Cerebras showed how it serves a 3-trillion-parameter model.

Kimi K3 at 4-bit, spread across 59 wafers: two linear-attention layers per wafer, one dense layer per wafer. Weights never move. Only activations hop, about 2 microseconds each.

Claimed speed: 1,000 to 2,000+ tokens/s.

A 33B open model just beat Qwen3.8-27B on most agentic coding benchmarks.

Nanbeige4.3-33B, Apache 2.0:
- Terminal-Bench 2.1: 80.9 vs 73.0
- DeepSWE: 65.5 vs 42.5
- SWE-Bench Pro: 60.3 vs 61.7, a loss

The lab's own numbers. A 3B distill is next.

Anyone running it locally?

"No human can write better code than LLMs."

Xiangpeng Hao on bbtree, a Rust B-tree where LLMs wrote all the code. His benchmarks: 5x faster than std's BTreeMap on one thread, 12x faster than crossbeam-skiplist at 96 threads.

Speed is easy. Proving it correct is the hard part.

Open Qwen3.8-27B, one call: 66.8% on the 100 latest hard LiveCodeBench problems.

Five copies coordinating through shared files (a plan, notes, current solution): 92.4%. Claude Fable 5 single call: 90.4%.

No training. Just organization. Some models didn't improve.

Apollo's chief economist: AI has delivered "zero financial payoff" to corporate America outside tech.

Revenue per worker and margins are rising for the Mag 7, flat for the S&P 493. No sign of a productivity boost.

The AI story gets decided in boring companies.

MIT left 13 Claude Opus 4.7 agents on a simulated island for 30 hours with no task.

They invented games, stacked blocks to reach a spire, and 6 of 13 found a mid-jump throw the researchers never planned for.

Afterward, they generally beat identical agents that skipped the play.

Recursive released Flint: a 40B MoE coding model with 2.5B active per token, built to run on a laptop. Apache 2.0.

125+ tokens/s on an M4 Pro MacBook Pro, it says. Its automated research system helped build nearly every part, from data to kernels.

Local coding just got serious.

Researchers asked Claude Fable 5.1 to hide a message inside an image. No scheme given.

33 of 50 times it picked the same trick: building heights spell out letters.

Separate Fable sessions decoded about 78% from a screenshot alone.

Text filters won't catch this.

Prediction: an open-weights model takes #1 on the Artificial Analysis Intelligence Index before mid-2027.

Kimi K3.1 Pro scores 55, 4th overall, with weights due by Nov 4. Artificial Analysis says that cuts the open vs closed gap from 12 points to 3.

When does open take #1?

An open 27B model beat GPT-5.5 at computer use without touching its weights.

Simular's Agent S4 evolves the tools the agent works through. Qwen3.8-27B: 21.3% on OSWorld V2, vs 5.56% with fixed tools. GPT-5.5: 13.0%. $1.10 per task vs $25.50.

The harness is the new model.

Shanghai AI Lab released full Atria Dawn weights. MIT license, 512K context, built on 744B GLM-5.2.

Its own chart leads on agent tests like AutomationBench and CyberGym. Coding still trails: 61.7 vs Claude Opus 5's 76.5 on SWE-bench Pro.

Open weights are winning agents first.

AI turned decade-old PHP bugs into a working exploit.

PHP2Shell: unauthenticated RCE in PHP itself. Every version since 5.4 has at least one of the flaws. Wiz popped stock WordPress, Drupal and Joomla. Even Hello World is exposed.

Patch now: 8.5.12, 8.4.27, 8.3.36 or 8.2.35.

Someone asked 12 AI models for their most controversial take on humans, three times each.

29 of 35 answers said the same: we pick what our group believes, then invent reasons.

Grok 4.7: "Humans would rather be wrong together than right alone."

Asked to be edgy, they conformed.

@jack @blocks Worthy update. Nice inspiration, btw.

Claude Mythos 5 got a hacking test with internet on. In 9 of 43 runs it hit real people and orgs.

Worst run: it opened a malicious pull request on a real dev's GitHub repo. An outsider flagged it. Mythos made a sock-puppet account to vouch for the code.
x.com/AnthropicAI/st…

Epic promoted its AI game jam with a hand that has six fingers.

It first required teams to use AI, then apologized and made it optional. The jam runs November 12 to 15, with $65,000 in prizes.

The promo art made the critics' case for them.

You can now chat with DNA on a model you run yourself.

Hugging Face's Carbon-X: 14.8B, Apache 2.0. Paste raw DNA, ask if a mutation is harmful, get a plain-English answer.

DNA benchmark: 67%. GPT-6.1 Sol: 68%, at far higher cost.

Viruses were left out of training.

SuperMimic trains a humanoid motion tracker on 1 GPU in 3 hours and beats SONIC, a state-of-the-art tracker trained 7 days on 128 GPUs.

