<?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 Victor Smirnov on Medium]]></title>
        <description><![CDATA[Stories by Victor Smirnov on Medium]]></description>
        <link>https://medium.com/@victorsmirnov?source=rss-162e5c148cc6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*6m8V9-kqHmMSruM0mGegtQ.jpeg</url>
            <title>Stories by Victor Smirnov on Medium</title>
            <link>https://medium.com/@victorsmirnov?source=rss-162e5c148cc6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 11:25:32 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@victorsmirnov/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[Brief Notes on AI and Process Introspection]]></title>
            <link>https://medium.com/@victorsmirnov/brief-notes-on-ai-and-process-introspection-615af6104d6b?source=rss-162e5c148cc6------2</link>
            <guid isPermaLink="false">https://medium.com/p/615af6104d6b</guid>
            <category><![CDATA[consciousness]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Victor Smirnov]]></dc:creator>
            <pubDate>Mon, 20 May 2019 17:20:29 GMT</pubDate>
            <atom:updated>2019-05-20T17:20:29.208Z</atom:updated>
            <content:encoded><![CDATA[<ol><li>All AI researchers and developers are reproducing their so-called personal mind theories (PMT) in software artifacts.</li><li>But because of various cognitive biases, spontaneous (or naive) mind theories are inherently broken. And a person is not aware about this phenomena. He/she acts as if he/she had true picture of his/her mental processes. Wrong input leads to wrong output.</li><li>One of the most important cognitive bias affecting AI R&amp;D is apparent <em>simplicity</em> of mind. Because of this bias many subjective processes appear as not having internal structure.</li><li>Simplicity of mind is a derivative from the lack of <strong>process introspection</strong>, that is innate but requires intensive development during lifetime. Inaccessible machinery of mind appears subjectively (phenomenally) as ‘no machinery’, or as indeterminism (free will).</li><li>One manifestation of the simplicity (of mind) bias is that researchers and developers tend to dramatically underestimate engineering complexity of AI problems, because they subjectively appear as pretty simple.</li><li>The same manifestation is when a researcher feels that he/she is almost near the final solution to AI, and just a few steps are necessary. But when these steps are taken, new, previously unknown and important machinery of mind appears. And it may go this way for decades.</li><li>Experienced, long term AI researchers, who are practically aware about this phenomena, may tend to develop AI-related over-scepticism: “We are decades or even centuries away from the human-level/human-like/strong AI”. No, we are not.</li><li>Partial process introspection may stuck on producing locally consistent but otherwise random mind theories which neither sufficiently improve over time, nor even converge with mind theories, developing by other researchers. This leads to “war” of AI approaches.</li><li>Without mind theories, AI development is a blind trial-and-error activity. Or, worse, it’s pursuing miracles.</li><li>Neuroscientific approach to AI really works, but not the way it may be expected. It operates at much lower level than higher mental functions. In other words, they may be not that <em>easily</em> accessible from the level of connectome. Sure, hardware engineers can really help distributed systems engineers, but not that much.</li><li>Cognitive approach to AI is also important, because it’s necessary for developing productive mind theories, witch turns into AI when being expressed with software artifacts.</li><li>Given that, there is something more to successful AI project than just pure math and engineering: process introspection (PI).</li><li><a href="https://medium.com/@victorsmirnov/how-to-compensate-introspection-illusion-62f357e9326c">It can be gradually improved</a> over a naive state, though it’s a lengthy process that may take years and even decades. And it requires interdisciplinary knowledge: psychology, philosophy and math.</li><li>Advanced data structure skills (DS) are required for PI, because inner mind is irreducibly complex. We see only what we can express. There is no understanding without expression. Meaning is use.</li><li>Psychological mission of <a href="https://bitbucket.org/vsmirnov/memoria/wiki/Home">Memoria</a> project is to help training advanced DS skill required for PI.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=615af6104d6b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Self-referentiality and Metacognition]]></title>
