<?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 Vini Bharat Bansode on Medium]]></title>
        <description><![CDATA[Stories by Vini Bharat Bansode on Medium]]></description>
        <link>https://medium.com/@vini.369.31052006?source=rss-46ae62b66e48------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*q986c-03rizUTeu_kD9Xmg.jpeg</url>
            <title>Stories by Vini Bharat Bansode on Medium</title>
            <link>https://medium.com/@vini.369.31052006?source=rss-46ae62b66e48------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 13:30:41 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@vini.369.31052006/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[Are We Teaching AI to Think Beyond Us, or Just to Think Like Us?]]></title>
            <link>https://medium.com/@vini.369.31052006/are-we-teaching-ai-to-think-beyond-us-or-just-to-think-like-us-416c91265b21?source=rss-46ae62b66e48------2</link>
            <guid isPermaLink="false">https://medium.com/p/416c91265b21</guid>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[alphagenome]]></category>
            <category><![CDATA[alphafold]]></category>
            <category><![CDATA[ai-in-drug-discovery]]></category>
            <dc:creator><![CDATA[Vini Bharat Bansode]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 08:02:04 GMT</pubDate>
            <atom:updated>2026-11-15T08:02:04.444Z</atom:updated>
            <content:encoded><![CDATA[<h4>Perhaps the biggest limitation of artificial intelligence isn’t its intelligence. It’s the world we have given it to learn from.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rG9T3WYhutjzO0zY.png" /></figure><p>I’ve been spending some time watching podcasts on the Google DeepMind channel, and a question has been bothering me ever since.</p><p>What if we’re making the same mistake with artificial intelligence that humans have been making for centuries?</p><p>Humans are not objective thinkers. We carry biases, consciously and unconsciously. Our education, geography, culture, experiences and the people we look up to shape what we consider worth exploring. Sometimes, we dismiss an idea before even giving it a chance, simply because someone we consider more knowledgeable has already dismissed it.</p><p>Think about how scientific recognition works. An idea coming from a prestigious institution can receive attention that an equally interesting idea from a less recognised institution might never get. Researchers from less represented countries and institutions can face barriers to visibility, funding and recognition. We cannot assume that every overlooked idea was a breakthrough, but we should question how many potentially valuable ideas never received a fair examination.</p><p>Now consider what we’re building with AI.</p><p>We want machines that can explore possibilities beyond human intuition. Machines that can challenge established assumptions, identify relationships we haven’t noticed and investigate paths that we might never have considered.</p><p>But where does their knowledge come from?</p><p>From data. From research papers, experiments, publications, observations and human-generated information. In other words, much of what we teach these systems is a reflection of what humanity has already explored, documented and considered important.</p><p>This doesn’t mean AI simply inherits every human prejudice, or that it cannot discover something genuinely new. It clearly can. AlphaGo demonstrated that a machine could develop strategies that surprised even the world’s best Go players.</p><p>But there’s a distinction worth making: <strong>exploring combinations within the knowledge we’ve accumulated is not necessarily the same as discovering what we’ve never thought to investigate.</strong></p><p>And perhaps that distinction matters more than we realise.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MxF37sJAiaI-Sm_h.jpeg" /><figcaption>What we know is partly shaped by what we have chosen to investigate.</figcaption></figure><h3>Are we confusing scale with progress?</h3><p>The AI industry has become increasingly obsessed with scale. More parameters, more training data, more computing power, larger models.</p><p>Transformers have been extraordinarily successful, and dismissing their achievements would be foolish. But have we become so invested in one direction that we’re overlooking other ways of building intelligence?</p><p>A larger model is not automatically a better scientist. A more complicated architecture is not automatically a better thinker.</p><p>The interesting question isn’t whether we can build a model with more layers or feed it another trillion tokens. It’s whether the additional complexity produces a better understanding of the problem.</p><p>We should be equally willing to investigate smaller models, different architectures, alternative learning mechanisms and approaches that don’t follow the current industry’s preferred direction.</p><p>Perhaps the next breakthrough won’t come from making today’s models bigger. Perhaps it will come from questioning the assumptions behind how we’re building them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UGP4hWAuU2fqyT9p.jpeg" /><figcaption>Scale is a strategy, not a definition of intelligence.</figcaption></figure><h3>The problem with calling everything AI drug discovery</h3><p>This is where things become particularly interesting to me as a biotechnology student.</p><p>One analogy from Matt Shlosberg’s article, <a href="https://mattshlosberg.substack.com/p/ai-drug-discovery-has-a-slightly"><em>AI Drug Discovery Has a Slightly Embarrassing Problem</em></a>, stayed with me.