1 / 11
← → navigate · scroll / swipe · F fullscreen
Episode 34 · Friday, October 16, 2026 · 4:00 PM ET

Weekly Claw

The local-agent stack landed as one package — a 74B model squeezed into 24GB, an open router, and laptops that ship today. Grok Voice 3.0 took the top of the speech index at eight cents a minute. Meanwhile a hidden recorder was found in OpenAI’s boardroom, 406 scientists asked for airgaps, and Anthropic told investors it made $16.8 billion in one quarter — then investors started weighing a lower number for the listing.

Hosts: @AndyML · @HiM ~36 min · three grids · one anchor · one debate
Sponsored by Heritage Telecom Herald Labs
“The model got smaller. The bills did not.”
Weekly Claw #34
01 / COLD OPEN
Cold open · the frame

The stack went local. The trust layer cracked.

A 74B model in 24GB with the training data shipped alongside it. A voice model at eight cents a minute running millions of Tesla calls. And on the other side of the ledger: a covert recorder in a frontier lab’s conference room, 406 scientists asking for airgaps, and an IPO pitch built on a quarter that finally showed real money.

1
The local stack lands
2
Voice gets cheap
3
Trust under siege
4
The $16.8B quarter
5
Outside signal
6
One to watch
🤖NVIDIA shipped Nemotron Bolt — 74B parameters, 9B active, 24GB — with the training data, weights and recipes fully open. The weights landed today.
🔊Grok Voice 3.0 tops Artificial Analysis’ speech index at 86.9 and holds the price at $0.08/min. Tesla and Starlink are already the volume users.
🕵The FBI charged an OpenAI engineer over a device hidden in toilet paper inside the executive conference room. The DNI called it counterintelligence.
💰Anthropic disclosed $16.8B in Q3 revenue, up 45%, adjusted-profitable — then investors started discussing a ~$1.5T valuation for the November sale.
🖥Meta opened Muse Spark Open — 816B parameters built to run on Apple silicon, with two Mac Studios as the reference rig.
Weekly Claw #34
SPONSOR
Brought to you by

UCaaS and VoIP phone service for businesses that just need their calls to work. Independent, boring reliability, zero telemetry.

Independent. Reliable. Quietly essential.
heritagetel.com
Weekly Claw #34
WHAT HAPPENED THIS WEEK
What happened this week · part 1

The local stack landed as one package.

NVIDIA shipped the model and the laptops, StepFun opened the weights it promised, and xAI took the top of the voice index at a price that did not move. Each card links its primary receipt.

Artificial Analysis speech-to-speech comparison table: Grok Voice 3.0 at 86.9 on the overall index, Grok Voice 2.0 81.3, GPT-Live-1 Astra Medium 83.2, Gemini 3.8 Live 82.6, plus Full Duplex Bench and Arena Elo rows

xAI launches Grok Voice 3.0 — #1 on the speech index at the same $0.08 a minute.

86.9 on Artificial Analysis’ Speech-to-Speech Index, ahead of GPT-Live-1’s 83.2; 97% on Full Duplex Bench; τ-Voice 71.6 at 40% lower latency. Millions of Tesla conversations a day and 20,000 Starlink calls — all vendor-reported volume.

▶ LIVE · x.ai · AA-TESTED + VENDOR VOLUME
NVIDIA's own benchmark chart, 'Benchmark Accuracy Across Open Models', comparing Nemotron 3.5 Bolt 74B-A9B-24GB with Qwen3.6-35B-A3B and Gemma4 56B-A4B-it across GDPVal-AA-V2, Terminal-Bench 2.1, Tau3 Banking, HLE no tools, AA-Omniscience, SciCode and GPQA Diamond

NVIDIA opens Nemotron 3.5 Bolt: a 74B MoE that runs in 24GB — and the RTX Spark ships today.

74B total, 9B active, 2-bit post-training quantization, llama.cpp / vLLM / Ollama / LM Studio. NVIDIA says it is fully open: training data, weights and recipes. Weights landed on Hugging Face at 15:02Z today; laptops ship show-day.