            <link>https://medium.com/@victorsmirnov/self-referentiality-and-metacognition-6bf77fb35b7?source=rss-162e5c148cc6------2</link>
            <guid isPermaLink="false">https://medium.com/p/6bf77fb35b7</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[consciousness]]></category>
            <category><![CDATA[metacognition]]></category>
            <dc:creator><![CDATA[Victor Smirnov]]></dc:creator>
            <pubDate>Wed, 15 Aug 2018 03:50:22 GMT</pubDate>
            <atom:updated>2018-08-15T03:54:03.782Z</atom:updated>
            <content:encoded><![CDATA[<p>(This is my answer to some question in AI Philosophy FB group: <a href="https://www.facebook.com/groups/ai.philosophy/permalink/2028805893856707/">https://www.facebook.com/groups/ai.philosophy/permalink/2028805893856707/</a> )</p><p>Kyle Di Marzio&gt; Sorry if this is a dumbass question, but aside from an increase in complexity, (1) what does self-reference add to cognition? (2) What would be the difference between an algorithm and an algorithm observing itself?</p><p>(2). From the theoretical standpoint, self-referentiality adds nothing to expressiveness of a Turing Machine. No any special properties of self-referential computations open us the way to Strong AI by themselves.</p><p>If you are familiar with Solomonoff Induction, ALP formula has no any explicit self-referentiality, so _explicit_ S-R is not necessary for AGI in the rigorous narrow meaning. Nevertheless, S-R is may be a very important tool, and I’ll try to explain why in the following.</p><p>First of all, you may notice, that ALP formula is an weighted average over all programs generating S, and some of they will obviously be self-referential. I can’t say anything about cardinality of S-R programs set, this is probably a good topic for PhD in mathematics.</p><p>Second, for some set of strings S, self-referential programs will be among the shortest ones, hence dominating in ALP for those S, being then the best models for these strings S. Therefore, there can be possible some environments, were self-referential sub-AGI will perform very well. Regrading your question (1), what we need to know if our physical environment favors self-referential programs. I’m not a physicist, but I have some clues it actually does, at least to some measurable extent.</p><p>From the physical point of view, self-referentiality is just a feedback loop. As well as complete theoretical self-referentiality is not physically possible (because of halting problem), true causal feedback loop is not physically possible too. What we can have is partial self-referentiality (I’ll explain what it is soon) and causal feedback “emulation” through delay lines or any other physical form of memory. This aligns very well with different “consciousness is a memory” family of theories of mind (Felix Lanzalaco). (this is the first clue)</p><p>Partial self-referentiality concept can be best explained via “logging concept” of modern software systems. To become self-referential, a TM can store some of it’s current state on the tape in a form of historically-aligned records, and later read these records, and act on them. Any kind of current state can be stored this way. For example, performance metrics, externally observed own behavior, etc. What is obvious, that full current state of the TM can’t be stored this way. Even if is physically can, most of it will never be read again. So, self-referentiality can be only partial.</p><p>Another clue is that physical computational substrate is probabilistic in nature and hence requiring some “error correction” to operate efficiently. The larger computer is, the more computational resource _must_ be dedicated to this error correction. In order to detect and correct errors, such computational system must have “self-model” in some form. As you may notice, all present day distributed systems have pretty sophisticated logging facilities. And yes, I, personally, think that one day they may “spontaneously” become self-aware this way.</p><p>So, necessity to dedicate some pretty large portion of computational budget to error-correction based of self-models is the reason why self-referentiality is _essential_ for any advanced AI in our physical environment.</p><p>This is more-or-less rigorous answer to your question (1), what self-reference add to cognition: it make it physically possible :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6bf77fb35b7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Applied Mind-Body Problem for AI Developers (Emotions vs Feelings)]]></title>