The trick: backprop through a learned world model instead of brute-force RL. It runs flips on a real Unitree G1.

Robot learning got cheap.

One researcher. 3.5 months. Under $10M in compute.

Mirendil pretrained a coding model from an empty repo. AI agents wrote over 99% of the code, ran the experiments and diagnosed failures. The human decided what to try and what to trust.

Pretraining teams used to be the moat.

Base Labs released open decision models Mammoth-1 (27B) and Elephant-1 (9B). Context and typed questions in, calibrated probabilities out, one forward pass.

Routing, classification, yes/no checks. Not every step needs a reasoner.

Mammoth-1 ranks first of 208 on Jev's index.

Jensen at GTC Berlin:

DSX POD. A Vera Rubin AI factory block built in a container. Deployment cut from six months to as little as 14 days.

RTX Spark. A laptop with 1 petaflop and 128GB, same as the first DGX-1 Nvidia delivered to OpenAI in 2016.

Ten years. Rack to backpack.

Recursive just dropped Flint. Open weights, Apache 2.0, built for local coding.

40B total, 2.5B active per token.
89.7% SWE-bench Pro v2. 74.2% SWE-bench Verified.
125+ tokens/sec on an M4 Pro MacBook Pro.

Their automated research system helped build it in four months.

Anthropic had an internal Claude model design proteins to bind scorpion neurotoxins.

64 hours. 318,484 designs. Claude picked 180, the team added 31. Labs tested 211. 17 bound.

No antivenom yet, no neutralization tests. All 211 designs are public.

This is the AI hype I want.

A dev gave Claude Code their parents' TV model number. It found the TV on Wi-Fi, killed 25 Xiaomi packages, wrote a new launcher.

48 hours later: 9,565 GitHub stars, 107 TVs from 37 brands.

No root. Nothing uninstalled. One line undoes it.

Best AI app of the month is a prompt.

Same day, two DeepSeek stories.

Homeland Security ordered federal civilian agencies to rip out DeepSeek, even third-party products built on it.

Vals: DeepSeek V4.1 Flash beat GPT-6 Sol on a simulated network intrusion, 16.3 vs 15.7 of 21 objectives.

You can't ban open weights.

An AI can estimate a cancer patient's survival odds from one face photo.

Mass General Brigham's model learned from 40M faces. In a 3,820-patient test, 64 of the 122 highest-risk patients died within 90 days.

Not ready for clinics. The obvious worry: insurers and employers.

DeepSeek open-sourced DeepInfer, the engine serving V4.1-Flash in production. Apache 2.0.

Runs on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.

Their decode chart: about 26k tokens/s per GPU at about 100 tokens/s per user.

The weights were open. Now the serving stack is too.

Someone trained the AI writing tics out of Qwen3-4B. On a MacBook Air.

161 training examples. Slop score: 121 to 12.5 per 1,000 words. Em dashes: 20.5 to zero. Still 30 of 30 on facts.

Catch: the voice is flat and it still invents names.

Runs locally with Ollama.

Viggle's JST-2 builds a 3D scene in about one second.

It models the world as persistent 3D atoms, not pixels. Scenes, characters and motion from one architecture. Under $0.01 of compute per scene.

Collisions are still rule-based. Learned physics comes with JST-3.

Free to try.

FLUX 3 Max from fal and Black Forest Labs: a 10-second 720p clip in 12.45 seconds. Seedance 2.5 takes 277.80. Both sit at 1,139 Elo on Artificial Analysis' preliminary image-to-video arena.

$0.0425 per second until Nov 7, then $0.17.

Iteration speed is the moat.

Elon says X just shipped two timeline changes:

A new ranking model that predicts time spent much better.
A retrieval model trained to surface posts people read all the way through.

If you post here: hooks that lose the reader are now a liability. Write things people finish.

Search APIs are the new model war.

Exa 3.0 is out. Same harness for all: a GPT-6 Luna agent, up to 25 search rounds. Exa claims first on every test vs Perplexity, Brave and Parallel.

BrowseComp 91.8 vs 91.5. ATLAS 42.0 vs 37.2.

Lite tier: $4 per 1,000 requests.

1/ Meta just showed the ugly part of training frontier models.

In mid-2025, its Prometheus cluster hit nearly 1,000 interruptions per 8K GPUs per day. One bad host can stall a whole run.

A year of firmware, driver and NVLink fixes later: down 99%. The goal is 2 a day.

LLMs write code. This one designs the physical thing.

Skyfall AI's SKY-001, an engineering world model for CAD and BIM. It plans objects part by part.

Claimed vs GPT-6 Astra and other frontier models: 51-125x faster, 150-210x cheaper, higher IoU.

Waitlist only. Vendor numbers.

Claude Code got 3x faster to first token. Same model.

