AI Weekly 2026-W45
2026-11-7
| 2026-11-7
字数 4112阅读时长≈ 11 分钟
OpenAI released GPT-6 Cyber, its first vertical security model: ExploitGym rose from 32.1% to 48.5%, and it confirmed 948 privilege-escalation vulnerabilities in the Linux kernel and found 81 previously unknown high-severity vulnerabilities in Firefox. It treats "capability gains" and "staying in bounds" as two goals to optimize at once, reaching 100% on the out-of-scope alignment eval.
Fireworks already serves over 40T tokens/day in a single vertical, while Perplexity has pushed per-search pricing extremely low. Anthropic uses dreaming to reorganize external memory stores; NVIDIA replaces long context with frequent compaction. The past month brought a dense wave of open-source LLM releases, with three attention tracks running in parallel: GQA+SWA, MLA, and LatentMoE. Cactus compressed a multimodal model to tens of MB, running in milliseconds on a phone NPU. MIT used GPT-6 Astra to edit Qwen weights and raise math scores, and was called out for it being the old playbook of steering and pruning. VBL fixes planning state with a goal stack; AWM rehearses actions through read-only interfaces without touching the real environment. Together hands its inference engine to an agent that compiles continuously; ForgeKernel turns hardware discoveries into a cross-run knowledge base. OpenAI built Pando in-house and then acquired Pierre; GitHub rebuilt its storage layer — both are responses to agents writing code concurrently. NVIDIA opened Cosmos 3 weights and plugged into LeRobot; HuggingFace filled in data and distribution. Reflex uses only onboard RGB-D to let a G1 catch thrown boxes; ODE Policy learns action evolution with differential equations. Oracle's share sale and Marvell's acquisition of Lumentum raised dilution fears, then the financing landed the next day and the stocks rebounded. Democrats won the midterms, and data center approval worries dragged down power and AI infrastructure stocks across the board. The STAR 50 rose 9.92% in a single day, swinging wildly between the US chip selloff and domestic money flowing in.

Industry & Ecosystem

Inference Cost and Market Size

Nathan Lambert tries to quantify the open-model inference market in Interconnects (interconnects.ai) with a dashboard: Baseten, Fireworks and Together AI serve as bellwethers for open-source inference startups, layered with OpenRouter's open-model token share and compared in magnitude against the platform-wide token volume Google discloses in its earnings. The hard number available is that Fireworks already serves over 40T tokens/day in a single custom-model vertical, more than double this summer; the author estimates its total will pass 100T in the fall. The curve kinks upward after GLM 5.2 delivered an open-source agent step change. He also says plainly that every open-source adoption metric is extremely imprecise — platform noise, startup reporting conventions, imperfect public evals — and concludes with a "best approximation."
On the pricing side, look at Perplexity. Arav Srinivas announced Fast Search in the Decisions API: $0.001 per search, $0.02 per million input tokens for the decision model, with batch decisions and web-grounded fact grounding. For comparison, according to a breakdown by CloudZero, Perplexity's Agentic Research API charges model list price plus $0.005 per web search — Fast Search's per-call $0.001 is another order of magnitude lower.
One side estimates the open-source inference economy upward (tens to over a hundred T tokens/day), the other pushes a single search toward a tenth of a cent. Together the two lines show not how big the market is, but how fast unit costs are falling — fast enough that even the measurement conventions struggle to keep up.

Agents: Technology & Applications

Autonomous Cybersecurity Agents

On November 5 OpenAI released GPT-6 Cyber, its first vertical model. On capability, ExploitGym rose from 32.1% for GPT-5.6 Cyber to 48.5% (+51% relative); on constraints, the out-of-scope alignment eval avoids attacking services outside the authorized scope 99.93% of the time without a deployment prompt and 100% with one. Raising capability and scope constraints together is the most notable part of this release — the barrier to deploying security agents isn't whether they can find vulnerabilities, but whether the process of finding them can be bounded by authorization.
The accompanying engineering write-up A thousand fewer paths to root lays out the pipeline: the cyber model drives a scout agent to scan and an analyst agent to dig deeper, followed by four stages — vulnerability validation, exploitability assessment, and patch generation. A single run confirmed and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel, at an API cost of about $1.2 million. Several counterintuitive findings: stronger models need lighter orchestration, maintainers patch quickly, the model can discover new exploitation techniques, and the test infrastructure itself has vulnerabilities. According to The Hacker News, the same work also found 81 previously unknown high-severity vulnerabilities in Firefox, which Mozilla had tested repeatedly; Mozilla has shipped patches.
Access is split into two tiers: Daybreak Blue for validating vulnerabilities and reducing false refusals, and Daybreak Red for authorized penetration testing and deep exploit development, with plans to open it to open-source maintainers and small organizations. OpenAI's official account notes that security teams at Mozilla, Palo Alto Networks, Cisco and others are already using it.

