<?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 Arvind Saraf on Medium]]></title>
        <description><![CDATA[Stories by Arvind Saraf on Medium]]></description>
        <link>https://medium.com/@arvindsaraf?source=rss-1c9cdf8e2101------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Stories by Arvind Saraf on Medium</title>
            <link>https://medium.com/@arvindsaraf?source=rss-1c9cdf8e2101------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 17:10:46 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@arvindsaraf/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[Machine Learning Evolution & Software Engineering Implications]]></title>
            <link>https://arvindsaraf.medium.com/machine-learning-evolution-software-engineering-implications-f44bc2d61dfe?source=rss-1c9cdf8e2101------2</link>
            <guid isPermaLink="false">https://medium.com/p/f44bc2d61dfe</guid>
            <dc:creator><![CDATA[Arvind Saraf]]></dc:creator>
            <pubDate>Tue, 17 Feb 2026 11:39:42 GMT</pubDate>
            <atom:updated>2026-02-17T11:39:42.997Z</atom:updated>
            <content:encoded><![CDATA[<p>I recently spoke about the brief history and evolution of machine learning — from early statistical methods to deep neural networks and today’s foundation models — along with a deeper look at what actually makes neural networks work under the hood. More importantly, I explored what these shifts mean for the future of software, especially software engineering: how abstractions are changing, why ML systems behave differently from traditional code, and how engineers must rethink architecture, evaluation, reliability, and product design in an era where intelligence is becoming a core layer of the stack. Sharing the <a href="https://docs.google.com/presentation/d/e/2PACX-1vQt5LuIDWnUvZrpLVd8SeUf1K5bQk6GnMaKqqURce-1kueYxi_A4Wh2nIwDmmvE6DFL1Y8IEt1SIR26/pub?start=true&amp;loop=true&amp;delayms=3000">slides</a> here.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fdocs.google.com%2Fpresentation%2Fembed%3Fid%3De%26size%3Dl&amp;display_name=Google+Docs&amp;url=https%3A%2F%2Fdocs.google.com%2Fpresentation%2Fd%2Fe%2F2PACX-1vQt5LuIDWnUvZrpLVd8SeUf1K5bQk6GnMaKqqURce-1kueYxi_A4Wh2nIwDmmvE6DFL1Y8IEt1SIR26%2Fpub%3Fstart%3Dtrue%26loop%3Dtrue%26delayms%3D3000&amp;image=https%3A%2F%2Flh7-us.googleusercontent.com%2Fdocs%2FAHkbwyLvqMZBX39gsgiJkLs2Dm12QLQd3FYtqJiUU7rAeDDoMrzO7s1aAltda0Dk0mK_RHEssrC3N35_eF1bKa_6MveihqTbkRXCwJ76lXDb9CvMQkfxWnY%3Dw1200-h630-p&amp;type=text%2Fhtml&amp;schema=google" width="700" height="559" frameborder="0" scrolling="no"><a href="https://medium.com/media/bd34fd8f9386e3cb03adc8c3c3453584/href">https://medium.com/media/bd34fd8f9386e3cb03adc8c3c3453584/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f44bc2d61dfe" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Technology for Impact]]></title>
            <link>https://arvindsaraf.medium.com/technology-for-impact-9b1c2c2c2934?source=rss-1c9cdf8e2101------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b1c2c2c2934</guid>
            <category><![CDATA[digital-public-infra]]></category>
            <category><![CDATA[emerging-markets]]></category>
            <category><![CDATA[technology-impact]]></category>
            <dc:creator><![CDATA[Arvind Saraf]]></dc:creator>
            <pubDate>Sat, 25 Nov 2023 16:32:21 GMT</pubDate>
            <atom:updated>2023-11-25T16:32:45.065Z</atom:updated>
            <content:encoded><![CDATA[<h3>Purpose of the document</h3><p>This document looks broadly at how information technology can enable social impact at scale. This could be a government, a non-profit, an emerging market-focused large private incumbent, or a startup.</p><p>Please note:</p><ol><li>This is not meant to be a scholarly article, but my attempt to synthesize my experience in a structure, admittedly for myself, first. I’m sure more thorough &amp; rigorously vetted frameworks are out there — reviewing/enriching this document with their review is yet to be done.</li><li>Much of the context is India &amp; my experiences in it, which may also be representative of most emerging markets, differing though in the stage of technology for impact adoption.</li><li>Impact interventions are never about only technology. Technology is one of the legs or enablers — but intervention has to be more holistic, covering skilling, sensitization, policy, outreach, tie-ups &amp; more. When designing or implementing any intervention</li><li>Many technologist-led interventions have failed due to inadequate holistic focus. And many interventions, technology or otherwise, look promising until they are evaluated or rolled out. We do not touch upon the rigorous evaluation or counter-design of any intervention or counter-design. Instead, we only look at broad-stroke ideas to enable thinking.</li></ol><h3>Categorizing impact interventions</h3><p>Even though this document is about technology, it’s helpful to step back &amp; look at some likely possible intervention types. The categorization below is not exhaustive but helps put technology in perspective. A typical intervention is a combination of multiple.</p><p>Below are some categories:</p><ol><li>Core Sector Advances: Fundamental innovations in the sector that make it more accessible or better and thus more impactful. These could be newer forms or tools of pedagogy, diagnostic mechanisms &amp; more.</li><li>Skilling: Identifying any skills gaps, ways of training, or new strata of the workforce that may help better a benefit reach more people — eg ASHA workers in healthcare.</li><li>Capital: Interventions that lead to more effective or impactful capital access for activities that drive impact, either directly or by setting up technology platforms/innovations/scores that other organizations use to do so. The capital access could be in credit, insurance, or any other instrument.</li><li>Information Technology: Technology is an enabler &amp; gives options to leapfrog reach or resource/human/capital dependency. We’ll talk more about this in the following sections. While there’s a bit of information technology in almost every intervention, some are inherently purely enabled by technology. Some of these, such as a new AI model for weather forecasting for fishermen, require domain expertise as well — and thus are core sector advancements. In this document, we’ll focus primarily on the Information Technology interventions.</li><li>Linkages (Tie-ups &amp; Reach): Even existing interventions often struggle to reach the targeted beneficiaries. Innovation or efforts that help expand the reach, either on the beneficiary or provider side, fall in this category.</li><li>Policy: Good policy puts guardrails to scale what’s proven, often by the government, or ensures what’s done right doesn’t break.</li></ol><p>To illustrate these categories, below are some top-of-the-mind examples in a few sectors:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kp4Xhbzd-KMhS7oraWz9nQ.png" /></figure><p>Here are the impact organizations I have been associated with using the above framework:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6ofmpZb0lBsEUQih5Z1QIQ.png" /></figure><h3>Impact Technology</h3><h3>Benefits</h3><p>Technology can help individuals or impact organizations in some of the following ways:</p><ol><li>Help organizations reduce costs/wastage using digital tools to manage processes. Could be:</li></ol><ul><li>SaaS (open-source / paid) for manufacturing operations, etc</li><li>ML models — eg for automating processes, forecasts</li><li>Public datasets — to allow use in planning, eg weather forecasts, soil data</li></ul><p>2. Revenue and expansion opportunities by enabling new market partnerships, eg marketplaces, platforms, credit, and protocol standards.</p><p>3. Remove access barriers to services and skills, eg platforms, and direct digitization of credit.</p><p>4. Reduce quality inequity by automation &amp; AI — eg, use of AI for content creation, support for teachers, doctors, etc.</p><h3>Classifying Technology</h3><p>One could classify technologies as follows:</p><ol><li>Tools: Easy to use or deploy standalone software/tools + technology partners to support to make an organization more efficient and accessible:</li></ol><ul><li>Language or agent tools — added to the existing software, broadening the access in different languages &amp; modalities, both input &amp; output or given a chat / NLP interface.</li><li>Software tools — for adding or automating some of the existing processes and enabling other benefits of digitization. These could be privately offered, closed-source, or open-source. Irrespective, they should be easy to use, self-deployable (eg Shopify), multi-lingual (easy to tie with a language stack. NLP enabled), and must handle India reality (eg fragmented handicraft manufacturing in India)</li></ul><p>2. Standards: These are standard APIs, often REST, for the commonly expected interactions between independent software in a sector for them to transact. All compliant software can then transact with others, thus allowing for transactions between entities that could not earlier.</p><p>3. Platforms: Multiple entities register on a Platform, allowing themselves to be discovered. These platforms bring trust for otherwise independent entities to transact. The platforms themselves may create &amp; expose a standard or implement an existing one. By aggregating demand &amp; supply, these help scale smaller entities’ business at a large company efficiency.</p><p>4. Public digital assets — eg AI models, DataSets</p><ul><li>AI models — to be used in custom impact applications. These AI models could be domain or use-case-specific, solving problems of a context. Language models enable broader access to services &amp; tools. With sufficiently generic language &amp; generation AI models, the Tools can complemented with Generation AI models &amp; interfaced with Langauge/NLP models as below:</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IaHMaI1asbT0ok07-4I0Aw.png" /></figure><ul><li>Datasets — for training, benchmarking, and building solutions. Creating new AI models for country-specific problems needs data. Democrating this process requires this data to be exposed for external consumption as datasets.</li></ul><p>Startups pick some of these because of technology’s capacity to scale non-linearly &amp; India’s rapidly increasing technology adoption.</p><h3>Digital Public Infrastructure (DPI)</h3><h3>Why DPI?</h3><p>Indian economy is highly fragmented. There are pockets of innovation that don’t reach their due scale. DPIs allow interaction between separate entities. This aggregation enables wider innovation reach, economies of scale, global market-linkage optimizations, and more efficiency-inducing learnings.</p><p>But why aren’t market forces enough? The Indian customer market can be split into three fuzzily different customer segments.</p><ol><li>Higher-income: The problems &amp; likely solutions for this segment are similar to those in the Western world, and they are commercially feasible to provide because of higher paying capacity. The private sector can thus address them.</li><li>Middle income: It is a large growing market, and while solution providers may not be currently viable, they shall be so at scale over time. Thus, venture capital funds these.</li><li>Lower income: This segment’s low-paying capacity means that public entities must create standards and possibly even platforms.</li></ol><h3>Kinds of DPIs</h3><p>More DPI standards &amp; platforms need to be developed. One approach is to identify possible such standards and platforms from the customer segment categorization — look at problems venture capital-backed startups are doing for the middle-income segment, and ask:</p><ol><li>Can a standard allow individual companies to talk to each other — and thus get VC-backed company scale/efficiency?</li><li>Should the above be a standard or a live platform? ie, is there to be centrally kept for security or history/optimization (eg Aadhar records)?</li></ol><p>Below are some ideas:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ddGknICXYP9_GZWKFuarBg.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9b1c2c2c2934" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Translating Engineering wisdom to Entrepreneurship]]></title>