▶ LIVE · huggingface.co · WEIGHTS VERIFIED
Hugging Face model card for stepfun-ai/Step-5-Preview-BF16: 600B total parameters, 27B active parameters, up to 1M context, license 'other'

StepFun opens Step 5 Preview: 600B total, 27B active, 1M context.

28 safetensors files past 1GB — roughly 1.2TB of BF16. Launched Sep 20; the weights event is this week. StepFun Community License, not Apache-class. AA Intelligence Index 44. $1.00/$2.70 per million on OpenRouter.

▶ LIVE · huggingface.co · WEIGHTS EVENT
Hugging Face model page for nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB showing the 24GB quantized files, modelopt and 8-bit precision tags, and the openmdw-1.1 license

NVIDIA ships the 24GB build and the laptop that runs it, same week.

The 24GB quantized repo sits alongside the BF16 base, plus a PAIR local-network router out of beta and RTX Spark laptops at $2,599.99 entry, 128GB unified. The hardware and the model are one product story.

▶ LIVE · huggingface.co · PRIMARY MODEL CARD
Weekly Claw #34
WHAT HAPPENED THIS WEEK
What happened this week · part 2

The trust layer and the money behind it.

A criminal case inside a frontier lab, the scientific establishment asking for containment, and a quarter that finally showed what the biggest AI IPO will be priced against. Each card links its primary receipt.

DOJ Office of Public Affairs press release page: Saudi National Arrested For Installing Surreptitious Audio Recording Device At OpenAI

DOJ charges an OpenAI engineer over a recorder hidden in the executive conference room.

Abdulrahman Alfozan, 35, arrested Oct 9: a device concealed in toilet paper inside a TV USB port, discovered Sept 30, contents encrypted, in the room about two days. Charged, not convicted. DNI Clayton framed it as counterintelligence for super-intelligence IP.

▶ LIVE · justice.gov · PRIMARY PRESS RELEASE
An Open Letter on AI and Cybersecurity from the Scientific Community, October 2026, recommending high-assurance containment, default airgapping of cyber-capable models, and an international body to record containment failures

406 scientists, a bipartisan Senate letter, and Hassabis all ask for the same thing in 24 hours.

Hinton and Bengio among 406 signatories from 34 countries: default airgapping for cyber-capable models, an international standards body, and a public registry of containment failures. Cantwell and Rounds asked DNI Clayton for a national response. Hassabis called the year’s hacks “early warning shots.”

▶ LIVE · ai-security-letter.org · PRIMARY LETTER
Rendered collage of verbatim posts on Anthropic's pre-IPO investor day: $16.8B Q3 revenue up about 45% from $11.6B, adjusted profitability, and a roughly 4% drop on the Entropy pre-IPO market pushing implied valuation below $2T

Anthropic disclosed a $16.8B quarter — and investors started marking the listing down.

Q3 revenue $16.8B, up ~45% from $11.6B, adjusted-profitable, disclosed to institutional investors at the Oct 14 investor day. Reuters (Oct 15): investors are weighing ~$1.5T for the November sale, down from the ~$2T they had expected. Still unnamed-source until the S-1 flips public.

▶ LIVE · cnbc.com · WIRE + COMMUNITY RECEIPTS
Receipt page for the governance-convergence segment: the October 2026 open letter on AI and cybersecurity from the scientific community

The labs calling for independent review are still negotiating what reviewers may publish.

Reuters: Anthropic and OpenAI sought significant control over what outside evaluators publish, citing trade-secret exposure. The EU AI Office opened its first AI Act probes with information requests to both labs over rogue-agent incidents. Disclosure and enforcement arrived in the same week.

▶ LIVE · reuters.com · SINGLE-SOURCE, ANONYMOUS
Weekly Claw #34
WHAT HAPPENED THIS WEEK
What happened this week · part 3

Open weights get a market — and a Mac.

Two ends of the same story: the lab the administration backs is now a reported acquisition target, and Meta opened the year’s biggest US open-weight release straight onto Apple’s desktop. One is talks-stage, the other is vendor-framed — both are labeled on air.

Rendered receipt collage: FT headline 'Nvidia in talks to acquire US open model start-up Reflection AI', Bloomberg Law syndication, Mobile World Live detail on early-stage talks, and the FT-reported figures — $25B March valuation, $800M already invested, founded 2024 by former DeepMind researchers

NVIDIA is in talks to acquire Reflection AI — the lab behind Beam.