</p><p>He compares parts of the AI drug discovery industry to building a car, attaching it behind a horse and announcing that we’ve revolutionised transportation.</p><p>The horse is still doing most of the work. We’ve just made the carriage considerably more sophisticated.</p><p>I think there’s an uncomfortable question hidden in that analogy.</p><p>What exactly are we calling a successful drug discovery model?</p><p>Today, AI can help us predict protein structures, screen enormous chemical libraries, generate new molecular structures, and estimate how strongly a molecule might bind to a particular protein.</p><p>These are significant achievements.</p><p>But let’s take a step back.</p><p>A protein is a molecule that performs a particular function in our body. Many diseases involve proteins behaving abnormally or biological pathways becoming disrupted. A drug molecule might bind to a particular protein and alter its activity.</p><p>Sounds straightforward, doesn’t it?</p><p>Except binding to a protein doesn’t necessarily mean that we’ve developed a useful drug.</p><p>A molecule might bind beautifully in a computational simulation and still fail to produce the desired effect inside a living cell. It might be toxic, get broken down too quickly, fail to reach the right tissue, or affect other proteins unintentionally.</p><p>And even if it works in cells, it might not work in animals or humans.</p><p>There is an enormous difference between designing a promising molecule and developing a medicine that is safe, effective and suitable for patients.</p><p>The distinction between ligand design and drug design is important here. A ligand is simply a molecule that interacts with a biological target. A drug must do considerably more than interact with something.</p><p>This is not to say AI has failed in drug discovery. Some AI-discovered molecules have entered clinical trials, and the technology is already contributing to research. But clinical development remains the difficult test.</p><p>Perhaps we’re celebrating the number of molecules generated when we should be asking how many predictions survive experimental validation, how many candidates demonstrate meaningful biological effects, and ultimately, how many improve patient outcomes.</p><p>Generating a million possibilities is impressive. Knowing which one will actually work is another problem entirely.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/0*eLRT1tf2L6N8TE9J.png" /><figcaption>A promising binding interaction is only the beginning of drug discovery.</figcaption></figure><h3>Biology doesn’t care how impressive your model is</h3><p>The more I study biology, the more I realise how humbling it is.</p><p>Take the human genome, for example.</p><p>The Human Genome Project gave us a reference sequence of roughly three billion DNA letters. Earlier sequencing technologies, such as Sanger sequencing, made it possible to read DNA with increasing accuracy. Later developments, including shotgun and whole-genome sequencing approaches, made it possible to sequence genomes at scales that were previously unimaginable.</p><p>But sequencing DNA and understanding what it does are two very different things.</p><p>Think of DNA as an instruction manual. Some sections contain instructions for building proteins, while other sections help regulate when, where and how much those proteins are produced.</p><p>Only around 2% of the human genome directly codes for proteins. Much of the remaining non-coding DNA participates in regulation and other biological functions, although not every non-coding region has a known function.</p><p>And the same DNA can produce very different outcomes depending on the cell.</p><p>A neuron and a liver cell carry essentially the same genetic information, yet they behave completely differently. This happens because different genes are switched on and off in different cells.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kwJUW6GSXAchR6hq.png" /><figcaption>Same genetic information. Different cellular identities. Biology is far more than just reading the sequence.</figcaption></figure><p>Even a tiny change in DNA can affect how a gene is expressed, how RNA is processed or how much of a particular protein is produced.</p><p>This is where Google’s AlphaGenome becomes interesting.</p><p>In its simplest form, AlphaGenome attempts to predict what happens when we change a letter in a DNA sequence. It can analyse long stretches of DNA and predict several molecular properties, including gene expression, transcription and RNA splicing.</p><p>Its newer iteration, AlphaGenome 2, goes further by incorporating specific cell types and states into its predictions.</p><p>This is important because a genetic variation might have different effects in different cells.</p><p>And the most fascinating part isn’t simply the model itself. It’s what researchers are beginning to do with it.</p><p>In the Google DeepMind podcast, the researchers discussed a case involving a child diagnosed with rare epilepsy. The clinical team had spent approximately four to five years investigating the genetic cause without identifying the specific responsible variant.</p><p>Using the AlphaGenome Atlas, researchers ranked the suspected genetic variants. The highest-ranked candidate pointed towards a biologically plausible explanation involving a protein associated with the condition.</p><p>They then experimentally investigated the prediction and validated an abnormal RNA-splicing effect.