            <link>https://arvindsaraf.medium.com/translating-engineering-wisdom-to-entrepreneurship-aabfdf7f125c?source=rss-1c9cdf8e2101------2</link>
            <guid isPermaLink="false">https://medium.com/p/aabfdf7f125c</guid>
            <category><![CDATA[engineering-mangement]]></category>
            <category><![CDATA[founders]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <dc:creator><![CDATA[Arvind Saraf]]></dc:creator>
            <pubDate>Sat, 21 Oct 2023 06:47:40 GMT</pubDate>
            <atom:updated>2023-10-21T06:47:40.844Z</atom:updated>
            <content:encoded><![CDATA[<p>Some of us make great engineers or engineering leaders. Yet, that doesn’t necessarily translate into conventional entrepreneurial success. This note reflects on the critical tasks for the entrepreneurs in these early stages, ie Product / Market discovery/validation, that the Engineering leaders often get wrong.</p><h3>Engineering leadership to Entrepreneurship</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1WETXxaiC75NzNU-TC-_Og.png" /></figure><p>I have found significant challenges &amp; learnings for myself in steps (0–3). Steps (4–5) are relatively more straightforward if the earlier stages are done right.</p><p>The above must be contextualized within the entrepreneur’s specific goals — eg to make an impact or work in a particular sector to improve it. These decisions must be explicit and may override the industry or market selection.</p><h3>Summarising the challenges</h3><p>Summarising the challenges that come in the early stage product fit:</p><ol><li>Entrepreneur’s echo chamber:</li></ol><ul><li>Entrepreneurs wrongly believe that inspiration &gt;&gt; process</li><li>Fear of the uncomfortable or awkward discussion of being proved wrong</li><li>Entrepreneurial fear of exposing weakness in the initial conjectures on which they hired, raised money</li></ul><p>2. Lack of process or team:</p><ul><li>Market sizing or sentiment is hard</li><li>Product validation is complex &amp; needs structure. Not many organizations &amp; teams have that process</li><li>No one to discuss with as a single founder</li></ul><p>Differentiating between an ideal solution &amp; what customers may be willing to pay for can take time and effort. It’s essential to:</p><ol><li>Validate the problem, i.e. the customers are aware of &amp; call this the problem, have been frustrated by or tried to solve it &amp; can see the implications of translating it to pay for it. This is not inspiration &amp; must be validated with the customers, however good.</li><li>Validate the solution, ie is better than the status quo to overcome the inertia (i.e. Kunal Shah’s Delta-4 framework). Sometimes, a high delta product then up-prioritizes a problem.</li></ol><p>Hence, Team &gt;&gt; Sector/Market &gt;&gt; Problem &gt;&gt; Solution (Product):</p><ol><li>(Founding) Team: Needed for a successful learning process of customer discussions/ideation. If none, get advisors/investors early.</li><li>Sector: Some sectors have inefficiencies, significant &amp; growing, and/or have VC money chasing them — some secondary research &amp; primary research with some close VCs may be good enough.</li><li>Problem: The process above is to be validated in customer or expert discussions.</li><li>Solution: The specific product or solution to the problem to be proposed by the founders/startup.</li></ol><p>A good, cohesive team in a happening sector can find an important enough problem &amp; the right solution, even over multiple iterations.</p><p>We have been too good a founder to take a less market-friendly idea further than most would have managed, to prevent us from giving up on them &amp; see that they may not have been worth the time anyway.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aabfdf7f125c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Affective computing stack?]]></title>
            <link>https://arvindsaraf.medium.com/notes-from-acii-2023-5335170f20df?source=rss-1c9cdf8e2101------2</link>
            <guid isPermaLink="false">https://medium.com/p/5335170f20df</guid>
            <category><![CDATA[emotionai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[affective-computing]]></category>
            <category><![CDATA[multimodal]]></category>
            <dc:creator><![CDATA[Arvind Saraf]]></dc:creator>
            <pubDate>Fri, 20 Oct 2023 13:53:49 GMT</pubDate>
            <atom:updated>2023-10-21T06:38:12.908Z</atom:updated>
            <content:encoded><![CDATA[<p>The Affective Computing + Intelligent Interaction <a href="https://acii-conf.net/2023">Conference 2023</a> happened at MIT in September.</p><p>Some key themes from the conference that came up in the conference:</p><ol><li>Health: Use of Affect for mental health, other health condition diagnosis &amp; early alarms.</li><li>AR/VR, chatbots, esp applications: VR &amp; chat-based applications across domains, also whether the LLMs or indirectly show anything resembling emotions.</li><li>Ethics &amp; Regulations: Europe’s AI laws, biases in emotion by gender, demographic, etc, amongst the foundation models.</li><li>Datasets &amp; tools: Lack of good Affect data, especially for multi-modal analysis &amp; benchmarking. Datasets created using foundation models + human review from movies, etc. Forums setup. Multimodal input &amp; multi-person video/interaction annotation tools gaps — and some research tools open-sourced.</li><li>LLMs, CLIP, etc.: Use of foundation models in affect detection (e.g., convert images to text), evaluation of emotion awareness of these tools — use foundation models as a proxy for emotions in dataset creation.</li><li>Multimodal: Ability to handle multiple input types, cameras, and audio, with gaps, fusion of model outputs, and transfer learning from single modal models.</li></ol><p>These themes can be mapped to the following rough stack:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*swH_F1nzJqZkU8meok_bvA.png" /></figure><p>Most of the research was furthering the highlighted row — affect detection or prediction models, incorporating &amp; continuously monitoring ethical guidelines, and the tooling around it — all in anticipation of these being critical to future applications. Even startups &amp; industry products using Affective computing in their functionality replicate these.</p><p>Unlike the foundation models, which are now highly reusable &amp; taken for granted, perhaps there’s a need for an Emotion AI toolkit — with the above models available as an offering? If such a kit existed, would you incorporate it into your application — or replace your Emotion model with this?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5335170f20df" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Guardrails]]></title>
            <link>https://arvindsaraf.medium.com/regulating-ai-1aa732d8f82e?source=rss-1c9cdf8e2101------2</link>
            <guid isPermaLink="false">https://medium.com/p/1aa732d8f82e</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[bias]]></category>
            <category><![CDATA[ethical-ai]]></category>
            <dc:creator><![CDATA[Arvind Saraf]]></dc:creator>
            <pubDate>Thu, 31 Aug 2023 14:09:06 GMT</pubDate>
            <atom:updated>2023-09-06T06:59:06.223Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>Artificial intelligence is becoming more prevalent in daily life. As a new yet powerful technology, it risks being deliberately or inadvertently misused to harm an individual or a group. Rights could easily be challenged should some guardrails not be implemented in these systems. This document proposes some ideas on guidelines/guardrails a conscious organization can put around it’s AI systems or be a part government regulatory framework for AI. We quickly look at some existing standards &amp; tools.</p><h3>Understanding AI</h3><p>AI systems can be broadly classified into multiple categories. A typical release of an AI system deploys a new ML model, which usually follows certain standard processes. The kinds of AI issues depend on the type of the AI system, and the lifecycle offers an intervention point to prevent these issues.</p><p>Below are some broad categories of AI, perhaps non-exhaustive, that most AI systems can be classified into. Please note that this is not a standard published classification, but categories based on the kinds of prevalent AI output, input &amp; risks, to better align with the objective of managing risks. An AI system may be, even likely to be, some combination of the below categories, rather than falling into one category.</p><h4>Discriminative AI</h4><p>These systems use the organization’s internal data to either build further value or expose the data to the customers. No external data outside the system or publicly available may be used. Examples:</p><ol><li>E-commerce product recommendation system: Uses product metadata and/or users’ prior behavior on the e-commerce site to recommend newer products of interest.</li><li>Customer support chatbot, for say order status and product details. Uses the order &amp; product catalog as the source of the truth. Even if the language inference &amp; response are trained from a public large language model (LLM), the information being exposed is still company-internal.