<?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 Prathibha Bharathi Alam on Medium]]></title>
        <description><![CDATA[Stories by Prathibha Bharathi Alam on Medium]]></description>
        <link>https://medium.com/@prathibhabharathialam?source=rss-a68f2febe208------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*vc_yvn4oCs2qHXUeoa8NOw.png</url>
            <title>Stories by Prathibha Bharathi Alam on Medium</title>
            <link>https://medium.com/@prathibhabharathialam?source=rss-a68f2febe208------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 06:01:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@prathibhabharathialam/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[What Spock, Data, and Sheldon Teach Us About Leadership]]></title>
            <link>https://medium.com/@prathibhabharathialam/what-spock-data-and-sheldon-teach-us-about-leadership-364ed2922ebf?source=rss-a68f2febe208------2</link>
            <guid isPermaLink="false">https://medium.com/p/364ed2922ebf</guid>
            <dc:creator><![CDATA[Prathibha Bharathi Alam]]></dc:creator>
            <pubDate>Tue, 19 May 2026 04:03:35 GMT</pubDate>
            <atom:updated>2026-05-24T20:38:36.223Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mFMkMlNqcIIstN_WSawwrA.png" /></figure><p>Our family has a routine of watching an episode of Star Trek every week, and every week I find myself wondering what makes Spock, Data, and Sheldon Cooper so indispensable despite their glaring social flaws.</p><p>Though they span different shows and eras, they all represent a version of the same archetype: the Hyper-Rational Anchor. What defines this archetype is a mind so thoroughly optimized for internal consistency that social performance becomes a low-priority process.</p><p>What makes them so compelling is relentless consistency. When the world around them descends into emotional chaos, alien attacks, or social drama, these characters do not waver. They process the world through a fixed, explicit framework. Spock has Vulcan logic, Data has his ethical subroutines, and Sheldon has his Roommate Agreement and rigid daily schedules.</p><p>We confuse good leadership with being liked, as if teams need endless charisma and warmth. But in a chaotic, fast-moving world, what people actually need is solid ground.</p><p>The mechanics of how these characters operate show us how trust and alignment actually get built.</p><p><strong>Predictability builds psychological safety</strong></p><p>Sheldon Cooper is an exaggerated, comedic version of this archetype. There is a specific kind of relief in someone who never surprises you. You don’t spend energy anticipating him. You don’t manage him. His reactions are always derivable from first principles and that predictability, in a world full of people performing moods, is extraordinary. Leonard knows exactly how Sheldon will react to any given situation. Spock and Data earn the deepest trust from their respective crews for the same reason: they are incorruptible.</p><p>For leaders, consistency is more important than charisma. If your team has to guess what mood you are in before they bring you bad news, they will start hiding the bad news. When your reactions are consistent, your team spends less energy managing emotions and more energy solving problems.</p><p><strong>Make your operating framework explicit</strong></p><p>Most interpersonal friction on teams happens because of unstated expectations. People assume others know how they want to work.</p><p>These characters never leave their frameworks implicit. Sheldon literally writes them down in binding contracts, while Data and Spock simply state their operational parameters out loud. Great leaders do something similar. They write down how they operate so nobody has to guess. When the rules of engagement are visible to everyone, you eliminate the politics of guessing what the boss wants.</p><p><strong>Separate ego from the data</strong></p><p>When Data proposes a solution and Captain Picard points out a flaw in it, Data does not get defensive. He instantly recalculates based on the new variable. Spock will calmly state, “I was in error,” without a shred of embarrassment.</p><p>This is one of the hardest skills for a human leader to master because we tie our ideas to our identities. The Hyper-Rational Anchor treats being wrong as a simple data update. A leader who can accept contradictory evidence without bruising their ego creates a culture where the best idea wins, regardless of who suggested it.