemical means. The modular architecture, functional separation and hemisphere dualism extend our mental capacity, reducing the energy requirements at the same time. What is more important at the moment gets more resources.</p><p>We have our mind — a kind of software equivalent running on the brain’s hardware. It generalizes for us the stimuli from the environment, changing colorless, odorless and tasteless atoms into something meaningful. It creates our perceived reality and because of that doesn’t have many problems with object separation.</p><p>It also provides a mental simulator that can teleport us into the past or imagined future. Evaluate actions that were done or are planned to be. Experience the effects once again or simulate in a test environment to pick the satisfying option or learn on mistakes.</p><p>The simulator is so good that it can generate dreams full of places, people and objects that never existed in the real world.</p><p>The unprocessed emotions of trauma or simply failure remain with us until we resolve them — find an explanation or a better solution to avoid pain or increase your chances of success. Then new knowledge is saved and emotional burden is reduced. It’s just a built-in learning mechanism spread over time.</p><p>We have a planning and scheduling module, the ability to set goals in line with personal values. Biochemistry regulating our nervous system, transmitting and modulating information flow by expanding or limiting the set of processed information.</p><p>Our brain is a self-organizing system based on the most proven strategies of the past, which can be customized through interaction with the world and under the influence of the culture in which we grew up.</p><p>Do you still think that even the most sophisticated neural network optimized for one task can be transformed into AGI?</p><p>We need much more.</p><p>In the brain, different networks constantly communicate with each other, modify themselves in real-time with synaptic plasticity and are influenced by many subcortical components and glial cells.</p><p>We have almost perfected the Narrow AI, but to experience the interaction with AGI — we will need something completely different.</p><p>It will not be easy to do. But we already have pieces of information necessary to do that, that are distributed in scientific books and papers that not many Machine Learning researchers look at.</p><p>On the other hand, neuroscientists more skilled in understanding the brain may lack capacity to translate these information into technology.</p><p>Is it possible that we already have the solutions hidden in plain sight?</p><p>Even if we will build a technology capable of achieving human-like intelligence — do you remember how many years it takes children to reach maturity?</p><p>Technology may be closer than we think — if we will look in the right direction. But we will have to wait longer for humanoid robots or virtual agents who will be like us. After this time, the more difficult part will be done and we will write articles about the ASI revolution, which will significantly benefit from the raw computing power of our next generation processing units or quantum computers.</p><h3>What experts have to say?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*bEuLcgEvNsAm5xz4xYhcUw.png" /><figcaption>Geoff Hinton, the “godfather” of deep learning</figcaption></figure><p><em>“My view is to throw it all away and start again”</em></p><p><em>“Max Planck said, ‘Science progresses one funeral at a time.’ The future depends on some graduate student who is deeply suspicious of everything I have said.”</em></p><p><em>“I suspect that means getting rid of back-propagation. I don’t think it’s how the brain works. We clearly don’t need all the labeled data.”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*K7uR6Q4BA2XU2Vqc3Xfqag.jpeg" /><figcaption>Demis Hassabis, CEO of DeepMind</figcaption></figure><p><em>“Deep learning is an amazing technology and hugely useful in itself, but in my opinion, it’s definitely not enough to solve AI, [not] by a long shot”</em></p><p>Read my previous article to get more information:</p><p><a href="https://medium.com/swlh/why-is-current-deep-learning-technology-a-dead-end-for-artificial-general-intelligence-56f781c36d82">https://medium.com/swlh/why-is-current-deep-learning-technology-a-dead-end-for-artificial-general-intelligence-56f781c36d82</a></p><p><strong>Maciej Wolski</strong>, founder of AGICortex — realistic AGI architecture and PoC technology, a company creating a dedicated AI chip and visual tools for autonomous, explainable AI. Before that, I spent a few years working on the AI R&amp;D projects.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c8e51481051e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why is current deep learning technology a dead end for Artificial General Intelligence?]]></title>