Agent Memory and Context Management

Three companies give three answers to the same question: where an agent's memory lives.
Anthropic chose external storage. A engineering blog post describes Claude's "dreaming": it reads the existing memory store and past session transcripts and produces a reorganized memory store — duplicates merged, outdated or overturned entries replaced with the latest values, new insights surfaced. According to Anthropic's research page, it defines agent memory narrowly: knowledge gained from experience and retained across sessions, explicitly excluding weight changes, in-context learning within a single conversation, and compaction.
NVIDIA goes the other way. You Don't Need Long Context argues for frequent context compaction instead of accumulating tokens — even resetting context after every agent action — preserving accuracy on six long-horizon benchmarks while cutting inference cost by up to 8.8x.
The third path is the filesystem. MFM (Multimodal Filesystem Memory) keeps weights frozen and makes "memory" the place where learning happens: find the experiences that matter now, turn successful actions into reusable skills, trace failures to correct what has been learned, and expose knowledge gaps to guide further exploration. In a Minecraft demo it went from bare hands to diamond tools, entered the Nether, fought Blazes, and completed 21 of 24 Ender Dragon milestones.
All three share the premise of not touching weights; they diverge only on the memory medium: context window, external store, filesystem. Anthropic reorganizes periodically, NVIDIA discards immediately, MFM is a persistent layer that can be retrieved and rewritten.

Foundation Models & Multimodality

Open Model Releases and Architecture Evolution

Sebastian Raschka calls the past month "Open October" — 10 open LLM architectures released. A few representatives: Kolibri-1 78B (German team, GQA + sliding window attention), Ling 3.1 Flash 560B (Kimi Delta Attention + gated MLA), Muse Spark Open 816B (Meta's first major open-weight release since Llama, adding LatentMoE on top of GQA + SWA), Nemotron 3.5 Super VL 120B (Mamba-2 layers with 8 attention layers, also with LatentMoE), and Mistral Large 4 1T (MLA + sparse attention).
Two more concrete ones landed this week. Ling-3.2-flash is natively multimodal, 560B total / 25B active parameters; in its release evals it scores 92.21 on OmniDocBench 1.5 (Opus 5 max: 90.56), 43.71 on ZEROBench Sub (39.52), and 85.86 on Vision2Web (80.20). ERNIE 5.5 Flash Base is 280B total / 14B active, pretrained on 33.5T tokens with a 128K training context, extendable to 1M for fine-tuning and inference; the architecture is 37 layers of compressed attention plus 6 layers of MLA-DSA, Multi-Router Latent MoE, and ECHO to reuse the last layer as a speculative draft. According to Baidu's official page, vLLM added Day-0 support; under the same settings it scores 82.3% on HumanEval+ and 40.1% on LiveCodeBench, versus 69.5% / 31.9% for DeepSeek-V4-Flash-Base.
Taken together, attention-layer choices are converging into a few camps: GQA + SWA is the conservative combination, the MLA family (DeepSeek lineage) is mainstream, and LatentMoE and hybrid linear approaches (Mamba, short convolutions) are the exploratory directions. Sparsification is spreading from attention into MoE routing.

