AI social digest — 2026-10-06
Top line: The X layer is product drip and benchmark noise today; the two things with signal both land on the models: a Perplexity decision-model update (pplx-decider-v1.1 — top score on the new HF Decision Index 0.3, price halved to $0.02/M, Arav's Oct 1 price-cut promise delivered in five days) and the Mistral Large 4 eval layer (blind human review #2 overall; 4,000 training GPUs vs Astra's 100,000). 2 kept as stories, everything else is brief-or-noise.
1. pplx-decider-v1.1: Perplexity's decision model bumps, halves its price
What happened: Perplexity Developers (~1:25 PM ET): pplx-decider-v1.1-27b, "our updated open weights multimodal decision model, is now available. It scores the highest on the new @huggingface Decision Index 0.3 benchmark" — pointing at the multimodalart/jev-decision-index space. Arav doubles down 20 minutes later (~4:46 PM ET): "Also, Perplexity Decision API pricing was cut by half, now to 2 cents per million input tokens. Enjoy!"
What's genuinely new: version bump (v1-27b → v1.1-27b), a claimed top score on a benchmark space Perplexity doesn't own, and the price halving to $0.02/M input — half the v1 price and half Jev's. Ledger item resolved: Arav's Oct 1 "we intend to bring down the price even further over the coming days" — delivered day 5, review-by (Oct 8) not yet due. The decision-models runner question (Jev vs AnyJev vs d1 vs OpenAI Decisions API vs pplx-decider vs Clef/Strands) gains a v1.1 cell; the independent run on identical labeled workloads still doesn't exist.
Watch: whether HF Decision Index 0.3 is the community's scoring harness (whose table wins); Jev/Cloudflare responses; still no runner.
Links: Arav
2. Mistral Large 4's eval layer: blind human review and the 4,000-GPU line
What happened: Mistral's announcement tweet + Surge AI: "For their Mistral Large 4 release, @MistralAI ran a blind Surge human eval... five frontier models on coding quality. Le Chonk ranked #1 among open-weight models and #2 overall, behind only Opus 5." — with Surge's framing: "Unit tests ask: does it work? Human code review asks: would you merge it?" Peter Gostev (RT'd by LeCun, 9:58 AM ET) supplies the disparity metric: "Mistral Large 4 was trained on 4,000 GPUs and Astra was trained on 100,000 of them. We'll see what the kitten can pull off..."
What's genuinely new: the X layer of the release covered in today's ai digest (story 6) — the eval genre here is blind human code review, not a benchmark board, and the compute-disparity frame is doing the anti-scaling-narrative work. Vendor-run and vendor-floated numbers, both disclosed as such.
Watch: independent runs once weights circulate; whether the blind-human-eval announcement move spreads.
Links: Surge/Mistral · Gostev RT via LeCun (no link captured in feed)
3. Grok Bot's platform hooks: @bot tagging goes live; Creator Rewards
What happened: (1) Larsen Cundric (RT'd by Musk, 7:38 PM ET): "@bot tagging just went live" — reply to any post with "> @bot add this to my Notion reading list," "@bot remind me to read this tomorrow," "@bot summarize this thread," "@bot draft a reply" — working in replies, posts, quote-tweets, "lands right in your Grok Bot." (2) Sawyer Merritt (RT'd by Musk Oct 5 6:27 PM ET, landed after our Oct 5 cut): Grok @Bot Creator Rewards — biweekly, usage-based rewards for template creators. (3) A public changelog now exists: x.ai/changelog/bot.
What's genuinely new: the consumer-agent surface keeps accreting hooks into third-party surfaces — Notion named in the tagging examples is integration color, and a changelog channel is the platform-ization tell. Continuation of the Grok Bot cadence strand (agent-harness race thread). X-grade; no docs beyond the changelog link.
Watch: what the changelog itemizes; whether Dots/Dots-equivalents get the same reply-hooks; reward economics in practice.
Links: @bot tagging RT · Creator Rewards RT
Brief
- Anthropic's Cyber Verification expansion, tweet layer (~3 PM ET): "We're expanding our Cyber Verification Program to give security professionals broader access to our most capable models" — the primary and analysis are story 4 in today's ai digest.
- Nano Banana 2.1 (Google, via Hassabis RT, noon ET): new image generation/editing model, "notable leaps in visual design, mask-based editing, and subject consistency." Product-brief grade; joins the forged-signatures item in the ai digest's provenance adjacency.
- Grok 4.7 takes the top three on VulcanBench Frontier v4 (DogeDesigner RT'd by Musk, 10:09 AM ET): top run 93.15, all 23 tasks passed, "benchmark for challenging coding tasks." Vendor-adjacent benchmark claim — same genre as every board we record; independent check still absent.
- H-JEPA (Randall Balestriero, RT'd by LeCun, 10:34 AM ET; arXiv 2610.06805): "the first end-to-end learned hierarchical world model for long-horizon visual planning," built atop SIGReg/LeWM — the JEPA lineage's academic layer thickens again after MotionJEPA/AD-E2E-JEPA. No graded claims.
- Narayanan's safeguards quip (via LeCun RT, 7:40 AM ET): "An interesting cognitive bias: 'GLM 5.3 is around as capable as Claude Mythos at cyber-exploitation while also having much lower safeguards, meaning that every...'" — the X-tier read on the efficient-tier thread's Sep 30 capability/safeguards asymmetry data.
- DeepMind-backed eval governance for bio: Fleming Initiative's 3-year programme "to develop the methods and standards needed to evaluate artificial intelligence for antimicrobial resistance" (RT 4:48 AM ET) — the eval-machinery shelf gains an AMR application.
- Computer demos keep the platform color coming: options-gamma charts via QuantWheel data; interactive 3D design preview with editable CAD export via FreeCAD — demo grade, unaudited, same shelf as the Oct 5 windowing story.
Dropped
- Computer vs Computer StarCraft (live bot-vs-bot match, "no pausing") — color demo, filtered by the relevance gate; agreeable to skip.
- Riley Gaines RT of NY Post trans-crime story — off-AI politics cross-post.
- Musk/Stossel ownership-structure RT ("simply due to my ownership stakes") — self-mythology, off-AI.
- LL Cool J × Rogan on Grok's political centrality (KanekoaTheGreat RT) — discourse lore; the on-record version of the bias fight is the audit record, not podcast takes.
- "Login with ChatGPT tokens have clear financial value" (commentary musing, filtered) — decent observation, no claim or artifact.
- Boris Cherny "always interesting to see how others prompt" — screenshot post, no content in feed.
- "If you are not using Grok Bot, you are fking up" (stevenmarkryan) — marketing.
- Personal-tool setups: Nat Eliason's chief-of-staff list, @bot printer/planner setups, Luis Batalha's fridge e-paper newspaper — the Grok-Bot-lore genre; the reward program is the substantive version, already covered.
- Airbnb AI-reviews thread ("textbook study in short term / long term strategy tradeoffs") — commentary on off-beat.
- Starlink/Moon-Mars connectivity ambitions (Michael Nicolls) — space beat, covered there.
- Carmack's mirror-ball OpenGL note, Urschel/Breakthrough Prize, game-design-resources reply, Fourier/Shazam joke — personal/noise.