Demis Hassabis reportedly tested Mark Zuckerberg's reasoning during a private dinner — using the Wason selection task, a classic logic puzzle that most people fail.
Steve Hou surfaced the anecdote from a longer profile: Hassabis steered conversation toward VR, AR, and 3D printing to gauge whether Zuckerberg was thinking about intelligence itself or just chasing the next platform cycle. The subtext is a familiar one — Hassabis has always positioned DeepMind as pursuing general intelligence, not just product features, and the dinner reads as a quiet competitive assessment disguised as shop talk.
Sebastian Mallaby's "The Infinity Machine" on Demis Hassabis and DeepMind is drawing early praise as essential reading for understanding the mind shaping AI's direction.
Daniel Rock and Kevin Gee highlighted the book's depth in capturing not just where AI is going but the specific kind of intellect driving it. Mallaby, who wrote the definitive book on venture capital, brings a biographer's rigor to a subject that usually gets hagiographic treatment.
Garry Tan revealed he uses /codex adversarial reviews — firing CEO, design, and engineering reviews from GStack — to sharpen thinking when building with Claude Code as his primary agent.
The trick is using one AI to audit another's output: a meta-layer that catches the assumptions a single-agent workflow misses. Minh-Phuc Tran noted he's switching from Claude Code to Codex after repeatedly typing "double check if you missed anything," suggesting the tools are converging toward built-in adversarial self-review.
Developers are frustrated by Claude Code's tightened rate limits, with 267 posts trending on the topic in just 2 hours.
The rate limit squeeze is forcing power users to optimize their workflows — which is exactly what aditya's viral article addresses, arguing most people waste 70% of their Claude budget by treating it like a faster ChatGPT instead of structuring prompts for efficiency.
MIT CSAIL shared a "Build Any App: The Technical Co-Founder" mega-prompt for Claude — a five-phase framework that treats AI as a co-founder rather than a code generator.
The prompt by Miles Deutscher includes instructions that sound more like a co-founder agreement than an engineering spec. It tells Claude to challenge assumptions and separate "must have now" from "add later" — the kind of product discipline that usually requires a human partner to enforce.
Kevin Simback asked the question every Claude Code team is debating: gstack or Compound Engineering — and concluded the answer is both.
The emerging pattern is layered tooling: Claude Code for generation, gstack for review, Compound Engineering for orchestration. The Claude Code ecosystem is fragmenting into specialized roles faster than anyone expected.
Ishan teased something useful for Claude Max plan subscribers, drawing 110 likes with zero details.
The Claude Max plan — Anthropic's premium tier — is becoming a status signal in developer circles, with an ecosystem of tools and tips building around the cohort willing to pay top dollar for AI access.
A single developer rewrote vLLM in Rust (rvLLM), spent 22 hours and $1,780, and hit 10,291 tokens/second — beating Python vLLM up to N=256 with 20x faster startup.
The article claims 31x improvements on some operations. Rust's zero-cost abstractions meeting LLM inference's embarrassingly parallel workload is the upstream enabler. If a weekend project can outperform the standard Python inference stack, it raises uncomfortable questions about the technical debt baked into current AI infrastructure.
Prince Canuma announced mlx-vlm v0.4.2 with Meta's SAM3 model — adding realtime mask-only label drawing to Apple Silicon.
The MLX ecosystem for on-device AI continues to mature quietly. SAM3 running locally on a MacBook is the kind of capability that was cloud-only six months ago.
Jonathan Ross argued that the scaling laws paper is the most misread paper in AI — people conclude AI is slowing down, but they're missing something.
The post drew 460 likes and 82K views, suggesting the "AI is hitting a wall" narrative has enough believers that a counter-argument resonates. The scaling debate is becoming the field's central fault line.
Karpathy's AI persuasion experiment hit trending with 4,367 posts — exposing how effectively LLMs can shift human opinions.
The experiment demonstrated that LLMs can be more persuasive than humans in structured argumentation, raising questions about AI's role in shaping belief at scale. This sits at the intersection of capability research and governance that rarely gets this kind of viral attention.
Trending · 9h · 4,367 posts
Josh proposed a hypothesis that could reshape AI evaluation: for every unverifiable task, there exist verifiable proxy tasks where hillclimbing generalizes.
Brendan Dolan-Gavitt prompted this with "there are verifiable rewards everywhere for those with eyes to see." If the hypothesis holds, the "alignment is impossible because we can't verify the important things" argument may have a workaround — train on verifiable proxies that transfer.
Haider predicted ARC-AGI 4 could be solved sooner than expected — if agents achieve superhuman logical deduction on ARC 3, the skills transfer directly.
The ARC benchmark remains François Chollet's signature contribution to AI evaluation. Chollet himself was trending today arguing about the nature of intelligence — a signal that the benchmark debate is intensifying.
An arxiv paper on "Arrows of Time for Large Language Models" asked whether autoregressive LLMs exhibit a temporal directionality bias.
Do LLMs reason about past and future symmetrically, or do they inherit a directional arrow of time from training data? The paper drew 195 likes from a technical audience, touching a genuine open question in the field.
arxiv:1511.06434 · 195 likes · 9.3K views
A viral thread positioned Alec Radford — the researcher who built GPT-1 when no one else at OpenAI believed in pretraining — as the most important and least famous figure in AI.