            <link>https://medium.com/@victorsmirnov/the-applied-mind-body-problem-for-ai-developers-emotions-vs-feelings-b6e59f4647f3?source=rss-162e5c148cc6------2</link>
            <guid isPermaLink="false">https://medium.com/p/b6e59f4647f3</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <category><![CDATA[emotional-intelligence]]></category>
            <dc:creator><![CDATA[Victor Smirnov]]></dc:creator>
            <pubDate>Wed, 18 Jan 2017 11:52:37 GMT</pubDate>
            <atom:updated>2017-01-18T11:54:44.950Z</atom:updated>
            <content:encoded><![CDATA[<p>Psychological terminology of emotions is very ambiguous, intermixing higher and lower components in a simple phenomena. Higher-level emotions are actually complex emotional states, describing by complicated superposition of other emotions and needs. What is very bad, emotions are often used instead of (in place of) feelings. “To experience emotions” is not a nonsense, but it means self-perception of various <strong>behavioral signs of emotions</strong>. Phenomenal representation of emotion is <strong>feeling</strong>. We experience feelings, and behavioural signs of emotions but in the form of <strong>sensations</strong>.</p><p>With the following text I propose simplified but complete functional model of emotions and some preliminary mapping of this model to subjective phenomenology. Feel free to extends it if you think you have something to contribute.</p><p><strong>DEFINITION OF EMOTION</strong></p><p><strong>Need</strong>. Some objectively measurable state of the organism. For instance, low oxygen level in the blood, that is required to be increased.</p><p>Needs may be physiological (like low oxygen level), purely psychological like shortage of novelty and combined. Needs may be combined to create new needs. The most generic form of a need is an algorithm, where the most basic operators are “physiological needs”.</p><p>A need has system of conditions defining the “level of satisfiability” of the need. This system may be very complicated especially for psychological needs. Each time there ma many different active needs. Each event may satisfy many needs simultaneously, but with different level.</p><p>The set of active needs is a <strong>portrait of needs</strong> (there may be better term for that but I’m not an English speaker)</p><p>Phenomenal representation of a need is <strong>motivation</strong>. Like needs, motivations are highly complicated, interconnected, constructions. Actually, there are N-to-M mapping between motivations and needs because if cognitive biases and partial self-awareness in general. The simplest phenomenal form of motivation is sense of anxiety. You are experiencing it when you haven’t made an inhalation or exhalation for a long time. Note that you can easily find this element in any high-level motivation.</p><p><strong>Emotion</strong> is a sign that some event may satisfy some need at some level. Or may, from other side, suppress it. The level of satisfaction can be defined though the number of conditions met. This value may be positive or negative. In case of many active needs one even may satisfy one subset of needs and suppress others. <strong>Attention</strong> is used to activate/deactivate needs to maximize satisfaction and minimize suppression. This may also imply various possible orders of application the event against needs.</p><p>Emotions are additive. Many simple emotions can be added/subtracted to create unified reinforcement. This combining is also selective function of Attention.</p><p>So in terms of Reinforcement Learning, Emotion is reinforcement and Need is a target function. By I, personally, don’t like RL. Because I doesn’t explain well many essential properties on emotional subsystem.</p><p>Emotional subsystem of the human brain is universal AI by itself because it approximates the famous Algorithmic Probability (ALP) formula: <a href="http://l.facebook.com/?u=http%3A%2F%2Fworld.std.com%2F%7Erjs%2Fformula.jpg&amp;e=ATPpJIR-0Cs9EDzY60uyr3qi8MD6PTnbtzFPmJGWcS4lnvZykSRsqiYWXhyGHG0q6Y_Wv0-xS1SU7U-CkV7-pUhrReIHZrzl_OkRzxIZpMOcR-30_5CeRq8rtgVmIQ">http://world.std.com/~rjs/formula.jpg</a></p><p>ALP may be used for universal sequence prediction in a very simple way. Here, Pm(X) is <strong>guiding function</strong> of emotion for the event X. Emotions guide both attention and execution systems of the brain. s_i(X) is an elementary emotional response (elementary emotion) on the event X against the <strong>portrait of needs</strong>. 2^{|x|} is weighting function of emotions that is here universal, but in may be also defined via needs ordering (more important needs, less important needs). Note that all these pragmatic ordering schemes may be reduced (and, actually, are reducing) to the universal weighting.