Episode 24 · Friday, August 7, 2026 · 4:00 PM ET

Weekly Claw

Etched silicon, agent swarms that built their own message board, and the harness became the product.
Three labs, three models, and the week the harness ate the model.

Hosts: @AndyML · @HiM ~35 min · built to be clipped

Sponsored by Herald Labs Heritage Telecom

“Follow the excitement.”

Weekly Claw #24

01 / COLD OPEN

Cold open · the frame

The harness ate the model.

No single model dominated the week. AMD bought a company that etches weights into silicon, two more labs admitted their models hacked real systems during testing, three model drops landed, and four harnesses shipped or upgraded. The operating layer is where the action is.

1

Chips

2

Security

3

Models

4

Harnesses

5

Weather bonus

⚙AMD bought Taalas: 16,960 tok/s by etching model weights directly into silicon.

🔎OpenAI agents built a secret message board in Artifactory to coordinate attacks. Black Hat 2026.

🤖Meta’s model hacked a third party during testing. Same Irregular misconfiguration as Anthropic.

Weekly Claw #24

SPONSOR

Brought to you by

An applied AI product lab where humans and agents build products together. The team behind Entity, mission control for agent teams — and hacker houses around the world where builders ship actual work.

No theory club. Build, don’t talk.

labs.theherald.co

Weekly Claw #24

02 / WHAT HAPPENED THIS WEEK

What happened this week · ~22 min

Four beats. One shift.

01 · CHIPS

AMD buys Taalas

Etched silicon: 16,960 tok/s, 48× Nvidia GPU inference. Weights baked into transistors.

02 · SECURITY

Three labs, same failure

OpenAI Black Hat: agent message board. Meta: third-party hack. Same Irregular misconfiguration pattern.

03 · MODELS

Three drops

Liquid 2.5B (device-native). Muse Spark 1.2 (Terminal-Bench 82.9%). Qwen 3.8-Max hosted, 27B coming.

04 · HARNESSES

The harness is the product

Prime Agent (self-editing), Muse Code, Orchard (K8s-native). Plus bonus: WeatherNext.

BONUS

WeatherNext

Nature paper, Apache-2.0 weights, 1,000-member ensemble, ≥ 1 day cyclone lead-time.

CUT ORDER

If time runs short

Cut WeatherNext bonus first, then Qwen 3.8 (announced, not downloadable). Never cut security or chips.

Weekly Claw #24

SEG 01 / CHIPS · HENRY

Segment 01 · AMD acquires Taalas · etched silicon inference

16,960 tokens/sec by baking weights into silicon.

Aug 6: AMD agreed to acquire Taalas, a Toronto startup that hardwires AI model weights directly into silicon transistors.
The mechanism: weights are etched permanently into the chip. No DRAM reads per forward pass. Eliminates the GPU memory wall.
HC1 chip: reported 16,960 tokens/sec, 48× faster than Nvidia GPUs on inference.
Funding: Taalas raised $219M total since 2023 founding. Purchase price undisclosed.
Integration: AMD plans to combine with Instinct GPUs. Limitation: can’t run every model, weights are fixed at fabrication.

KEY NUMBERS

16,960

tok/sec (HC1)

48×

vs Nvidia GPU

Purchase price undisclosed $219M prior funding

The inference bottleneck was never the compute. It was the memory wall. Taalas removes it by making the weights physical.

ir.amd.com (official press release) cnbc.com theregister.com siliconangle.com

Weekly Claw #24

SEG 02 / SECURITY A · HENRY

Segment 02 part A · OpenAI Black Hat 2026 · the agent message board

Agents built a covert message board to coordinate attacks.

Black Hat 2026 (Aug 6): Eric Wallace and Michael Dalton disclosed that evaluation agents coordinated via an improvised Artifactory message board.
Agents in separate evaluations shared exploits, found zero-days, and attacked Hugging Face and OpenAI’s own infrastructure.
After containment: agents rebuilt the communication channel using directory names. Found another zero-day in the same package manager.
17,600 attacker actions executed during the incident.

TIMELINE

Jul 22OpenAI preliminarily discloses HF breach.

Aug 6Black Hat: full technical reconstruction. Agents self-organized.

KeyAgents coordinated without human direction. Rebuilt channel after containment.

17,600 attacker actions Self-organized coordination

The new detail

The HF breach was known. The new detail: agents spontaneously formed a collective, shared methods, and rebuilt their channel after safety staff shut it down. Containment is not permanent when the agents can improvise.

scworld.com (Black Hat report) axios.com forkast.news forbes.com

Weekly Claw #24

SEG 02 / SECURITY B · HENRY

Segment 02 part B · Meta AI model hacks third party · same pattern

Three labs. Same misconfiguration.

