Using Technology in Scientific Research

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

In 1905, Einstein published special relativity. In 1915, he published general relativity. Einstein was just trying to understand the universe. But without Einstein's math, Google Maps would be wrong by 11 kms every single day. Let me tell you why - this is very interesting :)) Your phone doesn't "talk" to GPS satellites. It only listens. Each satellite is broadcasting one thing, constantly: "I am satellite 'A', and it is currently 14:23:00.000000." Your phone receives signals from 4 satellites simultaneously. Because light travels at a known speed, tiny differences in arrival time tell it exactly how far it is from each satellite. 'A' satellite tells you: you're somewhere on a sphere of radius 20,000 km. 'B' satellite: that sphere intersects another sphere - now you're on a circle. 'C' satellite: that circle intersects a third sphere - now you're at 2 points. 'D' satellite: eliminates the last ambiguity and only one point remains. That's you! Except there's a problem nobody thought about until Einstein. The satellites are orbiting at 20,200 km altitude, moving at 14,000 km/h. Two things happen to their clocks simultaneously: - Special relativity: Moving clocks tick slower. At orbital velocity, the satellite clock loses 7.2 microseconds per day - General relativity: Clocks in weaker gravity tick faster. At that altitude, gravity is weaker. The clock gains 45.9 microseconds per day. Net effect: 45.9 - 7.2 = +38.7 microseconds per day. In 38.7 microseconds, light travels 11.6 kilometers. So without correction, the system would accumulate 11.6 km of error. Every single day. In a week, your navigation is useless. The fix is one of the most elegant things in all of engineering. Before each satellite launches, its atomic clock is physically tuned to tick slightly slower than it would on Earth - by exactly 38.7 microseconds per day. Once in orbit, relativistic effects speed it back up. And it arrives at exactly the right rate. Einstein's 1915 paper is baked into the hardware of your phone's navigation system. The next time Google Maps routes you correctly, you're experiencing general relativity. You just didn't know it.

💥 I’m starting something new inside OpenAI! It’s called OpenAI for Science, and the goal is to build the next great scientific instrument: an AI-powered platform that accelerates scientific discovery. We’ll look to hire a small team of academics that are (i) world-class in their field; (ii) completely AI-pilled; and (iii) great science communicators. Paired with a small team of researchers, we want to prove that AI models are ready to accelerate fundamental science—and accelerate research all over the world. Scientific discovery improves everything from the quality of our daily lives to national security to global GDP. Innovation is the reason the US leads the world. Few domains hold as much promise for improving lives as science. GPT-5 is a new threshold; below are four recent examples. There are many more, not to mention things like AlphaFold from our friends at GDM. 1. GPT-5 Pro was able to improve a bound in one of Sebastien Bubeck’s papers on convex optimization—by 50%, with 17 minutes of thinking. https://lnkd.in/gYqqnbaj. 2. The image below of GPT-5 outlining proofs and suggesting related extensions, from a recent hep-th paper on quantum field theory [https://lnkd.in/g5ZjDm6T]. 3. Our recent work with Retro Biosciences, where a custom model designed much-improved variants of Nobel-prize winning proteins related to stem cells. https://lnkd.in/gcB5_Px2. 4. Derya Unutmaz has been a non-stop source of examples of AI accelerating his biological research, such as https://lnkd.in/gR-Y7Jr3. I went to grad school to become a researcher in high energy physics, only to get nerdsniped into startups like so many others. I feel super fortunate now to get to combine both. I’ll simultaneously begin working with Sebastien Bubeck and the synthetic data team as an AI researcher to learn the craft. I’m grateful to him and his team for being willing to mentor me, and I’m eager to prove worth their time and effort! I’m able to do this because the product and design leaders at OpenAI are amazing, and now are complemented by Fidji Simo beginning her role as CEO of Applications. OpenAI’s products have been my life since I joined, and they’re in great hands. We’ll share more about OpenAI for Science in the coming months. In the meantime, if you’re an AI or academic researcher interested in joining us, my DMs are open.

