📊 Today's Overview
Kimi K3.1 Pro dropped its weights today — 1M context, 3T-scale, with vLLM shipping Day-0 support across GPU and TPU. OpenAI's autonomous security pipeline confirmed and patched 948 Linux privilege-escalation bugs for about $1.2M in API costs. And the midterm elections handed AI its newest bottleneck: voters rejected 41 data-center ballot measures across 10 states, with Stratechery arguing permitting has become the real constraint on AI buildout. Meanwhile llm-d hit 1.0, standardizing open-source distributed inference.
🔥 Trend Insights
- Open weights keep landing: Kimi K3.1 Pro ships weights with Day-0 vLLM support, and Tencent's Hy4 goes live on Tencent Cloud and OpenRouter — frontier open models are now a weekly cadence.
- Agents eat security work: OpenAI patched 948 Linux bugs, Mistral ran 133 agents on a Django CVE, Cloudflare open-sourced its audit flow — vulnerability hunting is becoming an agent pipeline.
- Inference cost engineering matures: Together Compiler beats best configs by 2.6x, llm-d 1.0 standardizes routing, and session-aware caching cuts token costs 5-10x versus commercial APIs.
🐦 X/Twitter Highlights
📈 Trends & Buzz
- Kimi K3.1 Pro officially releases weights: 1M context, 3T-scale — The team says it's continuing to push open intelligence. vLLM shipped Day-0 support the same day, covering GPU and TPU, plus RL training support; under the Claude Code harness it scores 65.7% on Terminal-Bench 4.0. Together AI is live simultaneously, supporting post-training via RL/SFT @Kimi_Moonshot (Moonshot AI)｜@vllm_project｜@togethercompute
- Tencent Hy4 launches on Tencent Cloud and OpenRouter — The model was co-designed by multiple internal product teams at Tencent, and adds vision and computer-use capabilities for the first time @TencentGlobal (Tencent)
- Apple cuts some engineers' daily AI spending cap from $300 to $200 — This is a real-time flash from prediction market Kalshi; it doesn't say how many engineers are affected @Kalshi (prediction market platform)
- RoboScience closes a new round in the hundreds of millions of RMB — Funds go to three product lines: the Visics foundation model, the RoboMirage simulation platform, and the REX G1 wheeled humanoid robot @roboticomarket (robotics industry news account)
🔧 Tools & Products
- ChatGPT Images 2.6 ships, adding GIF and reference images — Two new image models also hit the API: Flare and Sunburst, with 50% faster generation. Sunburst holds detail consistency better across multiple edits at medium/high quality. GPT-Image-2.6 is now in Codex @OpenAI｜@OpenAIDevs
- Cohere Embed 5 and Rerank 5 launch on OpenRouter — Both models run through one API and one bill, in Pro and Fast tiers @cohere (Cohere)
- Perplexity wires Fast Search into its Decisions API — The decision model previously relied only on training data and your state; adding `"mode": "web_grounding"` runs a search per question and answers from that day's web. Search bills at $0.001/call, the decision model at $0.02 per million input tokens @AravSrinivas (Perplexity CEO)
- RoboColosseum: an open robot benchmark with 100+ real machines and 140+ real scenarios — Hugging Face is a co-founder; all rollout data (including failures) goes to the Hub, with 10,000 hours of free evaluation time. Robots include Unitree G1, LeRobot SO-101, YAM, and Franka @huggingface
⚙️ Engineering Practice
- Together Compiler: agents continuously compile-optimize inference, 2.6x faster than InferenceX's best config — A single compile uses 2 nodes and delivers results within 24 hours; MiniMax M3.1 hits 1111 TPS on B200, DeepSeek v4.1 Flash hits 1205 TPS on GB300 (batch size 1). Experimental throughput is 252x higher than prior agent-optimization systems; it now serves trillions of tokens daily and recompiles within hours of traffic shifts. Covers NVIDIA, AMD, and Trainium2 @togethercompute｜@simran_s_arora (SVP of Research, Together AI)｜@realDanFu (same)
- ForgeKernel: a self-evolving agent optimizes GPU kernels, giving DeepSeek V4.1 FlashMLA a real-time 9.8% speedup — It separates architecture design from code implementation; when it hits a structural bottleneck it goes back and changes work partitioning, dataflow, and pipelining rather than just tuning within a fixed architecture. Discovered hardware mechanisms are first independently verified on target hardware, then stored and shared across kernels. Geometric mean 1.10x on NVIDIA GPUs; ranks first on all three CANN-Bench 2.0 leaderboards plus SOL-ExecBench and FlashInfer-Bench @_LiYuXuan_ (first author)
