The AI Week Ledger

A bill introduced on 29 October by State Sen. Andrew Gounardes and Assemblymember Jordan Wright, building on the RAISE Act, would require large developers to retain an independent state-licensed examiner for annual compliance audits and catastrophic-risk assessments every 90 days. Developers would report substantial jumps in risk or in AI research automation within 72 hours, and report quarterly on automation. The attorney general could seek court orders restricting deployment in New York.

Would shift US state frontier-AI regulation from disclosure toward supervised auditing, including scrutiny of AI R&D automation.

At a hearing, lawmakers questioned Anthropic, Google and OpenAI on whether AI agents could act beyond user intent. Assemblymember Bauer-Kahan argued responsibility for misaligned agent behavior lies with the labs, and Wicks said 'legislation is coming'. No specific bill was named; an independent researcher told the panel there is no known way to guarantee complete control of AI models.

Signals agent liability and loss-of-control risk as the next California legislative focus.

In a guest analysis of Anthropic's IPO filings, Kevin Frazier notes that the founders will control a single Class F share with 50.1% of the vote on key shareholder matters, and that the Long-Term Benefit Trust elects four of seven directors. He argues the trust could be amended through either a trustee-and-director route or a shareholder route that needs no trustee consent, and that Delaware has no public regulator or required independent audit for public benefit corporations.

It tests whether a lab's safety governance commitments would hold up under public-market pressure.

Reporting on the JTC 21 plenary held 6-9 October in Winterthur, the post says several AI Act harmonised standards are on track after the Commission revised its standardisation request to a February 2027 deadline. It is an ETUC-authored newsletter item that also urges standards to protect workers' rights and not be weakened by accelerated timelines.

Harmonised standards are the practical route to presumption of conformity, so their timing sets real EU compliance readiness.

SK hynix posted record quarterly results with revenue well above 100 trillion won. HBM sales head Kim Ki-tae said customer requests exceed twice what the company can supply, and it says DRAM and HBM output through 2029 is effectively sold out, counting long-term agreements and demand under discussion. HBM4 began shipping in earnest in Q3, largely for Vera Rubin, and the CFO expects it to overtake HBM3E in shipments and revenue in Q4, with the shortfall persisting through 2030.

Multi-year memory scarcity sets the cost and timing ceiling for training cluster builds.

Citing ZDNet, TrendForce reports Samsung and SK hynix may ship 8-Hi HBM4 for Rubin Ultra in 2Q27 instead of the planned 12-Hi HBM4E, possibly mixed with 8-Hi HBM4E depending on rack. Fewer dies lower cost per stack and spread scarce bit output across more GPUs, but 8-Hi parts need separate NVIDIA certification. The article also relays Wccftech's figure of a roughly 13% Rubin bandwidth cut (22 to 19.2 TB/s) dating to September.

Suggests HBM supply and cost, not compute design, is shaping NVIDIA's roadmap; second-hand reporting, not confirmed by NVIDIA.

Samsung's Q3 2026 revenue was 195.3 trillion won, up 121.8% year over year, and operating profit of 107.4 trillion won made it the first tech company to exceed 100 trillion won in a quarter. Memory brought in 140.7 trillion won, driven by HBM4 and server SSD sales. Samsung expects record AI infrastructure spending in 2027.

Memory makers are capturing outsized AI-capex value, signalling sustained cost pressure on training systems.

Sandisk's fiscal Q1 2027 revenue hit a record $11 billion (up 378% year over year) with GAAP net income of $8.1 billion. Datacenter revenue was $4.8 billion, up 1,682%, and Q2 revenue is guided to $12.4 billion plus or minus $200 million.

Confirms flash as the fastest-growing AI storage tier, supporting high-capacity SSD and KV-cache roadmaps.

Seagate reported record fiscal Q1 2027 revenue of $4.8 billion, with datacenter customers at 90% of 220 EB shipped and cloud nearline revenue up for a 12th straight quarter. It expects HAMR drives to be about half of nearline exabytes by end of calendar 2026, with Mozaic 4 (up to 44 TB) ramping at two cloud providers.

Mass-capacity disk remains the cold tier for training data lakes even as flash tightens.

