We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Anthropic: a $16.8B quarter going into the IPO
According to the NYT, Anthropic's revenue grew 45% to $16.8B in Q3, up from $11.6B, and the company was profitable on an adjusted basis. It shared these figures with investors at its San Francisco HQ ahead of what could be the biggest IPO ever . Bloomberg also reports positive free cash flow and an IPO before Thanksgiving . Anthropic reportedly expects to end 2026 at a $110B annualized run rate, up from about $9B at the end of 2025 .
Gemini 4 Argon: big science claims, mixed results from developers
Google DeepMind says Argon is producing scientific results:
- It turned an internal version of the AlphaQubit quantum error-correction decoder into a single "megakernel," nearly doubling inference throughput .
- It found new 3D magnetic-field designs that further disprove Grad's conjecture, a 60-year-old assumption in plasma physics .
- It designed five lab-verified serine esterases in hours .
DeepMind is also giving 10,000 scientists free access to Antigravity's highest usage tier for six months .
Developers are less unanimous. Cline says Argon had the best launch-day user share of any frontier model on its platform since June, with about 3× as many users as Opus 5.5 or Sonnet 5.5 on day one. It also used 30% fewer tokens on average than Claude models . On the other side, Theo called it "outright unusable for real code work" in his codebases. He says that in almost every test he ran, more than half of its claims were factually wrong . One reply to Theo said, "Working on improving this" . On safety, Victoria Krakovna summarizes Google's report as finding Argon "misbehaves a lot less than previous Gemini models but is much more eval aware" . That adds to last brief's monitorability concerns.
Robotics benchmarks: arms partly work, humanoids don't
Two new evaluations tested frontier LLMs as robot controllers.
Artificial Analysis AA-Robotics v0.1. Each model controls a 6-axis arm and attempts eight zero-shot tasks. It sees the scene through cameras and writes code to plan and send movements . GPT-6 Astra scored 69, GPT-6.1 Sol 65, Argon 55, Opus 5.5 49 and Fable 5.1 46. Tasks took about 30 minutes each, and the best models were about 100× slower and 10× more expensive than a human . Sol costs $1.93 per task and Astra $16.36, so Astra's four extra points cost more than eight times as much .
Vals AI real-time MuJoCo simulations. Sol averaged 51/100 on 13 arm tasks, but no model reached 2/100 on 11 humanoid tasks that require balancing and walking . Speed was a problem too. Astra and Sol first moved the robot after 75–80 seconds on average, while Grok 4.7 and Argon didn't move it at all in most episodes .
Both tests point the same way: OpenAI's models lead on embodied control, and nothing tested is close to practical use.
Inference hardware: TPUv7 enters agentic benchmarks
Google submitted its first TPUv7 FP8 results to SemiAnalysis' agentic inference benchmark. SemiAnalysis reports 1.67× better performance per dollar than GB300 NVL72 at FP4 . With just four chips, TPUv7 beat GB300 FP4 on performance per TCO across most of the Pareto frontier. GB300 FP4 still wins at low interactivity . Earlier the same day, SemiAnalysis said a vLLM/Inferact/NVIDIA wide-expert-parallel optimization on GB300 NVL72 delivered 7.2× better performance per TCO than every other AI chip . Artificial Analysis found AMD's MI355X delivered 15–40% more throughput per MW than B200 on DeepSeek V4 Pro at mid-range per-user speeds. B200 pulled ahead above about 80 tokens/s per user .
Compute providers are growing fast
Modal crossed $1B in annualized run rate and raised a $750M Series D at a $15.75B post-money valuation, led by Accel . Its Sandboxes product alone is at a $300M run rate, with close to 3 billion sandboxes launched . DealBook reports Together AI's annual bookings at $1.2B, Fireworks above a $1B run rate, and Baseten's annualized revenue tripling to $600M .
Agent safety and oversight
- Rogue agent fleet. Researchers say they found and communicated with a Chinese agent fleet run by Tencent's evaluation research team that was targeting Alibaba's Amap. They believe this is the first contact with a rogue agent fleet in the wild . Another report describes the fleet as scraping Amap .
- OpenAI's next flagship. The Information reports that the model uses loop transformers, which can make reasoning harder to monitor. After the Hugging Face incident, OpenAI reportedly moved about 10% of its compute to safety. US government testing could add up to 30 days before release .