Small Models and Efficient On-Device Deployment

Three paths at different scales point to the same conclusion: multimodality is no longer an expense only large models can afford.
The most extreme is Mirage (Cactus): a 22MB vision-language model, a 7.8M encoder plus a 40.8M decoder, 2-bit quantized, doing scene description, object detection, OCR, segmentation, spatial relations, information extraction and decision-making in a single forward pass — 74ms on a Samsung Galaxy S26 NPU and 107ms on CPU, supporting seven languages (English, German, French, Spanish, Italian, Dutch, Polish). According to the Cactus blog, feeding Mirage's structured output and full trace on 256 VQAv2 questions to a text-only Qwen3-0.6B raised the latter's score from 9.3 to 31.3 (out of 100), with neither model fine-tuned.
The middle tier is Inkling 2 Small (Thinking Machines), with an AA intelligence index of 31, fine-tunable and runnable on Tinker now, with full weights coming soon. According to TLDR, it is an MoE with 594B total / 17B active parameters, with image input and variable effort.
The vertical tier is OCRia-1: a 4B-parameter document parsing model, averaging 0.87 on the four-dimension ParseBench versus 0.75 for the next-best system, and 83.5 on olmOCR; according to the technical report, it is more accurate than all compared systems on tables, charts and text formatting.
Each path makes a different trade-off — Mirage sacrifices general language ability for size, Inkling keeps general ability by using MoE to shrink active parameters, OCRia gives up generality to do only documents. What they share is turning "can it run on-device" from an engineering compromise into a product premise.

Research Frontier

Weight-Space Programming

An MIT team used GPT-6 Astra to directly edit the weights of Qwen3.5-0.8B: changing only 0.016% raised MATH-500 from 7.9% to 48.1% and tripled instruction following, all with no gradients and no training. According to VentureBeat, the authors named this Weight-Space Programming and argue that frontier models can natively read and write neural network weights.
The rebuttal comes from Sean Goedecke: identifying what game a small network is playing relies on labeled input channels plus an L2-norm heatmap of the first-layer weights, reproducible in three lines of Python; fixing Qwen's math ability was essentially the model getting stuck in a loop after 32k tokens, and Astra used the standard steering/abliteration playbook — write loop/no-loop text pairs, measure weight contribution, prune — which human researchers do too. His core analogy: editing weights directly is like editing compiled binaries — feasible but not the optimal path, and the right way for LLMs to help AI research is to improve training.
A third view is more mundane. Zeroth-order weight perturbation uses vanilla evolution strategies for post-training, pushing Qwen3.5-4B from 12.1% to 21.1% on Terminal-Bench 2.1, above the released Tmax RL checkpoint's 18.2%, with a population size of only about 30 — no credit assignment, no critic, no masking.
Three ways of intervening on weights — read-modify-write, critiquing its limits, search-based perturbation — put together, the real question isn't "can an LLM read weights" but, once weights are treated as a programmable object, which parts are actually old methods under new names.

Planning Control and Simulated Rehearsal for Long-Horizon Agents

Two industry papers, both answering what to do after a long-horizon agent makes a mistake.
VBL (CityU HK + Huawei Noah's Ark) is a training-free extension of ReAct: it maintains task decomposition as a bounded goal stack and runs dual-view joint validation when a goal is activated, execution fails, or physical state changes — whether the active goal still contributes to the final goal, and whether it is feasible in the current state — combined with typed failure attribution that triggers four explicit operations: Push, Pop, retry, BacktrackTo. With three backbones (Qwen3.5-9B, GPT-4o, DeepSeek-V3.2) on three benchmarks (AppWorld, ALFWorld, ScienceWorld), every combination beats ReAct and reflection-style control; on AppWorld it improves by up to 22.07 points. Ablations show the validation and diagnosis mechanisms are complementary — removing either drops several settings below ReAct.
AWM (Microsoft + UW-Madison) handles the other half. Test-time scaling usually assumes multiple candidate trajectories can run in parallel, but enterprise workspaces and cloud services are shared, persistent and hard to replicate, and actions may carry irreversible side effects. AWM turns next-observation prediction into multi-turn evidence gathering: for each hypothesized action, an agentic simulator queries the real environment through read-only interfaces, reasons over the returned evidence, and constructs observations with executable computation — all without changing shared state. On τ-bench, AppWorld and OfficeBench it beats single-shot simulation baselines by 40.07% on average and trained world-model baselines by 12.36%.
VBL fixes planning state; AWM fixes "rehearse without touching the environment." Both share an implicit premise: the bottleneck in long-horizon tasks isn't the quality of a single decision, but whether the system can retreat in an orderly way after failure.