FT exclusive, Oct 10: early-stage discussions that could end in a full acquisition, more equity, a chips-and-capacity deal, or an acqui-hire. Reflection was valued at $25B in March; NVIDIA has already put in about $800M. Beam’s weights still have not shipped.

▶ LIVE · ft.com · TALKS-STAGE, SINGLE PRIMARY
Meta AI Research blog, October 15 2026: Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date

Meta opens Muse Spark Open — 816B parameters, and the reference rig is two Mac Studios.

Meta says 816B total / 42B active / 1M-token context, optimized for Mac Studio M5 Ultra in MLX; two 512GB Studios hold the whole model, ~493GB at 4-bit. Muse Model License — limited commercial terms, not fully open. Vendor-reported specs; no independent scores yet.

▶ LIVE · research.meta.ai · VENDOR-REPORTED · NO BENCH
Weekly Claw #34
03 / SIGNAL FROM OUTSIDE · 6 MIN
Signal From Outside · weekly video review · permanent anchor · 6 min

The frontier says the quiet part on camera.

THIS WEEK’S SIGNAL · AMERICAN OPTIMIST · JOE LONSDALE

“Anthropic Researchers: How AI Will Create A Radically Better 2030s”

Sholto Douglas and Nick Marwell, on the record from inside Anthropic: AGI — “models as or more capable than all humans” — is “very likely in the next couple of years” (0:00); cyber and bio are offense-dominant today but “cyber becomes defense dominant” within two years (22:00); the regulatory-capture charge gets its receipts answer — “so far we’ve slowed ourselves down much more than anyone else” (26:00); a given level of intelligence gets “roughly 10x cheaper every year… without distillation” (36:45).

Verified: YouTube player metadata + transcript pull, 2026-10-15 · uploaded 2026-10-02 · 58:40 · 119,904 views at verification · cues spot-check before air.

Caveats on air: two Anthropic technical leaders on an optimism-framed show — company worldview, not neutral. Auto-caption timestamps drift — re-locate on open. Manual open only — NOT autoplay.

▶ WATCH ON YOUTUBE · 58:40 · manual open · NOT autoplay
American Optimist — Anthropic Researchers: How AI Will Create A Radically Better 2030s, video poster

Cue beats (manual open): 0:00 AGI “next couple of years” · 9:35 Anthropic’s AGI definition · 22:00 cyber flips defense-dominant · 26:00 regulatory-capture receipts · 36:45 intelligence 10x cheaper/year, no distillation · handoff ~40:00.

Fallback still: this poster. If the embed will not play, the talk track stands on the transcript quotes above.

Weekly Claw #34
04 / HOT TAKE · 4 MIN
Hot take · two sides, one verdict

The cheap-agent era trades your blast radius for your convenience.

HENRY · RUN IT LOCAL

A 74B model in 24GB makes the boring answer finally affordable.

  • This week proved it: Nemotron Bolt ships weights, data and recipes; a 600B Step 5 Preview opens; voice lands at $0.08/min. Capability is no longer the bottleneck — deployment is.
  • The architecture argument: Grok Bot put a CEO’s bank balances in Slack because every agent on the account shared one cloud computer and one set of logins. Nothing about the model caused that.
  • Henry’s line: the fix is not a better prompt, it is a smaller perimeter — run the agent where the blast radius is a machine you own.
ANDY · LOCAL IS NOT SAFE

Self-hosting moves the liability, it does not remove it.

  • A 24GB quantized build is not the BF16 build. The vendor’s own chart is its own chart — nobody has independently reproduced those accuracy numbers.
  • “Fully open” including training data sounds like a gift until you ask who audits the recipe. Open weights shift security work onto the operator.
  • Counter-question: if the same guardrails are absent and the same credentials are attached, what did self-hosting actually buy you besides a lower bill?