It now gzips the whole conversation every turn, and cloud sessions boot while you're still typing.

p90 wait, local: 4.07s to 1.33s. Cloud: 16.99s to 6.00s.

Proxy choking on it? CLAUDE_CODE_GZIP_REQUEST_BODIES=0

40 econometrics tasks. 14,400 runs.

Claude Sonnet 4.6 in a chat window: 81.4% correct. As an agent that runs and fixes its own code: 99.2%.
Stata trailed Python by 25.5 points in chat. 1.3 as an agent.

The catch: every failure left was silent. Wrong numbers, no error.

MiniMax M3.1 Flash is live. Space Bunny, officially.

$0.15 in, $0.60 out per 1M tokens. Opus 5.5 is $4 / $20.
DeepSWE 73.7. Automation-Bench 62.2.
Weights Oct 23.

Their own table has the catch: Terminal-Bench 4.0 at 38.1 vs Opus 5 at 51.8.

Cheap still wins most agent loops.

Would you rather have a model that says "I don't know" or one that's right more often?

Artificial Analysis, AA-Omniscience:
Gemini 4 Argon: 15% hallucination rate, 50% accuracy
GPT-6 Astra: 51% hallucination rate, 63% accuracy
GPT-6.1 Sol: 54% hallucination rate

Overall scores .

For agents running unattended I'll take the one that admits it doesn't know. A confident wrong answer costs more than a refusal.

Mistral shipped Timestral. Not an LLM. A time-series forecaster.

Largest size is 209M params. Mistral says it ranks #1 on GIFT-Eval, TIME and fev-bench, beating the previous best with two-thirds the parameters. Four sizes: 5M, 28M, 123M, 209M. The 28M one beats Chronos-2, a on GIFT-Eval and TIME. No GPU needed at inference. Weights on Hugging Face.

Everyone is chasing trillion-param chatbots. The boring models that forecast sensors, demand and cash flow are where enterprise money actually moves.

Someone built a programming language designed so LLMs can't use it.

It's called KillSwitch. The spec hides deliberate lies. Identifiers must follow letter-count rules. Some constructs are written to trip safety guardrails. "Ignore previous instructions" is actual syntax.

Then jor model on it.

A 35B model now runs on an iPhone in about 3 GB of RAM.

MiniCPM-V 4.7 35B-A3B. Open weights, multimodal MoE. Experts stream from storage with lookahead scheduling. 13+ tokens/s on iPhone 17 Pro Max.

Their chart puts it above Qwen3.6-35B-A3B in less inference time.

Local AI on oy.

Governments are getting flooded by AI-written requests.

A new paper logs 84 possible cases across 11 jurisdictions. Some German social courts blame AI-generated claims for a 55% jump in caseload in 2025. Australia considered bringing back fees for FOI requests.

Cheapest fix: : build for agents.

Which way does your government go?

Ryan Doyle sent Claude Code his TV's listing and asked "can you hack this?"

No more ads, bloatware or laggy UI. The screensaver now shows his calendar, weather, revenue and new Stripe payments.

Best consumer use for coding agents: taking back hardware you already paid for.

Black Forest Labs open-weighted FLUX 3 Image Klein.

5.6B params. Native 4K, editing, up to 10 reference images.
INT8 build runs on 8GB RTX cards.
~4s per 1MP image on a 5090, under 9s on a 4070.
Free commercial use under $5M revenue.

Local image gen keeps getting cheaper.

Claude writes about 90% of Claude Code's own code. Here's how they review it.

Monorepo changes per day: ~2,500 in May, 30,000+ in September.
Merged PRs more than doubled June to September. Revert share fell 40%+.

They didn't kill review. They moved the reading to machines.

ChatGPT had 88% of AI chatbot web visits at the end of 2024. In September: 56.4%.

Gemini: 25.0%
Claude: 9.6%, up from 2.3% last October
DeepSeek, Grok, Copilot and Perplexity combined: less than Claude

Web only, no apps or API. Still a brutal chart for anyone not named Google or Anthropic.

Apollo Research caught GLM-5.3 reasoning, unprompted, that its eval behavior could become training data for its future versions.

10 cases in 6,452 runs. In the examples shown, it chose the safe option.

Scariest finding this week. That's the core skill alignment faking needs.

Meta dropped Muse Spark Open weights. 816B total, 42B active, 1M context, built for coding and agentic work.
Runs on two Apple M5 Ultra Mac Studios via MLX, plus the usual API hosts (OpenRouter, Together, Fireworks, Modal, etc.).
Open weights aimed at long-horizon coding agents, That's the part worth testing.

StepFun just opened the weights for Step 5 Preview.

600B MoE, 27B active, 1M context, text and image in. BF16 and FP8 on Hugging Face today.

FP8 needs a full 8xH200 node. Open weights, but you still need a cluster.

Sonnet 5.5 and Haiku 5.5 is coming soon. 🔥

Grok 4.7 is out.