📊 Today's Overview
Mistral dropped the open weights for Mistral Large 4, a 1T-parameter multimodal flagship that hits 19.7% on FrontierCode 2.0 and jumps Terminal-Bench 4.0 from 28% to 41% — Europe's strongest open coding model. OpenAI published a deep architecture writeup on Pando, its in-house Git backend now serving the entire monorepo, with push latency down nearly 5x and ~1PB moved daily. Meanwhile AI infrastructure stocks got hammered: Marvell fell 12.72% on its $115B Lumentum acquisition, Oracle dropped 8.76% on a $20B raise, and CoreWeave hit an all-time low.
🔥 Trend Insights
- Agent scaling breaks the toolchain: OpenAI had to rebuild Git itself (Pando) to survive Codex's concurrent writes — agent scale now stresses version control before models.
- Token efficiency as the new benchmark: Mistral Large 4 matches GLM-5.3 on coding agents while using 47% fewer tokens; raw scores alone no longer tell the story.
- Real-time, not reasoning, is the agent bottleneck: Realtime-Bench shows top models at just 36% under wall clock but 70-80% with a paused clock.
🐦 X/Twitter Highlights
📈 Trends & Buzz
- Anthropic discloses in amended S-1: models may be nearing the RSI policy threshold - "RSI" means "rapid, AI-driven acceleration in progress," a risk tier in Anthropic's Responsible Scaling Policy. The company says this is an internal estimate and crossing it isn't confirmed @MTSlive (AI news aggregator account)
- Runway, 8 years in, added over $100M in net-new ARR last quarter alone - Co-CEO Anastasis Germanidis said in an interview the company "accidentally built one of the best robotics models," and its video models can train robots and autonomous driving. After Sora launched, the team built a stronger model in three months @LubaYudasina (podcast host)
- Nextcloud picks Olmo 3 to power Nextcloud Assistant - Ai2 says Nextcloud chose an open model so users can inspect how the model was built; local deployment is only part of digital sovereignty @allen_ai (Allen Institute for AI)
- Ethan Mollick: large-scale replication of old LLM research with new models shows many common beliefs are outdated - Results show new models no longer fabricate citations, and hallucination rates and stereotyping both dropped; but models are still sycophantic and still lean left politically @emollick (Wharton professor)
- Figure's newly built F.03 robots leave the BotQ factory - Founder Brett Adcock posted footage of robots crossing the campus; no numbers or delivery recipients disclosed @adcock_brett (Figure founder/CEO)
🔧 Tools & Products
- Scenario open-sources 3D OS: six departments and 80 skills for coding agents - Covers modeling, rigging, Look Dev (materials and lighting), world building, and engine export, with each department handing artifacts to the next. Expert Tools can operate Blender, Maya, ZBrush, Unreal, Unity, Godot, and Roblox locally. Supports Claude Code, Codex, Cursor @Scenario_gg (3D content generation platform)
- Replit launches Cloud Agents: non-engineers can now edit company codebases - Submitted changes are reviewed and approved by engineers. Also supports creating sessions inside an Agent, and GPT-6.1 Terra is added to the model selector @Replit
- Runway ships a Cursor plugin: wire the Runway API into your app in one sentence - After installing the plugin, describe what you want to generate and Cursor writes the integration code @runwayml
- 3D World Simulator: an interactive world model at 50+ FPS on an RTX 5090 - Predicts in a unified 3D representation space, generating novel-view appearance, multi-view-consistent RGB-D, 3D geometry, and spatially aligned tactile signals @jwang771 (3D World Simulator author)
- Perplexity open-sources vision-language embedding model pplx-embed-v2-context, in 9B and 0.6B versions - Returns one int8 vector per document; when chunked, each vector still sees the whole document. Paper co-written with turbopuffer @perplexitydevs
⚙️ Engineering Practice
- Weight-Space Programming: GPT-6 Astra directly edits 0.016% of Qwen3.5-0.8B's weights - No gradients, no training; math ability improves up to 6x, instruction following 3x, and it fixes the 0.8B model's reasoning infinite loop. Yoav Goldberg's reaction: "This makes me want to know even more what OpenAI actually trained these models on" @rikiyatakehi (first author) | discussion @yoavgo