</li></ol><h4>Summarization AI</h4><p>These systems expose existing publicly available information, possibly combined, in a structured or summarized format. Primarily, no new creation may happen. Unlike Generative AI, these systems may use publicly available information. Examples:</p><ol><li>An LLM falls into this category because it presents information from the internet, but LLMs also do a fair bit beyond information compilation &amp; do generation too.</li><li>Any encoding system — e.g. CLIP text embedding models will fall into the category.</li></ol><h4>Generative AI</h4><p>Generative artificial intelligence or generative AI is a type of artificial intelligence (AI) system capable of generating text, images, or other media in response to prompts. Generative AI models learn the patterns and structure of their input training data and then generate new data that has similar characteristics. Examples.</p><ol><li>Image &amp; video diffusion models such as Dall-E, Midjourney, and Runway — fall into these categories.</li><li>ChatGPT. Combining existing information after a point starts becoming like creative generation introducing additional risks around copyright etc. — hence the rationale to differentiate from the Summarization AI.</li></ol><h4>Agent</h4><p>Most AI systems described above may show information, and even recommend a course of action to a user, but do not necessarily act on another system. An AI agent actually acts on the decision on another set of systems usually using the external systems API. This action without human vetting may introduce additional risks, hence the differentiation.</p><p>For the purpose of this classification, we call an Agent the logic acting on the external system, excluding any possible AI that goes into the decision-making. Examples:</p><ol><li>Chatbot that, basis allows users to place or cancel orders.</li><li>Financial instrument trading agent that makes the buy or sell transactions on the equities.</li></ol><h4>Embodied AI (robots etc)</h4><p>Embodied AI systems have physical form &amp; move in the environment, bringing in a different set of risks around safety, etc. Like with an agent, we exclude the AI logic from this definition — to enable us to segregate the risk of physical form.</p><p>It’s easy to see how the real physical systems combine multiple of these:</p><ol><li>Chatbot is an Agent using Discriminative AI.</li><li>Trading bot: Agent, Discriminative, Summarization</li><li>Siri: Summarization, Generative, and if it can pull up information say from your Excel sheets or email on your behalf, also Discriminative &amp; Agent.</li></ol><p>An AI system may output Text, Images, Video, Software code — or any combination of these. The control mechanisms of the AI systems depend on the format of the AI output too.</p><h3>AI lifecycle: ML ModelOps</h3><p>Unless a traditional software system where the decision logic is coded by the developers, an AI system learns the logic as a model from a sufficient number of examples (ground truth) to be fed to it. Any updates, corrections, and improvements typically involve the updation of the model, possibly even the type of model requiring code changes.</p><p>This model development, release, monitoring &amp; update process is fast becoming a mature workflow &amp; is called MLOps (drawing a parallel from DevOps for similar processes for software). The following MLOps stages give an indication of possible interventions to bring in AI regulatory interventions.</p><ol><li>Design: The AI problem definition — i.e. inputs to the model, the output, constraints, checks &amp; requirements are laid out. This is like the software requirements document.</li><li>Data collection/labeling: Machine learning learns from the data fed to it. This data could be fetched from public sources, and live software systems, or could be manually captured by human annotators (“labeled”) in annotation software. Sufficient correct &amp; representative data is vital to the right AI model being learned.</li><li>Model selection &amp; coding: Different AI models work for different kinds of problems. The right model, either publicly available/open-source, must be selected &amp; possibly modified, or a fresh new model or implementation must be done.</li><li>Training: The implemented model must be trained on the data collected to be usable. Training samples from the data collected, with the sampling process offering some intervention points to handle possible model issues.</li><li>Evaluation: AI models are probabilistic, not deterministic. Various metrics of model performance such as precision, and recall on a subset of collected data, sliced further into categories, are used for the statistical soundness of the model before it is put out in production.</li><li>Post-release: Any software system needs constant monitoring. The real-world problem that software or the ML is trained on may change, or some cases/situations not handled well. Constant monitoring mechanisms, the ability to report an issue &amp; quickly repair it, and periodically improve/retrain the models are important even after the initial model deployment.</li></ol><p>A good MLOps system should follow the above processes, with provision for reviews/evaluations at each, and the ability to revert back a model quickly should a bug or an issue be detected.</p><h3>Model deployment scenarios</h3><p>Basis the use case, the AI model may be trained on or run inferencing on the server or the customer device. These choices may matter for risks such as those around privacy.</p><h3>AI system requirements</h3><p>An AI system, like any software, must solve a certain use case. It should do so while respecting basic human principles &amp; rights, often enshrined in the law of the land. Regulatory frameworks bring in a structural framework to ensure these principles are not violated, and there are mechanisms to fix promptly if &amp; when they are.</p><p>An AI system should:</p><ol><li>Be non-misleading: An AI system should be correct (sufficiently, statistically) for all or individual sets of people (“unbiased”). Even indeliberate inaccuracies or biases can mislead the users.</li><li>Be harmless: An AI system be safe for humans, physically or mentally (non-toxic). This requires ensuring they are physically safe &amp; unbiased. Even an advertent flaw may lead to an agent not hoarding the resources to deprive humans or other systems of the resources (non-resource hoarding).</li><li>Respect for individual choices: An AI system often captures individual information that must be securely managed. Individual choice to control both, how this information is revealed or used (privacy), as well as what information the user should see (censorship) should be respected by the AI system.</li><li>Encourage creativity &amp; innovation, i.e. protection of copyright &amp; intellectual property.</li><li>Have adequate processes &amp; controls: No systems are perfect, but processes &amp; controls reduce the likelihood of errors &amp; time to fix any errors/issues seen. Auditable and accountable systems ensure mandated visibility/transparency of the system status, logic &amp; process. Formalized customer redressal process with SLAs allows unexpected issues to be addressed.</li></ol><p>Thus, an AI system must be the following:</p><ol><li>Correct</li><li>Unbiased</li><li>Safe</li><li>Non-toxic</li><li>Non-resource hoarding</li><li>With Censorship controls</li><li>Privacy-aware</li><li>Auditable &amp; accountable</li><li>Redressable for Customer complaints</li></ol><h3>Proposed AI guardrails</h3><ol><li>Any organization building AI systems begins by deciding the kinds of AI it is implementing, and the risks. One framework is in <a href="https://docs.google.com/spreadsheets/d/1IoI5T0zjpHhU7IdGfsc5Gw-o0YxI43Yca6fH-TxjFws/edit#gid=1135598880">this doc</a>.</li><li>Basis the risk assessment, some or all of the controls/reporting mechanisms must be implemented — along with the stated frequency as required.</li><li>For many controls, some existing industry standards or companies offering existing controls are reviewed &amp; a subset of those that be used is recommended, along with the best practices for using them. A preliminary work-in-practice review of existing industry standards is below.</li><li>Additionally, the org should implement:</li></ol><ul><li>Risk assessment periodically (at least Quarterly?)</li><li>Expose API endpoints to test out the system, as described below in the Auditability / Accountability control to an internal or external authority, if there is one.</li></ul><p>Exceptions can be made but must be called out explicitly in design — e.g. if using an attribute is core to AI’s functioning &amp; correctness, and doesn’t violate any principles (e.g. medical diagnostic AI will evaluate differently based on regional propensities for certain conditions, and this is acceptable)</p><p>We detail the interventions for the individual system requirements below. The issues &amp; mechanisms to control them are active research areas. One should expect this document to be state of the art, but must constantly be updated as technology advancements.</p><h3>Requirements &amp; interventions</h3><p><a href="https://docs.google.com/spreadsheets/d/1IoI5T0zjpHhU7IdGfsc5Gw-o0YxI43Yca6fH-TxjFws/edit#gid=1311258362">This document</a> lists the issues &amp; interventions in each type of AI.</p><p><a href="https://docs.google.com/spreadsheets/d/e/2PACX-1vTYK0Z4Xgc73nKirncNeqRijHMbOZhe3KR3N7K-scnkFwO7aK6wgmSyONzy4WRJYHmapG9p1KR1sQ9U/pubhtml">AI Regulatory framework - Google Drive</a></p><h3>Existing standards/tools to implement the AI goals</h3><ul><li><strong>Opting out from being used/scanned by (Gen) AI systems:</strong></li></ul><ol><li>Consent-respecting models: e.g. <a href="https://api.spawning.ai/spawning-api">Spawning APIs</a>— the AI systems have to respect</li><li><a href="https://spawning.substack.com/p/aitxt-a-new-way-for-websites-to-set?subscribe_prompt=free">Ai.txt </a>(inspired by robots.txt )</li><li>Other web mechanisms: Robots.txt to prevent scanning, noindex metadata tag on the pages, Terms &amp; Conditions (though typically not scanned)</li></ol><ul><li><strong>Standard options for media provenance:</strong></li></ul><ol><li>Coalition for Content Provenance and Authenticity (C2PA): An open technical standard (created by Adobe, Arm, Intel, Microsoft etc.) providing publishers, creators, and consumers the ability to trace the origin of different types of media. Handles Privacy and provenance (Attribution/Copyright) — even for AI-generated images</li></ol><p>2. <a href="https://developers.google.com/search/docs/appearance/structured-data/image-license-metadata#iptc-photo-metadata">Google’s IPTC photo metadata</a></p><ul><li>AI Generation: Possibly use the “Special Instructions” field to custom fields</li><li>Copyright Notice, Rights Usage Terms fields</li><li>Confidence Score: Possibly use custom fields</li><li>Source, Credit fields can be used for original sources used in AI training</li></ul><p>3. As a non-owning platform, allow companies to at least implement either of the 2 — specific fields to mention.