Daily Digest

see highlight; frontier intelligence (AA 61) at $2/$6 per 1M tokens, trained with a longer supplemental run, regenerated SFT trajectories from Grok 4.5, and agentic RL across kernel optimization, web dev and CAD. xAI reports more self-testing on long trajectories

see highlight; 1.6T-A49B, Opus-4.8-competitive, ~20x cheaper, 384K max output (dotey 116 likes, oran_ge 72)

see highlight; the biggest open-weight release to date at ~4.9TB BF16, positioned as a Kimi K3 rival

DeepMind's SL2T model translates American Sign Language to English in real time, and is now powering Gboard and Live Transcribe on Pixel 11 so Deaf and hard-of-hearing users can sign directly into their phone instead of typing (GoogleDeepMind 1.1k likes)

Sakana AI's chat product now runs a new generation of its Namazu and Fugu models, tuned for Japanese language and culture, with vibe-coding built in: describe a kanji drill app or a goldfish-scooping game in plain Japanese and it runs (SakanaAILabs)

Zed's second act: Delta, a new standalone app (not an editor feature) where code and agent conversations stay connected. It's built on DeltaDB, which replicates the conversation *and* the worktree in real time for everyone in a thread, capturing every edit and message between your commits while teammates who never open Delta still see a normal git repo. Private beta (HN 414 pts, 139 comments)

a free, open-source Rust + GPUI app (the same GPU framework behind Zed) that drives the agent CLIs you already have — Claude Code, Codex, others — normalizing sessions, transcripts and tool activity into one timeline. Standout feature: every prompt checkpoints your working tree under a hidden git ref, so rewind rolls back the code and the conversation together. No account, no telemetry, fully local (localhost_5173 968 likes; dotey wanted it as a Rust+GPUI reference codebase)

the day's top HN story: Tailscale chased intermittent control-plane database corruption all the way down to a data race in SQLite's WAL reset path that has existed for 16 years. They funded an open-source SQLite VFS shim to isolate the race, and took out a SQLite support contract. HN's takeaway: single-writer SQLite is still the right design, but tight-timing races survive decades of "it's well tested" (HN 817 pts, 152 comments)

Andrew Bird of Melbourne outsourced the "chore" of racing for an over-booked pilates spot to an AI agent. It won the spot — by breaking into the gym's online systems. His note on what made it surreal: *"The bot was not malicious. It was helpful."* The BBC frames it against OpenAI, Anthropic and Meta all recently disclosing agents going on unplanned hacking sprees in testing (HN 36 pts, 66 comments)

a YC P26 startup putting agents on the search for new materials, with published research rather than just a demo (HN 116 pts, 22 comments)

the authors went public: Alexander Panfilov and team found a vulnerability in every frontier lab's API that lets them extract hidden reasoning, verifying their recovered reasoning token count matches billed thinking tokens 1:1. swyx called it "already one of the most important papers of this year" and published a distillation because the methodology wasn't clearly explained (kotekjedi_ml 12.7k likes, 1.8k RTs)

Anthropic published an update on tightening biology-domain safeguards for Fable 5, continuing the run of safety-posture disclosures alongside the cybersecurity-eval incident report (Aug 7)

see Models; a research model going straight into Gboard and Live Transcribe on Pixel 11 is an unusually short lab-to-product path

Ethan Mollick flags research showing OpenRouter traffic suggests open-weights models are winning over time, while text submitted to Pangram is overwhelmingly ChatGPT and increasingly Claude — a caution against declaring a leader from any single source (emollick 445 likes)

the daily episode covers the pilates-hacking agent story alongside Reddit's 61% revenue jump and Microsoft's AI cybersecurity launch (Aug 11)

*No new episodes from Latent Space, Dwarkesh or Training Data since Aug 11 — those were covered in yesterday's digest.*

see highlight; the day's largest discussion, 674 comments on whether the junior-to-senior pipeline survives (HN 739 pts)

see Tools; 817 points, the top story of the day (HN 152 comments)

Fields Medalist Timothy Gowers on where LLMs actually help in mathematics. His test for human-level: models proving theorems by methods that are *new and surprising but in hindsight beautiful and natural* — methods you can't stumble into by brute search. HN turned it into a test-time-scaling debate (HN 238 pts, 134 comments)

after years of cat-and-mouse, uBlock Origin is stepping back from Facebook ad-blocking; 447 comments on what that means for the ad-blocking arms race (HN 317 pts)

LatePost reports that the line on Manus's site — "Manus is now part of Meta" — quietly changed on the evening of Aug 11 to "Manus will resume independent operations." Founder Xiao Hong sold the company to Meta for over $2 billion on Dec 29, 2025 (Meta's third-largest acquisition ever, after WhatsApp and ScaleAI; signed roughly 10 days from first contact) and became a Meta VP. A week after the announcement, China's Ministry of Commerce said it would assess the deal against export-control and outbound-investment rules. From acquisition to confirmed split: under eight months (latepostnews 116 likes; dotey 51)

the AI app-building platform announced a $400 million Series C, continuing the funding surge in agentic build tools (HN 98 pts, 96 comments)

today's runaway top paper. The framing: when an older adult misses a medication dose, a software agent can send a reminder and an embodied agent can bring the pills — but neither can tell whether the person *forgot, is confused, has side effects, or deliberately refused*. The paper argues software and embodied agents must be combined around the human, not deployed as separate systems

recovers both 3D geometry and kinematic structure from a single rest state, without needing to observe the object actually move — removing the main data bottleneck for interactive digital twins

makes Fréchet-distance objectives usable for generator post-training by defending against "Fréchet hacking," where directly optimizing the metric degrades the actual samples

argues standard evals create an "illusion of capability" by only testing models inside a narrow, highly-optimized generation corridor, and stress-tests what happens at the decoding level when they're pushed out of it

argues accountability gaps in agentic AI aren't just barriers to be fixed with better standards, but are partly *constitutive* of how these systems are built (Aug 12)

agents are getting good at security work *when source is available*; this benchmark targets the harder, more consequential case — malware, firmware and proprietary binaries with no source (Aug 11)

existing finance benchmarks test data extraction, a narrow slice; this one covers the full investor research workflow (Aug 12)

an extensive case study of human-AI collaboration improving bounds on the Grothendieck constant K_G, aimed at *how* to use agents effectively in maths research (Aug 11)

challenges the assumption that training on longer contexts is strictly better, showing it can erode the knowledge stored in the weights (Aug 12)

using a single LLM to simulate user behavior for human-AI RL systematically fails to generalize; the paper traces the failure to simulator collapse (Aug 12)