Ollama announced it is now the official local-model provider for OpenClaw — making the most popular open-source AI assistant fully runnable on personal hardware at zero cost.
The integration supports streaming, tool calling, and auto-discovery of local models. OpenClaw has seen staggering momentum: 6,700 commits in two months, with contributor Vincent Koc logging 413 contributions on his most active day. The provider-plugin architecture means Ollama joins Anthropic and OpenAI as first-class backends, but it's the only one that keeps every conversation and code snippet entirely on-device.
Simon Willison published the twelfth chapter of his Agentic Engineering Patterns guide — this one titled "What is agentic engineering?" — and Garry Tan reposted it, signaling that the VC ecosystem considers this required reading.
Willison's guide treats coding agents (Claude Code, OpenAI Codex) as a new medium with its own craft: the defining feature is that agents can both generate and execute code, iterating independently of turn-by-turn human supervision. Each chapter is designed to be updated over time rather than frozen at publication — a living reference, not a blog post.
A viral thread showed someone using Claude to declutter their entire house with just seven prompts — racking up 268K views in two hours, the highest-impression "AI in daily life" post of the day.
CooperBaggs documented the workflow: room-by-room photo uploads, Claude generating keep/donate/discard lists with reasoning. The subtext is that "agentic" doesn't always mean code — it increasingly means physical-world planning through conversation.
Aidan McLaughlin revealed that he uses auto/instant mode for 70% of his coding turns — suggesting experienced practitioners are increasingly trusting agents to run unsupervised.
The claim aligns with Willison's "agentic engineering" thesis: the workflow is shifting from prompt-and-review to delegate-and-audit, with human attention reserved for the 30% that needs it.
The viral story of a pet owner who used AI to design an mRNA cancer vaccine for his dog turns out to be more nuanced than the headline: the final vaccine design was created using Grok, not ChatGPT — and the correction matters.
Bo Wang flagged the attribution error most outlets repeated. Paul Conyngham, an Australian AI consultant with no biology background, used ChatGPT for initial research, AlphaFold to identify mutated protein targets, and Grok for the final vaccine design. His dog Rosie's tumor shrank 75%, and she regained enough mobility to chase rabbits at the dog park. But this is a single anecdotal outcome in one animal — not clinical evidence.
Joscha Bach surfaced a provocative thread from cancer researcher Patrick Heizer arguing that individualized mRNA cancer vaccines could probably cure many cancers for about $100K per patient — but ethics committees consider the cost too high.
Jeremy Howard reposted Bach's framing, which cuts to the regulatory paradox: the same technology that saved a dog is technically within reach for humans, but institutional risk calculus — not science — is the bottleneck. The juxtaposition with the Conyngham dog story is striking: a pet owner could move faster than an oncology department.
Bojan Tunguz called it "unbelievable": someone hard-coded a WebAssembly (WASM) interpreter directly into the weights of a transformer model — meaning the LLM could, in principle, run XGBoost natively inside itself.
The original post from @joemccann described the feat as encoding a WASM interpreter losslessly into model weights. The implication is that models are no longer just text predictors — they're becoming execution environments, capable of running arbitrary computation within their own weight space. This is a different kind of "tool use" than API calls: the tool is literally embedded in the model.
Bindu Reddy claimed we are "shockingly close to automating frontier model training" — and Jack Cole of MindsAI pushed back hard, calling the claim premature.
The exchange captures a genuine fault line: optimists see the meta-learning loop closing (AI improving AI), while practitioners point out that the last 5% of training — data curation, evaluation design, failure mode analysis — is precisely where human judgment is hardest to replace. The debate is worth watching because it's no longer theoretical: both sides are building systems that test the thesis.
François Fleuret posted that Claude Opus "found how to remove permanently the vertical slider" in a Jupyter notebook — and called it the closest thing to AGI he'd seen.
The humor lands because it captures something real: AGI discourse focuses on grand benchmarks, but the moments that feel most uncanny are when an AI solves the kind of small, irritating problem that has quietly defeated thousands of programmers for years. Fleuret, a respected ML researcher, was being wry — but not entirely joking.
Rohan Paul shared Terence Tao's latest reflections on AI in mathematics: AI excels at brute-force testing of conjectures, while humans still hold the edge in seeing structural patterns — but the boundary is shifting.