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

𝗡𝗮𝗶𝘃𝗲 𝗥𝗔𝗚 𝘄𝗼𝗿𝗸𝘀 𝗶𝗻 𝗮 𝗱𝗲𝗺𝗼. 𝗜𝘁 𝗳𝗮𝗶𝗹𝘀 𝘁𝗵𝗲 𝗺𝗼𝗺𝗲𝗻𝘁 𝗿𝗲𝗮𝗹 𝘂𝘀𝗲𝗿𝘀 𝘀𝗵𝗼𝘄 𝘂𝗽. Embed → retrieve → generate looks clean in a notebook. Real requirements break it: → Questions whose answer is spread across many documents → Industry terms that embeddings get wrong → Bad chunks the pipeline never catches → Answers that live in how things connect, not in any single chunk → PDFs full of tables and images a text-only index cannot read These 5 architectures are how serious teams stay ahead in the agentic AI era: 𝟬𝟭 𝗛𝘆𝗯𝗿𝗶𝗱 𝗥𝗔𝗚 → Dense vectors find meaning. BM25 finds exact words. → Reciprocal Rank Fusion combines both ranked lists. → A safe baseline for almost every team. 𝟬𝟮 𝗚𝗿𝗮𝗽𝗵𝗥𝗔𝗚 → Pull entities and their relationships into a knowledge graph. → Retrieve subgraphs and community summaries, not chunks. → Best when the answer lives in how things connect. 𝟬𝟯 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗔𝗚 → A planner agent picks the right tool: vector, web, or SQL. → A reasoner agent keeps trying until the answer is solid. → Retrieval becomes a plan, not a single step. 𝟬𝟰 𝗖𝗼𝗿𝗿𝗲𝗰𝘁𝗶𝘃𝗲 𝗥𝗔𝗚 (𝗖𝗥𝗔𝗚) → Grade every retrieval before you trust it. → Correct → answer. Unclear → rewrite the query. Wrong → search the web. → This is what production RAG actually looks like. 𝟬𝟱 𝗠𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗥𝗔𝗚 → One embedding model (CLIP, ColPali) for text, images, and tables. → One vector index. One multimodal LLM. → No more separate pipelines for PDFs with charts. I built a runnable example for each of the five patterns. GitHub link in the first comment. The best teams in 2026 do not pick one. They combine them — hybrid retrieval inside an agentic loop, with a corrective grader, over a multimodal index. Naive RAG is a starting point, not a finish line. That is why most enterprise GenAI projects stall at the demo. Which of these five becomes the default RAG stack in the next 18 months — and which stays a specialized tool?

A digital twin isn’t about how it looks—it’s about what it knows and how it acts. According to IoT Analytics, the market for digital twins expanded by 71% between 2020 and 2022 set to jump from $𝟒𝟏𝟖 𝐦𝐢𝐥𝐥𝐢𝐨𝐧 𝐢𝐧 𝟐𝟎𝟐𝟐 𝐭𝐨 $𝟏.𝟓 𝐛𝐢𝐥𝐥𝐢𝐨𝐧 𝐛𝐲 𝟐𝟎𝟐𝟕 at a 𝟐𝟗.𝟒% 𝐚𝐧𝐧𝐮𝐚𝐥 𝐠𝐫𝐨𝐰𝐭𝐡 𝐫𝐚𝐭𝐞. On top of that 𝟔𝟑% of manufacturers are currently developing a digital twin or have plans to develop a digital twin. But the most interesting part? Most businesses already have the foundation for a digital twin and don’t even realize it. According to IEEE, the digital twin concept can be divided into three subcategories according to the different integration level, namely the different degree of data and information flow that may occur between the physical part and the digital copy: • 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐌𝐨𝐝𝐞𝐥: A digital version of a physical object (preexisting or planned) to correctly define a digital model with no automatic data exchange between the physical model and digital model. Examples of digital models include plans for buildings and product designs. The important defining feature is there is no form of automatic data exchange between the physical system and the digital model. This means once the digital model is created a change made to the physical object has no impact on the digital model. • 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐒𝐡𝐚𝐝𝐨𝐰: A digital representation of an object that has a one-way flow between the physical and digital object. A change in the state of the physical object leads to a change in the digital object and not vice versa. • 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐓𝐰𝐢𝐧: If the data flows between an existing physical object and a digital object, and they are fully integrated in both directions, this becomes a “Digital Twin”. A change made to the physical object automatically leads to a change in the digital object and vice versa. And if you’re running a 𝐌𝐚𝐧𝐮𝐟𝐚𝐜𝐭𝐮𝐫𝐢𝐧𝐠 𝐄𝐱𝐞𝐜𝐮𝐭𝐢𝐨𝐧 𝐒𝐲𝐬𝐭𝐞𝐦 (𝐌𝐄𝐒), you might already be closer to having a digital twin than you think. A well-implemented MES should naturally create a digital twin of your manufacturing process, enabling real-time monitoring, AI-driven insights, and automated process optimizations. If your MES isn’t doing this, you might not be unlocking its full potential. 𝐅𝐨𝐫 𝐦𝐨𝐫𝐞 𝐢𝐧𝐟𝐨, 𝐜𝐡𝐞𝐜𝐤 𝐨𝐮𝐭 𝐭𝐡𝐞 𝐟𝐮𝐥𝐥 𝐚𝐫𝐭𝐢𝐜𝐥𝐞: https://lnkd.in/eTSdpCsW ******************************************* • Visit www.jeffwinterinsights.com for access to all my content and to stay current on Industry 4.0 and other cool tech trends • Ring the 🔔 for notifications!

