[LLM/AI for Science et al] 🤖 🦠 🧬
[I]: GPT‑4b micro: 50x Increase in Expressing Stem Cell Reprogramming Markers
OpenAI, in collaboration with Retro Biosciences, has reported a milestone for AI use in life sciences: using a custom, scaled-down GPT-4 model, they achieved a 50x improvement in stem cell reprogramming efficiency by redesigning proteins central to cellular rejuvenation. Their experimental model, GPT‑4b micro, was specially adapted for protein engineering and succeeded where years of traditional efforts fell short, rapidly engineering variants of the Yamanaka factors that enabled faster, more robust conversion of adult cells to stem cells—even in aged donors. Impressively, these AI-designed proteins not only improved cellular reprogramming rates but also enhanced DNA repair, pointing to real rejuvenation gains. As OpenAI notes, this collaboration demonstrates how tailored AI models, coupled with deep scientific expertise, can yield scientific breakthroughs in a fraction of the time.
[II]: LLM Agents Making Agent Tools
Very interesting work here: turning code repositories and scientific papers into tools, particularly LLM-compatible tools, to enhance the capabilities of LLM agents and reduce the technical burden on human researchers.
Link: https://arxiv.org/abs/2502.11705
AI Paper Explainer:
[III]: Asta: Agentic System that Advances Scientific Discovery
I have written in this newsletter about various Allen AI scientific agents, such as Paperfinder and ScholarQA. All these are now packaged into a new ecosystem, an agentic AI ecosystem called Asta. It also comes with Asta resources. “Asta resources are a set of tools, baseline agents, templates, and APIs that are fully integrated with AstaBench to provide a complete environment for developers to build, test, and refine scientific AI agents.” So you can build your own agent or model, and evaluate. I wish they do have video tutorials that go with their beautiful website to get folks started right away. It looks like the ecosystem for building science agents is shaping up nicely. More here.
[AI/LLM Engineering] 🤖🖥⚙
[I]: LangExtract: Google’s New Library for NLP Task.
Traditional NLP meets modern AI. Google recently released LangExtract, a new NLP library that streamlines traditional language processing tasks—like entity extraction, sentiment analysis, and text classification—by harnessing the power of LLMs and structured outputs. In this tutorial video, Sam Witteveen provides a hands-on walkthrough, highlighting how LangExtract leverages models such as Gemini to perform precise information extraction, complete with source grounding and robust output formatting. The library is designed for ease of use, supporting few-shot examples, long-context documents, and even visualization features, all while opening compatibility for both Gemini and open-source models. Compared to the classic workflow of fine-tuning BERT or DistilBERT for each task, LangExtract aims to make extraction, labeling, and data handling both faster and more scalable, potentially shifting the way companies manage large-scale NLP.
[II]: Context Rot
Chroma’s Kelly Hong spotlights “Context Rot”, the overlooked challenge of LLMs as input tokens grow. Despite the buzz around massive context windows in state-of-the-art models, real-world performance degrades when prompts get longer or more ambiguous. Hong’s team shows that even the best LLMs struggle to retrieve relevant information, reason through ambiguity, and filter out distractions as input length increases. Their findings reveal that longer context doesn’t equal better results and that LLMs’ outputs become less reliable, especially for complex or multi-step tasks. The takeaway? Effective “context engineering” (not brute-forcing more data) matters most: optimal results come from summarizing, filtering, and precise retrieval, not maxing out token limits.
[III]: Vibes Won’t Cut It
At AI Engineer World's Fair 2025, Chris Kelly of Augment Code delivered a compelling talk, "Vibes won't cut it," challenging the notion that AI-powered "vibe coding", letting AI generate code without deep oversight, can sustain production-ready applications. Drawing on decades of developer experience, Kelly argues that while AI tools are transforming how code is written, software engineering's essence remains firmly in human hands: context, critical decision-making, and code review are irreplaceable. He warns that generating more code with AI is not a virtue, every line adds maintenance risk, and that complex systems demand human understanding to ensure reliability, safety, and future-proofing. For teams adopting AI, he recommends clear standards, reproducible environments, robust testing, and well-defined tasks, emphasizing that AI should augment, not replace, the craftsmanship at the core of great software. In this rapidly evolving landscape, Kelly believes software engineering jobs aren't disappearing; rather, they're evolving, with context and judgment more valuable than ever.
[IV]: AI skills are redefining what makes a good developer.
From one of Andrew Ng's latest letters: In today's rapidly evolving tech landscape, possessing strong AI skills has become the defining factor for outstanding developers. While traditional computer science fundamentals remain crucial, there's a growing demand for talent that can leverage AI tools to engineer software more efficiently, build advanced applications using techniques like prompting and agentic workflows, and iterate rapidly. As businesses across industries race to adopt AI, those with expertise in both foundational CS and modern AI stand out, often surpassing even experienced developers who haven't kept pace with the latest advances. This shift creates both a wave of new opportunities for "AI native" engineers and a challenge for recent graduates whose skills may not yet match the demands of today's market. The future belongs to those who combine core programming knowledge with cutting-edge AI proficiency, shaping what it truly means to be a great developer in 2025 and beyond.
[V]: Using deepagents to build deep research
In my last newsletter, I wrote about the python library deep agents that (promise) to empower developers to build deep agents, intelligent assistants that go far beyond simple tool loops. Here is a video tutorial by Harrison Chase, using it to build deep research.