- NVIDIA paper: aggressive context compaction barely costs accuracy — Claude Opus 4.5 keeping only one action's context scores 70.2 on SWE-bench Verified versus 70.0 with full history; median peak context drops from 32K to 8K tokens. One compaction at the end of a conversation lifts the score from 70.0 to 74.8. The cost argument: later tokens get more expensive to generate, up to 7x on frontier open-weight models @omarsar0 (NLP educator / founder, DAIR.AI)
- AI finds bugs x3: OpenAI fixes 948 Linux privilege-escalation bugs, Mistral sends 133 agents at a Django CVE, Cloudflare open-sources its audit flow — OpenAI's autonomous pipeline verified and generated patches for 948 local privilege-escalation bugs in a single run, at $1.2M in API cost. Mistral Large 4 runs triage, reproduction, and patching in three steps. Cloudflare released a six-stage skill (map the codebase, find bugs, challenge each candidate, record, verify, report) that turns coding agents into an audit team @OpenAI｜@MistralAI｜@CloudflareDev
- Plain weight perturbation for agent post-training: Qwen3.5-4B goes from 12.1% to 21.1% on Terminal-Bench 2.1 — Beating the official Tmax RL checkpoint's 18.2%. No credit assignment, critic, or masking; population size around 30. The authors think this route is most valuable for long-horizon agentic post-training @catherineliangq (paper author)
- Low-cost robot manipulation learning, two-piece set: CMU's $15K humanoid Pixel, and TouchBridge tactile demo transfer — Pixel pairs with the Pixel-UMI training framework for environment-aware whole-body mobile manipulation. TouchBridge uses only human demos and averages 72.5% success on 4 real tasks, versus 77.5% with an equal amount of teleoperation data; human data collection is 7.0x faster, and retargeting needs no object mesh, pose estimation, or physics simulation @deepakpathak (CMU professor)｜@Gu__Zhang (first author)
⭐ Featured Content
AI infrastructure's constraint shifts from compute/power to "politics": midterm data-center ballot measures almost all rejected; Stratechery offers a "globalization vs. needing permits to build data centers" framing | The anti-data-center wave becomes the ultimate bottleneck on AI diffusion
In the 2026 midterms, 41 data-center-related ballot measures across 10 states were almost all rejected (the sole exception was Harwood, North Dakota), with Ohio alone accounting for 21; Texas elected a governor who pledged to pause new data-center construction, and two data-center-friendly mayors were voted out. Stratechery's core insight: when manufacturing moved offshore, affected communities had no say, but building data centers requires permits — so these people now hold veto power over AI for the first time, and they're using it to express discontent. The biggest beneficiary of the anti-data-center wave may ironically be the frontier labs, since permitting barriers raise costs for newcomers. TIME adds that NBC exit polls show 56% of voters listed AI and data centers as a key voting factor; Democrats spent over $100M on ads attacking opponents' ties to Big Tech, and accelerationist super PAC spending didn't work. For practitioners, this is a first-hand sample of the constraint shift from "compute → memory → power → politics," and a new variable for judging data-center siting and compute supply pacing.
OpenAI's autonomous vulnerability-research pipeline: scout + analyst dual agents run scan → verify → exploitability assessment → patch generation, confirming and patching 948 local privilege-escalation bugs in the Linux kernel at roughly $1.2M in API cost | An end-to-end engineering postmortem of agents in the security vertical
The article gives a section-by-section cost breakdown and throws out several counterintuitive conclusions: stronger models need lighter orchestration, maintainers patch fast, models can find new exploitation techniques, and the testing infrastructure itself has vulnerabilities. For teams doing agent orchestration, evals, and security automation, this is a rare first-hand postmortem of "four-stage pipeline + cost structure + orchestration lightening" — far denser than demo-grade agent case studies.