AI Infra & Engineering

Agent-Driven GPU Kernel Generation and Inference Engine Optimization

Together Compiler has an agent continuously compile and optimize inference for production traffic: 2.6x faster than the leading InferenceX configuration and up to 50% faster on customer workloads, reaching 1111 TPS in a single compile within 24 hours (MiniMax M3.1 / B200), at a service scale of trillions of tokens per day. Tri Dao says it is a two-year project that "distills most of the expertise in kernel and inference engine optimization into an agent loop."
ForgeKernel is more structured: it separates architecture design from code implementation, uses explicit architecture decisions to guide implementation, then feeds performance back to revise the decisions; across runs it verifies hardware mechanisms with independent probes and stores them in a shared knowledge base. According to the paper, across 27 operators and 29 input configurations on NVIDIA Hopper and Blackwell, 25 configurations match or beat expert references, at a geometric mean of 1.10x; on Ascend 910C, all 8 level-4 operators in CANN-Bench Season 1 beat the best public results, at a geometric mean of 25.1%. As of October 30, ForgeKernel-generated submissions rank first on all three CANN-Bench 2.0 leaderboards (910C, 950DT, 950PR) and in the SOL-ExecBench FlashInfer-Bench track.
The shared pattern is distilling expert experience into an agent loop, then using measurement to turn one-off discoveries into reusable knowledge. The bottleneck shifts accordingly: from "can it write kernels" to "how do you validate this knowledge base and keep it from going stale."

Code Hosting and Git Backends at Agent Scale

Two companies found at the same time that when agents write code concurrently, the first thing to buckle isn't CI — it's version control.
OpenAI built its own Git backend, Pando, now carrying all monorepo development, with median push latency cut nearly 5x, nearly 1PB of Git data transferred per day, and full compatibility with the Git protocol and clients — engineers' and agents' git push needs no change in habits. The design chooses causal consistency while meeting two goals at once: writes scale horizontally, reads run at full bandwidth. According to a write-up by MarkTechPost, its peak single-hour transfer exceeds 57TB, and a standard cloud VM can saturate a 10Gbps link.
GitHub's Project Rubicon comes at it from another angle: Git objects are immutable and content-addressed, so two writers can store simultaneously without negotiation; the only thing that truly needs consensus is the reference update, and that is small. On that basis it decouples storage, request serving and repository maintenance into three independently scaling parts. Internal tests support 5,000 concurrent agents and 7,200 pushes/s on a single repository; daily requests per repository rose from 44 million in March to 181 million in May.
Add one acquisition: OpenAI announced it will acquire Pierre Computer Company — the Trees for browsing files and Diffs for viewing changes in the Codex app come from this company, whose products also include an Edit mode with an inline diff editor, Highlights that update live as the agent streams output, and Code Storage that keeps code history in Git. The intent is clear: pull the whole chain — agent creates branch → human reviews code → deploy — into Codex.
The two approaches differ — OpenAI swaps the backend then fills in frontend tools, GitHub keeps the Git protocol and rebuilds the storage layer — but the judgment is the same: agent scale will push version control to the point of needing dedicated architecture design.

World Models & Embodied AI

World Models and the Open Robotics Ecosystem

NVIDIA released four 16B open-weight checkpoints under the Cosmos 3 family. According to the HuggingFace blog, all four are post-trained examples of Cosmos 3 Nano: Transfer-Auto generates multi-view, multi-sensor driving data (weather, lighting and time of day adjustable), Sim-Auto streams synchronized camera plus LiDAR observations, Sim-Transfer turns a physics simulator's control inputs and text into photorealistic video, and Sim-Bimanual predicts video from robot or camera actions. The OpenMDW1.1 license allows commercial use.
The same model family also entered the open robotics stack. NVIDIA Robotics announced that Cosmos 3 and Long-WAM are integrated into LeRobot 0.7, that RoboLab (built on Isaac Lab, covering 120 tabletop tasks) is used to evaluate policies, and that SONIC handles humanoid whole-body control with decoding running on Jetson. According to the LeRobot 0.7 release blog, Cosmos 3's role is to predict future video frames and actions from camera views and language instructions.
HuggingFace fills in data and distribution. The numbers from Clement Delangue: nearly 100,000 public robotics datasets (+48%), over 30,000 public robotics models (+41%), over a million LeRobot downloads; the open-source robots Microduck and Reachy mini are expected to sell 50,000+ and ship 20,000+ by year-end; RoboColosseum evaluation covers 140+ real scenarios, with every run (including failures) published openly on HF.
The division of labor is clear: NVIDIA provides models and the compute toolchain, HuggingFace provides data, distribution and an evaluation venue. Open weights are a precondition, but without data scale and evaluation infrastructure, weights alone don't make an ecosystem.