- SGLang HiCache brings KV reuse on top of CP/PP/speculative decoding/PD disaggregation - iFlytek's Astron MaaS CodingPlan workload, serving GLM-5.1/5.2, raised cache hit rate from 30% to 95% after connecting external storage; inputs often run 30K–200K tokens, up to 1M. Median time-to-first-token dropped under 3 seconds, down from latency so high the service was basically unusable @lmsysorg (SGLang open-source team)
- Latent Sparse Attention: up to 22.7x faster than sparse attention, 72.4x faster than full attention - At 10M-token context, finding the right tokens to attend to takes 98% of the attention layer's time; LSA compresses that step. The authors say accuracy matches full attention at ultra-long context @jcrwhittington (first author; collaborators include ETH's Torsten Hoefler)
- MIT CSAIL's OPAL: the optimizer reads dynamics during training and rewrites its own source - Driven by an AI agent, it rewrites source mid-training, inventing learning-rate schedules, suppressing loss spikes, and escaping overfitting. Improves both vision and language tasks @MIT_CSAIL (MIT Computer Science and AI Lab)
- h1: a Lean proof harness "embarrassingly simple" yet SOTA, plus 4 more Erdos problems solved - The team was working on asymmetric self-play test-time scaling; each time they simplified the baseline, the self-play gain vanished — and the leftover h1 beat most methods. The paper has been held on arXiv for 30+ days @KempeLab (NYU professor / KempeLab lead)
- Databricks: after agents took over the experiment loop, AI Functions research training runs grew 8x - Researchers now have agents propose changes, run experiments in parallel, evaluate, and decide next steps. At scale, startup latency and GPU reliability became bottlenecks, solved with AI Runtime @databricks (data and AI platform company)
⭐ Featured Content
Mistral Large 4 (Le Chonk) open weights: 1T native multimodal flagship focused on agentic coding, 19.7% on FrontierCode 2.0, +13 points on Terminal-Bench 4.0 | Europe's strongest open coding model for sovereign AI
The 1T-parameter flagship previewed three weeks ago now has open weights. Just by continuing to scale RL, it hits 19.7% on FrontierCode 2.0 (ahead of GLM-5.3, Kimi K3, DeepSeek V4.1 Flash) and lifts Terminal-Bench 4.0 from 28% to 41%. On the Artificial Analysis Coding Agent Index it roughly ties GLM-5.3 while using 47% fewer tokens. It adds Low/Medium/High reasoning modes for cost-performance tradeoffs, and emphasizes training from scratch on European infrastructure and self-hostability. Worth noting: "token efficiency" is becoming a new comparison axis for open coding models — but all figures are vendor-reported and need third-party retesting.
Sources: mistral.ai
OpenAI's in-house Git backend Pando serves the entire monorepo: median push latency down nearly 5x, ~1PB transferred daily, causal consistency satisfies both write scaling and read bandwidth | Agent scale crushes version control first
To handle the concurrent code-writing pressure from Codex, OpenAI built its own Git service, Pando, now fully deployed. The post starts from Git object storage basics, compares the tradeoffs of three-phase commit vs. WAL distributed Git models, and lands on causal consistency to satisfy both "horizontally scalable writes" and "reads that saturate bandwidth." It's fully compatible with the Git protocol and clients, so engineers and agents don't change workflows, and changes were contributed back upstream. For teams running large-scale parallel agent coding, this is a complete architecture decision chain for "agent scale will blow up version control first."
Sources: openai.com
Realtime-Bench reveals agents' real bottleneck is real-time responsiveness, not reasoning: strongest model passes just 36% under wall clock, 70-80% with a paused clock | A high-density metric for predicting where agents break
Import AI 479 highlights Realtime-Bench: 50 interactive game and web tasks, split into RT-Perceive (temporal state awareness), RT-Deliberate (adjusting thinking length to remaining time), and RT-Act (timing actions), with three clock conditions: wall clock, paused clock, and token clock. The results sting — GPT-6 Astra passes just 36% under wall clock, while high-reasoning models reach 70-80% with a paused clock, showing the bottleneck isn't reasoning but real-time response. For anyone building Computer Use or real-time interactive agents, this is a proxy metric closer to real failure modes than static benchmarks. Also includes a DIA space-warfare report and a short take on Genie 4.
