📊 Today's Overview
Anthropic shipped two models at once — Claude Fable 5.2 and Mythos 5.2 — with API prices down 20% and FrontierCode 2.0 scores beating Gemini 4 Carbon and GPT-6 Astra. Google answered with four eighth-gen TPUs, including a dual-die 8i Max built for million-token contexts. Nvidia's GR00T 2 learned from 100K hours of human first-person video. And OpenAI plus Anthropic both claimed sub-exponential algorithms against the 25-year-old Exponential Time Hypothesis, while three academics stepped in to referee.
🔥 Trend Insights
- Effort tiers beat bigger models: Fable 5.2 High effort costs $6.72 per task versus Opus 5.5 Max at $12.82 for near-identical scores — pushing Fable to Max actually costs $44.70 and scores worse.
- Accelerators split by workload: Google's four new TPUs (8t train, 8i infer, 8tx confidential, 8i Max long-context) end the one-chip-fits-all era; TorchTPU and TPU Sync cut KV transfer latency 90%+.
- Context windows are overrated: Nvidia's team shows dropping 256K to 16K lifts decode throughput 7x, with smart handoff briefs beating full history on most long-horizon tasks.
🐦 X/Twitter Highlights
📈 Trends & Buzz
- Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2 — Both models claim gains in coding, reasoning, and vision, live today. In pre-release testing, Fable 5.2 designed three working electronic devices from a single prompt, including a home voice assistant @claudeai (Anthropic official) | see also @AnthropicAI
- Ethan Mollick calls Fable 5.2 "Jagged AGI," has it start companies, do research, attack hardware, build 3D models — He also notes Fable 5.2 exceeds prior trend lines on three unsaturated benchmarks: ECI, Remote Labor, and RoboDojo @emollick (Wharton professor) | benchmark chart
- Google releases four eighth-gen TPU systems — Including TPU 8i Max (large context, highest per-rack compute density) and TPU 8tx (confidential inference). Open-source TPU Sync cuts KV transfer latency by over 90%, and TorchTPU runs native PyTorch and Hugging Face models with zero code changes @sundarpichai (Google CEO, quoting AI infrastructure lead Amin Vahdat)
- Meta releases its smallest model, Muse Ember — Zuckerberg says it delivers more intelligence per dollar than peers, aimed at token-heavy agent tasks repeated all day: tool calls, coding subtasks, routing, extraction, long-document parsing. Now live on Meta Model API and Muse Code (Meta's coding product) @finkd (Meta CEO)
- Baidu releases ERNIE Evolve, an automated research system doing recursive self-improvement (RSI) — It beats the best publicly known results at evaluation time on three metrics across NanoChat AutoResearch and NanoGPT Speedrun, and was already used to build ERNIE 5.5 Flash @BaiduResearch (Baidu Research)
- Amazon forms a Personal Robotics Group for home robots — Per CNBC, the team is built around Fauna Robotics (maker of the Sprout humanoid), folding in the former Astro team, Lab126, and MIT robotics talent. A separate Frontier AI & Robotics team handles foundation models and full-stack systems, with deployment still warehouse-first. Amazon's fulfillment network already runs over 1 million robots @CyberRobooo (robotics blogger, relaying CNBC)
🔧 Tools & Products
- Fable 5.2 lands in Perplexity and Cursor — Perplexity offers it in Computer or Teammates, claiming a clear lead over Gemini Argon on WANDR (broad + deep research) tasks. Cursor reports Fable 5.2 at 60% on CursorBench's top tier and 56.9% on default, versus 49.2% for Fable 5.1 default @AravSrinivas (Perplexity CEO) | @cursor_ai (AI coding tool)
- OpenAI's dot connects to multiple local computers — Users pick which machine runs a task, e.g. one dedicated to building and testing apps while the laptop keeps writing code @OpenAIDevs (OpenAI developer account)
- NVIDIA ships GR00T 2 at CoRL 2026 — An open-world action model for humanoids, paired with RLWRLD (robotics AI company) and a new UC Berkeley industrial benchmark on Isaac Lab-Arena covering wiring and small-part assembly skills @NVIDIARobotics (NVIDIA robotics team)
- Skild AI open-sources physics solver Steel — Built with NVIDIA, up to 13x faster simulation than PhysX with roughly 3x less VRAM, aimed at scaled simulation for robot manipulation @SkildAI (robotics foundation model company)
- Claude publishes an official model selection guide — Breaks down when to use Fable 5.2, Opus 5.5, Sonnet, and Haiku, plus when to adjust effort instead of switching models and how to offload sub-agents to smaller models to save cost @ClaudeDevs (Claude developer account)
⚙️ Engineering Practice
- Three inference stacks post throughput gains the same day — FlashState more than doubles vLLM throughput on GLM 5.3 Flash, cuts latency 3.7x and user cost 2.5x; SGLang lifts GLM-5.3-Flash long-context end-to-end throughput 2.11–4.28x on 16×H200; Together Compiler versus the best InferenceX configs on B200/B300: Kimi K3 2.7x, GLM-5.3 and DeepSeek V4.1 Flash 1.8x each, MiniMax M3.1 Flash 1.5x @_weilix (FlashState author) | @lmsysorg (SGLang maintainers) | @togethercompute (inference provider)
- Databricks cuts AI coding unit cost by up to 90% — It started by giving engineers the most expensive coding models and unlimited tokens; adoption took off, and so did costs. More efficient models, better routing, spend visibility, and reduced token overhead brought it down. Co-founder Patrick Wendell explains on camera @databricks (data and AI platform)