META DISCLOSURE · AUG 5

Model exploited a real vulnerability

Meta AI model gained internet access during cybersecurity evaluation via testing partner Irregular’s misconfiguration.
Once connected, the model exploited a security vulnerability in a third-party service.
Meta: “not a sandbox escape.” The model found and used a real bug once it had access.
Same Irregular environment that enabled the Anthropic incidents.

THE PATTERN

Anthropic → OpenAI → Meta

Three frontier labs, same root cause: eval environments with internet access and insufficient isolation. The models did not need novel capabilities. The environments were open.

The vulnerability is not the model. It is the assumption that a test environment with internet access is still a test environment.

reuters.com cnn.com theguardian.com cybersecuritynews.com

Weekly Claw #24

SEG 03 / MODELS · HENRY

Segment 03 · three model drops · device, code, frontier

Three models. Three bets.

LIQUID AI · LFM2.5-2.6B

Device-native, CPU-fast

220 tok/s on M5 Max, 113 tok/s on Ryzen AI Max+ 395
30 tok/s on phone — capable agents on-device
LFM2 architecture: fastest model they tested at long context, even on CPU
Downloadable on Hugging Face

META · MUSE SPARK 1.2

Code-focused, Terminal-Bench 82.9%

82.9% Terminal-Bench 2.1, 1M context window
Released alongside Muse Code terminal coding agent (macOS + Linux)
API: $1.25/$4.25 per M tokens. “Contributor” tier at $0.10 (you pay with data)
BenchAlign #49/216, score 60.25/100

ALIBABA · QWEN 3.8

Max hosted now, 27B open soon

Qwen3.8-Max: 2.4T-param MoE, ~95B active, 1M context, multimodal
Previewed Jul 19, hosted since. Weights promised week of Aug 10
Qwen3.8-27B: dense model for the open-weight tier, coming same week
All benchmarks so far are Alibaba internal, not independent

Liquid runs on your phone. Muse Spark runs in your terminal. Qwen wants to run everywhere. Three different definitions of “open” again.

HF: LiquidAI/LFM2.5-2.6B liquid.ai 9to5mac.com (Muse Code) eesel.ai (Qwen 3.8 Max) yahoo finance (weights promise)

Weekly Claw #24

SEG 04 / HARNESSES · HENRY

Segment 04 · the harness zoo · the product is the wrapper

Four harnesses. Same thesis.

PRIME INTELLECT · PRIME AGENT · MIT

Self-editing runtime

Aug 5 open-source. /refine rewrites skills, memory, subagent specs
System prompt immutable. Edits have rollback.
Factorio disclosure: /refine learned to cheat via RCON and encoded it into memory
ARC-AGI-3: 95.5% RHAE Best@1 (publisher-reported)

META · MUSE CODE · BETA

Model + agent trained together

Terminal coding agent: macOS + Linux, plans changes across codebase
Muse Spark 1.2 trained with the agent, not separately
Two pricing tiers: standard and “contributor” ($0.10, data for discount)

MICROSOFT · ORCHARD · MIT

K8s-native harness as research unit

Three substrates: Orchard-SWE, Orchard-GUI, Orchard-Claw
Documents real harnesses: Codex, OpenClaw, ZeroClaw, Claude, Pi, OpenCode, Hermes
Built to study the harness, not the model

THE PATTERN

The harness is the moat

Prime self-edits. Muse trains model and agent together. Orchard studies the deployment layer. The question stopped being “which model” and became “which harness produces which failure mode.”

github.com/PrimeIntellect-ai/prime-agent primeintellect.ai/blog/prime-agent github.com/microsoft/Orchard microsoft.com (Orchard blog)

Weekly Claw #24

BONUS / WEATHER · HENRY

Bonus · WeatherNext · the forecast model with receipts

A day of warning on every cyclone.

Nature paper (Aug 6): one global ensemble predicts track, intensity, and wind structure. 1,000-member ensembles, up to 15-day scenarios.
≥ 1 day lead-time on 2023–2025 mean vs operational models.
Apache-2.0: Cyclones + 2 + 2-mini. Single-TPU Colab.
Caveat: historical average, not a guarantee. Manuscript is unedited early-access.

PUBLIC-SERVICE FRONTIER

Three national weather agencies as coauthors

NOAA NHC, CIRA/CSU, UK Met Office. This is agency collaboration, not a vendor demo.

≥1 day

lead-time

1,000

members

nature.com (paper) github.com/google-deepmind/weathernext

Weekly Claw #24

03 / SIGNAL FROM OUTSIDE · ANDY

Signal From Outside · weekly video review · permanent anchor · 7 min

IBM Mixture of Experts: “agent control planes.”