Sources: importai.substack.com
Legora CTO on building legal agents: zero customer traces visible, so 3 engineers + 3 lawyers built a simulated environment for eval — one release lifted citation accuracy from 70% to 99% | Agent eval methodology when data is invisible
Latent Space interviews Legora CTO Jacob Lauritzen, who reached $100M ARR and 1800+ customers in 18 months. The core tension: legal agent traces belong to law firm customers, so the team sees none of them. The fix was a simulated environment built by 3 engineers and 3 lawyers mimicking real client work (e.g., a 10K-file data room for due diligence, requiring the agent to flag the right red flags), rewriting traces into redacted versions, narrowing failures into 10-20 failure modes, and using feature flags to attach cheap LLM checks in production. Also: a 100-person legal engineer pool (3-4 years at big firms + CS degrees), a full-stack lock-in on Azure, post-training open weights for narrow tasks like legal retrieval, and citations with highlighted source-PDF bounding boxes on all generated content. The two interviews overlap; read them together.
AWS breaks down agent cost structure: a task isn't one call but a search→read→refine→verify→decide loop, with SeqKD distilling small models to cut costs | A cost-reduction recipe for taking agents from demo to production
Using a "shopping assistant agent" as an example, it explains agentic app cost structure: each step bills separately, and at scale per-task cost eats ROI. The fix is sequence-level knowledge distillation (SeqKD): a frontier teacher generates multi-step trajectories, a programmatic grader keeps only those that actually solve the task, and they're encoded into SFT data to train a small model — no teacher logits or RL infrastructure needed, so closed-source API teachers work too. The post notes a special value of agents — the copied sequence is the behavior itself, so distillation transfers both reasoning and multi-step actions. Suited to teams blocked by inference cost and pushing agents to production, though it locks you into the SageMaker vendor stack.
Sources: aws.amazon.com
"LLMs can directly read and write neural network weights" taken apart point by point: identifying small networks via first-layer weight L2-norm heatmaps (three lines of Python), and "fixing" Qwen math is really steering/abliteration | A critical framework for spotting inflated capability claims
Responding to a paper claiming GPT-6 Astra can natively edit 0.8B Qwen weights to boost math ability, the author dismantles each "superhuman" conclusion: identifying what game a small network plays relies on labeled input channels plus a first-layer weight L2-norm heatmap, reproducible in three lines of Python; "fixing" Qwen math is really the model looping after 32K tokens, and Astra uses a steering/abliteration trick — "write looping/non-looping text pairs → measure weight contribution → prune" — that human researchers do too. Core point: directly editing weights is like editing compiled binaries — doable but not optimal; the right way for LLMs to help AI research is improving training, not magic weight-tweaking. You come away with a reusable framework for judging inflated claims.
Sources: seangoedecke.com
Four physical bottlenecks for VLA on real robots: hundreds of visual tokens from multiple cameras crush the control loop to 5-10Hz, fixed-view blind spots, 2D RGB with no metric depth, camera ISP auto-exposure causing pixel jumps | A failure-mode checklist + paper index for embodied deployment
The article argues VLA looks like it solved robot manipulation in demos, but real robots expose four bottlenecks: multi-camera high-res images produce hundreds of visual tokens, and a 100-200ms transformer forward pass crushes the control loop to 5-10Hz, too slow to react to slipping objects; fixed third-person and wrist cameras each have blind spots (occlusion / lost target), and current policies often fixate on one; 2D RGB lacks metric depth, causing grasp misalignment; and commercial camera ISP auto-exposure/gain/sharpening cause pixel jumps under lighting changes, breaking out-of-distribution policies. It then walks through five works including FastVLA (language-vision attention pruning, 35-59% fewer FLOPs, 20% lower real-robot latency) and FocusVLA (dynamically routing between static and wrist cameras by phase), all with arXiv IDs and real-robot numbers — a failure-mode checklist and paper index for VLA engineering.