Tao's observation is characteristically precise: the division of labor isn't "AI can't do math" but "AI does different math." The upstream context is the SAIR Foundation's new Mathematics Distillation Challenge, which Tao himself is involved with — an attempt to systematically transfer mathematical intuition into model training data.
Surya Ganguli, the Stanford neuroscientist, posted about his son's homework introducing Euler's formula V − E + F = 2 — and the thread became a gentle masterclass in mathematical wonder.
Ganguli's delight at watching his child encounter the formula for the first time resonated well beyond the math community. The formula connects vertices, edges, and faces of any convex polyhedron — a result so clean it feels like a glitch in reality. A reminder that this feed carries a strong contemplative thread alongside the AI signal.
Parimal traced the Rose of Venus — the geometric pattern traced by Venus's orbit relative to Earth — back to Bhaskara II's 12th-century mathematical texts, connecting orbital mechanics to ancient Indian astronomy.
The pattern, an eight-petaled flower emerging from planetary motion, is one of those results that makes you suspect mathematics was discovered, not invented. Parimal also shared detailed metallurgical analysis of Nataraja Panchaloha idol-making, arguing the precision of ancient South Indian bronze casting has been consistently underestimated.
Brian Roemmele teased a project using AI to perform "unique research on this ancient manuscript" — with no further details but an image that drew 274 likes and 20K views purely on mystique.
Roemmele has been building toward something at the intersection of AI and historical document analysis for months. The tease format (image + "more soon") is itself a signal: the audience trusts him enough to engage with an incomplete story.
The eGangotri Project shared the J. Bhattarai Collection of 212 digitized manuscripts, continuing the quiet work of making South Asian textual heritage searchable and preservable.
In a feed dominated by what's next, this is about what came before — and the thread connecting it to AI is precisely Roemmele's: these manuscripts become exponentially more valuable when machine reading can extract and cross-reference their contents.
eGangotri · 4h
Zishta showcased artisan soapstone Kalchatti cooking pot production in Tamil Nadu — and Karthi asked how much stone leaches into food over time, and why that might be preferable to stainless steel.
The thread is a microcosm of a recurring theme in this feed: traditional knowledge that was dismissed by industrial modernity is being re-examined with fresh eyes, sometimes aided by the same analytical tools that power AI research.
Tim Sweeney, the Unreal Engine creator, has quietly purchased over 50,000 acres of North Carolina forest for conservation — and the tweet revealing this drew 142K likes and 8.3 million views, the most-engaged post in this digest window by a factor of 80x.
The ratio (8.3M views for a conservation story vs. 96K for Grok's latest video demo) suggests that the tech audience's appetite for "billionaire does something unambiguously good" far exceeds its appetite for product updates. Sweeney hasn't publicized the purchases; Aakash Gupta's thread compiled the evidence from land records.
Sam Altman, answering "what's the biggest mistake young people make," said listening to old people — arguing that traditional career advice "is probably not going to work as well" in the current moment.
The clip, shared by Rohan Paul, is more interesting as a cultural artifact than career guidance: the CEO of OpenAI explicitly telling young people to discount received wisdom is either refreshingly honest or self-servingly disruptive, depending on where you sit.
Bojan Tunguz posted a single sentence — "We need to relearn how to be bored" — that quietly accumulated engagement in a feed otherwise optimized for novelty and dopamine.
No elaboration, no thread, no link. The format is the message.
VictorTaelin wrote a philosophical reflection on humility and boundaries in relationships — one of those posts that reads like a private journal entry published to 15K people.
The 229 likes suggest it resonated with an audience that processes emotional complexity through the same feed where they track transformer architectures. The two activities may be more related than they appear.
This Sunday evening feed has an unusually high ratio of contemplation to announcement. Tim Sweeney's forest, Surya Ganguli's son's homework, Bojan Tunguz on boredom, VictorTaelin on humility — the signal is that even in the fastest-moving technical community, the weekend creates space for people to share what they actually care about, not just what they're building. The AI stories (OpenClaw, WASM-in-weights, mRNA vaccines) are genuinely remarkable, but the feed's emotional center of gravity this evening is elsewhere.