[AI X Industry + Products] 🤖🖥👨🏿💻
[I]: a16z Partner 2025 AI Productivity Stack
In this video from a16z, partner and AI investor Olivia Moore shares her top 10 AI tools that form her 2025 productivity stack. From using Comet as an all-in-one AI browser and workflow hub, to Julius for data analysis, Happenstance for next-gen professional networking, and Granola for seamless AI-driven meeting notes. Moore also highlights Gamma for dynamic slide and document creation, Willow for voice-powered editing, Superhuman for smart email management, Overlap for automated video clipping, Krea for creative AI generation, and ChatGPT for research-heavy tasks.
I have tried most of the apps on her list, or a replica of them. But the one I find perhaps most interesting (also new to me) is Happenstance.
[II]: Google’s “Nano Banana”
Google ‘recently’ unveiled Gemini 2.5 Flash Image, nicknamed Nano Banana, its latest image generation model that blends speed with creativity. Beyond sharper visuals, it offers natural-language editing, multi-image fusion, and consistent character rendering—making it easy to transform a selfie, pet photo, or sketch into polished outputs like collectible figurines, toy-style box art, or stylized portraits. The result: a playful yet powerful tool that’s already fueling viral trends and giving creators new ways to turn imagination into shareable, lifelike images.
[III]: Replit Agent 3
I've used Replit in the past, most recently in May of this year. My experience with it was not great; it simply didn't perform as needed for my project. Perhaps my project was too complex for the system. They recently released Agent 3, and I hope to give it another try soon. Their pitch is the following: Agent 3 is its most autonomous AI assistant yet, it can build full-stack apps from natural-language prompts, run for standalone sessions up to ~200 minutes, and test/fix its own work (via browser simulations) without constant supervision. (Replit Blog) It can also generate other agents and automations (Slack bots, Telegram bots, workflows, etc.), letting you automate repetitive tasks and third-party integrations. Here is the launch video.
[IV]: Alterego Wearable
AlterEgo is a non-invasive wearable “silent speech” interface that lets you communicate with devices (or people) without speaking out loud, instead it picks up subtle neuromuscular signals in the jaw, face, and throat when you internally vocalize words, processes them via machine learning, and sends you feedback through bone conduction audio. This is the most futuristic product on this newsletter, I think. You can watch the demo here, it’s very impressive! Here is a paper about it, from back in 2018.
[AI + Commentary] 📝🤖📰
[I]: A guide to understanding AI as normal technology.
After reading "AI 2027" several months ago, I wrote a deep summary, and review of the ideas. I soon stumbled on the AI as normal technology camp, and I have kept up with the debate ever since. There has been a new installment in the debate in the manner of a well written essay from the AI as normal technology camp. In it they continue to argue that rather than seeing AI as some extraordinary, world-bending force demanding unique interventions, real impact comes not simply from technical advances in the lab, but from how, and how thoughtfully, AI gets deployed in the real world. They argue that treating AI like any other powerful general-purpose technology gives society many opportunities to steer both its benefits and risks. Although this viewpoint might appear almost tautological, it stands in stark contrast to the superintelligence paradigm dominating tech debate. The duo urges readers, especially those recalibrating expectations after releases like GPT-5, to focus less on speculation or hype, and more on the hard, systemic work of deploying, governing, and adapting to these transformative tools. Resilience, not prediction, is the watchword for the age of normal AI.
[II]: Costco Era of Software.
Rex Woodbury explores the “Costco era” of software—an age defined by mass-produced, AI-generated code, where the cost and effort to build products has plummeted. As technology becomes increasingly commoditized, Woodbury argues that design is now the ultimate differentiator. Using examples from healthcare, he contrasts the dull UI of incumbent Epic’s AI with the vibrant, intuitive approach of newer entrants like Abridge. He highlights that while big players can win on scale and price, beautiful, user-centered design still offers startups a real edge. Drawing from everyday design lessons and the animation world’s “Spiderverse effect,” the essay insists that clarity, simplicity, and taste will set the best products apart in a world flooded with software. Ultimately, in a landscape crowded with “bulk” creation, it’s great design that signals quality, wins users, and drives enduring value.
[III]: Video Coding is the Worst Idea of 2025
Dave Farley, renowned software engineering expert, takes aim at the latest industry trend—vibe coding—in his much-discussed video, “Vibe Coding Is The WORST IDEA Of 2025.” Farley argues that while the allure of casually chatting with AI to produce a working app may sound futuristic, it fundamentally misunderstands what makes good software engineering. He warns that relying solely on AI coding tools, without a deep grasp of problem decomposition, design discipline, and robust engineering practices, can lead projects astray or even to failure. Farley insists that the heart of true software development lies not just in writing code, but in careful specification, version control, incremental changes, and especially automated testing. His message is direct: If we skip the rigor and precision of traditional practices, the next wave of code could be harder to fix, riskier to change, and in the long run, a setback for the craft of programming, AI or not.
[IV]: The Future of Software Business Models.
Here, Eric Flaningam writes about how the rise of AI-generated code is reshaping software business models, arguing that as software becomes increasingly cheap and commoditized, its value shifts away from standalone products. Drawing on historical trends, the newsletter outlines seven future-proof business models—from using software to boost hardware sales and offering vertically integrated solutions, to focusing on services, payments, platforms, advertising, and the infrastructure underpinning this shift. Flaningam emphasizes that while software alone is less likely to command premium prices in the AI era, opportunities abound for companies that leverage integration, services, platforms, and creative monetization of “free software,” making this the best time for innovators to build competitive businesses on top of these new foundations.
🎙 Podcast on AI and GenAI
(Additional) podcast episodes I listened to over the past few weeks: