AI In Scientific Research

AI adoption is one thing, transformation is something else entirely.    Few companies have been bold enough to fundamentally change how they are organized and how decisions are made. At Pfizer, we are determined to break through that inertia and be the most AI-forward company in our industry because of what this technology will do for patients.    In R&D, as I said on our earnings call this week, our ambition is to build an AI-native organization where every insight, from target discovery through medical evidence, continuously informs the next decision.    We are making great progress, so I'd like to share more about how we think about it overall: why our 177 years of data is our alpha, why I chose a federated model over a central team, and why we are certifying every eligible colleague, including me, in AI fluency.

Loop engineering is replacing yourself as the person who prompts the agent. You design the system that does it instead. My latest free deep dive: https://lnkd.in/gV_b9AsR ✍ A loop can be thought of a recursive goal where you define a purpose and the AI iterates until complete. I believe this may be the future of how we work with coding agents. However, its still early and you absolutely have to be careful about token costs. For two years, the way you got something out of a coding agent was simple: write a good prompt, read what came back, write the next prompt. You held the tool the whole time, one turn after another. That part is changing. Peter Steinberger of OpenClaw put it like this: "You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." Boris Cherny, head of Claude Code, said the same thing differently: "I don't prompt Claude anymore. I have loops running that prompt Claude." A loop is five things: → Automations - a heartbeat that runs discovery on a schedule so you're not the one going around checking → Worktrees - isolated branches so parallel agents don't collide on the same files → Skills - project knowledge written down once so the agent doesn't re-derive your conventions from scratch every run → Plugins/connectors - MCP connections to your real tools, so the loop opens PRs and updates tickets instead of just telling you what it would do if it could → Sub-agents - one agent writes, a different agent checks. The one who wrote the code is too nice grading its own homework. Plus one more thing that sounds too simple to matter: a memory file. Markdown, a Linear board, anything that lives outside the conversation. The agent forgets between runs. The repo doesn't. What surprised me is this isn't a bespoke scripting problem anymore. A year ago building a loop meant a pile of bash only you could maintain. Now the pieces ship inside the products. Claude Code and Codex both have all five. But three problems get sharper as loops get better, not easier: The loop changes the work, it does not delete you from it. And three problems actually get sharper as the loop gets better, not easier. Verification is still on you. A loop running unattended is also a loop making mistakes unattended. Your understanding still rots if you allow it. The faster the loop ships code you did not write, the bigger the gap between what exists and what you actually get. Thats comprehension debt and a smooth loop just makes it grow faster unless you read what the loop made. And the comfortable posture is the dangerous one. When the loop runs itself its very tempting to stop having an opinion and just take whatever it gives back. I called that cognitive surrender. Designing the loop is the cure when you do it with judgement. Build the loop. But build it like someone who intends to stay the engineer, not just the person who presses go. #ai #softwareengineering #programming

Coding agents are accelerating different types of software work to different degrees. When we architect teams, understanding these distinctions helps us to have realistic expectations. Listing functions from most accelerated to least, my order is: frontend development, backend, infrastructure, and research. Frontend development — say, building a web page to serve descriptions of products for an ecommerce site — is dramatically sped up because coding agents are fluent in popular frontend languages like TypeScript and JavaScript and frameworks like React and Angular. Additionally, by examining what they have built by operating a web browser, coding agents are now very good at closing the loop and iterating on their own implementations. Granted, LLMs today are still weak at visual design, but given a design (or if a polished design isn’t important), the implementation is fast! Backend development — say, building APIs to respond to queries requesting product data — is harder. It takes more work by human developers to steer modern models to think through corner cases that might lead to subtle bugs or security flaws. Further, a backend bug can lead to non-intuitive downstream effects like a corrupted database that occasionally returns incorrect results, which can be harder to debug than a typical frontend bug. Finally, although database migrations can be easier with coding agents, they’re still hard and need to be handled carefully to prevent data loss. While backend development is much faster with coding agents, they accelerate it less, and skilled developers still design and implement far better backends than inexperienced ones who use coding agents. Infrastructure. Agents are even less effective in tasks like scaling an ecommerce site to 10K active uses while maintaining 99.99% reliability. LLMs' knowledge is still relatively limited with respect to infrastructure and the complex tradeoffs good engineers must make, so I rarely trust them for critical infra decisions. Building good infrastructure often requires a period of testing and experimentation, and coding agents can help with that, but ultimately that’s a significant bottleneck where fast AI coding does not help much. Lastly, finding infrastructure bugs — say, a subtle network misconfiguration — can be incredibly difficult and requires deep engineering expertise. Thus, I’ve found that coding agents accelerate critical infrastructure even less than backend development. Research. Coding agents accelerate research work even less. Research involves thinking through new ideas, formulating hypotheses, running experiments, interpreting them to potentially modify the hypotheses, and iterating until we reach conclusions. Coding agents can speed up the pace at which we can write research code. (I also use coding agents to help me orchestrate and keep track of experiments.) [Truncated for length; full text: https://lnkd.in/gCnqy_4e ]