</p><p><strong>Pure logic requires a human counterpart</strong></p><p>None of these characters are the Captain. But take them out of the show and there is no show. The title sits with someone else. The story moves because of them. I have always believed that people who operate with genuine agency and ownership are leaders regardless of where they sit in a hierarchy. Spock and Data are proof of that. They need the intuition and moral philosophy of their captains, just like Sheldon relies on Leonard to translate social cues. Pure logic is brittle because it cannot account for the messy, irrational realities of human behavior.</p><p>It’s also a lesson in team design. If you are a highly analytical, data-driven leader, you need to intentionally surround yourself with empathetic, intuitive operators. If you are a visionary, emotional leader, you need a Spock in your inner circle to ground your ideas in reality. The deeper lesson these characters offer is also about what a fully committed mind looks like, one that operates from first principles, updates without ego, and never needs to perform certainty because it has actually earned it.</p><p>I keep coming back to these characters not because they are aspirational in the conventional sense. They are not warm or easy or likable on first meeting. But there is something in how they operate that I recognize and admire. A mind that knows how it works. That doesn’t perform. That updates without drama. I’m not sure that’s a leadership lesson so much as it is a way of being that I find genuinely beautiful.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=364ed2922ebf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fine-Tuning in the Era of Strong Foundation Models]]></title>
            <link>https://medium.com/@prathibhabharathialam/fine-tuning-in-the-era-of-strong-foundation-models-e9fe2d672723?source=rss-a68f2febe208------2</link>
            <guid isPermaLink="false">https://medium.com/p/e9fe2d672723</guid>
            <dc:creator><![CDATA[Prathibha Bharathi Alam]]></dc:creator>
            <pubDate>Sun, 22 Feb 2026 22:09:31 GMT</pubDate>
            <atom:updated>2026-02-22T22:14:53.158Z</atom:updated>
            <content:encoded><![CDATA[<p>In the past few years, language models have become much stronger out of the box, mostly because of instruction tuning and alignment.</p><p>Today’s large models are not just trained to predict the next token. After pretraining, they are further trained on thousands of tasks framed as natural language instructions. They see prompts like summarize this, classify this, extract entities, answer this question, explain your reasoning. Alignment training then nudges them toward responses that humans actually prefer, such as clear structure, direct answers, and fewer obvious hallucinations.</p><p>When you combine large-scale pretraining with instruction tuning and alignment, you end up with a model that already behaves like a general-purpose assistant. You can give it a task description it has never explicitly seen before, and it often produces something sensible. The zero-shot baseline is strong because the model has already been exposed to a wide range of task formats during training.</p><p>In computer vision, the workflow has historically looked different. You pretrain on ImageNet and then fine-tune on your dataset. Even with stronger modern architectures like Vision Transformers or CLIP, adapting to a specific dataset remains common. In video, models such as VideoMAE or SlowFast are typically adapted to the downstream benchmark. Pretraining gives useful representations, but strong task-specific performance usually requires additional training.</p><p>In language, instruction tuning moves part of that adaptation earlier in the pipeline. The model has already learned to interpret tasks as instructions. For many general use cases, especially those that resemble common NLP benchmarks, prompting alone can take you far. In this context, fine-tuning becomes less of a default step and more of a deliberate choice.</p><p>When you start working in narrower domains, especially areas like healthcare or legal work, you quickly notice where the generalist model starts to fall short. For something like clinical summarization, getting the terminology right is often more important than producing something that simply reads well. In finance or compliance workflows, you cannot tolerate hallucination or format drift. And, in many of these environments, sending sensitive data to an external API is simply not an option.</p><p>There is also a cost dimension. Large aligned models accessed through APIs are powerful, but inference costs accumulate at scale. Distilling behavior into a smaller model and deploying it locally can provide better control over latency, data flow, and long-term economics.</p><p>We are no longer forced to fine-tune for every task. The foundation models are strong enough that prompting and retrieval often provide a solid starting point. When domain precision, privacy constraints, cost structure, or reliability requirements demand more control, targeted fine-tuning is still an effective tool.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e9fe2d672723" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Distilling GPT-4 into BART: Cost-Efficient, Privacy-Aware Clinical Summarization]]></title>