Joseph Francis kicked it off sarcastically ("None of this guy's stuff has been peer reviewed. Loser.") while showing Radford's Scholar page: 68,880 citations for GPT-3, 57,679 for CLIP. Flowers followed up noting every LLM from every lab traces back to Radford's lonely bet on pretraining. A reminder that the most consequential research decisions often look like career suicide at the time.
Two gravitational forces tonight: the Iran crisis pulling in geopolitical anxiety, and Claude Code's ecosystem pulling in builder energy. The Alec Radford thread is the quiet highlight — a reminder that the most important bets in AI were made by people whose names most people still don't know.
Viktor Oddy combined NanoBanana + Kling 3 + Claude to build "Aethera" — a landing page so polished it drew 1.3K likes and 107K views.
The stack is becoming a standard creative pipeline: AI image generation (NanoBanana), AI video (Kling), and AI orchestration (Claude). What took a design agency a week now takes a solo creator an afternoon.
Paige Bailey showcased NanoBanana use cases from UCLA's Glitch Hacks hackathon — decomposing images into step-by-step drawing tutorials.
The application turns complex images into beginner-friendly drawing guides — using AI not to create art but to teach the process of creating art. A clever inversion of the usual generative workflow.
An AI film studio claimed to make films for governments and billion-dollar companies, with a Seedance 2.0 guide drawing 1K likes and 201K views.
EP separately noted the deals in AI filmmaking are reaching a scale that surprised even insiders. The space is crossing from hobby to industry faster than AI art did.
Levelsio made code-generated designs interactive — turning every AI idea into a launchable app — then asked: why not auto-generate every idea, add Stripe, and launch automatically?
His observation that AI designs all converge on brown tones and serif fonts when asked to avoid cliche is small but revealing about AI's aesthetic blind spots. The avoidance of one cliche simply creates another.
A crypto analyst connected dots nobody else was drawing: Iran mines Bitcoin at $1,300/coin, sells same day. The US bombed their grid for 27 days. BTC pumped $71K to $74K. Then Trump paused strikes — and mining machines came back online.
The thread claims BTC sell pressure resumed as Iran's grid recovered. Whether the causal chain holds or not, it frames the Iran conflict through a lens that traditional geopolitical analysis ignores entirely — energy infrastructure as both military target and crypto mining substrate.
Rajdeep Sardesai reported Elon Musk joined the PM Modi-Trump phone call amid the Iran conflict (per NYT) — raising questions about a private citizen on a wartime diplomatic call.
The Modi-Trump call itself is significant — India has been walking a diplomatic tightrope on Iran — but Musk's presence is the detail that makes it extraordinary. A Gulf diplomat separately told Okonkwo the strategic picture is bigger than anyone is seeing.
India went into full security mode: Modi met all state Chief Ministers, Ajit Doval conducted a high-level security meeting, Rajnath Singh met other ministers.
The thread drew 7.2K likes and 472K views, indicating the India-focused audience is watching the Iran escalation with intense concern. The coordinated ministerial response suggests India is preparing for scenarios beyond the current strikes.
Gaurav Sabnis found a Diego Rivera mural in Mexico featuring Dr. Pandurang Sadashiv Khankhoje — a Nagpur-born agronomist who became Mexico's "Norman Borlaug of corn."
A cabbie recognized "Nagpur" and pointed Sabnis to the mural in Mexico's Education Secretariat. The physical record of an Indo-Mexican scientific exchange that most Indians have never heard of — the kind of story that only surfaces through serendipitous conversation.
Varsha Singh posted a video of an 1800-year-old Devi Mahatmya scripture from Oxford University, drawing 5.9K likes and 62K views.
The artifact connects to the broader question of how Indian manuscripts ended up in European university collections — a topic that reliably generates both awe and indignation on Indian Twitter.
Nokia collects a hidden royalty on every smartphone sold — owning 20K+ patent families covering the foundational math for 4G, 5G, and 6G.
Parimal's thread drew 2.1 million views, turning an obscure fact about patent licensing into viral content. Nokia's transformation from phone maker to patent licensor is one of the most underreported business stories in tech.
Alex Kontorovich highlighted a hierarchy of success: Speak > Write > Have good ideas — in that order — citing MIT's Patrick Winston "How to Speak" lecture.
The implication is uncomfortable for researchers and builders: the quality of your ideas matters less than your ability to communicate them. Winston's 2019 lecture, with 18M+ views, keeps circulating because the insight never stops being relevant.
Vinod Khosla argued the biggest risk right now is not taking risks — quoting a CEO letter (via Alfred Lin) about leading through compounding change.
The post drew 478 likes and 149K views, suggesting the VC class is bracing for a period where conservative strategies become the most dangerous ones.
Bret Weinstein drew an unexpected parallel between prompting LLMs and prompting students: "Answer the question I should have asked you rather than the one I did."
The Wyndo prompt he quoted — "Interview me until you have 95% confidence about what I actually want, not what I think I should want" — captures the gap between stated and actual requirements that kills most projects.
Project Hail Mary sits there with 4.8 million views, evidence that this audience still looks up. And Mathematica's Fourier series animation — a square conjured from rotating circles — drew 130K views, a reminder that beauty in math never needs a marketing budget.