            <link>https://medium.com/swlh/why-is-current-deep-learning-technology-a-dead-end-for-artificial-general-intelligence-56f781c36d82?source=rss-c9620c68a273------2</link>
            <guid isPermaLink="false">https://medium.com/p/56f781c36d82</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Maciej Wolski]]></dc:creator>
            <pubDate>Thu, 06 Feb 2020 15:39:04 GMT</pubDate>
            <atom:updated>2020-02-25T21:06:14.497Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zYNKZq15VlawGX0K" /><figcaption>Photo by <a href="https://unsplash.com/@randylaybourne?utm_source=medium&amp;utm_medium=referral">Randy Laybourne</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3><strong>Introduction</strong></h3><p>To not question things is to agree to stay in the same place. Often during that process, your mind can go to wrong directions. But still, you can learn a lot, during the exploration of uncharted territories of the human potential.</p><p>Excuse me that I will move away from the main topic for a moment, but first I want to share something.</p><p>More than 10 years ago I started to learn in a hard way what is the power of continuous effort. I suffered from an unusual health ailment that made basic things problematic like walking, breathing and sleeping. And believe me, this was just a small part of the whole experience.</p><p>I did anything I could to overcome it. But no one could help me, even the most esteemed professors. The proposed treatments made things only worse, some negative effects of the physical intervention will stay with me until I will die.</p><p>Without any other choice I have decided to learn about human physiology and biochemistry. It took me 10 years to fully recover, but I saw the incremental progress all the time. And the effort was totally worth it.</p><p>That taught me a lot about questioning things and my problem-solving potential. After thousands of hours of learning, I found the most precious piece of information in a very unpopular medical book, written on the other side of the world. I connected the dots and for the last 4 years, I enjoyed perfect health again. I also discovered how to extend my mental and physical performance.</p><p>Some people like doing risky things and experiencing an adrenaline rush, ignoring or not feeling the fear at all. I don’t feel now any limitations in expanding my knowledge in any chosen area. What stands between me and my goal is just time.</p><p>Artificial Intelligence is one of these areas. Once again I have spent countless hours to analyze a hard subject. And after years of research and practical experiments, I am ready to share some of my conclusions. The rest is a foundation of AGICortex — the realistic architecture of Artificial General Intelligence with an early stage proof of concept available.</p><p>So before we start, one additional comment. If what you will read in a moment will make sense — all conclusions were possible to make because of decades of hard work of all involved researchers. And I don’t have any doubt that each human generation can improve the work of the previous one.</p><p>I don’t pretend to have all the answers, but I am quite confident about the direction.</p><p>Having a reliable background in biochemistry, while working since 2007 in the software development industry just helped a lot to open new perspectives of thought.</p><h3>1) Backpropagation is perfect for Narrow AI, but terrible for AGI</h3><p>Backpropagation is the fundamental algorithm of most currently popular Deep Learning applications. It allows the neural network to improve and find the optimal shape. And I can’t imagine right now a better solution to a single problem. But when we try to teach a network another task or just utilize a significant portion of new data some problems may arise, trained abilities partially disappear in a process called “catastrophic forgetting”. This really makes continual learning hard.</p><p>And isn’t intelligence about constant adaptation?</p><p>Of course, there are some ideas about how to overcome that problem, like slowing the modification of parts crucial for previously learned tasks. But to fit new data the number of neurons must be increased.</p><p>The brain has many more neurons than our networks so it is not a problem, right?</p><p>Well, yes but it is also much more energy-efficient than our technology. Like hundreds of thousands of times.</p><p>You see, the brain is modular and can switch on and off large parts of itself. You have billions of neurons in your head, but you only use a small portion of them at a time.</p><p>Only those that are useful. Backpropagation uses all neurons at each iteration.</p><p>Brain’s compartmentalized architecture allows it to learn different things, while still exchanging information between different regions.</p><p>So why not create multiple deep networks and connect them? Yes, there were already such attempts, but advanced intelligence is much more than that — what you will at least partially discover in the rest of this article.