Everyone's asking: Will AI kill India's GCC advantage? We went and found out. We used a framework developed by Prof. Ashish Nanda and Prof. Das Narayandas, building on David Maister's original work, that maps work across four stages: Rocket Science (cutting-edge R&D creating future opportunities), Grey Hair (complex problem solving drawing on deep experience), Procedures (skilled execution of well-understood processes), and Commodities(standardized, repeatable tasks with minimal variation). The AI disruption gradient runs in exactly that order — from low/enabler at the top to existential threat at the bottom. By analyzing over 1.7 million job descriptions with the help of Draup and thousands of job postings across GCC locations and letting AI score each role against this framework, we quantified real commodity automation exposure by geography. The result? India sits at 17.7% commodity automation risk — closer to HQ locations (13.4%) than to peers like the Philippines (40.1%) or Costa Rica (41.1%). This is promising. But it is not a reason to relax. Knowledge diffusion is relentless as work naturally flows left to right on this framework. What is Rocket Science today becomes Grey Hair tomorrow, Procedure next year, and Commodity soon after. If a GCC stands still, its work mix quietly degrades underneath it. The automation exposure doesn't stay at 17.7%. It drifts higher, invisibly, until it doesn't. Like a shark, a GCC has to keep moving — or die. The real question for every GCC leader isn't "Are we safe from automation today?" It's "How fast are we moving up the value chain — and is it faster than knowledge diffusion is pulling us down?" Zinnov

While waiting for DeepSeek V4 we got two very strong open-weight LLMs from India yesterday. There are two size flavors, Sarvam 30B and Sarvam 105B model (both reasoning models). Interestingly, the smaller 30B model uses “classic” Grouped Query Attention (GQA), whereas the larger 105B variant switched to DeepSeek-style Multi-Head Latent Attention (MLA). As I wrote about in my analyses before, both are popular attention variants to reduce KV cache size (the longer the context, the more you save compared to regular attention). MLA is more complicated to implement, but it can give you better modeling performance if we go by the ablation studies in the 2024 DeepSeek V2 paper (as far as I know, this is still the most recent apples-to-apples comparison). Speaking of modeling performance, the 105B model is on par with LLMs of similar size: gpt-oss 120B and Qwen3-Next (80B). Sarvam is better on some tasks and worse on others, but roughly the same on average. It’s not the strongest coder in SWE-Bench Verified terms, but it is surprisingly good at agentic reasoning and task completion (Tau2). It’s even better than Deepseek R1 0528. Considering the smaller Sarvam 30B, the perhaps most comparable model to the 30B model is Nemotron 3 Nano 30B, which is slightly ahead in coding per SWE-Bench Verified and agentic reasoning (Tau2) but slightly worse in some other aspects (Live Code Bench v6, BrowseComp). Unfortunately, Qwen3-30B-A3B is missing in the benchmarks, which is, as far as I know, is the most popular model of that size class. Interestingly, though, the Sarvam team compared their 30B model to Qwen3-30B-A3B on a computational performance analysis, where they found that Sarvam gets 20-40% more tokens/sec throughput compared to Qwen3 due to code and kernel optimizations. Anyways, one thing that is not captured by the benchmarks above is Sarvam’s good performance on Indian languages. According to a judge model, the Sarvam team found that their model is preferred 90% of the time compared to others when it comes to Indian texts. (Since they built and trained the tokenizer from scratch as well, Sarvam also comes with a 4 times higher token efficiency on Indian languages.

Today in Cell, we published new research showing how AI can help accelerate cancer discovery. With GigaTIME, we can now simulate spatial proteomics from routine pathology slides, enabling population-scale analysis of tumor microenvironments across dozens of cancer types and hundreds of subtypes.   Developed in partnership with Providence and the University of Washington, our hope is that this work helps scientists move faster from data to insight, revealing new links between genetic mutations, immune activity, and clinical outcomes, and ultimately improving health for people everywhere. https://lnkd.in/dSpPdtzz