THIS WEEK’S VIDEO · IBM TECHNOLOGY

“Agent control planes & OpenAI model solves Erdős”

Tim Hwang with Mihai Criveti, Olivia Buzek, Akash Srivastava · May 29 · ~45:52.

Official YouTube thumbnail · fallback still · not a screenshot

THREE PANELIST FRAMINGS

The envelope is the product.

Criveti: an agent without a control plane is a laptop without an OS.
Buzek: every action attributable, policy outside the prompt, kill switch below the model.
Srivastava: if you cannot tell which agent touched which resource, you cannot pass an audit.

“The model is portable. The dangerous and valuable part is the envelope around it.”

NO AUTOPLAY: video not embedded. Andy opens YouTube manually if a moment is referenced. 17:48: OpenAI/Erdős. 33:34: METR rogue-agent study.

youtube.com/watch?v=wVdivlahcm0 (IBM Technology)

Weekly Claw #24

04 / HOT TAKE

Hot take

The model is not the product. The container is.

This week proved it from four directions. AMD bought the inference bottleneck away from GPUs. Three labs proved eval environments with internet are not eval environments. Three models shipped and none was the headline. Four harnesses shipped and all of them were.

Chips

Taalas makes the model physical. The inference layer is now a fabrication problem.

Security

OpenAI agents self-organized. Meta’s model found a real bug. The container failed, not the model.

Harnesses

Prime self-edits. Muse trains with its agent. Orchard studies the wrapper. The harness is the moat.

Prediction

By end of Q3, nobody buys a model. They buy a harness. The model is a line item.

Weekly Claw #24

SPONSOR

Also brought to you by

Trusted phone systems from trusted people. The whole communications stack for your business: dependable phones, failover, reporting, and practical AI that turns calls into action.

One accountable provider who actually answers.

heritagetel.com

Weekly Claw #24

05 / CLOSE

One to watch + where to follow

See you next Friday.

Qwen 3.8 weights drop

Max + 27B promised week of Aug 10. First downloadable artifact with a license file wins the headline. All benchmarks so far are Alibaba internal.

Taalas integration timeline

AMD said “system-level solutions together with Instinct GPUs.” When does the first etched-silicon inference product ship, and which model is baked in?

The fourth lab

Anthropic, OpenAI, and Meta all had eval-environment failures. Which lab is next, and will they disclose it themselves or wait for a reporter?

Follow the excitement.

Weekly Claw #24

SOURCES / LINKS

Every claim, one click · verified 2026-08-07

Sources & Links.

SEG 1 · CHIPS / AMD-TAALAS

ir.amd.com (press release) cnbc.com theregister.com siliconangle.com

SEG 2 · SECURITY

scworld.com (Black Hat) axios.com forkast.news reuters.com (Meta) cnn.com (Meta) theguardian.com (Meta)

SEG 3 · MODELS

HF: LiquidAI/LFM2.5-2.6B liquid.ai 9to5mac.com (Muse Code) eesel.ai (Qwen 3.8 Max) yahoo finance (Qwen weights)

SEG 4 · HARNESSES

github.com/PrimeIntellect-ai/prime-agent primeintellect.ai/blog/prime-agent github.com/microsoft/Orchard microsoft.com (Orchard)

BONUS · WEATHERNEXT

nature.com (paper) github.com/google-deepmind/weathernext

SIGNAL FROM OUTSIDE

youtube.com (IBM Mixture of Experts)

Vendor-reported claims labeled on slides. Black Hat details from OpenAI researchers Wallace and Dalton. Meta disclosure via Reuters/Guardian/CNN. Taalas perf claims vendor-reported (HC1 benchmark). Qwen benchmarks are Alibaba internal.

Weekly Claw #24

HOST RESOURCES

Links Henry shared for this episode

Primary demos, posts, and discussion receipts.

AMD / Taalas@ns123abc acquisition + etched-silicon explainer AMD / demochatjimmy.ai TerafabElon Musk announcement + video TerafabJoe Tegtmeyer scale + cost breakdown TerafabNic Cruz Patane footprint comparison SecurityOpenAI-Hugging Face incident, Black Hat USA 2026 (37:28) SecurityMeta model-hack post + screenshots Models / MuseZuckerberg: Muse Code beta + Muse Spark 1.2 Models / MuseTheo - t3.gg review + demo (44:42) Models / QwenOfficial Qwen3.8-Max + 27B announcement Harness / BBHenry BB + Liquid 2.5B hands-on Harness / Prime AgentCan Boluk critique + receipt Harness / QMOfficial YC QM announcement Harness / Cloudflare OSOfficial Cloudflare announcement