</p><h3>2) Unsupervised learning</h3><p>The brain does not need thousands or millions of examples to learn something. It also does not need a supervisor all the time as our deep networks.</p><p>It clearly signals that the method of learning is quite different — more incremental and complex. You can learn something, even if you don’t have clearly defined label or mapped outcome.</p><p>Of course, we did not build flying machines by imitating wing moves — but the birds served as a proof that when we introduce a stronger force that will overcome gravity — the objects heavier than air can fly.</p><p>Even if our AGI will not learn exactly like the brain does, it will be able to do it in an unsupervised way.</p><h3>3) Cognitive maps vs routes</h3><p>Imagine for a moment a solution to your problem as a cognitive route from the input data to the desired output. This is a cognitive route and deep learning strives for mastery in this discipline.</p><p>Deep learning is a form of associative memory between inputs and outputs.</p><p>In the past, people found places by moving along the directions provided by their friends or family.</p><p>„Move straight through the forest, until you will find a river — cross it, turn left and stop near the weird curved tree. Go to the hill with three big rocks on top.”</p><p>Worked fine until somebody cut off the tree or somehow moved the rocks. The additional drawback was that this knowledge had a very narrow application — to a single task.</p><p>What we do now instead of relying on such hints? We create maps that are reusable — we can find multiple routes from various start points to unlimited end points.</p><p>Imagine now building a knowledge representation that will allow you to get to any place in the physical or mental space.</p><p>This is the difference between Narrow and General AI.</p><p>And it has serious implications for neural architectures as well.</p><h3>4) Predictive processing</h3><p>Our brain always actively predicts what will happen in the next seconds. And constantly updates its reality model.</p><p>If you ever wondered why humanity is characterized by such significant curiosity — here is the answer.</p><p>Besides our urge to avoid the pain or get rewards for fulfilling the basic needs, we have also another driver.</p><p>When our basics are good, we want to improve our cognitive map — a mental model of the world. Whenever there is a lack of understanding — we want to know why something is as it is.</p><p>Trust me, I know something about it because I am a bit extreme case. I never stop until I find a satisfying explanation. So I feel this desire quite strongly.</p><p>This predictive processing saves our mental energy but also motivates us to become better and better, expand our knowledge and skills.</p><p>Improve our cognitive map of knowledge and skills to be used at wish in the future.</p><p>Predictive processing is also the source of our intuition. But more about this in one of the next points.</p><h3>5) Energy-efficiency</h3><p>Let’s dig deeper into this topic. Our body is self-assembling biotechnology, including the brain. The energy was something really scarce in the previous centuries and millennia — something you really had to respect. And not just spend however you wish.</p><p>Our brains say „no” to as many things as they can. When your brain processes a continuous data stream it is filtering the inputs through the thalamus and sending to neocortex only those parts of information that are important.</p><p>When a result of unconscious mental processing is perceived as good enough, it is not bothering your consciousness. When you have an important task to do, a large-scale network called DMN (Default Mode Network) is partially disabled and CEN (Central Executive Network) is taking over.</p><p>You don’t have full control over your mind. You can learn how to override this, but it is a part of a different story.</p><p>So, your brain has 86–100 billion neurons, depending on the study. Most of them are densely packed in a part called the cerebellum, responsible mostly for movement of your body.</p><p>Around 16 billion are in the neocortex which is uniquely developed in primates, especially humans what resulted in our special capabilities.</p><p>Deep Learning utilizes all neurons all the time, that is why we don’t have networks with such size. Besides, it is so energy-inefficient that even processing units with operation frequency counted in billions of Hz, do not allow us to compete with the brain, which operates usually at just 10–40Hz.</p><p>Something is seriously wrong…</p><p>Intelligence is not only about introducing a stronger force, like more raw computing power. It will not be the same as with flying machines.