</p><ul><li><strong>Toxicity:</strong></li></ul><ol><li>Text:</li></ol><ul><li><a href="https://www.perspectiveapi.com/how-it-works/">Perspective API</a> (using Jigsaw by Google) to rate text content, filter &amp; score on toxicity.</li><li><a href="https://github.com/unitaryai/detoxify">Detoxify</a> — open source model for toxic comment classification.</li></ul><p>2. Image and video toxicity detection tools — to identify or create requirements for</p><ul><li><strong>Bias</strong>:</li></ul><ol><li><a href="https://www.iso.org/obp/ui/en/#iso:std:iso-iec:tr:24027:ed-1:v1:en">ISO/IEC TR 24027:2021</a> Information technology — Artificial intelligence (AI) — Bias in AI systems and AI aided decision making.</li></ol><ul><li><strong>Red team datasets (useful for Toxicity, Bias):</strong></li></ul><ol><li><a href="https://github.com/facebookresearch/ParlAI/tree/main/parlai/tasks/bot_adversarial_dialogue">Meta’s adversarial dialogue dataset</a></li><li><a href="https://huggingface.co/datasets/Anthropic/hh-rlhf/tree/main/red-team-attempts">Anthropic red team attempts</a></li><li><a href="https://huggingface.co/datasets/allenai/real-toxicity-prompts">AI2’s Real toxicity prompts</a></li></ol><ul><li><strong>New training approaches to reduce Toxicity:</strong></li></ul><ol><li>Anthropic’s Constitutional AI approach by creating a standard &amp; create a model for Harmfulness evaluation additionally as proposed by the approach using <a href="https://github.com/anthropics/hh-rlhf">Reinforcement learning human feedback (RLHF)</a></li><li>Meta Research’s <a href="https://dynabench.org/">Dynabench</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1aa732d8f82e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Becoming multi-cloud & on-prem]]></title>
            <link>https://arvindsaraf.medium.com/becoming-multi-cloud-on-prem-fdb09e0f3900?source=rss-1c9cdf8e2101------2</link>
            <guid isPermaLink="false">https://medium.com/p/fdb09e0f3900</guid>
            <category><![CDATA[multi-cloud]]></category>
            <category><![CDATA[on-premise]]></category>
            <category><![CDATA[cloud-agnostic]]></category>
            <dc:creator><![CDATA[Arvind Saraf]]></dc:creator>
            <pubDate>Thu, 25 May 2023 07:02:12 GMT</pubDate>
            <atom:updated>2023-06-13T09:52:27.078Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why be multi-cloud or go on-premise?</h3><p>Customer data may be sensitive about their data being captured by the product software. If the data is too much to travel on the internet, needs fast enough processing, or a simple pre-installed box is easier to sell than enabling a bunch of services on the customer cloud, then hardware running the software that can be deployed in the customer premises (on-premise or on-prem) may be a requirement.</p><p>So, the following 2 questions must be answered about the product:</p><ol><li>Is the product offering a hosted SaaS, or is it a product the customer can host within their cloud? Providing support in a customer environment needs significant re-engineering &amp; support.</li><li>If customer-hosted, what clouds are important to support? Is on-prem hosting needed? Porting to every additional cloud provider is effort and taking it on-prem — much more so.</li></ol><p>For the rest of this document, we assume that the answer to question (1) is to make it customer host cloud hostable. The last section assumes an on-prem requirement for question (2).</p><h3>Engineering steps</h3><h4>Simplify stack footprint</h4><p>A simple stack is always better. And when one has to move across public clouds, complexity can be doubly expensive. Going multi-cloud might be a time to relook at the stack, and evaluate the cost benefits of any stack simplification.</p><p>Here are some suggestions:</p><ol><li>Unless there’s a good reason, use a single network, single Kubernetes cluster, and fewest infrastructure components &amp; databases.</li><li>Currently, Docker &amp; Kubernetes are accepted as de facto standards. Being dockerized &amp; kubernetized is practically the first ask by a cloud or platform vendor.</li></ol><h4>Automate deployment</h4><p>The deployment should be scripted, for ease of customer setup. Scripting early ensures there are no information gaps within the Engineering team — code is the most accurate document! It also enables fast iterations.</p><p>Terraform has modules for most common publics clouds, allows for modulization, and due to its internal graph state being maintained, handles upgrades/partial installations optimally. Thus, it works well for this automation.</p><h4>Become cloud agnostic</h4><p>Public clouds differ in their stack, APIs &amp; SDKs. Knowing which clouds one may have to run the software on in the foreseeable future helps plan. Different cloud-specific software versions are unmaintainable. Good engineering &amp; modularization can ensure a largely single largely cloud-agnostic code base, with cloud-specific variations minimized.</p><p>While one may never be fully cloud-agnostic, the below framework helps:</p><ol><li>Use cloud-agnostic abstraction for all the services being used from the cloud — ie Kubernetes, MySQL, object storage etc. Do not use any cloud-specific SDK in the application code.</li></ol><ul><li>Wherever wrapper libraries exist that work across cloud-specific implementations (eg minio client for GCS, S3, etc), use them.</li><li>If an open-source self-hosted version of the service works as well as the cloud-provided version (eg MySQL instead of GCP CloudSQL), use that instead.</li><li>If not, use a wrapper class around the cloud-specific SDK. This wrapper should be cloud agnostic. Push the cloud-specific configuration to settings or secrets out of the application code.</li></ul><p>2. Even infrastructure-as-code (IaC, likely terraform) should be modularized. Perhaps have a common terraform folder, and a cloud-specific folder linking to the implementation in the common folder or overriding it as needed:</p><ul><li>List down the cloud-provided services from the current cloud, and for every cloud option, see if there’s an equivalent in the target cloud.</li><li>If most application services are self-hosted in Kubernetes, the infrastructure code is small — and most deployment can be in helm charts.</li><li>Is the service usage clean &amp; are the cloud abstractions consistent? Can the usage be single terraform module, with cloud-specific implementations? Eg network to cover GCP VPC &amp; AWS network, or Kubernetes to cover GCP GKE &amp; AWS EKS?</li></ul><h4>Prioritization</h4><p>The above steps feed off each other, forcing more clarity, structure &amp; cleanup among them. The sequence of the steps may depend on the organization&#39;s business priorities &amp; the current state of the code. For example:</p><ol><li>If the company needs to release a customer deployable on the currently hosted cloud should prioritize scripting/automating the deployment, then simplification or cloud agnosticity.</li><li>A company in the early stages of development with a clear multi-cloud requirement should prioritize simple stack &amp; cloud agnosticity.</li></ol><p>Irrespective of the prioritization &amp; phases, do minimize any throwaway work.</p><h4>Port to a cloud</h4><p>Porting a specific cloud then includes:</p><ol><li>Mapping &amp; then porting the application services to the cloud-specific equivalents, wherever needed. Eg AWS EKS instead of GCP GKE. In the absence of an equivalent, is there a compelling existing open-source alternative? If so, will the move to the open-source alternative be a step towards cloud-agnosticity?</li><li>Port the IaC code. Terraform providers exist for most clouds, but the terraform module interfaces defined above must be implemented for the cloud using the provider. Or the terraform code itself overwritten.</li></ol><p>This porting effort progressively decreases as the system becomes more kubernetized.</p><p>The exception is when going on-premise, detailed below.</p><h3>Going on-prem</h3><p>Public clouds have a range of services that on-premise stacks do not give it out of the box. The first step is to identify the stack &amp; the hardware one’ll port the application software on — this stack is then installed on the identified hardware. The service gaps vs public cloud are installed separately, filled, the application ported (any code &amp; IaC code changes) &amp; tested.</p><h4>Identify the hardware &amp; the stack</h4><p><strong>Stack</strong></p><p>Public clouds offer compute &amp; disk virtualizations (Virtual machines &amp; Persistent volumes respectively) &amp; the ability to launch Kubernetes clusters, amongst others. At the very least, the one needs a stack offering these to run on the on-premise hardware.</p><p>While multiple Kubernetes variants exist, the following 2 seem the most common &amp; well-integrated:</p><ol><li>Redhat OpenShift</li><li>VMware ESXi, vCenter, VMware Tanzu (multiple flavours)</li></ol><p>On-prem port is a large engineering undertaking. The stack should be well considered using the below factors:</p><ol><li>Stack maturity:</li></ol><ul><li>Stability &amp; Support: A more stable, established or well-used stack means fewer surprises, and the ability to resolve the issues using the community knowledge. Is there a support partner or an expert partner to help with the unknowns?</li><li>Gap vs services needed: The alternatives for filling the infrastructure service gaps are documented to work on only some virtualization OS or Kubernetes software. Does the chosen stack support them?</li><li>GPU support. Most stacks may not support GPUs well — ie multiple GPUs on a single physical node or a VM, GPU multiplexing or time-slicing. This should be an important factor if the application uses GPUs.</li></ul><p>2. Costs &amp; overhead: In addition to the engineering effort, there are other costs:</p><ul><li>Licensing cost: Solutions like VMware &amp; some Loadbalancer options (eg AVI) are licensed.</li><li>Compute Overhead: The virtualization &amp; Kubernetes stack brings its own overhead. Though hard to measure, a quick review of existing online literature will help with a quick evaluation.