Robot Control Policy Learning

Reflex (University of Washington, USC, Stanford, NUS, UNC, AI2) lets a Unitree G1 catch boxes thrown by a person using only onboard RGB-D. The robot doesn't catch with its hands; it moves its feet and torso, forms a support surface with its forearms and chest, absorbs the momentum and stays upright. Training runs in three stages in Isaac Lab: first a state-based teacher (PPO plus AMP-style rewards learned from human motion references), then PPO to learn to infer box dynamics from delayed, incomplete observations (position, velocity, orientation, size, observation age) and compress them into a latent that conditions the controller, and finally the controller is frozen and the same latent is recovered from a 25-frame history of egocentric RGB-D plus camera self-motion (using latent matching, an auxiliary box-state loss, heatmap supervision and DAgger). Rendering stays out of the RL loop; vision training needs only about 1.6 million rendered steps. The notable result: the policy driven by camera images outperforms the policy given perfect simulator state.
ODE Policy swaps a different premise: it uses neural ODEs to learn how actions evolve, rather than predicting actions directly. It was validated on five robot types — single-arm, dual-arm, dexterous hand, humanoid whole-body, and aerial. For VLA (π0.5), success rate after one epoch is 96.0% vs. 75.5%; on a 4.5B World Action Model it is 82.6% vs. 24.6% with action generation about 4x faster; RL humanoid whole-body motion learning is also more efficient.
One compresses observations into a latent and then controls; the other treats actions as solutions to continuous differential equations. Both are leaving the main road of "directly fitting actions," which shows robot policy representations are not yet settled.

Capital Markets

Benchmarks this week: Hang Seng TECH +8.61%, CSI 300 +1.73%, STAR 50 +7.96%, ChiNext +2.14%, Nasdaq-100 ETF (QQQ) -4.98%, Semiconductor ETF (SOXX) -71.04%
  • Marvell Technology (MRVL · Compute chips): -28.11% this week, +42.93% vs. benchmark; biggest day -16.43% (11-05)
  • Micron (MU · Compute chips): -22.01% this week, +49.03% vs. benchmark; biggest day -17.46% (11-05)
  • Vertiv (VRT · Power & cooling): -21.35% this week, -16.37% vs. benchmark; biggest day -14.02% (11-05)
  • Zhongji Innolight (300308 · Cloud & infra): -18.99% this week, -21.13% vs. benchmark; biggest day -12.69% (11-03)
  • Jinpan Technology (688676 · Power & cooling): +17.56% this week, +9.60% vs. benchmark; biggest day +10.82% (11-04)
  • Kunlun Tech (300418 · AI apps): +16.51% this week, +14.37% vs. benchmark; biggest day +8.69% (11-02)
  • WUS Printed Circuit (002463 · Cloud & infra): -16.18% this week, -17.91% vs. benchmark; biggest day +9.99% (11-04)
  • Nebius Group (NBIS · Cloud & infra): -15.99% this week, -11.01% vs. benchmark; biggest day -14.94% (11-05)
  • CoreWeave (CRWV · Cloud & infra): -15.44% this week, -10.46% vs. benchmark; biggest day -9.90% (11-04)
  • Arm Holdings (ARM · Compute chips): -15.22% this week, +55.82% vs. benchmark; biggest day -13.08% (11-05)
Top 10 by absolute weekly move (of 59 tracked); weekly change is this week's last close vs. last week's.

AI Infrastructure Financing and Dilution Fears

On November 2, Oracle (CNBC) launched a $10 billion stock offering and a $10 billion senior notes offering the same day to finance AI expansion; the stock fell 8.76% on dilution and debt concerns. The same day CoreWeave (Investing.com) fell 8.6% to a record low, with its own leverage high and its bonds yielding about 15%. Marvell (Yahoo Finance) announced an all-stock acquisition of Lumentum for $115 billion, with LITE shareholders receiving 4.5 shares per share and holding about a third of the combined company; fears of heavy dilution and integration risk sent it down 12.72%.
The turn came the next day. On the evening of November 2, Oracle priced the $20 billion stock and bond offering; the market read the completed financing as removing balance-sheet uncertainty through FY2030, and on November 3 CoreWeave rose 7.16% and Oracle rebounded 5.85% (TIKR). The same set of assets moved in two directions within three days, which shows the market's disagreement isn't about "whether to build AI infrastructure" but "who carries the balance sheet."