</p><p><a href="https://www.technologyreview.com/s/613630/training-a-single-ai-model-can-emit-as-much-carbon-as-five-cars-in-their-lifetimes/">https://www.technologyreview.com/s/613630/training-a-single-ai-model-can-emit-as-much-carbon-as-five-cars-in-their-lifetimes/</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/623/0*SElsprTK4aaWdOXi.png" /><figcaption>Google’s TPUs (Tensor Processing Units)</figcaption></figure><h3>6) Multi-sensory data representation</h3><p>Language is one of the hardest problems in AI. We can now create systems so good in text generation, that it can be perceived as written by humans. But is there any understanding behind it or just a heavy statistical relationship between words?</p><p>Before a child learns language it has experienced a lot of multi-sensory stimuli. Every meaning is not characterized by the neighboring words in some dataset — but a rich set of experiences that include vision, sound, smell, taste, touch, emotions and developed common sense.</p><p>Only then on top of those experiences, we put a correct label, replacing whatever child called eating, a need to go to the bathroom or anything else.</p><p>We can classify unknown objects by similarity of their attributes to already known classes. Even if we don’t know the exact name.</p><p>A sound can easily induce emotional reaction and smell bring back memories.</p><p>Because it is all connected, even if processed separately.</p><p>We need to reflect that in the neural architectures.</p><h3>7) Embodied experience</h3><p>How to explain the world to AI, which knows it only from fixed datasets of still images or text without all the background context.</p><p>How to teach it the meaning of the word „gravity” if it never experienced it.</p><p>If we want machines with human-like abilities, we must recognize that at least our bodies are biological machines. But I am pretty sure that brain is too.</p><p>One of things that make us humans is our rich sensory experience in a body. Our mind renders the world from colorless, soundless and tasteless atoms, interpreting the incoming signals. At the same time creating a reality model that allows us to understand.</p><p>Only then we know that whenever we throw something on Earth it falls down. But a dot on the digital map moving downwards is not affected by the gravity.</p><p>Because digital and physical world may have very different rules.</p><h3>8) The continuous stereo data stream</h3><p>We have this duality in our bodies — double eyes, ears, hands, legs, brain hemispheres.</p><p>It helps us to perceive and interact with the world in a new way, confronting one side with the other. Stereo vision helps us to measure depth in the visual scene, sound allows to locate the source and duality in the brain helps us to confront different sides of our mental capacity: hard logic with creativity.</p><p>At the same time, continuity makes us sure that the person who was your friend a second ago, still is the same and you don’t need to process his visual or vocal attributes to recognize it.</p><p>You can easily divide the background from the moving or newly introduced objects. The same is for sounds.</p><p>Your brain actively predicts what is going on and saves energy, improving confidence and learning in real-time.</p><h3>9) Non-random initialization</h3><p>Neural networks are randomly initialized because gradient-based methods used to train them to break down when all values are the same.</p><p>But do they need to be random? Our brain is not.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/760/1*w_Em91wam196ANfaUfvd5g.png" /><figcaption>Human brain’s semantic space</figcaption></figure><p>What explains why we share similar semantic space in our brains.</p><p>A neural network can be optimized and forced to get the desired form. This is not a problem in a supervised environment. But random initialization can become one if we will have autonomous AI in the real world, where we would like to have the process of an adaptation done in the similar way each time, regardless of time and location.</p><p>I believe that in the future we will use Neuroevolution and pre-trained weights more.</p><p>We think about self-programming machines. Languages are good for humans, they perfectly handle the numbers as the native speakers.</p><h3>10) The emotional state as a general rating system</h3><p>We may think of emotions as unique only to humans and even inferior to the hard logic.</p><p>But it is a general rating system in the body. We can quickly evaluate our state through emotions, distinguishing between energized or even euphoric attitude and depressed low-energy states.</p><p>Whenever we need to make a decision it is based on our feelings. Even if we will think a long time about it, evaluating different options — in the end, we pick what „feels” best.</p><p>We already tried to mimic that with reinforcement learning, but it is only part of the story.</p><p>Emotions can be quantified because they are constructed with various levels of neurochemicals like serotonin, dopamine, noradrenaline and others.