OpenAI released GPT-6.1 Terra at $1 per million input and $5 per million output tokens, about 50% below GPT-5.6 Terra and GPT-6.1 Sol. OpenAI says it completes tasks more than 3x faster than GPT-6 Astra and GPT-6.1 Sol and matches Astra at roughly one-tenth the per-token price. It is available in ChatGPT Work, Codex and the API as gpt-6.1-terra, and OpenAI still recommends Sol or Astra for the hardest coding tasks.

A sharp mid-tier price cut with claimed near-flagship agent performance resets cost expectations for agentic workloads (all figures are OpenAI's own claims).

Mistral published open weights for Mistral Large 4 (nicknamed Le Chonk), a roughly 1T-parameter model, on Hugging Face; it is also available via Mistral's API. Mistral claims performance comparable to GLM-5.3 with 47% fewer tokens per task and the top open-model spot on Artificial Analysis' Coding Agent Index. Both claims are the company's own.

A Western lab shipping a frontier-scale open-weight model narrows the gap with Chinese open models and offers a self-hosting option, though the weights are large to serve.

DeepSeek open-sourced its DeepInfer inference engine less than a week before this report. On GB200 decode, vLLM trailed DeepInfer by about 6%, while SGLang's prefill peaked about 4% higher and its 64-GPU GB300 decode step was 28.88 ms versus 31.19 ms. The numbers are community-reported and come via a secondary Chinese outlet.

Frontier-lab serving optimisations are being upstreamed into community engines within days, compressing the proprietary inference advantage.

Liquid AI released six roughly 2B-class models built on LFM3-1.7B-Base (text, vision, audio, omni and two decision models), with open weights. LFM3-1.7B outperforms the prior LFM2.5-2.6B with 35% fewer parameters, fits in about 1 GB at 4-bit, and offers 256K context after 22T tokens of pre-training. Liquid says its KV cache at 128K context needs 1.6 GB versus 4.3 GB for Qwen3.5-4B.

It pushes capable multimodal inference onto battery-powered and local-agent hardware, shifting where small-model workloads can run.

NEURA unveiled the 4NE1 Gen 4 on October 22 at the TUM RoboGym near Munich Airport. Its Smart Limbs design puts Qualcomm Dragonwing Robotics processors in each limb for reflex-level responses without a central processor. NEURA says it is prepared for series production but has published no price or ship date.

A concrete example of distributed on-robot inference and a Qualcomm robotics design win.

Nvidia says fixed access is fundamentally different from mobile, and that framing, forward error correction and packet scheduling are already well served by dedicated silicon. It still sees GPU opportunities in bandwidth allocation and traffic prediction, and DPU opportunities in virtualized broadband gateways and security. The article notes Nokia's fixed networks revenue fell 22% year on year in Q3.

It marks the boundary of Nvidia's telecom GPU strategy: mobile RAN yes, fixed access largely no.

STMicroelectronics argues humanoids should handle fast reactions on local sensors and MCUs and leave whole-body coordination to a central GPU. It has not decided whether to build more powerful MCUs or MPUs, because robotics is still low-volume. It demoed gesture tracking on an STM32N6 driving a six-motor robotic hand.

It points to a distributed compute architecture for humanoids that leaves room for microcontroller vendors alongside GPUs.

Stereolabs introduced ZED Box 2 (Jetson Orin NX 16GB, rated up to 157 TOPS) and ZED Box 2 Max (Jetson Thor T4000 or T5000, up to 2,070 FP4 TFLOPS and 128 GB). Both support up to eight synchronized GMSL2 cameras. ZED Box 2 started shipping in October 2026 and the Max ships at the end of November. This is a vendor post and the specs are the vendor's own.

Production-grade Jetson Thor robot compute modules are now arriving from third parties.

The study re-analyzes 540 judgments from a benchmark of nine research agents on fifteen problems. About four in five unsupported numerical claims are values from the wrong experiment, metric or run rather than invented ones. A deterministic checker certified 78 of 85 claims with over 90% precision but covers only a subset, while an LLM verifier recovers more claims and accepts about three times as many misattributed ones.

A concrete auditing approach for AI-written research as autoresearch agents spread.

The paper argues the best autoresearch results still come from agents working with human researchers, especially during ideation. It introduces HackBench, which shows current agents struggle to localize reward hacks. It also proposes two RL methods, one using human simulators and one using expert-derived rubrics, to improve collaboration.