Sources: harpreetsahota.substack.com
A self-built 30-case trap benchmark tests agent skills: today's Claude dodges nearly all traps without a skill, and skills only help where model knowledge is weakest | The empirical case for whether skills still matter
The author wrote an agent skill (each rule from a paper or official doc) to make coding agents more reliable at data analysis, built 30 trap cases (subtotal rows getting summed, join fan-out duplicating orders, average of averages, p99 getting averaged, Simpson's paradox, etc.), ran with/without skill on Haiku 5.5 / Sonnet 5.5, and scored with Sonnet as an LLM judge over three votes. Results: without a skill it already hit 58/61 and 60/61; the skill only caught scattered misses. Three reusable findings: skills only help where model knowledge is weakest (a post-cutoff TypeScript 7 upgrade task went from 10/18 to 17/18); suspect your own test cases before trusting a baseline failure; and a skill that doesn't activate doesn't exist — changing the description from "list data types and trap names" to "explain when to load" raised trigger rate from 7/20 to 20/20. Small sample (one run per arm); verify the conclusions yourself.
Sources: dev.to
📈 Market Movers
Trading day: 2026-11-02
Market: Semiconductor ETF -3.79%, STAR 50 +3.11%
Semiconductor ETF: AI bubble worries keep building; SOXX broke below its July low and entered a bear market, with capital flowing out. The selloff triggered by Intel's disappointing guidance on Oct 30 continued, and semis kept sliding on valuation re-rating.
STAR 50: Market-stabilization measures arrived, reversing the stocks-and-bonds double selloff and easing liquidity concerns. After Friday's US selloff, Asia-Pacific steadied, with semiconductor AI hardware leading; STAR 50 ETF volume surged to its highest since mid-July. Strong Q3 net profit growth on the STAR board also provided support.
Marvell Technology (MRVL) -12.72% | Semiconductor ETF -3.79%, relative -8.94% | down 3 days in a row, -31.89% cumulative
Announced an all-stock acquisition of Lumentum for $115B; LITE shareholders get 4.5 shares per 1 share and will hold about one-third of the combined company, sparking heavy dilution and integration-risk concerns; shares plunged.
Shenling Environment (301018) +9.30% | ChiNext +1.87%, relative +7.43%
(Limited evidence) The company disclosed on Oct 31 that its Shenling convertible bond (1B yuan) will list on the Shenzhen Stock Exchange on Nov 6; listing expectations lifted the stock, helped by ChiNext's 1.87% gain, rebounding after several days of pullback.
Sources: ipo123.com.cn
Oracle (ORCL) -8.76% | Nasdaq-100 ETF -1.36%, relative -7.40% | down 3 days in a row, -17.52% cumulative
Oracle launched a $10B new share offering and a $10B senior notes offering the same day to fund AI expansion, triggering dilution and debt concerns; shares fell sharply. Its investor day had just committed to an additional $20B financing cap, and the market worries about the pace of fundraising.
CoreWeave (CRWV) -8.60% | Nasdaq-100 ETF -1.36%, relative -7.24% | down 3 days in a row, -21.55% cumulative
Oracle's $10B bond offering reignited worries about AI infrastructure debt; CoreWeave's own leverage is high, its bonds yield around 15%, and a technical head-and-shoulders pattern added pressure — shares hit an all-time low.
AMD (AMD) -7.52% | Semiconductor ETF -3.79%, relative -3.73% | down 3 days in a row, -19.29% cumulative
Amazon's disappointing earnings plus Intel's PC demand warning kept weighing; the market frets over semiconductor valuations and AI capex, chips sold off again, and AMD fell 7.52%.