</p><p>Our fight or flight response depends on the high level of dopamine which motivates us to pick action quickly. But mostly to the level of noradrenaline: if it is low we experience fear and try to escape, but when it is high we find the courage to fight or at extreme levels we experience anger or even violent madness.</p><h3>11) Digital neuromodulators</h3><p>In the same manner, digital neurochemicals can lead to the autonomous AI allowing it to switch on and off large-scale neural subnetworks.</p><p>In the human brain high level of neuromodulator acetylcholine increases the activity of neurons related to memory, internally directed cognition, thinking, reasoning.</p><p>While dopamine increases the importance of external cognition and faster picking the good enough action.</p><p>Orexin modulates energy availability and is increased when we are most awake and decreasing when we are going to sleep or have an immune response. Digital orexin could help optimal energy expenditure in case of devices not connected to a power source all the time, like autonomous robots.</p><p>Our body uses chemical substances to inform itself and auto-regulate. I think there is a big potential in doing similar things in the area of Artificial Intelligence.</p><h3>12) Artificial intuition</h3><p>For Deep Learning it is only one way to process data. From input to output, a very reactive approach. But we have something really powerful — intuition.</p><p>Our neural units can become pre-activated when all conditions match, but we have not yet experienced the final effect of something in reality. We just feel that it will happen.</p><p>The same mechanism fills our consciousness with thoughts that came out of nowhere, but the mind thinks they will be useful.</p><p>Intuition can help us prepare, but also makes data processing easier. If the context is right — we can lower the thresholds for the object or sound recognition.</p><p>This mechanism sometimes is wrong and we make mistakes with recognizing other people as our friends or hearing words that were not said.</p><p>But most of the time it saves a lot of energy and probably saved a lot of human lives, sending a warning about potential danger.</p><h3>13) Hidden brain — glial cells</h3><p>What for many years was considered merely a filler in the brain — glial cells may have a really strong impact on its operation. We have more glial cells than neurons. They support them, providing nutrients and taking care of the garbage — toxic metabolic byproducts and external dangers by triggering the immune response.</p><p>The current state of the art neural networks completely ignore them. But they literally control the neurons.</p><p>Besides doing the maintenance, they influence the neuronal spikes and possibly also measure the data prediction error.</p><p>Astrocytes are not electrically excited for a short glimpse of time but chemically stimulated for a much longer period. And they can communicate globally the state of things among themselves, not only where more resources are needed, but how to improve the operation of the whole system.</p><h3>14) Subcortical components</h3><p>Current neural networks not only ignore astrocytes, but also subcortical components.</p><p>Besides neocortex, we have a rich supply of modules that support it in data processing. Thalamus, hippocampus, striatum, amygdala to name a few of them.</p><p>And they fulfill really important roles.</p><p>Without the hippocampus and entorhinal cortex, we would not have our memories and it would be hard to orient in physical space. Thalamus filters data and relays it to the correct parts of the brain. Striatum and amygdala modulate responses to the input data.</p><p>And there is also one really interesting part — claustrum. Scientists found that when electrically stimulated it serves as a consciousness on/off switch. (<a href="https://www.newscientist.com/article/mg22329762-700-consciousness-on-off-switch-discovered-deep-in-brain/">https://www.newscientist.com/article/mg22329762-700-consciousness-on-off-switch-discovered-deep-in-brain/</a>)</p><h3>15) Causal reasoning</h3><p>One of the capabilities that makes us really powerful is causal reasoning. We can find the probable cause of the effect in our brains, by doing mental simulations — imaging or reminding the steps of the process.</p><p>Correlation does not mean causation, as expressed in a famous saying.</p><p>Statistics is not enough, we need relations, rich context information and multi-sensory experiences.</p><h3>16) Mental simulator</h3><p>As mentioned at the beginning of the article, our brains literally renders our reality from atoms that do not have any color, smell or sound.</p><p>Because it is so good at it — it can also simulate things that did not happen. This is the reason why our imagination exists, but also why our dreams happen.</p><p>It allows us to experience and learn without any damage in reality. The mental simulator is also a foundation of our conscious experience.</p><p>We created abstract things, existing only in our minds. We imagined many inventions before making them happen. It was truly our source of evolutionary advantage.