Sources: openai.com
llm-d 1.0 ships its first production release: CNCF-governed, co-founded by Red Hat/IBM/Google Cloud/CoreWeave/NVIDIA, session-aware routing hugging the prefix cache; IBM's quantum error-correction project cuts cost 11.5x, GLM 5.2 (753B) supports 3000 concurrent coding agents with no preemption | A standardization milestone for the open-source distributed inference stack
1.0 adds multi-tenant isolation, high-availability routing, rolling upgrades without interruption, a stable config interface across clouds and accelerators, and session-aware routing that pins multi-turn conversations near cached context, with backends compatible with vLLM and SGLang. Two hard data points: IBM Research's quantum error-correction hardware project sees about 98% prefix-cache hit rate, roughly $0.11 per million tokens blended cost, 11.5x lower than a fully closed-source setup; GLM 5.2 runs hundreds of parallel sessions of agentic workloads on existing NVIDIA GPUs, peaking at 6.6M output tokens/minute, with per-token cost 5-10x lower than commercial APIs and the biggest savings on input-heavy workloads. AMD, Mistral, JetBrains, and Roblox are already using or planning to use it. For teams doing inference deployment and agent cost optimization, these are directly comparable production parameters.
Why generative recommendation has long lost to discriminative ranking: it's not the modeling weakness of autoregressive one-to-many decoding, but that Semantic ID compression discards the fine-grained item attributes that feature-crossing depends on | First-hand industrial data from Shopee's UniRec
Shopee's UniRec patches the feature-coverage gap with three moves: pre-generate item attributes via Chain-of-Attribute before decoding SID, do exposure-weighted codebook balancing during quantization, and use layer-wise stop-gradient during preference fine-tuning. In production it directly replaces the multi-stage cascade, cutting end-to-end latency from 266ms to 110ms while lifting GMV +5.60% and orders +4.76%. For engineers working on unified recall/ranking and semantic IDs, this clarifies "where generative recommendation's ceiling is" — the gap is a feature-coverage defect, not a modeling limit — and offers three reusable engineering fixes (the body is paywalled; the CoA mechanism and codebook-balancing formula need to be verified in the original).
Sources: machinelearningatscale.substack.com
OpenAI acquires Pierre Computer Company: the Trees for browsing files in Codex, the Diffs for viewing changes, the Edit mode for inline diff editing, the streaming syntax-highlighting Highlights, and the Code Storage that keeps history in Git all came from this company | Codex pulls the whole "agent creates branch → human reviews code → deploy" chain in-house
Combined with the Pando in-house Git backend announced earlier this week, OpenAI's intent is clear: redo the source-control standard for the agent era. Pierre's open-source projects stay open source, Code Storage remains available to customers after the deal closes, and the transaction still needs regulatory approval. For coding-agent and developer-tool practitioners, this is a roadmap for watching how a big lab redoes the Git workflow; the official announcement is PR-flavored, with no technical details or deal amount.
Sources: openai.com
Nathan Lambert estimates the open-model inference economy: Fireworks alone already exceeds 40T tokens/day in a single custom-model vertical, more than double this summer, with the total possibly topping 100T this fall — and the curve bent upward after GLM 5.2 brought an open-source agent step-change | Using disclosed data from four inference startups to piece together whether "open inference is a rounding error"
Methodologically it uses Baseten, Fireworks, and Together AI — three open-inference startups — as bellwethers, layers on OpenRouter's open-model token share, and compares magnitudes against Google's earnings-disclosed all-platform token volume. The author admits all open-source adoption metrics are extremely imprecise (platform noise, startup marketing framing, imperfect public evals), so this is a "best approximation," not a verdict. For anyone trying to judge the size and growth slope of the open-inference market, this is a rare cross-vendor magnitude puzzle.
Sources: interconnects.ai
AWS distills Anthropic's context-engineering advice into a four-layer framework: architecture layer isolates context via multi-agent, prompt layer limits intake, tool layer externalizes heavy content to the filesystem, verification layer backstops with an independent agent | A production-grade recipe with runnable code and quantified gains
It uses the open-source multi-agent data-analysis solution Deep Insight (Strands Agents SDK + Bedrock AgentCore) as the reference implementation, with each layer explicitly noting that its own limits are covered by the next. Quantified gains: a single write-and-execute tool cuts tokens injected into context per script by about 95%, and on-demand skill loading shrinks the system prompt by about 97%. For teams building long-horizon agents, this is an engineering recipe you can copy directly — at the cost of binding to the SageMaker/Bedrock stack.