</li><li>Team skills: Familiarity with virtualization stack is usually a system admin skill than DevOps, and may not be well available in cloud-native organizations. Bringing in an experienced consultant can help save time on experiments &amp; trials.</li></ul><p><strong>Hardware</strong></p><ol><li>Cloud is elastic, on-prem hardware isn’t. Kubernetes should hopefully prioritize this fixed compute by deprioritising &amp; delaying the batch jobs.</li><li>Basis the cloud setup &amp; expected load, estimate the compute, memory &amp; storage needed — closer to the max load that may be needed. A production-grade setup should have (n+1) nodes put together in a single rack — where n can be usually 2–5. This calculation can estimate the size of each node.</li><li>The virtualization stack &amp; networking is more involved in a multi-node setup. For quick POC deployments — a single node meeting the consolidated compute requirement may be good enough.</li></ol><h4>IP / networking</h4><p>Below are the IP addresses in an Application service running on Kubernetes cluster on hardware:</p><ol><li>Physical nodes, routers</li><li>VMs. This includes some service gap-filling VMs — eg for DHCP, DNS, artefact-repository, etc</li><li>Kubernetes pods &amp; services</li><li>External IPs for load balancers</li></ol><p>The hardware on-prem deployable box should largely be portable, without needing any of the internal network changes. Hence a few best practices:</p><ol><li>Create a separate private LAN in a preferably less common (and hence unlikely to conflict) IP range — eg 192.168.23.0/24 or so.</li><li>Allocate the IP ranges explicitly in the block amongst the 3 categories above. Some IP ranges like the Kubernetes pods need not be on the same network but are better managed by Kubernetes.</li></ol><h4>Mapping the services</h4><p>Below are some of the common services to be missing in the virtualization stack often used by the applications. We also list the recommended alternatives:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tgsgGC22qMXikq8Wim1ZWg.png" /></figure><h4>Port, test &amp; test on-site</h4><p>Expect significant code &amp; IaC code rewrites. Different public clouds &amp; on-prem stacks are very different — there will be innumerable surprises. Don’t expect things to work the first time — Murphy’s law will kick in. Budget at least time for testing &amp; fixing them as it took to port, and test it in the customer environment.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fdb09e0f3900" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Modelling the Mind]]></title>
            <link>https://arvindsaraf.medium.com/modelling-the-mind-e4237435f4b1?source=rss-1c9cdf8e2101------2</link>
            <guid isPermaLink="false">https://medium.com/p/e4237435f4b1</guid>
            <category><![CDATA[cognition]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[vipassana]]></category>
            <category><![CDATA[neuroscience]]></category>
            <dc:creator><![CDATA[Arvind Saraf]]></dc:creator>
            <pubDate>Sat, 20 May 2023 14:07:24 GMT</pubDate>
            <atom:updated>2023-06-13T10:35:37.577Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>This document conjectures a modular approach to the description of the mind as a reinforcement learning agent that incorporates the subconscious actions of the mind, with a reward function (“happiness score”) that covers aspects of mental peace &amp; moral alignment. Open problems, research questions &amp; possible approaches are added. It is a research proposal on the above-proposed framework.</p><p>Many of these pieces about understanding the human brain are actively researched. We put them together as a working model, along with some novel components, especially subconscious thinking &amp; incorporating ethics. In the MIT hacking tradition, the model is expected to change as more data points emerge. The model thus has helped articulate some open research problems at the intersection of cognition, machine learning (ML) &amp; neuroscience.</p><p>A better understanding &amp; modelling of our mind &amp; what makes us happy perhaps allows for better treatments &amp; parallel advancements in computing/ML, leading to an overall better life.</p><h3>Related work</h3><p>Language has been long studied &amp; documented[1] to understand how humans learn[2], often serving as representative debates[3–6]. Symbolic AI[7] &amp; neuron-inspired Connectivist approaches[8–10] have been studied, including comparisons[11] &amp; attempts to reconcile or combine them[8,12–14]</p><p>Early attempts to model the human mind included knowledge graphs[15]. ACT[16,17] is a flexible modular software implementation of brain functionality with memory &amp; intention systems used by a control production system.</p><p>Sensory capabilities have been studied in neuroscience with recent attempts to create computational models[18–22], incorporating multi-sensory inputs[23–25].</p><p>Human learning is much more data efficient than artificial neural networks[26], leading to conjectures of a mix of very specialized innate/evolutionary structures &amp; learning models working together in the human brain[27]. These could be innate probabilistic engines of space &amp; physics[28], faces[29], numbers/quantities [30], etc with the human using a hacker-like approach to building his or her working model of the world[32,33], possibly probabilistic[31]. This work has cross-leveraged work in neuroscience identifying functionally specific parts of the brain[32].</p><p>The mind has been often modelled as a Reinforcement learning (RL) agent[33]. Multiple sensory/motor or cognitive subsystems with unified learning have been captured in models of hierarchical learning[34,35]. The brain learns continually unlike common batch learning RL agents. Continual learning research has also modelled the brain’s plasticity[36] to prevent catastrophic forgetting[37–39]. Brain imaging has suggested replay of prior experiences for lower dimensional generalized learning happening in the brain, especially during sleep via a Complementary learning system (CLS)[39]. Memory is essential to short-term &amp; long-term learning[40–42].</p><p>With multimodal sensory inputs and multiple computing engines, an attention mechanism in the brain has been long proposed[43]. The transformer approach[44] has been modified to work for multimodal channels for tasks such as action recognition[45,46]. The Limbic system has been linked with human emotions [47] &amp; modelled [48,49].</p><p>Ancient Eastern literature[50] is rich in the subconscious mind. Freud proposed a 3 part mind model — covering the conscious, subconscious &amp; unconscious[51]. The role of unconscious cognitive reappraisal in emotion regulation has been shown[52,53]. Mindfulness and cognitive therapy have been studied in Psychology[54], with a review of various scales therapies[55]. Various surveys &amp; measures have been measured &amp; reviewed in the context of these mindfulness therapies such as Vipassana[56,57]. Neuroscientists have studied the subconscious activity in the brain using response times to attribute face recognition tasks[58] or brain activity indicators post such meditation practices[59,60].</p><h3>Proposed Human mind model</h3><p>Herein, we define the human mind to be part of a human being that reacts to inputs, decides on what to do &amp; determines the general fuzzy sense of happiness or contentment. The definition is more functional than an anatomical reference to the brain. The mind is modelled as:</p><ol><li>A continually learning reinforcement learning (RL) agent.</li><li>Receives sensory inputs from the experience of the external world as inputs &amp; physical responses (movement of arms, legs, body parts, speech etc) as output.</li><li>Attempts to maximize the reward function (“happiness score”) — by combining different components, whose weights are itself learnt.</li></ol><p>The mind is multimodal with multiple sensory inputs, hierarchical (across senses, systems &amp; timescale), continual learning RL agent with controlled habit formation (plasticity).</p><p>The individual components of the system listed below should explain how the above requirements come together. These modules are deliberately kept high level to allow for current state-of-the-art individual implementations to be incorporated. Possible implementation ideas of individual modules are described for clarity but need not be the only module implementation.</p><h3>Sensory (input)</h3><p>The Sensory module takes the raw sensory inputs (ie light, sound) from the external world, and processes them into higher-level entities (eg scene, action, words) for the further subsystems to consume. This module likely may have different components for different sensory input types, eg:</p><ol><li>Visual component, perhaps like a convolutional neural network (CNN) for optical inputs, with the ability to focus [46,61].</li><li>Auditory component, perhaps like some of the research[62]. An area of work is to see if Fourier neural operator[63] can be adapted to this model.</li><li>Somatosensory (touch) with possible neural maps to map the region of the body[64–66]</li></ol><p>The actual processing of different senses may be implemented in different parts of the brain.</p><p>The Attention module serves a hierarchical attention[45] signal to the Sensory module, like value (“key”) that is used by this module to selectively activate different modes or even regions (ie focal area of view[46], or parts of the body to feel the sensation in).</p><h3>Compute</h3><p>This module contains intelligent components or engines[32] that capture the conscious decision-making, reasoning &amp; learning part of the brain. These could be the physics engine[28], numbers/quantities engine [30] and any other innate construct/models.</p><p>There are some design decisions:</p><ol><li>The Complementary learning system (CLS)[35] that learns generalization in our brains, often in sleep is kept, in this layout as part of this module.</li><li>It can be debated if the external state-based context/attention vector should be an input to this, or if the attention control is already captured in the Sensory module output into this. However, there are cases of us trying to focus attention on something not in our explicit senses (eg trying to do a maths problem in our head while watching a movie). To accommodate for this, a subset of hierarchical Attention vector output goes into this module.</li><li>In the current model, memory subsystems[40–42] are captured here. This may also be using some constructs from ACT-R[67].</li></ol><p>The output subvectors are:</p><ol><li>Reward function (Direct reward output subvector) allowing for the back-propagation like online learning structure to the various engines/submodules. The inferencing could be probabilistic[31] using a hacker-like policy[68]. experiences[68,69].