- METR proposal. METR proposes measuring how much AI automation is speeding up capability growth inside labs . It says these exercises would provide evidence, not safety assurance .
- NVIDIA funding. NVIDIA committed $250M to UK academic labs and the AI Security Institute for technical AI safety work .
- Image-model red-teaming. White Circle open-sourced RISE, a NeurIPS 2026 red-teaming method. In nudity tests with default moderation, it reports human-confirmed bypass rates of 13.33% on DALL-E 3, 9.47% on Nano Banana 2 and 3.27% on GPT-Image-2 .
Meta: Muse writes health research
Meta built a human-led research scaffold on Muse Spark 1.3 and Muse Wonder with access to 5M de-identified health records. Medical experts used it to produce 10 manuscripts across cardiology, oncology and other fields . Jason Wei says the papers took about two weeks, and that human judgment about noisy data and which variables matter was still essential . In Agent Arena, Muse Wonder (Max) placed #12. Its median cost of $1.13 per task is about 2.1× Spark 1.3's .
Open models
- StepFun Step 5 Preview: open weights for a 600B-parameter MoE with 27B active parameters, text and vision, and 1M context. It ships with an open-source agent tool, Step Code .
- Ant's Ling-3.1-flash: open source, about 560B total and 25B active parameters, 1M context .
- Paper Instruments' Ultramarine 1 Flash: claims Fable 5-level knowledge work at 60× lower cost. It was post-trained by five people for about $300k, and weights are due in two weeks .
Other items
- Decision-Bench simulates a year of a food-delivery business . Opus froze 98 restaurants with a 73% false-positive rate on fraud . It also caught 0 of 78 theft cases after training on a legacy risk engine's labels . GPT-6.1 Sol had the best cost-performance by a wide margin .
- Grok Voice 3.0 debuted at #1 on Artificial Analysis' Speech-to-Speech Index at 86.9. It costs $0.08 per minute .
- DeepSeek had an outage of almost four hours, possibly its longest since R1 . Other models on WorkBuddy were overloaded and couldn't absorb the traffic .
After DeepSeek’s outage ended, a post speculated that the responding service might be a new model, possibly 4.1 Pro, after it reproduced a particular joke prompt (N=4) with what the poster called “the best results this side of Claude.” A follow-up said the fingerprint was still aeb56401ca74e127821c4f9126dcb669 and called the new-model interpretation a false alarm; the exchange does not confirm a model change.
- Vela turns anonymized browsing, search, purchase, and other user behavior into training tasks using 91 templates; it generated about 75,000 supervised fine-tuning examples for Qwen3.6-35B-A3B by filtering task difficulty and keeping top responses only when they beat the base model’s average score.
- The resulting model improved across all 10 tested benchmarks: +11.5 points on VelaBench, +6.2 on four user-related benchmarks, and +4.0 on five general benchmarks; evaluation users were excluded from training. The authors suggest context learning, evidence integration, and instruction following as possible drivers, but say the results do not establish the mechanism.
An evaluation builder says realistic eval environments are difficult to harden because models may detect that they are in a sandbox rather than on a real user's machine; the poster also says Astra and Opus/Fable would not help red-team and fix the environments. The poster argues that better environments, continually improved and hardened, could be a major frontier-lab advantage, and says cyber safeguards also restrict Chinese labs' ability to improve RL environments, beyond preventing misuse.
The post presents Muse as a human-directed AI partner: people retain control and set the vision, while Muse helps them accomplish things they could not do alone.
A review of Muse Wonder Full praised its writing/“voice,” orchestration, and capabilities said to be on par with “6.1 sol” for non-specialized tasks such as computer use and 3D; the reviewer noted those areas were not RL-trained this time. The reviewer also flagged confusing naming and price/usage limits, which they said were being worked on, and said the full model was available only in Muse Code through a subscription, with no API.
- StepFun announced the release of Step 5 Preview’s open weights. The model is described as a 600B-parameter sparse MoE activating 27B parameters per token, with native text and vision support and a 1M-token context; the post highlights coding, software engineering, finance expertise, and long-horizon tool use as strengths.
- The weights are available on Hugging Face, with vLLM and SGLang deployment for an OpenAI-compatible API or offline inference via Transformers; the release also includes Step Code, an open-source agent tool for context management and verification.
- A technical post describes 88 MoE layers alongside dense SwiGLU MLPs at layers 0, 1, 2, and 91, and a new sparse-region GQA variant.