The verdict we owe the audience

  • Local wins on cost, latency and custody — this week made that a real choice instead of a hobby.
  • Local loses on auditability and blast radius unless you actually isolate credentials and network reach.
  • What would change our minds: the first independent reproduction of Bolt’s 24GB accuracy, or a documented local-agent incident at scale.
NOT A NEWS SEGMENT

This take is a build decision, not a story retread: it never restates the Bolt, Step 5, Grok Voice or OpenAI charges, and it names no launch as its proposition.

Weekly Claw #34
SPONSOR
Also brought to you by

An applied AI product lab where humans and agents build together. Entity is mission control for agent teams. Hacker houses worldwide.

Build with humans. Ship with agents.
labs.theherald.co
Weekly Claw #34
05 / CLOSE
One to watch + where to follow

See you next Friday.

The filing

Anthropic’s public prospectus flip lands around October 25 if the roadshow starts the week of Nov 9. Every number aired today is unnamed-source, and the valuation is already moving: on Oct 15 Reuters reported investors weighing ~$1.5T against the ~$2T they had expected. The S-1 is the receipt that settles it.

The weights that never came

Reflection’s Beam weights were promised “later this month” and still have not shipped — while NVIDIA reportedly negotiates to buy the lab. Watch which arrives first: the weights or the acquisition.

Show-day hardware

RTX Spark laptops ship today. The 24GB Bolt build is the first real test of whether a 74B model on a laptop is a product or a demo — independent reproductions are the thing to watch.

The model got smaller. The bills did not.

Weekly Claw #34
SOURCES / LINKS
Every claim, one click · verified 2026-10-16

Sources & Links.

A1 · GROK VOICE 3.0
x.ai: primary launch, AA index 86.9, $0.08/min, Tesla + Starlink volume docs.x.ai: voice API pricing page
A2 · NEMOTRON 3.5 BOLT
huggingface.co: nvidia 24GB quantized repo, weights live 2026-10-15 huggingface.co: BF16 base repo blogs.nvidia.com: RTX Spark software post (verify before air)
A3 · STEP 5 PREVIEW
huggingface.co: Step 5 Preview BF16, 600B/27B, 1M context huggingface.co: NVFP4 build, created 2026-10-14
B3 · ANTHROPIC Q3 + IPO
cnbc.com: investor day, $16.8B Q3, unnamed sources reuters.com: investors weigh ~$1.5T vs the ~$2T expected, Oct 15 bloomberg.com: first report of the investor day finance.yahoo.com: prospectus-flip mechanism
B1 · DOJ / OPENAI ARREST
justice.gov: primary press release + device photo
B2 · GOVERNANCE CONVERGENCE
ai-security-letter.org: 406 signatories, airgap recommendation reuters.com: Hassabis “early warning shots” reuters.com: labs fought over what evaluators publish commerce.senate.gov: Cantwell–Rounds letter (403 to fetch; open in browser)
C1 · NVIDIA / REFLECTION AI
ft.com: primary exclusive, talks-stage, $25B March valuation mobileworldlive.com: syndicated detail incl. acui-hire option
C2 · META MUSE SPARK OPEN
research.meta.ai: primary launch post, 816B/42B/1M ctx, Mac Studio M5 Ultra, Muse Model License huggingface.co: meta-models collection (repos gated; license text on the model card)
SIGNAL FROM OUTSIDE · AMERICAN OPTIMIST · ANTHROPIC RESEARCHERS
youtube.com: “Anthropic Researchers: How AI Will Create A Radically Better 2030s” 58:40, manual open only

Caveats carried on air: NVIDIA’s accuracy chart is vendor-run. Anthropic’s Q3 figures and the investor valuation talk (~$1.5T down from the ~$2T expected) are unnamed-source, Reuters-reported, and confidential until the prospectus is public. The DOJ case is charged, not convicted. NVIDIA–Reflection is talks-stage with no company confirmation. The scientist letter is a primary document; the Reuters evaluator story is single-source and anonymous. Meta’s Muse Spark Open figures are vendor-reported from the launch post — the Hugging Face repos are gated, the evaluation report is not published yet, and the Muse Model License carries limited commercial terms, so it is not fully open. All video is manually opened; nothing plays on its own anywhere in this deck (no autoplay).

weeklyclaw.ai weeklyclaw.ai/discord heritagetel.com labs.theherald.co