Midterm Elections and Data Center Approvals

On November 3, Bernstein said election-day fears of a data center construction slowdown were overblown; GE Vernova rose 6.49% and Vertiv rose 5.23%. On November 4 Democrats won the midterms (including the Texas governor's seat), and the fear reversed: Vertiv fell 12.89% and GE Vernova fell 9.3% (Yahoo Finance); CoreWeave fell 9.9% (Stocktwits) with its bonds falling too, compounded by uncertainty ahead of its November 10 Q3 earnings and high-debt concerns. Within a week, pricing for both power and compute-leasing assets flipped from "fears overblown" to "approvals constrained" — policy expectations became the most sensitive variable in this supply chain.

Sharp Swings in A-Share Semiconductors

On November 2, market-stabilization measures came out and STAR Market Q3 earnings rose sharply year over year; with semiconductor AI hardware leading, the STAR 50 rose 3.11% (24TopNews). On November 3, the overnight plunge in US AI chip stocks (Marvell -12.72%, Intel -8.80%, AMD -7.52%) mapped onto A-shares: Cambricon fell 10.81%, its market cap dropping below 500 billion yuan (Sina Finance), MetaX -11.01%, Moore Threads -8.95%, WUS Printed Circuit -9.19%, and the STAR 50 closed down 3.81%. At the stock level, Zhongji Innolight (Sina Finance) fell 12.69% on November 3 after Goldman Sachs cut its 2026-2028 revenue and net profit forecasts, compounded by a circulating Nomura report that was debunked.
On November 4 everything reversed. The STAR 50 rose 9.92% (21 Caijing), with main funds pouring in and a wave of limit-ups across the semiconductor supply chain; Zhongji Innolight (The Paper) rose 11.07% after clarifying on a call that there were no widespread order cuts and Citi maintained its Buy rating; Moore Threads (Sina), together with Qutrunk and Zhipu AI, released a domestic heterogeneous inference Token Factory solution and rose 13.6%. Over three trading days the STAR 50 went from -3.81% to +9.92%, driven entirely by external mapping and fund flows rather than changes in fundamentals.

📌 Notable This Week

AI on the Ballot — Stratechery / After the midterms, the constraints on AI diffusion shift from compute, memory and power to politics: data centers need permits, and opponents have a veto for the first time.
ViBench 2.0 — Replit / A long-horizon software-building benchmark: 17 full-stack products, 241 incremental requirements, 544 browser test plans; the strongest model delivers a complete product in only 65% of builds.
Clocktower — Dwarkesh Patel / Spent about $20,000 of his own money turning Werewolf into a multi-agent benchmark, 386 games; GPT-6 Astra beats Opus 4.6 about 92% of the time, a gap of roughly 430 Elo.
MoE training handbook — Thomas Wolf / 800,000 GPU-hours of experiments, about 70,000 words, 55 interactive figures, covering practical questions like expert count, all-to-all hiding and router choice.
llama-modchip — GitHub / Runs 284B DeepSeek-V4-Flash on 64GB of DDR5 with zero GPUs; AVX-512 kernels raise prefill from 25.05 tps to 44.36 tps.
Weekend at Bernie's II — nesbitt.io / Before-and-after snapshots of 6,106 critical open-source packages: 421 zombie repos revived, and the share of commits with AI disclosure rose from 3.3% to 7.9%.
PDD Kit — Microsoft Research / Proof-Driven Development: intent is first turned into a formal Verus spec, agent code is proven to satisfy it for all inputs, and parts it can't cover fall back to testing.
Gemini Omni 1.2 Flash — Google / Video model update, about 30% cheaper, supports multi-turn conversational editing; Ultra subscribers can generate and edit videos up to 15 seconds.
  • AI
  • 周报
  • AI Tech Daily - 2026-11-08RecSys Weekly 2026-W45
    Loading...