I’ve worked on AI my whole life because I’ve always believed it could unlock the ability to answer some of the biggest and most intractable problems in science. Our first big science breakthrough happened five years ago when we announced our solution to the protein structure prediction problem: AlphaFold 2. It has been incredible to see its impact since then. More than 3 million researchers across 190 countries have used this tool for disease understanding, drug discovery and more. And it was an honour of a lifetime for our work to be recognised last year with a Nobel Prize. One of our greatest ambitions is for AI to aid in accelerating drug design and help cure all diseases. This is what led me to found Isomorphic Labs, which is already making amazing progress. We’ve also expanded AlphaFold to predict the interactions of all of life’s molecules. But AlphaFold represents more than a solution to a biological puzzle. It demonstrated how AI can crack ‘root node’ problems - where a single breakthrough unlocks entire new avenues of research. It is a critical step towards a long-held dream of mine: building a virtual cell. Imagine running ‘in silico’ experiments orders of magnitude faster than in a wet lab. Scientists could rapidly test hypotheses, model complex pathways and see how a drug affects a cell. It would be an incredible boon not only for fundamental biology but also for medicine. Although for me, AlphaFold was never just about biology. It was the first major proof point for a much larger thesis: that AI could be the ultimate tool for advancing science. By processing data or helping us come up with new hypotheses, I think AI will help us tackle some of humanity’s greatest challenges and answer fundamental questions about the universe. From materials design to fusion energy to mathematics, I believe we’re on the cusp of a new golden age of discovery. We’re just getting started. Read more about AlphaFold’s impact: https://lnkd.in/eNeqxqQp

Our team at Google DeepMind has been collaborating with Terence Tao and Javier Gómez-Serrano to use our AI agents (AlphaEvolve, AlphaProof, & Gemini Deep Think) for advancing Maths research. They find that AlphaEvolve can help discover new results across a range of problems. As a compelling example, they used AlphaEvolve to discover a new construction for the finite field Kakeya conjecture; Gemini Deep Think then proved it correct and AlphaProof formalized that proof in Lean. AI-powered Maths research in action! Paper: https://lnkd.in/eSjggApg

It is interesting that the new DeepSeek AI v3.1 talks about the UE8M0 FP8 data format, which is nothing but the logarithmic number system (LNS), meaning it has only exponent and no mantissa. While DeepSeek v 3.1 didn't entirely train on that format, we have a multiplicative weights update (Madam) for training entirely in LNS format that was done several years ago while at NVIDIA It yields maximum hardware efficiency with no accuracy loss https://lnkd.in/gu3V4BN8 Logarithmic number system achieves a higher computational efficiency by transforming expensive multiplication operations in the network layers to inexpensive additions in their logarithmic representations. In addition, it attains a wide dynamic range and can provide a good approximation. Also, logarithmic number system is biologically inspired, and there is evidence that our brains use such a format for storage. However, using standard SGD or Adam optimization for training in logarithmic format is challenging, and requires intermediate updates and optimization states to be stored in full precision (FP32). To overcome this, we proposed Multiple Weights update (Madam) that instead updates directly in the logarithmic format and leads to good training outcomes. Our LNS-Madam when compared to training in FP32 and FP8 formats, LNS-Madam reduces the energy consumption by over 90% and 55%, respectively, while maintaining accuracy.

At the weekend, the The Wall Street Journal published a feature on McKinsey & Company’s AI transformation, presenting the firm’s adoption of new tools as bold and forward-thinking. But look more closely, and the picture is far more defensive than disruptive. The article highlights McKinsey’s deployment of 12,000 AI agents and a move to outcomes-based pricing, now covering around 25% of its work. Framed as innovation, this reads more like a late-stage response to structural pressure: a quiet pivot away from the old playbook of long, people-heavy engagements. What the article doesn’t contextualize is how fundamentally the Consulting model is being rewritten. Graduate hiring is collapsing as delivery teams become leaner and AI-fluent. Modular teams built around productised IP, outcome-based pricing, and nearshore hubs are replacing the traditional pyramid. Many of the challenger firms we’ve benchmarked are much further ahead, already embedding sector-specific AI solutions into their core offerings and building recurring revenue streams from subscriptions and managed services. The real transformation is happening not in how firms decorate the old model with AI tools, but in how they replace it. That means every manager leading blended teams of humans and machines. It means turning proprietary tools into licensable products. It means capturing and recycling internal knowledge to create an “insight flywheel” that scales without adding headcount. It’s not about bots that write in your tone of voice, it’s about whether your firm can deliver faster, more repeatable outcomes without relying on brute force. McKinsey has brand strength and institutional capital, no question. But the Consulting firms winning in this new era aren’t just experimenting with AI, they’re building businesses around it. The real question now isn’t whether AI will reshape consulting, that’s already underway. It’s who will have the conviction to redesign their operating model fast enough to lead the next era.