</p><p>It needs to be present in the AGI architecture.</p><h3>17) Incremental learning</h3><p>Incremental learning is the total opposite of training the neural network with the fixed dataset.</p><p>It allows the machine to learn new things all the time and update existing knowledge to improve itself.</p><p>Sure, we can always re-train the model with the new data but it is very energy inefficient and with larger neural architectures will be problematic to do all the time.</p><p>Although intuitively incremental learning will lead to so-called overfitting, there are possible techniques to overcome that. And it is much easier with enough data that will be incrementally added to the memory.</p><p>Human beings are not masters in everything, they are good with things they had enough experience. And can update their mental models at any time to pick better solution.</p><h3>18) Master algorithm</h3><p>I believe that we will find the master algorithm, that will unlock the key to advanced Artificial Intelligence. The data pre-processing will be different, so-called hyperparameters (elements that are not subject to training) will be different for each case — but the learning algorithm may be the same in the whole Artificial Cortex. Deciding what is more and less relevant, what should be remembered and what forgotten.</p><p>Plus what should be remembered as a negative example.</p><h3>19) Hardware — dedicated processing units</h3><p>Processing units for general purposes like CPUs are not so efficient as dedicated ones. To achieve efficiency required by complex cognitive architectures we absolutely need much more parallelism.</p><p>Because of many layers of abstraction, when you are doing something on your computer it is translated between all those layers like graphical user interface, frameworks and libraries, operating system, programming languages to machine code. And it takes time.</p><p>Multiply it by billions of operations each second and you have a full understanding of what is going on.</p><p>Programming languages are good for prototyping and experimenting, adapting the solution to own needs. But the crucial part needs to be in the processing unit, just as each CPU has Arithmetic Logic Unit inside. Because computer computes.</p><p>We need to implement our key algorithms in hardware. Many companies are already doing that, but are they going in the right direction?</p><h3>20) Self-assembling vs constructed</h3><p>We don’t understand the brain fully. It is very complicated. There are contradictory research results about brain structure. We cannot rely only on neuroscience to build our AGI.</p><p>But we can understand its high-level functions. What it does and more or less how.</p><p>Because the brain is a self-assembling processing unit — it does a lot of things because of its biological and physical needs and limitations.</p><p>Many observed actions or attributes may be related to the fact that they are necessary for this structure to work and survive. Not because it is required for high-level intelligence.</p><p>Having a correct balance between neuroscience, computer science and mathematics — should allow us to finally build the machines with human-like capabilities.</p><h3>Summary</h3><p>With all those technologies around that have developed tremendously throughout our lives, especially in recent years… it is easy to forget that in this field we are still in the STONE AGE, comparing to what we will achieve as humanity in the future.</p><p>Our computation era has just begun. And changing the perspective opens the mind.</p><p>What can bring us somewhere, usually is not the same that will push us forward so much again.</p><p>Evolution usually makes giant leaps forward when someone does something slightly different from the others — expanding the range of the available options.</p><p>And I believe that we can make large progress by looking at different ways of making Artificial Intelligence possible.</p><p>You can read my next article here:</p><p><a href="https://medium.com/@maciejwolski/from-neuroscience-to-the-autonomous-general-intelligence-c8e51481051e">https://medium.com/@maciejwolski/from-neuroscience-to-the-autonomous-general-intelligence-c8e51481051e</a></p><p><strong>Maciej Wolski</strong>, founder of AGICortex — realistic AGI architecture and PoC, a company creating a dedicated AI chip and visual tools for autonomous, explainable AI. Before that, I spent a few years working on the AI R&amp;D projects.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=56f781c36d82" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swlh/why-is-current-deep-learning-technology-a-dead-end-for-artificial-general-intelligence-56f781c36d82">Why is current deep learning technology a dead end for Artificial General Intelligence?</a> was originally published in <a href="https://medium.com/swlh">The Startup</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>