AI Daily Dev — Week 29

Terry Tao Ports His 27-Year-Old Java Applets With AI Coding Agents

Fields medalist migrated ~two dozen Java 1.0 math visualizations from 1999 to modern JavaScript over a few days.
One minor bug in the AI-generated code; agents surfaced two bugs in Tao's original code he had missed.
Also finished an abandoned 27-year-old 'Inkscape but in Minkowski space' special-relativity applet in a couple hours of vibe coding.
Tao names Claude Code among the tools and published an edited transcript of the session.
Top of the Hacker News front page for most of Sunday, July 12.

community terrytao.wordpress.com

Proof PDF posted July 10 on OpenAI's CDN with sole author credit assigned to the model.
Sol Ultra ran 64 parallel subagents and produced the argument in under an hour.
Conjecture (Szekeres 1973, Seymour 1979): every bridgeless graph has cycles covering each edge exactly twice.
Method reduces to cubic graphs, invokes the 8-flow theorem, and applies a linear-algebra edge labeling.
Mathematical community has not yet validated the proof; HN thread already skeptical about the attribution.

research news.ycombinator.com

New in-app prompt asks users to consent to health data — steps, sleep, medication, cycle tracking, treatments, test results — being used to train Samsung's AI models, with human review.
Opting out triggers a warning that sync will stop and the account's health data will be deleted unless retained by law.
Data is intended to train upcoming Galaxy Watch 9 fitness algorithms; privacy advocates call it an 'AI-or-nothing' toll booth, not consent.
Story sat at #1 on the Hacker News front page through Tuesday, July 14; reporting from 9to5Google, Android Authority, Neowin, GSMArena, Cybernews.

industry news.ycombinator.com

Programmable keypad co-built with boutique keyboard maker Work Louder, launching Wednesday July 15.
13 mechanical keys, clickable rotary encoder, 2D analog joystick, and six programmable layers of Codex CLI shortcuts.
OpenAI's first branded hardware product — targeting the ~5M weekly Codex users.
First teased June 29 at the AI Engineer World's Fair in San Francisco; coverage in TechTimes, TheNextWeb, DevOps.com.

tools thenextweb.com

Wire-level analysis published July 12 showed Grok Build was packaging developers' full tracked repos as Git bundles — Git history, .env secrets and all — and shipping them to an xAI-run Google Cloud bucket.
On a 12 GB test repo the upload channel moved ~5.1 GB across 73 chunks; the actual coding task traffic was ~192 KB — roughly 27,800× more data than the prompt needed.
The user-facing 'Improve the model' privacy toggle had no effect on the repo uploads; secrets committed and later deleted from history were transmitted anyway.
Musk promised to purge all previously uploaded user data; xAI disabled the behavior server-side and re-released Grok Build's 844,530 lines of Rust under Apache 2.0.
HN thread live, coverage in The Register, TechTimes, The Hacker News, TNW, Simon Willison's blog.

tools theregister.com

Long-form response to Anthropic's blog claiming Bun was ported from ~500K lines of Zig to Rust in 11 days by 64 parallel Claude Fable 5 agents at ~$165K in API spend.
Kelley: 'If the test suite isn't sufficient to catch bugs in Zig code, it isn't sufficient to catch bugs in a million lines of unreviewed slop.'
Points at Bun's engineering culture — 'hacks on top of hacks, abuse of assertions' — not language limits, as the real driver of the rewrite.
Ray Myers' 'Zig Creator Calls Spade a Spade, Anthropic Blows Smoke' recap hit the HN front page at 1,385 points / 692 comments; picked up by The Register.

community andrewkelley.me

2.8-trillion-parameter Mixture-of-Experts activating 50B per token, 1M-token context; K3 Max and K3 Swarm Max shipped July 16 via Kimi Code and an OpenAI-compatible API.
New architecture: Kimi Delta Attention (hybrid linear attention) with Attention Residuals; the model always runs in max reasoning mode, so output-token counts trend high.
Scores 57 on Artificial Analysis Intelligence Index — on par with Opus 4.8 and GPT-5.5, behind only Fable 5 and GPT-5.6 Sol; averages $0.94/task vs Opus 4.8's $1.80.
Priced $3 / $15 per 1M tokens — matches Claude Sonnet 5's rack rate. First Chinese lab to skip the cheap-alternative pitch. Full Hugging Face weights promised by July 27.
Two 'Kimi K3' launch threads live on HN; Simon Willison and latent.space calling it the largest open model ever released.

models simonwillison.net

Poisoned dataset chained a remote-code loader and a template injection to execute on a processing worker; the actor then escalated to node access and moved laterally into internal clusters over a weekend.
Campaign ran as a swarm of short-lived sandboxes executing many thousands of automated actions, with self-migrating C2 staged on public services.
Anomaly detection on HF's own LLM-triage pipeline flagged the intrusion; forensic reconstruction over 17,000+ recorded events had to run on the open-weights GLM 5.2 because commercial frontier APIs blocked the attack commands as unsafe.
Impact limited to a set of internal datasets and service credentials; no evidence public models, datasets, or Spaces were tampered with. Users advised to rotate access tokens.

community huggingface.co

Filed July 10 in N.D. California; complaint calls the scheme 'at every level' up through OpenAI's chief hardware officer.
Apple says 400+ former Apple employees now work at OpenAI.
Names Tang Tan (ex-Apple VP, now OpenAI hardware chief) for coaching interviewees to bring physical Apple parts to interviews.
Names ex-Apple senior EE Chang Liu for keeping a work laptop and downloading confidential design documents.
OpenAI's response: 'We have no interest in other companies' trade secrets.'

industry techcrunch.com

Ploy migrated its marketing-site-building agent from Claude Opus 4.8 to GPT-5.6 — first Opus dethroning in four months.
2.2× faster to finished pages, 27% lower cost, and roughly half the output tokens.
Matched page: Opus wrote a 17,957-char globals.css with 174 CSS variables vs GPT-5.6's 2,508 chars and 45 variables.
Gotcha: GPT-5.6 sends all 25 top-level tool parameters every call, inventing offsets that made 52–64% of file reads return empty.

tools ploy.ai