            <link>https://medium.com/@prathibhabharathialam/distilling-gpt-4-into-bart-cost-efficient-privacy-aware-clinical-summarization-f0082716a99f?source=rss-a68f2febe208------2</link>
            <guid isPermaLink="false">https://medium.com/p/f0082716a99f</guid>
            <dc:creator><![CDATA[Prathibha Bharathi Alam]]></dc:creator>
            <pubDate>Sun, 22 Feb 2026 20:30:54 GMT</pubDate>
            <atom:updated>2026-05-24T19:17:19.875Z</atom:updated>
            <content:encoded><![CDATA[<p>I recently ran a focused experiment on biomedical summarization using the EBM-NLP dataset, which contains 4,993 PubMed abstracts annotated with PICO elements. The question I wanted to answer was: do we really need a large decoder-only model for high-quality summarization, or can a smaller encoder–decoder model perform just as well with targeted fine-tuning?</p><p>I began by evaluating several off-the-shelf summarization models. DistilBART-CNN produced fluent summaries, but they lacked clinical depth and did not consistently preserve the PICO structure. After benchmarking multiple candidates, including Pegasus-PubMed, Pegasus-XSum, Pegasus-Large, and long-input models such as BigBird-Pegasus, facebook/bart-large-cnn stood out on qualitative inspection of a sample of outputs. Its summaries were more structured and better retained clinically relevant details compared to the other candidates. This selection was manual rather than metric-driven, which is a limitation worth noting.</p><p>Rather than relying on a large model at inference time, I used GPT-4 via the OpenAI API where it adds the most value: generating high-quality supervision. I created structured, PICO-aligned gold summaries for exactly 100 documents from the dataset. Because API-generated labels are expensive, the dataset was intentionally constrained. With only 100 examples, every summary had to be precise. There was no room for noisy supervision.</p><p>I then fine-tuned facebook/bart-large-cnn using HuggingFace’s Seq2SeqTrainer. Conceptually, this resembles a student–teacher setup, though with an important constraint: because GPT-4’s logits are not accessible via the API, the transfer happens at the output text level rather than through soft probability distributions. This is more precisely <em>black-box distillation</em>. GPT-4 acts as a labeler, and BART learns to approximate its outputs via standard sequence-to-sequence loss. Once fine-tuned, the smaller model runs efficiently and does not require external API calls for production use.</p><p>Evaluation was critical. I used ROUGE-1 to measure lexical overlap, but abstractive summarization requires semantic evaluation as well. I therefore used BERTScore F1, which measures contextual similarity using transformer embeddings. The results showed moderate ROUGE-1 scores (0.51–0.64) but high BERTScore F1 (0.88–0.94). BERTScore precision exceeded 0.90, with recall between 0.87 and 0.92. This indicates strong semantic alignment even when phrasing differs. In clinical summarization, semantic fidelity matters far more than surface overlap. The fine-tuned model preserved meaning while producing outputs that differed in phrasing from the source.</p><p>When comparing the baseline BART Large CNN model to the fine-tuned version, improvements were observed across all BERTScore metrics. Precision, recall, and F1 all increased after fine-tuning on GPT-4-generated summaries. What surprised me was how effectively that signal transferred into a much smaller encoder–decoder model with only 100 labeled examples.</p><p>This experiment reinforced an architectural point that often gets overlooked. Decoder-only models are exceptional generalists. They are trained to predict the next token and scale extremely well. But summarization is fundamentally a transformation task. It requires reading the entire source, forming a representation, and generating a compressed version that remains faithful to that source. Encoder–decoder models like BART are structurally aligned with that objective. The encoder processes the full input bidirectionally, and the decoder generates conditioned on that representation. That separation can provide stronger grounding for summarization tasks.</p><p>There is also a practical advantage in privacy-sensitive domains such as healthcare. Fine-tuning a strong encoder–decoder model allows inference to run locally. Once trained, production systems do not need to transmit protected health information to external APIs for each request. This reduces compliance risk and lowers recurring costs. A large model can still be used selectively for label generation or evaluation, but inference can remain within a controlled environment.