</li><li>Action subvector, which is input to the Motor output &amp; the Subconscious modules. Our calculated responses over time become either form a part of a subconscious reaction.</li></ol><p>Perhaps many of the multiple demand system[70] regions are being used by this module.</p><h3>Subconscious</h3><p>Ancient Eastern literature[50] speaks of the Subconscious as important for our focus &amp; happiness. Loosely, this module captures the union of Freud’s subconscious &amp; unconscious mind[71]. It is currently hard to reason about &amp; hence hardly modelled separately from the Compute module. The current proposed model is based on intuitive experiences &amp; loose study of psychology &amp; meditation.</p><p>The development of innate fear or hatred for certain words, for example, or situations due to prior experience will be a function of this module. Not all decisions by humans are logically well-deducted. Many actions result from subconscious impressions, perhaps innate or developed early in our life that we may not recall forming. This is differentiated from a low dimensional generalized complementary learning system (CLS) often in the sleep state[34].</p><p>Some traits:</p><ol><li>We form moral judgements about any situation or experience.</li><li>We are stressed or disturbed when situations or experiences challenge our current value system.</li><li>Our ability to focus is adversely affected when we are disturbed. Hence the state of the Subconscious affects our Attention.</li><li>The subconscious seems likely involuntary (ie less in our control). However, some deep meditation practices aim to bring in more control of the Subconscious feeling. Hence, the Attention input to this module.</li></ol><p>This Subconscious module thus takes as input a subset of Attention, Sensory module output &amp; Compute module’s Action output, and outputs a combined vector of Subconscious reaction &amp; Moral judgement. This goes into both the Reward function capturing the above traits described below.</p><h3>Attention</h3><p>This module uses the previous state, computed action &amp; sensory input to determine what different systems should focus on. This attention output should be multimodal (ie whether to focus more on visual vs somatosensory)[45,46] &amp; hierarchical (ie focus vision at what position, or touch sensation in which body part)[43]. This module may perhaps output a key[44] to be used to weigh in the respective input to the following modules.</p><ol><li>Sensory module: weights for optical, auditory or somatosensory input processing, and within touch — the part of the body to focus on. These have been proven to work well with action detection[46].</li><li>Compute module: helps select the context &amp; hence the weights of compute engine/subsystem to activate. For example, even for visual inputs — if the images are inferred to be numbers, the maths engine may get more activated, but if these are moving objects, perhaps the physics engine. Or even if the input is largely visual (say a movie) but one is trying to solve a mental maths problem, the context may be explicitly weighted towards the maths engine.</li><li>Subconscious module: perhaps though a weak control that is strengthened with some mindfulness.</li></ol><h3>Reward function</h3><p>The mind has been modelled as a Reinforcement learning (RL) agent with a novel reward function with multiple components.</p><ol><li>Direct reward — similar to artificial RL agents, ie the result of logically deducted action (Compute module output) against the Sensory input. This reward function is a direct result of the Actions by the Compute system, with hierarchical[34,35], continual[37–39] &amp; complementary learning[39].</li><li>A novel reward function to unconscious cognitive reappraisal in emotion regulation [52,53]. Various measurements have been created &amp; reviewed in psychology[72]. Two values are proposed:</li></ol><ul><li>Conscience reward. Each human may have a certain set of social moral values (also gradually learnt, but here captured as input). Our subconscious forms a moral judgement of every input/situation, and those in line with our Moral values lead to happiness. This may be modelled as the distance between the normalized Moral values &amp; Moral judgement vectors.</li><li>Equanimity reward. Humans learn continuously in response to external inputs. However, some experiences or actions significantly challenge internal mental state/world view/ or world view, leading to reactions such as anger, guilt, etc. One expects calmness to take in less mental energy, and perhaps energy in the context of a neural network may be correlated with the number of neurons firing. This can be the sum of the output of each neuron in this module.</li></ul><p>Thus, the reward function proposed is a weighted sum of the 3 components, with the weights to be learnt by the model. More work is needed on this cognitive reappraisal reward to learn from or align with the existing mindfulness measurements[72], possibly to even serve as ground truths to train the revised above Reward function.</p><h3>Motor (output)</h3><p>The Motor module works like a decoder, taking in some encoded Action &amp; Subconscious reaction vectors, and translating the output of the neural network into the movement of visible external reactions &amp; words.</p><h3>Summary</h3><p>This is summarized in the diagram below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*231QKwSgJZWlMd4veCtT-g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zSMrgRs-iOyfkldwaqxRiQ.png" /></figure><h3>Learning in various stages</h3><p>Humans learn continually in real life, with hierarchical[34,35] &amp; continual learning [37–39], and complementary learning systems (CLS)[39]. Memory is essential to short-term &amp; long-term learning[40–42].</p><p>Some notable points:</p><ol><li>Learning is faster in children than in adults, perhaps due to the biological construction of the body. Let’s model this as an age-dependent learning rate.</li><li>Some setups, eg schools, leverage the high learning ability in childhood to primarily focus on academic results as Direct Rewards for Compute module learning.</li></ol><p>The learning happens through an equivalent of back-propagation methods through the above modules, suitably adapted for the above Reward function &amp; components in the individual modules.</p><p>However, an adult may get constant reinforcement of his existing neural pathways through Hebbian plasticity interfering with Continual learning[73]. How does an adult then make a significant shift in learning?</p><ol><li>Unlike artificial intelligence trained on one system &amp; inferring on another, humans use the same hardware (brain/body) for learning &amp; inferencing. The hardware (ie disengaged from work and/or family) must be freed up for faster learning &amp; removing the Hebbian plasticity interference.</li><li>Any machine learning training needs visibility into the model weights, especially for an RL agent — the reward function components. The Direct Reward is more obviously felt or conceived.</li><li>What if the learning has to be in the Subconscious component?</li></ol><ul><li>The Sensory inputs should be minimized, deactivating the Compute module &amp; corresponding Direct Reward. The RL pathways then focus on Equanimity &amp; Conscience rewards, including mindfulness practices to somewhat measure them.</li><li>However, Equanimity &amp; Conscience reward components should be made measurable, even if approximately.</li><li>Any lectures/talks may help move the Moral compass for the person.</li></ul><p>This explains why mindfulness or corporate retreats, coaching sessions, wellness or spiritual retreats often position them away from daily work setup and encourage or include some aspects of introspection or meditation. Let’s do a quick detour on one such practice of Vipassana.</p><h3>Vipassana: Looking within</h3><p>Vipassana is an ancient Indian mindfulness practice[56]. A popular form is Goenka Vipassana, now practised usually as a 10-day silent &amp; disconnected retreat in a monastic setup. The core practice is focused on feeling the breath &amp; body sensations for almost 10 hours every day — thus creating more awareness about it. This is accompanied by some talks/lectures.</p><p>The active part of the Mind model during Vipassana thus is below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1003/1*CqeLdqvtJN5dgPc0J5fPEg.png" /></figure><p>One controls the mind’s attention on the Somatosensory senses, heightening their awareness — while talks inculcate moral values.</p><p>There are other aspects of the course about the senses being reflective of what goes on in the subconscious mind, etc — but these may not be currently scientifically tenable or may need more work.</p><p>Studies have indicated that the participants have gone back with heightened brain activation levels in response to the body&#39;s senses post Vipassana [56,57].</p><p>The practice thus meets all the criteria for an Adult’s relearning, triangulating with Reward &amp; Subconscious models proposed above.</p><h3>Key Ideas &amp; Research Questions</h3><p>A few novel ideas are proposed in this document. We list them along with areas of future work within them:</p><ol><li>Full system model. The document proposes a single unified computational model of the brain. It is hard to validate &amp; expected to grow in complexity, but the following work can help move towards it:</li></ol><ul><li>Detailing of the full model, with memory, and temporal states.</li><li>Detailed comparison with, learnings from or even integration with ACT-R models.</li><li>An open-sourced well-defined interface &amp; working prototype/simulator, with a community &amp; research push to build upon, iterate or port individual components &amp; add to it.</li><li>Datasets &amp; benchmarks for the consolidated model, including leveraging existing ones. Attempting full system training will help resolve more issues, and lend specificity.</li></ul><p>2. Subconscious module &amp; Reward function. This is one of the novel components. Current work has been limited to psychological understanding, measurement surveys, and some brain imaging exercises. Research is needed on:</p><ul><li>Review mindfulness or happiness measures, modifying the model accordingly.</li><li>Computation network structure of the subconscious module, and learning — esp with the weights of the reward components.</li><li>Validating with neurosciences research on the model.</li><li>Developing the module prototype, and datasets, and conducting actual training.