posts · 2026-07-16

After 487 days, Roc's 300K-line Rust compiler rewrite in Zig reached feature parity. A demo game now compiles to a 31KB wasm binary, less than half the original size. The new compiler adds hot code loading during dev and reproducible cross-compilation. The team says this was a redesign, not a direct port, so direct Rust-vs-Zig comparisons don't apply. No formal release yet; v0.1.0 is planned later this year.

Federal filings show Anthropic CEO Dario Amodei donated $1M in May to Public First, a super PAC that advocates for AI safety regulations. The post body is blocked by Reddit's network security, so no further details are available—how the money will be used or whether Amodei has commented publicly remains unknown.

The author uses a song lyric to unpack how LLMs reverse our intuition about intent. Tokens are generated step by step, and direction emerges during generation rather than from a pre-formed thought. Anthropic's research shows Claude plans rhymes ahead when writing poetry, but the explanation it gives afterward is another generated continuation—not a faithful transcript of internal state. The post frames this as 'throw first, then draw the bullseye,' and warns that confident model explanations are themselves new throws.

Tiangong Short Drama Workbench uses a director Agent to auto-parse scripts, plan blocking and camera positions, targeting the persistent face-swap and position-drift issues in AI short dramas. It packs film-grade prompt templates, 720° panoramas, and a 3D director console for controllable production. Three works already launched on DramaWave, hitting seven-figure USD revenue in 7 days.

Cloudflare launched Precursor on July 13, 2026, a session-level behavior detection system inside Enterprise Bot Management. Instead of relying on browser fingerprints or single-request signals, it continuously collects pointer movement, keystroke timing, focus changes, and page visibility, then cross-checks whether the whole session is internally consistent. The shift matters because Playwright-driven agents now run inside real browsers, making old shortcuts like checking navigator.webdriver unreliable. Precursor injects scripts into HTML responses, compresses event streams, and evaluates them at the edge, updating Bot Score and challenge clearance as behavior changes. Cloudflare states it does not record actual keystrokes or bind signals to user identities. The post also covers the FP-Agent paper, where an XGBoost classifier distinguished seven browser agents in a controlled lab setting, but notes that doesn't translate to production accuracy. Independent accuracy rates, false-positive data by user group, and accessibility testing are not yet public. Official materials conflict on rollout scope. Behavior detection alone doesn't solve agent identity, user delegation, or business authorization—Precursor only addresses runtime risk.

Tencent released IQ1_M quants for the 295B Hy3 MoE model, fitting into a single 96GB GPU at ~85.5 GiB. Actual BPW is ~2.4, not 1-bit. With 64K q8 KV cache, only ~2 GiB headroom remains—official config is marked tight. Agent and coding abilities dip slightly vs BF16; multi-step reliability is untested. Community GGUF hits ~23–25 tok/s on M3 Max, but 8K prompt prefill takes ~200s. Worth experimenting if you already own the hardware, not a reason to buy it.