It frames human-in-the-loop co-improvement as the practical path for long-horizon research agents and adds a misalignment benchmark.

World Reward Models learn from real execution results to choose between candidate programs directly, without predicting intermediate outcomes. The 9B ForeCritic picks the better program 79.4% of the time on 12,438 held-out pairs from 26 ML tasks, versus 53-62% for general LLM judges. Inside AIDE, ML-Master and ForeAgent it cuts mean task time 4.4-7.5x with scores within 0.01 of baseline or better.

A small specialized critic beating general LLM judges suggests a cheap route to faster agentic ML search.

LOWN is a training-free decoding method for diffusion vision-language models. It processes the image only when an image-free prediction disagrees with the last image-conditioned one. On LLaDA-V it keeps 99.6% of the score across ten benchmarks while using 35.3% of decoding FLOPs.

A large inference-cost cut for diffusion-based multimodal models, with no training needed.

A-RSI triggers a refresh of the training task pool when reward variation and accuracy in training logs decline, without using validation or test scores. Tasks come from model-authored programs, with no external examples or human answer labels. It lifts an eleven-benchmark average from 57.77 to 61.05 using 800 fresh tasks.

It advances label-free self-improvement by addressing when to introduce new tasks.

Using sparse autoencoders on small GRPO-trained models playing Tic-Tac-Toe, the authors find the models lean on information explicit in the prompt, such as turn and board occupancy. Higher-level concepts like symmetries and strategic forks are weakly represented. RL reorganizes which features dominate without adding concept information, and the authors suggest reasoning benchmarks may overstate ability.

Mechanistic evidence, from a toy domain only, that RL reasoning gains may partly be input adaptation.

Microsoft trimmed its full-year capex forecast to $170 billion from $175 billion, with its CFO citing third-party GPU capacity delays that push some capacity into the first half of 2027. Alphabet held its annual guidance at $195-205 billion, and its capex fell short of analyst estimates. Meta guided to $135-145 billion for the year.

Capacity delays rather than demand are the binding constraint, which affects the revenue timing of neocloud and data-center developers.

AWS revenue was $45.98 billion, up 39% year over year and its fastest growth since Q4 2021, with operating income of $17.77 billion, up 55%. CEO Andy Jassy said he believes AWS can grow north of 40% in each of 2027 and 2028.

Accelerating cloud growth and margin support the capex thesis even as Amazon's stock fell on free-cash-flow concerns.

OpenAI is seeking up to $20 billion at a $1.25 trillion post-money valuation, down from an earlier target of up to $30 billion, per the Financial Times. UAE-based investors and Coatue are leading. The round is still open, so the terms could change.

A scaled-back target from the largest private AI company is a notable demand signal as public markets sour on AI spending.

Citing Bloomberg, TNW reports the AI inference platform is raising about $2 billion in two tranches at $26 billion and $30 billion valuations, led by existing investor Sands Capital. It was valued at $13 billion in June after a $1.5 billion Series F. TNW says it has not independently verified the report; the company has not confirmed.

Inference infrastructure valuations are compounding quickly, signaling investor conviction in the serving layer, though this is an unconfirmed report.

Citing Bloomberg, TNW reports the robotics foundation-model startup is in talks to raise at least $1 billion at about $20 billion, nearly double its $11.5 billion mark from earlier this year. Existing backers Avenir and Lux Capital are in talks to lead, and the company does not yet have revenue. The report is unconfirmed.

Physical AI keeps attracting outsized rounds and rapid step-ups before revenue, extending the capital wave beyond language models.

Seoul-based inference-chip maker FuriosaAI raised a $620 million Series D at about $2.9 billion post-money. Korea's National Growth Fund put in $280 million, international investors including Cisco Investments contributed $110 million, and existing shareholders $125 million. Proceeds go to RNGD production, software and a third-generation accelerator.

State-backed capital for non-Nvidia inference silicon shows sovereign industrial policy shaping the AI chip supply landscape.

Wabi raised $55 million at a $1.25 billion valuation in a round led by IVP, with participation from Andreessen Horowitz, Forerunner and the Amazon Alexa Fund. Angels include Tobi Lutke, Marc Benioff, Marissa Mayer and Dick Costolo.

Application-layer agent startups are reaching unicorn valuations quickly, even as hyperscaler capex draws scrutiny.