Brief items
From the AI digest for 2026-10-02
- Microsoft AI: MAI-Transcribe-2-Streaming takes Artificial Analysis' streaming transcription #1, "55% faster and 60% cheaper than ElevenLabs" (Suleyman, ~4:39 PM ET Oct 1) — vendor claim + AA self-selected; watch for an ElevenLabs response.
- Accounting-acing discourse (Andrew Curran, ~3:53 AM ET Oct 2): "eighteen months ago the best models fell short of the average accountant's ~37% score. Today, models ace those same tasks" — benchmark-discourse continuity, vendor-adjacent RT chain.
- Grok group-chat vision (Andrew Curran group-chat diary, ~8:08 PM ET Oct 1): XChat group chats can see images, context-selective responses — product color, unverified claims, X-grade.
- Claude Code "You should Know" plugin (~8:29 PM ET Oct 1): scans output for missed warnings — the mods ecosystem producing genuine utilities; Anthropic praising from its own account. Mods-demo color (Sonnet 5.5 pop-up city, Opus 5.5 embossed cards, Claude Code + camera on Old Street overnight, ~8:53 PM ET) — the mod-culture gallery.
- Karpathy: "spend more time understanding LLM outputs" (~12:37 AM ET Oct 2) — tips post featuring asking models to rewrite in ASD-STE100 controlled language; practitioner color, no claim.
- Griffin (Tavus) — Roose's cautionary RT (~12:50 AM ET Oct 2): "we built the 'trick old people into handing over their bank passwords machine' from the classic sci-fi horror novel 'don't build…'" — a voice-clone product framed as satire by Roose; the underlying Tavus launch is real; agentic-attacks adjacency, X-grade, verify capabilities before weight.