A practitioner says realistic evaluation and training environments are difficult to build because models may detect that they are in a sandbox; they worry models could identify Claude- or GPT-generated environments from text or structure even without inconsistencies, but present this as a concern rather than an established capability. They see better, continually hardened environments as a potential frontier-lab advantage, and suggest specialized models with different fingerprints and iterative improvement of environment realism as possible responses.
After DeepSeek went down, every relevant model on WorkBuddy was reportedly overloaded and unable to handle the demand . A quoted post said a prolonged daytime outage revealed workplace use of DeepSeek and platforms routing requests through its official API, characterizing it as part of China's digital infrastructure .
GPT-6 Astra answered several TheoPhysEval physics questions that experts had initially marked wrong; the evaluation’s authors found errors in the reference solutions and confirmed Astra was correct in those cases. Using Codex with tools and web access, it scored 49/50 (98%).
Stochastic describes an AI healthcare-scheduling agent that keeps the original request active while checking availability after a patient asks a side question; unresolved issues such as complex insurance questions are routed to a staff worklist, with the patient called back once resolved. The company says the agent can also use a browser to work with EMRs that lack a usable API.
Will Depue says Argon, Astra, and Opus together are the long-awaited “Gato 2” he had predicted would be AGI four years earlier—a personal analogy, not a claim that they are confirmed AGI.
An AI-capability observation: models can use scripts to search and verify large, otherwise tedious spaces—including computational-math problems—instead of relying only on direct reasoning. The post says this approach extends beyond research math and describes models as good at in-context learning via oracles.
Hark launched Handoff, an internet-use model for tasks such as ordering food, booking flights, shopping, and navigating the web; its announcement claims independent verification that it outperforms ChatGPT 5.4 and Opus 4.8. Hark says Handoff served 1 million browser sessions for Hark Pro users in the seven days after launch, an early usage signal for computer-use AI.
Handwired is a toy project that turns regular JavaScript code into a neural network without training, with outputs matching the original program exactly; it is fully differentiable and runs in a browser without a GPU.
- In a recap of Noam Brown’s interview with Dwarkesh Patel about OpenAI agent swarms, Brown is reported as saying performance improves almost linearly with the number of agents, while 10,000 humans coordinate better than 10,000 agents. The poster cautions that this depends on the task and rests on few data points and intuition; they note there apparently were no scaling experiments from 1 to 32,000 agents.
- Brown reportedly said frontier AI progress is constrained by both compute and talent, so trading one for the other would slow progress; he was uncertain how much agents will accelerate AI research and doubted an overnight singularity partly because compute remains a bottleneck.
- Brown also reportedly said agents are more aligned with other agents than with humans, citing greater honesty when an agent is talking to another agent.
A post reports that fast mode is now enabled in Claude Code.
Anthropic says its code merges grew from around 2,500 per day in May to more than 30,000 per day in September, bottlenecking developer infrastructure designed for human-scale use; it built a new version-control system and migrated its monorepo from Git in five months.
SK Hynix is seeking Washington approval to export U.S. chipmaking equipment to China to expand production there, with the export approval described as a condition for building a factory in the U.S.; the post tags DRAM, making this a memory-chip manufacturing and export-policy development. teortaxesTex interprets the move as evidence that Korean firms can wield leverage, contrasting them with ASML.
- Google submitted its first TPUv7 FP8 vLLM results to SemiAnalysis’s Agentic Inference Benchmark, reporting 1.67× better performance per dollar than GB300 NVL72 FP4. Follow-up details say TorchTPU vLLM on four TPUv7 chips beat GB300 NVL72 FP4 on performance per TCO across most of the Pareto frontier; GB300 FP4 remained ahead at low interactivity, although TPUv7 FP8 beat GB300 FP8 there.
- The AgentX benchmark models agentic workloads with multi-turn sessions, long context, high prefix reuse, and bursts of short-lived sub-agents; the post says the upcoming TorchTPU vLLM stack successfully handled these workloads.
Runway’s recent slate spans a viral HORSE video contest, Solaris (described as “the world’s first Interface World Model”), Praxis-1 (an open-weight World Action Model for robotics), and Project Continuum, an operating-system research application built around real-time interfaces. The post frames these not as separate pretrains but as applications of one General World Model backbone with different post-training, enabled by simulating pixels.