</p><p>The broader takeaway is that smaller models are not relics. They are specialized tools. When architecture aligns with task and supervision is high quality, moderately sized encoder–decoder models can produce strong, domain-specific summaries. I used GPT-4 sparingly to generate high-quality training labels, then distilled that capability into a more efficient model that performs well without ongoing API costs. Scale is powerful, but it is not the only path to performance.</p><p>A note on scope: this was an exploratory experiment rather than a rigorous research study. A comprehensive analysis would include statistical significance testing, clinician-led human evaluation, detailed cost–performance comparisons, and a larger training set. The 100-example constraint was driven by API costs. However, the central insight remains: encoder–decoder architectures with targeted fine-tuning can deliver strong results on specialized summarization tasks without the cost and operational complexity of large decoder-only models at inference time.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0082716a99f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Modern Models Reason]]></title>
            <link>https://medium.com/@prathibhabharathialam/how-modern-models-reason-61830cf702f6?source=rss-a68f2febe208------2</link>
            <guid isPermaLink="false">https://medium.com/p/61830cf702f6</guid>
            <category><![CDATA[reasoning]]></category>
            <category><![CDATA[generative-model]]></category>
            <category><![CDATA[llm-reasoning]]></category>
            <dc:creator><![CDATA[Prathibha Bharathi Alam]]></dc:creator>
            <pubDate>Sun, 09 Nov 2025 22:52:29 GMT</pubDate>
            <atom:updated>2025-11-10T05:32:11.505Z</atom:updated>
            <content:encoded><![CDATA[<p>When people see AI systems solving complex problems such as writing code, analyzing data, or explaining probability, it is easy to assume they are thinking. They are not, at least not in the human sense. What they are doing is closer to generative reasoning, a hybrid process that bridges pattern recognition and deliberate problem solving.</p><p>At their core, these models are autoregressive language models. They learn to predict the next token in a sequence based on everything that came before it. With scale, structure, and fine-tuning, that simple mechanism begins to capture richer patterns that extend beyond language into the shape of reasoning itself. When trained on vast corpora of world knowledge across text, images, and video, a model observes how humans describe, decompose, and solve problems across domains. It learns the generative processes of problem solving: how a line of reasoning unfolds, how an explanation builds, and how cause and effect connect across disciplines. Over time, it forms internal representations that reflect the statistical regularities of how humans reason and communicate ideas step by step.</p><p>In more advanced systems, this generative reasoning is augmented by tools. When asked to perform a computation, the model can write code, execute it in a separate environment, and fold the results back into its reasoning. This approach, sometimes called tool-augmented modeling, mirrors how humans think, calculate, and interpret.</p><p>While the model itself remains a probabilistic generator, the surrounding system behaves like a reasoning stack: a layered architecture that blends generative understanding with symbolic execution. The base model provides intuition and linguistic coherence, while the tools provide precision and grounding. Together, they create a process that approximates reasoning by alternating between generation and verification.</p><p>When such a system solves a multi-step problem, it generates hypotheses, tests them through execution, and refines its output based on results, all within a probabilistic framework. It does not think as humans do, but it is learning the statistical patterns of how we reason, validate, and build understanding.</p><p>These systems represent a computational realization of long-standing ideas about how structure emerges from statistical regularities. Probability encodes preference, and learning discovers those preferences from data. What has changed is scale. We now have machines that can explore probabilistic structures at speeds far beyond human cognition, built upon the principles that information theory and statistical physics first established.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=61830cf702f6" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>