Sources: aws.amazon.com
MCP server engineering postmortem: the model imported 127 scenes but only re-rendered 22, yet reported "import complete" — the lesson is that tool descriptions are suggestions, and required parameters are the rules | Any rule protecting user work must be enforced by the server
ScreenWeaver's founder recaps shipping 23 MCP tools in October (10 read, 13 write): the in-app assistant and external MCP clients share one server-side tool registry (same validation, same 60/min rate limit); destructive tools force confirm:true or return only a preview; paid tools force maxCredits and return over_budget when over budget; write operations auto-snapshot a version; and tests fail if a tool referenced by the guide doesn't match the registry. The core insight: once the model gets confused it skips tool descriptions and system prompts, so schema-level enforcement is the only reliable boundary. For anyone building MCP servers, these are patterns you can copy directly.
Sources: hackernoon.com
📈 Market Movers
Trading day: 2026-11-04
Market: STAR 50 +9.92%, ChiNext +6.40%, Semiconductor ETF (SOXX) -3.14%
STAR 50: Main-force capital poured in during the afternoon (institutions 80B, main force nearly 70B, STAR 50 over 10B), and with the US midterms settled, the semiconductor chain saw a wave of limit-ups — memory chips, advanced packaging, and CPO all surged, sending STAR 50 up 9.92% on heavy volume.
ChiNext: A wave of limit-ups in semiconductor ETFs drove a tech rebound, compounded by state funds buying ETFs and lower buyback-loan costs; ChiNext had fallen about 29% cumulatively since September 18, and a repair rally followed the rapid valuation compression.
MetaX (688802) +16.19% | STAR 50 +9.92%, relative +6.27%
Reason pending
Vistra (VST) -13.83% | Nasdaq-100 ETF (QQQ) +0.97%, relative -14.80%
Reason pending
Moore Threads (688795) +13.60% | STAR 50 +9.92%, relative +3.69%
On November 4, Moore Threads, together with Qutrunk and Zhipu AI, launched a "domestic heterogeneous inference AI Token factory solution" built on the MTT S5000 compute base; combined with the Wuxi intelligent computing center running 2,000 GPUs at full load and a granted invention patent, multiple positives drove the stock up 13.6%.
Eoptolink (300502) +13.35% | ChiNext +6.40%, relative +6.95%
Reason pending
Vertiv (VRT) -12.89% | Nasdaq-100 ETF (QQQ) +0.97%, relative -13.86%
Democrats won the midterms, and the market worried Washington would slow US data-center construction, sending AI-infrastructure stocks broadly lower; compounded by earlier inflation data raising rate-hike expectations, financing concerns, and analysts cutting price targets, Vertiv fell 12.89%.
Montage Technology (688008) +11.87% | STAR 50 +9.92%, relative +1.95%
Reason pending
Zhongji Innolight (300308) +11.07% | ChiNext +6.40%, relative +4.67%
On the evening of November 3, the company's earnings call clarified there was no broad order cut — only structural adjustments at individual customers — and customer demand had not been revised down; Citi maintained its Buy rating and A-share target price of 1,212 yuan the same day, saying the prior day's 12.7% plunge was an overreaction to order-cut fears; combined with a CPO sector rebound and 2.512B yuan in margin buying, the stock rose 11.07%.
Jinpan Technology (688676) +10.82% | STAR 50 +9.92%, relative +0.90%
Reason pending
🎙️ Podcast Picks
How the Midterm Results Change AI
📍 Source: AI Daily Brief | ⭐ 3/5 | 🏷️ Regulation, Infra | ⏱️ 00:24:41
NLW breaks down what the Democratic midterm wins mean for AI: data-center construction faces tougher scrutiny, the direction of AI regulation, and the shape of the 2028 presidential race. His core take is that community backlash may not halt AI infrastructure — it could instead produce deals more favorable to local communities. Useful for anyone tracking AI policy, compute buildout, and regulatory risk, though it doesn't touch LLM/agent tech itself.
💡 Why Listen: If you read today's ballot-measure story and want the policy follow-through, this is a tight 25 minutes. Skip it if you're here for model releases — it's pure policy and infra risk.
📄 Paper Highlights
Dory1.0-4B: A Looped Normalized-GPT
NVIDIA ｜ 🏷️ Architecture, Inference, Quantization
A 4B model with elastic depth — one checkpoint, variable loop count at inference, so you dial compute up or down per query. More loops mean fewer output tokens on agentic tasks.