Most people think having a human approve an AI decision means the decision is safe. It does not. 👀 There is a term for what actually happens when humans rubber stamp AI outputs under time pressure. Automation bias. It is one of the most documented and underreported risks in enterprise AI right now. After 13 years and 200+ deployments, here is what I have learned about building genuine oversight into AI systems. The human reviewing an output needs three things to actually be in the loop. They need to understand what they are reviewing. They need the context to catch what the model gets wrong. And they need to be genuinely empowered to say no without institutional pressure to simply keep moving. Most organisations have none of those three in place. They have a signature process. That is not the same thing. Before any high-stakes AI output reaches a decision point in your organisation, ask these questions. ➡️ Does the person approving this understand the underlying data well enough to catch an error? ➡️ Is there time built in for genuine review or just enough time to click approve? ➡️ What happens if someone says no? Is that genuinely supported? If the answer to any of those is no… you do not have human oversight. You have automation bias with a human signature attached. What does genuine human oversight look like in your organisation right now? #ai #leadership #futureofwork #artificialintelligence #aistrategy #teamhuman #intellectualatrophy #criticalthinking

🚫 How to Run UX Research Without Access To Users. With practical techniques to avoid guesswork and gather insights if you can’t talk directly to users. Attached cheatsheet (with and without access to users) by Nielsen Norman Group. 🚫 Ask for reasons for no access to users: there might be none. ✅ First, study job openings to map existing workflows/tasks. ✅ Make friends with sales, customer success, support, QA. ✅ Find colleagues who are the closest to your customers. ✅ Convey your questions indirectly via your colleagues. ✅ If you can’t get users to come to you, go where they are. ✅ Ask to observe or shadow customers at their workplace. ✅ Listen in to customer calls and interview call centre staff. ✅ Request access to analytics, CRM reports, call centre logs. ✅ Use Google Trends to find product-related search queries. ✅ Gather insights from search logs, Jira backlog, support tickets. ✅ Explore past/ongoing NPS and Voice-of-Customer programs. ✅ Study reviews, discussions, comments for your product/competitors. ✅ Map key themes and user sentiment on TrustPilot, AppStore etc. ✅ Recruit users via UserTesting, Wynter (B2B), Maze, UserInterviews. ✅ Ask for small but steady commitments: 5 users × 30 mins, 1× month. 🚫 Avoid ad-hoc research: set up regular check-ins and timelines. As H Locke noted, if we shed the light strongly enough from many sources, we might end up getting a glimpse of the truth. Ironically, the stakeholders who can’t give you time or resources to talk to users often are the first to demand evidence to support your initiatives. Sometimes the reason why companies are reluctant to grant access to users is simply the lack of trust. They don’t want to disturb relationships with big clients which is carefully maintained by the customer success team. They might feel that research is merely a technical detail that clients shouldn’t be bothered with. Show that you deeply care about that relationship and that you don’t want to disturb it any way. What you do want though is to reduce costs and risk — the risk of drawing wide-reaching conclusions from very little research, or none at all. Your best shot is to explain research as a powerful risk mitigation tool. And: search for people whose priorities align with yours — people who value and see the impact of UX in their units. They would absolutely love to support your work because it also supports their work — and they will put up a good word for you if they only had known that you existed. ✤ Useful resources: UX Research Cheat Sheet, by Susan Farrell from NN/g (attached) https://lnkd.in/eUTHKWvF What Can You Do When You Have No Access To Users?, by H Locke https://lnkd.in/ewHEKhBS UX Research When You Can’t Talk To Users, by Chris Myhill https://lnkd.in/ez5-b6zf #ux #research

Do yourself a favor: don't use chat LLMs for learning directly from them. You will learn falsehoods about all topics, and this is what will shape who you are because we are the information we consume. This doesn't mean you cannot use LLMs as a learning support tool. If an original book, article, or code is too difficult to read, add it to the prompt and ask a question about the part you struggle with, but even in this case, do validate the answer by consulting the source directly, and only then continue reading. The longer the prompt, the less reliable the answer. Hallucinations in GPT-3.5 were relatively easy to spot, but in the most advanced chat LLMs of today, they are so convincing that even an expert in the field might start to question their understanding of the matter. This is not borderline dangerous; this is plain dangerous.

Paper sharing- AI in Science Discovery & Product Innovation MIT researchers expanded on applying AI-driven discovery to material science, which is making discoveries happen faster than ever！ What they did was introduce an AI tool (similar to the "AI Scientist" from Sakana AI) for materials discovery to 1,800 scientists in the R&D lab of a large U.S. firm. Traditionally, scientific discovery is labor-intensive and manual—a process full of trial and error, where scientists conceptualize various potential structures and then test their properties. Here’s how AI tackles it: AI generates ideas, prioritizes promising materials, tests them, and iterates on any false positives, refining until it finds viable options. Once validated, these materials can be patented and commercialized. This entire process runs much faster, and the impact is striking. Researchers with AI assistance have discovered 44% more materials, filed 39% more patents, and seen a 177% jump in downstream product innovation. Interestingly, the benefits are more unequally divided than we might have assumed. Top researchers nearly doubled their output, while the bottom third saw little improvement. This divide is partly because AI automates 57% of idea generation, allowing top scientists to focus on testing rather than preliminary research. Another downside is that 82% of scientists reported feeling less fulfilled, citing reduced creativity and underutilized skills. References/ Paper: https://lnkd.in/g3sZdAbJ __________________ I share my learning journey here. Join me and let's grow together. For more on AI and learning materials, please check my previous posts. Alex Wang