</li><li>A text chat or a robot that talks along with the corresponding happiness/reward signal, like MIT Kismet[74] will be great!</li></ul><p>3. Refining the other submodules proposed. Part of this is needed to put the full system together (1), but these are independent research problems. Some open questions are called out in the individual module described above.</p><p>4. Neurosciences research &amp; mapping. This includes:</p><ul><li>Creation of test cases/activities/subjects for imaging.</li><li>Comparison with the current fMRI / imaging studies, mapping to specific areas of the brain, including imaging studies of those practising mindfulness</li></ul><p>5. Mapping into the hardware. One of the drivers of neural network success is the ability to map the RL &amp; Perception networks to matrix operations now done very efficiently in hardware. As the full system is put together or newer modules are inferred, can all of it be mapped efficiently in hardware? Or are new computing architectures needed?</p><h3>References</h3><p>1. Kak SC. The Paninian approach to natural language processing. Int J Approx Reason. 1987;1: 117–130.</p><p>2. Chomsky N. Principles and parameters in syntactic theory. Explanation in linguistics. 1981. Available: <a href="https://cir.nii.ac.jp/crid/1574231873788085632">https://cir.nii.ac.jp/crid/1574231873788085632</a></p><p>3. Pinker S, Prince A. On language and connectionism: analysis of a parallel distributed processing model of language acquisition. Cognition. 1988;28: 73–193.</p><p>4. Ervin SM. LANGUAGE AND TAT CONTENT IN BILINGUALS. J Abnorm Psychol. 1964;68: 500–507.</p><p>5. Cotterell R, Mielke SJ, Eisner J, Roark B. Are All Languages Equally Hard to Language-Model? arXiv [cs.CL]. 2018. Available: <a href="http://arxiv.org/abs/1806.03743">http://arxiv.org/abs/1806.03743</a></p><p>6. Corkery M, Matusevych Y, Goldwater S. Are we there yet? Encoder-decoder neural networks as cognitive models of English past tense inflection. arXiv [cs.CL]. 2019. Available: <a href="http://arxiv.org/abs/1906.01280">http://arxiv.org/abs/1906.01280</a></p><p>7. Spelke ES, Kinzler KD. Core knowledge. Dev Sci. 2007;10: 89–96.</p><p>8. Zador A, Escola S, Richards B, Ölveczky B, Bengio Y, Boahen K, et al. Catalyzing next-generation Artificial Intelligence through NeuroAI. Nat Commun. 2023;14: 1597.</p><p>9. Fodor JA, Pylyshyn ZW. Connectionism and cognitive architecture: a critical analysis. Cognition. 1988;28: 3–71.</p><p>10. Rosch E. Principle of categorization. Cognition and Categorization. 1978 [cited 15 May 2023]. Available: <a href="https://cir.nii.ac.jp/crid/1570291224974644864">https://cir.nii.ac.jp/crid/1570291224974644864</a></p><p>11. Dennett DC. The Logical Geography of Computational Approaches: A View From the East Pole. In: Brand M, Harnish RM, editors. The Representation of Knowledge and Belief. University of Arizona Press; 1986.</p><p>12. Smolensky P, Goldrick M, Mathis D. Optimization and quantization in gradient symbol systems: a framework for integrating the continuous and the discrete in cognition. Cogn Sci. 2014;38: 1102–1138.</p><p>13. Martie L, Rosenberg J, Demers V, Zhang G, Bhardwaj O, Henning J, et al. Rapid Development of Compositional AI. arXiv [cs.SE]. 2023. Available: <a href="http://arxiv.org/abs/2302.05941">http://arxiv.org/abs/2302.05941</a></p><p>14. Hamrick J, Mohamed S. Levels of analysis for machine learning. arXiv [cs.CY]. 2020. Available: <a href="https://baicsworkshop.github.io/pdf/BAICS_6.pdf">https://baicsworkshop.github.io/pdf/BAICS_6.pdf</a></p><p>15. Newell A, Shaw JC, Simon HA. Elements of a theory of human problem solving. Psychol Rev. 1958;65: 151–166.</p><p>16. Anderson JR, Matessa M, Lebiere C. ACT-R: A Theory of Higher Level Cognition and Its Relation to Visual Attention. Human–Computer Interaction. 1997;12: 439–462.</p><p>17. Taatgen NA, Lebiere C, Anderson JR. Modeling paradigms in ACT-R. Cognition and multi-agent interaction: From cognitive modeling to social simulation. 2006; 29–52.</p><p>18. Yamins DLK, DiCarlo JJ. Using goal-driven deep learning models to understand sensory cortex. Nat Neurosci. 2016;19: 356–365.</p><p>19. Siu CR, Murphy KM. The development of human visual cortex and clinical implications. Eye Brain. 2018;10: 25–36.</p><p>20. Cowley BR, Smith MA, Kohn A, Yu BM. Stimulus-Driven Population Activity Patterns in Macaque Primary Visual Cortex. PLoS Comput Biol. 2016;12: e1005185.</p><p>21. Kaas JH, Hackett TA, Tramo MJ. Auditory processing in primate cerebral cortex. Curr Opin Neurobiol. 1999;9: 164–170.</p><p>22. Kanwisher N, Khosla M, Dobs K. Using artificial neural networks to ask “why”questions of minds and brains. 2023. Available: <a href="https://www.sciencedirect.com/science/article/pii/S0166223622002624">https://www.sciencedirect.com/science/article/pii/S0166223622002624</a></p><p>23. Calvert G, Spence C, Stein BE. The Handbook of Multisensory Processes. MIT Press; 2004.</p><p>24. Doehrmann O, Naumer MJ. Semantics and the multisensory brain: how meaning modulates processes of audio-visual integration. Brain Res. 2008;1242: 136–150.</p><p>25. Zhaoping L. Peripheral and central sensation: multisensory orienting and recognition across species. Trends Cogn Sci. 2023;27: 539–552.</p><p>26. Fong RC, Scheirer WJ, Cox DD. Using human brain activity to guide machine learning. Sci Rep. 2018;8: 5397.</p><p>27. Yildirim I, Wu J, Kanwisher N, Tenenbaum J. An integrative computational architecture for object-driven cortex. Curr Opin Neurobiol. 2019;55: 73–81.</p><p>28. Xue H, Torralba A, Tenenbaum JB, Yamins DLK, Li Y, Tung H-Y. 3D-IntPhys: Towards More Generalized 3D-grounded Visual Intuitive Physics under Challenging Scenes. arXiv [cs.CV]. 2023. Available: <a href="http://arxiv.org/abs/2304.11470">http://arxiv.org/abs/2304.11470</a></p><p>29. Liu J, Harris A, Kanwisher N. Perception of face parts and face configurations: an FMRI study. J Cogn Neurosci. 2010;22: 203–211.</p><p>30. Barth H, Kanwisher N, Spelke E. The construction of large number representations in adults. Cognition. 2003;86: 201–221.</p><p>31. Griffiths TL, Chater N, Kemp C, Perfors A, Tenenbaum JB. Probabilistic models of cognition: exploring representations and inductive biases. Trends Cogn Sci. 2010;14: 357–364.</p><p>32. Kanwisher N. Functional specificity in the human brain: a window into the functional architecture of the mind. Proc Natl Acad Sci U S A. 2010;107: 11163–11170.</p><p>33. Mnih V, Kavukcuoglu K, Silver D, Rusu AA, Veness J, Bellemare MG, et al. Human-level control through deep reinforcement learning. Nature. 2015;518: 529–533.</p><p>34. Botvinick MM. Hierarchical reinforcement learning and decision making. Curr Opin Neurobiol. 2012;22: 956–962.</p><p>35. Eckstein MK, Collins AGE. Computational evidence for hierarchically structured reinforcement learning in humans. Proc Natl Acad Sci U S A. 2020;117: 29381–29389.</p><p>36. Gütig R, Aharonov R, Rotter S, Sompolinsky H. Learning input correlations through nonlinear temporally asymmetric Hebbian plasticity. J Neurosci. 2003;23: 3697–3714.</p><p>37. Parisi GI, Kemker R, Part JL, Kanan C, Wermter S. Continual lifelong learning with neural networks: A review. Neural Netw. 2019;113: 54–71.</p><p>38. Russin J, Zolfaghar M, Park SA, Boorman E, O’Reilly RC. A Neural Network Model of Continual Learning with Cognitive Control. arXiv [q-bio.NC]. 2022. Available: <a href="http://arxiv.org/abs/2202.04773">http://arxiv.org/abs/2202.04773</a></p><p>39. van de Ven GM, Siegelmann HT, Tolias AS. Brain-inspired replay for continual learning with artificial neural networks. Nat Commun. 2020;11: 4069.</p><p>40. Rodríguez-Flores TC, Palomo-Briones GA, Robles F, Ramos F. Proposal for a computational model of incentive memory. Cogn Syst Res. 2023;77: 153–173.</p><p>41. Sharma S, Chandra S, Fiete I. Content addressable memory without catastrophic forgetting by heteroassociation with a fixed scaffold. 2022; 19658–19682.</p><p>42. Ruff DA, Xue C, Kramer LE, Baqai F, Cohen MR. Low rank mechanisms underlying flexible visual representations. Proc Natl Acad Sci U S A. 2020;117: 29321–29329.</p><p>43. Posner MI, Petersen SE. The attention system of the human brain. Annu Rev Neurosci. 1990;13: 25–42.</p><p>44. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al. Attention is all you need. Adv Neural Inf Process Syst. 2017;30. Available: <a href="https://proceedings.neurips.cc/paper/7181-attention-is-all">https://proceedings.neurips.cc/paper/7181-attention-is-all</a></p><p>45. Shi Q, Fan J, Wang Z, Zhang Z. Multimodal channel-wise attention transformer inspired by multisensory integration mechanisms of the brain. Pattern Recognit. 2022;130: 108837.</p><p>46. Wang L, Zang J, Zhang Q, Niu Z, Hua G, Zheng N. Action Recognition by an Attention-Aware Temporal Weighted Convolutional Neural Network. Sensors . 2018;18. doi:<a href="http://dx.doi.org/10.3390/s18071979">10.3390/s18071979</a></p><p>47. Kuremoto T, Ohta T, Kobayashi K, Obayashi M. A Functional Model of Limbic System of Brain. Brain Informatics, International Conference, BI 2009, Beijing, China, October 22–24, 2009, Proceedings. 2009;5819: 135–146.</p><p>48. Parsapoor M. An introduction to brain emotional learning inspired models (BELiMs) with an example of BELiMs’ applications. Artif Intell Rev. 2019;52: 409–439.</p><p>49. Assunção, Patrão, Castelo-Branco, Menezes. An Overview of Emotion in Artificial Intelligence. IEEE Transactions on Artificial Intelligence. 2022;3: 867–886.</p><p>50. Wikipedia contributors. Samskara (Indian philosophy). In: Wikipedia, The Free Encyclopedia [Internet]. 15 Sep 2022. Available: <a href="https://en.wikipedia.org/w/index.php?title=Samskara_(Indian_philosophy)&amp;oldid=1110482043">https://en.wikipedia.org/w/index.php?title=Samskara_(Indian_philosophy)&amp;oldid=1110482043</a></p><p>51. Sigmund F. Quelques considérations pour une étude comparative des paralysies motrices organiques et hystériques (1893). Résultats, idées, problèmes. 1999;1.</p><p>52. Hopp H, Troy AS, Mauss IB. The unconscious pursuit of emotion regulation: implications for psychological health. Cogn Emot. 2011;25: 532–545.</p><p>53. Sander D, Delplanque S. Unconscious emotional processing. Food Qual Prefer. 2021;92: 104177.</p><p>54. Bishop SR, Lau M, Shapiro S, Carlson L, Anderson ND, Carmody J, et al. Mindfulness: A proposed operational definition. Clinical Psychology: Science and Practice. 2004;11: 230–241.</p><p>55. Khoury B, Lecomte T, Fortin G, Masse M, Therien P, Bouchard V, et al. Mindfulness-based therapy: a comprehensive meta-analysis. Clin Psychol Rev. 2013;33: 763–771.</p><p>56. Zeng X, Oei TPS, Ye Y, Liu X. A critical analysis of the concepts and measurement of awareness and equanimity in Goenka’s Vipassana meditation. J Relig Health. 2015;54: 399–412.