A couple days ago, I attended the IBM Tech Exchange Dev Day at Georgia Tech on Open Source AI. I ended up learning more than I thought I would. From Docling to Generative Computing, my notes: theepsilon.substack.com/p/from-gran...

The latest pulse, Issue 28: GPT-5👨🏿‍💻, Comet🌐, and How to Publish for Machines📄 theepsilon.substack.com/p/pulse-iss...

It's already clear that knowledge works are going to undergo a massive restructuring (to say the least) over the next decade; many (most notably in coding) are already well under way. And knowledge work does not preclude running a lab & doing science.

My latest AI reading list ε Pulse: Issue 27 -> Context Engineering👷🏿, Grok 4📒, and a Foundation Model for Tumor Microenvironment 🎗 theepsilon.substack.com/p/pulse-iss...

What happens when an AI runs a business? Anthropic put Claude in charge of a real-world vending machine for their "Project Vend" experiment.

Folks at Cognition argues against building multi-agent AI systems. The author argues that decomposing tasks for multiple agents creates more problems (like context sharing) than it solves. More here: cognition.ai/blog/dont-b...

Anthropic engineers' take on building multi-agent research system to tackle complex, open-ended questions. A lead agent delegates tasks to sub-agents who research in parallel in an orchestrator-worker agentic pattern

Fine essay here on Bio-ML and Noetik (a startup building foundation models of tumor microenvironment) The thesis: The future of bio-ML is using rich, multimodal data to build foundation models for complex diseases, rather than focusing on simpler problems. www.owlposting.com/p/joining-n...

My latest academic work on applied NLP is now published in Analytical Chemistry: "A Large Language Model–Powered Map of Metabolomics Research" Paper: pubs.acs.org/doi/10.1021... App: metascape.streamlit.app/

my notes on topic coherence measures in NLP, over at my AI blog. theepsilon.substack.com/p/topic-coh...

Andrej's analogy game is a hard. I really enjoyed listening to his talk on software in the age of AI. Highly recommend, if you haven't watched it already. www.youtube.com/watch?v=LCE...

We traded calloused hands for blue-light screens, thinking knowledge work would end human suffering. It eased plenty no doubt; yet we swapped worn-out bodies for sedentary-induced metabolic ‘fallouts’ (obesity, diabetes, and the like.), and all the suffering downstream. Progress with a hidden bill.

I came across this nice talk the other day: www.youtube.com/watch?v=uFP... Building effective agents with mcp via an agent library "mcp-agent"

This is a good (video) podcast on Mech Interp, highlighting recent publications from Anthropic: www.youtube.com/watch?v=9YQ... Their research aims to move beyond treating LLMs as black boxes and instead reveal their "intermediate thought processes" via attribution graphs.

What I am reading this summer: www.epistemeengine.com/p/what-i-am...

Sam Altman on the "Gentle Singularity": blog.samaltman.com/the-gentle-...

Which startup are your investing in? Cogito, The Prince, Form Room, Proletariat Perks, or Sisyphus

The other day, I used a prompt to write a prompt that then writes another prompt. I suppose you can just do stuff these days.

instagram.com/epistemeengine [1]: www.nature.com/articles/s4... [2]: jamanetwork.com/journals/ja...

It takes some time for it to sink in but properly viewed, productivity isn’t about doing more; it’s about ignoring nearly everything, and that precisely, is what makes it incredibly hard.

🧵👇🏿

Powerful action space, decentralization, privacy & security. A clear-eyed argument for personal, local, private AI agents by soumith chintala www.youtube.com/watch?v=jMo...

Interesting, and unsurprising findings here on genAI usage at work: www.axios.com/2025/05/29/... The speed of change is so much that, I wonder how many "leaders" are able to keep up, talk less of writing useful guidelines. Keeping up is *literally* a day job.

instagram.com/epistemeengine [1]: www.worldobesity.org/about/about... [2]: www.who.int/news-room/f...

The next wave of commerce would be agentic. From malls to website to apps, and now agents, and this essay argues correctly for the case (with a more developed thesis): www.digitalnative.tech/p/agentic-c...

We are at the verge of a full-blown prompt economy.