</p><p>For anyone unfamiliar with splicing, it is the process by which cells cut and join RNA segments before using the instructions to make proteins. If this process goes wrong, the resulting protein can be altered or dysfunctional.</p><p>The model did not magically cure the child. It helped researchers identify and investigate a previously unresolved genetic clue.</p><p>That distinction matters.</p><p>The researchers also discussed how identifying the underlying mutation could eventually open possibilities for personalised treatments, including antisense oligonucleotides, which are the short pieces of synthetic RNA designed to influence how other RNA molecules are processed.</p><p>This is the kind of AI application I find genuinely exciting: not a machine replacing biology, but a machine helping biologists ask better questions.</p><h3>Perhaps AGI is easier than understanding a cell</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*PtrORzQbLxvUnYVF.jpeg" /><figcaption>The complexity of a living cell.</figcaption></figure><p>One of the most interesting ideas discussed in the AlphaGenome 2 podcast was the possibility that we might achieve artificial general intelligence (AGI) before we achieve a convincing virtual cell.</p><p>Think about that for a moment.</p><p>We are trying to build machines capable of reasoning across different domains, solving complex problems and performing tasks that humans can perform.</p><p>Yet accurately simulating a living cell, understanding how its components interact and predicting what happens when we perturb one part of it remains an enormous challenge.</p><p>A cell isn’t simply a collection of independent molecules. It is a constantly changing system involving thousands of interacting components, feedback mechanisms, chemical reactions, environmental signals and regulatory processes.</p><p>Knowing the structure of a protein doesn’t mean knowing exactly how it behaves in every cellular environment.</p><p>Knowing the sequence of a gene doesn’t mean understanding every consequence of changing it.</p><p>And knowing that a molecule binds to a target doesn’t mean knowing what will happen to an entire organism.</p><p>This is why I think the idea of a virtual cell is so fascinating.</p><p>Imagine being able to introduce a genetic mutation into a computational cell, observe how its molecular machinery responds, predict which pathways become disrupted and test potential interventions before moving into the laboratory.</p><p>We are nowhere near a complete, reliable simulation of this kind. But getting closer could fundamentally change how we study diseases and develop personalised treatments.</p><p>It would also change what we consider progress in AI-driven biology. Instead of asking how many structures a model can generate, we could ask how accurately it predicts the behaviour of a living system.</p><h3>Maybe the real problem is what we choose to explore</h3><p>There is one more question I keep coming back to.</p><p>Are we fighting for more data because data is genuinely the missing piece, or because acquiring more data has become the easiest way to justify the next investment, the next infrastructure build and the next larger model?</p><p>Data matters enormously. AlphaGenome itself is a good example of what becomes possible when years of publicly available genomic and experimental research are brought together.</p><p>But data is not automatically knowledge. And knowledge is not automatically understanding.</p><p>If our datasets disproportionately represent certain populations, tissues, diseases or experimental conditions, our models can become better at predicting those well-represented areas while remaining less reliable elsewhere.</p><p>Even the AlphaGenome 2 researchers acknowledge that the availability of data across different cell types can influence model performance.</p><p>The solution isn’t to abandon data or stop building larger models. It’s to question what is missing, what is underrepresented and what assumptions we have quietly accepted.</p><p>Perhaps we should spend more time investigating questions that don’t attract funding because they aren’t fashionable, architectures that aren’t receiving attention because they aren’t transformers, and biological mechanisms that don’t fit neatly into the current drug discovery pipeline.</p><p>And perhaps we should stop treating a failed prediction as wasted effort. Sometimes discovering that our biological hypothesis was wrong is more valuable than generating another thousand molecules based on it.</p><p>I don’t think AI needs to think without any influence from humans. That might not even be possible, considering how we design, train and evaluate these systems.</p><p>But I do think we should build systems that can challenge the assumptions embedded in their training, expose uncertainty, explore alternative hypotheses and learn from experiments that contradict their predictions.</p><p><strong>The purpose of artificial intelligence should not be to make human thinking faster. It should be to help us discover where human thinking has been insufficient.</strong></p><p>We have spent centuries accumulating knowledge about the world. Perhaps the next challenge isn’t accumulating more of it, but learning how to question what we already believe we know.</p><p>And if AI can help us do that, I think we will have built something far more meaningful than another impressive model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*E8wEsDoFq55jUbiR.jpeg" /><figcaption>Perhaps the next breakthrough lies in asking a question nobody thought was worth asking.</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=416c91265b21" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>