</p><p>57. Zeng X, Oei TPS, Liu X. Monitoring emotion through body sensation: A review of awareness in Goenka’s Vipassana. J Relig Health. 2014. Available: <a href="https://link.springer.com/article/10.1007/s10943-013-9754-6">https://link.springer.com/article/10.1007/s10943-013-9754-6</a></p><p>58. Hoshiyama M, Kakigi R, Watanabe S, Miki K, Takeshima Y. Brain responses for the subconscious recognition of faces. Neurosci Res. 2003;46: 435–442.</p><p>59. Delgado-Pastor LC, Perakakis P, Subramanya P, Telles S, Vila J. Mindfulness (Vipassana) meditation: effects on P3b event-related potential and heart rate variability. Int J Psychophysiol. 2013;90: 207–214.</p><p>60. Cahn BR, Delorme A, Polich J. Occipital gamma activation during Vipassana meditation. Cogn Process. 2010;11: 39–56.</p><p>61. Cadena SA, Denfield GH, Walker EY, Gatys LA, Tolias AS, Bethge M, et al. Deep convolutional models improve predictions of macaque V1 responses to natural images. PLoS Comput Biol. 2019;15: e1006897.</p><p>62. Chambers JD, Elgueda D, Fritz JB, Shamma SA, Burkitt AN, Grayden DB. Computational Neural Modeling of Auditory Cortical Receptive Fields. Front Comput Neurosci. 2019;13: 28.</p><p>63. Li Z, Kovachki N, Azizzadenesheli K, Liu B, Bhattacharya K, Stuart A, et al. Fourier Neural Operator for Parametric Partial Differential Equations. arXiv [cs.LG]. 2020. Available: <a href="http://arxiv.org/abs/2010.08895">http://arxiv.org/abs/2010.08895</a></p><p>64. Blatow M, Nennig E, Durst A, Sartor K, Stippich C. fMRI reflects functional connectivity of human somatosensory cortex. Neuroimage. 2007;37: 927–936.</p><p>65. Brecht M. The Body Model Theory of Somatosensory Cortex. Neuron. 2017;94: 985–992.</p><p>66. Lassagne H, Goueytes D, Shulz DE, Estebanez L, Ego-Stengel V. Continuity within the somatosensory cortical map facilitates learning. Cell Rep. 2022;39: 110617.</p><p>67. Borst JP, Anderson JR. A step-by-step tutorial on using the cognitive architecture ACT-R in combination with fMRI data. J Math Psychol. 2017;76: 94–103.</p><p>68. Rule JS, Tenenbaum JB, Piantadosi ST. The Child as Hacker. Trends Cogn Sci. 2020;24: 900–915.</p><p>69. Ding M, Xu Y, Chen Z, Cox DD, Luo P, Tenenbaum JB, et al. Embodied concept learner: Self-supervised learning of concepts and mapping through instruction following. 2023; 1743–1754.</p><p>70. Duncan J. The multiple-demand (MD) system of the primate brain: mental programs for intelligent behaviour. Trends Cogn Sci. 2010;14: 172–179.</p><p>71. Freud S. The Unconscious. Penguin UK; 2005.</p><p>72. Bergomi C, Tschacher W, Kupper Z. The Assessment of Mindfulness with Self-Report Measures: Existing Scales and Open Issues. Mindfulness . 2013;4: 191–202.</p><p>73. Zenke F, Gerstner W. Hebbian plasticity requires compensatory processes on multiple timescales. Philos Trans R Soc Lond B Biol Sci. 2017;372. doi:<a href="http://dx.doi.org/10.1098/rstb.2016.0259">10.1098/rstb.2016.0259</a></p><p>74. Wikipedia contributors. Kismet (robot). In: Wikipedia, The Free Encyclopedia [Internet]. 1 Apr 2023. Available: <a href="https://en.wikipedia.org/w/index.php?title=Kismet_(robot)&amp;oldid=1147693935">https://en.wikipedia.org/w/index.php?title=Kismet_(robot)</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e4237435f4b1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The ethics of layoffs]]></title>
            <link>https://arvindsaraf.medium.com/the-ethics-of-layoffs-1f61e95a7564?source=rss-1c9cdf8e2101------2</link>
            <guid isPermaLink="false">https://medium.com/p/1f61e95a7564</guid>
            <category><![CDATA[layoffs]]></category>
            <dc:creator><![CDATA[Arvind Saraf]]></dc:creator>
            <pubDate>Sat, 06 May 2023 16:35:53 GMT</pubDate>
            <atom:updated>2023-05-06T16:35:53.724Z</atom:updated>
            <content:encoded><![CDATA[<p>I recently had to participate in some actions on layoffs and had to reconcile with the ethics of it.</p><p>A healthy organization must continuously let go of non-performers or misfits for its own good. However, when employees are let go off for reasons beyond their act or performance but because of a company’s status, it is a layoff. Organizations choose to lay off people, and the ethical correctness of it must be evaluated in terms of its benefit vs other choices (“consequential utilitarianism”):</p><ol><li>Does the action lead to net long-term happiness for the people combined?</li><li>Have those to benefit from it not “unearned” the benefit?</li></ol><p>A publicly incorporated company can possibly justify laying off 6% of its workforce (for larger benefit or protection of existing millions of shareholders or hundreds of thousands of employees), but should a private startup that is unable to raise money lay off 70% of its team? Would that startup be better off paying generous severance to the team than working on an exit plan that can potentially reward its small group of investors &amp; subset of management (which one may claim, have failed so far — given the current state of the startup)? Looks like Google may not be as evil after all.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1f61e95a7564" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Social enterprise vs System change agent]]></title>
            <link>https://arvindsaraf.medium.com/social-venture-heropreneurship-1aca8e262372?source=rss-1c9cdf8e2101------2</link>
            <guid isPermaLink="false">https://medium.com/p/1aca8e262372</guid>
            <category><![CDATA[impact-investing]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[social-ventures]]></category>
            <dc:creator><![CDATA[Arvind Saraf]]></dc:creator>
            <pubDate>Sat, 06 May 2023 15:57:08 GMT</pubDate>
            <atom:updated>2023-05-24T17:00:57.608Z</atom:updated>
            <content:encoded><![CDATA[<p>Creating social impact may not be the most creative or glamorous pursuit, and may in fact conflict with it, as <a href="https://ssir.org/articles/entry/tackling_heropreneurship">this article on Heropreneurship</a> (by <a href="https://www.linkedin.com/in/danielapapi/">Daniela Papi-Thornton</a>) mentions. It’s easy to reflect on my prior experience &amp; see why.</p><p>Back at MIT in 2001, CK Prahlad’s “Fortune at the Bottom of Pyramid” was frequently talked about as a new paradigm in social development. One needed to create a new innovative solution and scale it up as a venture. This new “Social enterprise” promised a chance to combine the intellectual high of creativity, the glamour of entrepreneurship &amp; the conscience of social impact. I joined Swasth India as a co-founder (2008–2011). Swasth provided healthcare for the poor as a social impact business.</p><p>Here are the questions I reflect on:</p><ol><li>Entrepreneurs believe they can solve problems better than others, else we may not even venture. “<a href="https://en.wikipedia.org/wiki/Dunning%E2%80%93Kruger_effect">Dunning-Kruger effect</a>” has merit. Smart founders are expected to learn quickly &amp; catch up. Low immersion, however, can lead to poor starting points or complete repeat or dismissal of existing organizations. Both are hard to correct.</li><li>The existing organizations have survived for years on frugality. In the larger view, it’s more capital efficient to spend on experiments &amp; fail quickly. Yet, both duplication of work of existing organizations &amp; a spend on experiments can make existing organizations see you as an outsider.</li></ol><p>It’s important to think these through.</p><p>Somewhere in our execution, we transitioned to being social change agents a few years later, setting up a non-profit — and moved most of our work there.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1aca8e262372" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Engineering design principles]]></title>
            <link>https://arvindsaraf.medium.com/engineering-design-principles-1120422a9ec2?source=rss-1c9cdf8e2101------2</link>
            <guid isPermaLink="false">https://medium.com/p/1120422a9ec2</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[tech-debt]]></category>
            <dc:creator><![CDATA[Arvind Saraf]]></dc:creator>
            <pubDate>Fri, 05 May 2023 02:04:33 GMT</pubDate>
            <atom:updated>2023-05-05T02:13:57.090Z</atom:updated>
            <content:encoded><![CDATA[<p>Good engineering, like any other team within an organization, has to further the organization’s goals. In this case, in particular, the goal primarily is to build a product that is bug-free, robust, and sufficiently scalable. Based on the stage of the company, it should have fast enough first-time development and allow for quick iterations.</p><p>This requires the engineering team to make various tradeoffs between conflicting priorities, e.g. code quality and robustness vs development speed. Tech debts are inevitable. Good engineering then is about minimizing bad tech debts and taking good tech debts.</p><p>The slides share some of my learnings and thumb rules I have used in building software:</p><ol><li>Tech debts (more <a href="https://arvindsaraf.medium.com/tech-debts-2439d70c3f9c">detailed article</a>).</li><li>Engineering principles (and how to use them). Often teams tend to overdo principles such as generalizability, simplicity &amp; reusability.</li><li>Frameworks &amp; tools</li><li>Scalability &amp; Optimization. Is building a highly scalable solution necessary at any stage? And when does one optimize? More <a href="https://arvindsaraf.medium.com/scalable-software-systems-9929697a1064">here</a>.</li><li>Engineering team metrics. Like any business function, there are proven &amp; established metrics for engineering team performance.</li></ol><p>Happy to invite feedback/additions. You can directly comment in Google <a href="https://docs.google.com/presentation/d/1MtEKGo_9RLkhmQBs8fMa94x1PeNbxkZrdjSJN_7J9vQ/edit#slide=id.p">slides here</a>.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fdocs.google.com%2Fpresentation%2Fembed%3Fid%3D1MtEKGo_9RLkhmQBs8fMa94x1PeNbxkZrdjSJN_7J9vQ%26size%3Dl&amp;display_name=Google+Docs&amp;url=https%3A%2F%2Fdocs.google.com%2Fpresentation%2Fd%2F1MtEKGo_9RLkhmQBs8fMa94x1PeNbxkZrdjSJN_7J9vQ%2Fedit&amp;image=https%3A%2F%2Flh3.googleusercontent.com%2Fdocs%2FADP-6oHt33zJ6rwqM3Sx_VmEurYmGBKwnK5ma_cJA4K0k1wBwxOz7GrwKb3hu2MEiTLojU44r_bgBxq_Wcdmia62maAappgaTlxBy59AA_KEsKy-%3Dw1200-h630-p&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=google" width="700" height="559" frameborder="0" scrolling="no"><a href="https://medium.com/media/d10da794f09452373864c593f090159c/href">https://medium.com/media/d10da794f09452373864c593f090159c/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1120422a9ec2" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>