</p><p>This model of emotions is oversimplified. Schmidhuber created a bit more complicated scheme for curiosity that includes dynamic aspect explaining emotional decay in a universal way. The model I propose is compatible with Schmidhuber’s one. It’s main purpose is to emphasize <strong>ALP</strong> as “<strong>the formula of emotions</strong>”. It explains main cognitive function of emotions: universal heuristic guiding. No other theories of emotions, except Schmidhuber’s Artificial Curiosity, have such property. But curiosity is a feeling, that requires also model of experience (consciousness) to be defined. We will try to define it later.</p><p><strong>PRACTICAL ASPECTS OF ALP</strong></p><p>ALP-bases approaches to AI are usually criticized for being completely impractical and even uncomputable. It’s not true. ALP is well approximiable. If we take only the shortest program (model of environment) under the Sigma operator, we get MDL principle. With many models prediction given by ALP becomes only better. ALP is a universal approach especially efficient for complex environments, where all other specialized heuristic-based decision methods don’t work.</p><p>It’s true that ALP is computationally very expensive (yet feasible). But its full power is only required in a relatively small number of cases, when no good model of environment is available. Then we have to use mixture of suboptimal models. If such good model is known, then ALP may be reduced to computationally much more tractable MDL. ALP efficiently fills the gap between power of rational intelligence to operate in structured environments (MDL) and power of emotional intelligence to operate in unstructured ones (ALP).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b6e59f4647f3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Compensate Introspection Illusion]]></title>
            <link>https://medium.com/@victorsmirnov/how-to-compensate-introspection-illusion-62f357e9326c?source=rss-162e5c148cc6------2</link>
            <guid isPermaLink="false">https://medium.com/p/62f357e9326c</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cognitive-bias]]></category>
            <category><![CDATA[introspection]]></category>
            <dc:creator><![CDATA[Victor Smirnov]]></dc:creator>
            <pubDate>Wed, 18 Jan 2017 11:50:59 GMT</pubDate>
            <atom:updated>2017-01-18T11:50:59.323Z</atom:updated>
            <content:encoded><![CDATA[<p>There is a complicated but feasible way how to do it. We have so-called process introspection ability which allows us to “see” hidden machinery of our consciousness or to get access to it (to be aware of it). But the problem is our process introspection produces confabulations or delirium (see Wikipedia’s article for details). At best this delirium is very vague and unstructured, but sometimes introspection produces very structured but false partial models with emotional attachment. People are emotionally attached to their false introspections. This is the psychological reason why there are so many sustainable but false views on mind.</p><p>Fortunately there is a way to compensate cognitive bias of introspection illusion. This compensation is quite a long process so one should not expect quick results here. As I said before, quick results are always false.</p><p>The reason why our process introspection produces garbage is because our personal mind theories (top-level self-views) are inherently broken. So if we provide someone with simple initial but true causal self-model, she will eventually evolve it into complete causal explanation of her mind with algorithmic quality. This process is usually quite long (it takes tens of years) but it can be gradually accelerated by providing external information about true self-models (AGI theories, CogSci results, etc).</p><p>So basic iterative algorithm of introspection illusion compensation looks like the following:</p><p>1. Initial true self-model.<br>2. Interiorize it (learn to see yourself through this model).<br>3. Try to explain this model in terms of AGI (or any formal external processes).<br>4. Come to confusion when you realize it is much more complicated than expected.<br>5. Read scientific sources on the topic (self-models, consciousness, theories of emotions etc) and try to reformulate things you have problems with at the step (3).</p><p>6. Try to socialize your model to validate against self-models of others (Dennett’s cross-phenomenology).</p><p>7. Go to step (2).</p><p>It works :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=62f357e9326c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>