1/ I recently completed Alan Lightman's recent book, "The Miraculous from the Material." I enjoyed reading it but not without some qualms. I had to write a video essay + some philosophical commentaries, contrastingly titled, "The Material from the Miraculous" 🎥: www.youtube.com/watch?v=5hC... 🧵

Boethius: A person is an individual substance of a rational nature. Made with Veo 3. This model will supercharge video essayists' workflows.

Apparently LangGraph agents can now be exposed as MCP tools. Exciting to see how the entire agent engineering ecosystem is rapidly maturing. doc: langchain-ai.github.io/langgraph/c... video: www.youtube.com/watch?v=AR4...

I first covered MCP when it was released in late 2024 and revisited at least once afterwards. Since it's gained popularity in the AI community, I thought it deserved a dedicated Pulse entry for a deeper dive. ε Pulse: Issue 25 → MCP Special + theepsilon.substack.com/p/pulse-iss...

1/ Finished up "AI 2027" recently It is my recommended reading for anyone with even the slightest interest in AI forecasting and safety. My summary & reflection: theepsilon.substack.com/p/the-intel... Some quotes from the scenario: 🧵 ai-2027.com/

I can think of exceptions, but in general the intuition here is 🎯

On AI writing detection: my intuition is that with increasing access to cheap, powerful AI (+ improved prompt eng. education): false negatives > false positives. (i.e., it will become easier to miss AI-generated writing than to misidentify genuine human writing.),

An obvious hallmark of highly productive people today: avoiding *distractions & constant task-switching like a plague. My latest video essay explores this. www.youtube.com/watch?v=PiB... The essay is an excerpt from my book Wander: epistemeengine.gumroad.com/l/wander

Around the Web Issue 36: The Next Pope. 🦠Recipe for a Cell. 🔰 AI Assistant for Avid Readers. 💻 A.I. Is Changing How Folks Builds Start-Ups. 🌎Ten Charts That Show How the World is Changing www.epistemeengine.com/p/around-th...

I'm currently reading a fine book on LLM finetuning by Voigt Daniel. Sometimes, jumping from one tutorial to the next can be ineffective. I’ve found this book helpful for reviewing all the key aspects of the subject. I plan to share my notes on my blog once I’m done reading.

Anyone who has spent less than 30 mins probing Gemini 2.5 Pro or o3 would know that these models are not nothing, agentic at core and extremely capable (2.5 Pro is my favorite, by a small margin).

1/ When I first tried out GPT-3 circa September 2022 (via the Playground), a few months before ChatGPT came out, it was clear to me that this is fundamentally a technology that's going to upend the state of affairs.

AI systems are getting a massive upgrade, but it’s not making the splash it deserves—too much at once, and too many ghiligi-style memes :) Here’s a great podcast episode on AI-enabled biology and medicine, highlighting some of the latest advances. www.youtube.com/watch?v=Jo3...

I wrote this app prototype to solve a personal habituated pain point.  To stop reading in a circle, find knowledge gaps, read on-demand, and dig in.  Missing Pages: Towards an AI agent for avid readers.  Back Story: www.youtube.com/watch?v=4wn... Waitlist: tally.so/r/wgOJvM

Forget secrets hidden from others. In a sense, the biggest secrets are the ones we hide from ourselves. I say more in my latest (short) video essay: What Are You Hiding? www.youtube.com/watch?v=BeO...

I have now had sufficient play time with Gemini 2.5 Pro. It is too good. While I believe we will still have better models, these things are hard to imagine.

My latest AI newsletter -> ε Pulse: Issue #24 AI Paper Finder📄, TxGemma💊, and the Future of SaaS 📱 theepsilon.substack.com/p/pulse-iss...

Dario Amodei predicted a "compressed 21st century," where AI-driven biology accelerates 50-100 yrs of progress into just 5-10 yrs. I was skeptical when I first read this prediction, and I remain so, although less so than before, largely because of some recent advances.

Described as the USB-C port for AI apps, MCP is the open standardized protocol that has gained lot of attention lately. It connects AI apps (IDE, Claude Desktop) 🔌 to various data sources, tools, etc. If you say MCP is like a magic wand for LLMs, you might just be right.

Some Computer Use Agents (LLM that can *use* mouse and keyboard like you and me.) dev tools I stumbled on recently. The space seems to be maturing pretty well. Amazon's Nova Act SDK: github.com/aws/nova-act LangGraph CUA: openai & scapybara: www.youtube.com/watch?v=ndC...