🎙️ Podcast Picks
Building Legal Agents When You Can't See the Data - Jacob Lauritzen, CTO of Legora
📍 Source: Latent Space | ⭐ 4/5 | 🏷️ Agent, Product, Research | ⏱️ 1:07:21
Legora CTO Jacob Lauritzen explains how to build and test legal agents when the team can't see the traces of what agents actually do for lawyers — the data belongs to law firm clients. The conversation centers on eval, testing, and observability design under privacy constraints, and how to keep agents reliable without real feedback. Highly relevant for anyone building vertical-industry agents under compliance and data-isolation limits.
💡 Why Listen: A rare look at agent eval when you literally can't see your own product's behavior. If you're shipping in a regulated vertical, this is your playbook.
Ep 96: Nebius CTO on Racing the Hyperscalers Before AWS Wakes Up
📍 Source: Unsupervised Learning | ⭐ 4/5 | 🏷️ Infra, LLM, Interview | ⏱️ 01:01:08
Nebius CTO Danila Shtan talks with Jacob Effron about running an AI cloud at the frontier: why open-model inference is far from optimized, why 95% of customers run modified models rather than vanilla, and how prefill/decode disaggregation is more talked about than done. They dig into how chip scarcity kept AWS/Google from winning AI cloud, the real bottlenecks of the AI buildout, whether we're overbuilding, and how the company's engineering org puts no limits on AI token usage. Practical reference for anyone watching inference infra and compute supply.
💡 Why Listen: Straight talk from someone actually running GPU clouds. The "95% run modified models" stat alone will change how you think about inference serving.
How to Deslopify Your AI
📍 Source: AI Daily Brief | ⭐ 3/5 | 🏷️ Product, LLM | ⏱️ 00:42:42
NLW and Nufar Gaspar discuss avoiding AI slop — low-quality AI output — with the core idea of shifting the thinking burden from receiver back to producer. Practical tips include forming your own point of view, teaching the model your taste, and setting effective review checkpoints. NLW also shares how he kept control of narrative perspective while using AI to make his podcast How We Got Here. Useful for anyone doing content production with LLMs, though it's more workflow habits than model technique.
💡 Why Listen: Less about models, more about not shipping garbage. Quick listen if AI-generated content is part of your job.
Why Leibniz Couldn't Find the Mind Inside the Mill (Joscha Bach)
📍 Source: ML Street Talk | ⭐ 3/5 | 🏷️ Research, Agent, Interview | ⏱️ 01:37:07
Joscha Bach proposes the Machine Consciousness Hypothesis: consciousness is a biological learning algorithm, and self-organizing learning machines might discover it the same way. They discuss why there's no Turing test for consciousness, how chess programs are already zombies, and how faking a mind is sometimes harder than building one — weaving in Leibniz's mill, Searle, Nagel's explanatory gap, IIT and its pseudoscience controversy, Gödel's formal-system limits, and Mike Levin's morphogenesis and mind-isomorphism hypothesis. Conceptually stimulating for anyone thinking about the nature of agents and the boundaries of intelligence, though more philosophy than engineering.
💡 Why Listen: If you want a mental stretch rather than a how-to, Bach is always worth an hour and a half. Just don't expect implementation tips.
📄 Paper Highlights
Agentic World Modeling for Agent Test-time Scaling without Environment Replica
Microsoft, University of Wisconsin-Madison ｜ 🏷️ Agent Framework, Agentic Workflow, Tool Use
Reframes next-observation prediction as multi-turn evidence gathering through safe read-only queries, letting agents explore in parallel on shared, non-resettable environments — a real deployment blocker for test-time scaling.
Verify--Backtrack Loop: Structured Precondition Recovery for Long-Horizon Language Agents
City University of Hong Kong, Huawei Noah's Ark Lab ｜ 🏷️ Agent Framework, Agentic Workflow, Reasoning
Turns recovery from free-form reflection into structured planning-state updates with a bounded goal stack and dual-horizon verification — training-free, and it beats ReAct by up to 22 points on AppWorld.
Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention
ETH Zurich, Google DeepMind ｜ 🏷️ Architecture, Inference, Scaling
Hierarchically clusters keys in a learned latent space so queries score clusters instead of every key — 22.7x faster prefill at 10M tokens while holding accuracy, a practical path to ultra-long context.