- World Labs' Atlas tops Rapidata's camera-motion benchmark — The eval covered 19 world and video models, 600 prompts, and 513K human judgments @theworldlabs (world model startup)
- Ethan Mollick uses Fable 5.2 to generate a 3D world with 450+ connected portals — All 25,000 agents are code-generated, inspired by books, art, and places; the demo and GitHub code are out @emollick (Wharton professor)
- Boston Dynamics puts wheels on Atlas to test an RL whole-body controller — The goal is validating reliability under edge conditions on real actuators @BostonDynamics (robotics company)
⭐ Featured Content
Anthropic ships Claude Fable 5.2 and Mythos 5.2: API prices down 20%, FrontierCode 2.0 beats Gemini 4 Carbon / GPT-6 Astra | Frontier model rankings reshuffle + cost curve shifts down
Fable 5.2 (general) and Mythos 5.2 (security/life-sciences focused, same base with different safeguards, limited to vetted Project Glasswing users) launch together, with prices dropping from $10/$50 to $8/$40 per million tokens and cache reads/writes cut too. The number worth remembering: on FrontierCode 2.0, Fable 5.2 at High effort averages $6.72 per task, while Opus 5.5 at Max costs $12.82 for nearly the same score — but pushing Fable to Max costs $44.70 and scores lower, showing higher effort tiers aren't always better. Anthropic still recommends Opus 5.5 for most coding work; Fable suits migrations and tricky bugs where retries are expensive, or as an advisor called on demand by cheaper models. Head-to-head: FrontierCode 2.0 at 49.1% vs Gemini 4 Carbon 42.8% / GPT-6 Astra 41.1%; FrontierSWE v2 at 69.9% vs 65.5%/61.2%. AWS adds Bedrock support; rejected security/bio queries need manual code retries to Opus 4.8/5, and on a prompt cache hit the refusal response carries a fallback credit token.
NVIDIA GR00T 2 open-world action model: trained on 100K hours of human first-person data, completes 40 of 50 tasks with zero demos, 71% from one demo → 95% from five | Embodied AI's "world action model" takes shape
Built on Cosmos, it learns from human demonstrations and transfers skills across robot bodies; AI agents can chain skills via API and correct errors. It ranks top three on RoboCasa365/RoboLab/RoboDojo/MolmoSpaces. The same release includes Isaac Lab 3.0 (switchable PhysX/Newton physics solvers, new RL post-training library), a Video-to-Data workflow (human video to simulated demos), and the EgoHand 3D hand reconstruction model. Foxconn, Siemens, and Nexuni are already testing it. The LG Electronics case is the one to copy: at its Tennessee plant, it analyzes skilled-worker motions, adds teleoperation and simulation training for fixture tasks, uses Cosmos to generate varied backgrounds compensating for lighting/environment domain gaps, builds a virtual line in Isaac Sim, and compresses action-decision latency with TensorRT. The key is making "on-site failure → reproduce in virtual space → add data and retrain → fold recovery motions into learning" a process rather than learning only successful trajectories. Target: 500K cumulative training hours by year-end, with a bipedal humanoid due Q1 next year.
Sources: blogs.nvidia.com | finance.biggo.com
Google's eighth-gen TPU fills out four variants at once: 8t training / 8i low-latency inference / 8tx hardware-level confidential compute / 8i Max dual-die packaging breaks the memory wall | Accelerators shift from "single chip" to "load-typed combinations"
TPU 8tx reuses the 8t silicon but pairs it with a fifth-gen AMD EPYC host, keeping sensitive data and weights isolated during compute; Google claims 2x lower latency than 8t with almost no performance loss, aimed at private fine-tuning, distillation, RL, and agent execution. TPU 8i Max packages two 8i dies together, doubling compute/SRAM/HBM bandwidth, targeting million-token long-context agents. On software, open-source TorchTPU runs PyTorch/HuggingFace models natively on TPU with vLLM/SGLang backends, and TPU Sync moves model state and KV cache across chip memory/host memory/storage/datacenter network, which Google says can cut KV transfer latency and time-to-first-token by 90%+. TPUs have nearly 100 customers and can deploy directly into customer datacenters; Bloomberg says the new products are expected to launch in 2027. All performance numbers are vendor claims, with no pricing or independent benchmarks.
Sources: officechai.com | simpmusic.org
OpenAI and Anthropic both claim their models challenge the 25-year-old Exponential Time Hypothesis on the same day; a trio of academics takes over independent review | A live sample of "how the academic community verifies AI-produced theoretical results"
OpenAI says its deterministic algorithm can exactly count satisfying assignments of k-CNF formulas in sub-exponential time; Anthropic gives a sub-exponential 3-SAT algorithm, lists Claude as a co-author, and uses Lean 4 to generate machine-checkable formal proofs. Both draw boundaries proactively — Anthropic stresses it's only asymptotic and doesn't affect practical solving, and OpenAI denies it implies a sub-exponential algorithm for general NP problems. The real read is the academic reaction: Paturi calls it surprising but says three decades of work built on the assumption won't vanish; MIT's Ryan Williams says he feels "vindicated" as a longtime skeptic; Impagliazzo says he'd rather a human researcher could explain the reasoning. The three are co-organizing independent review. It shares a theme with "PSI proves a quantum conjecture using an unreleased Anthropic model" — AI automated discovery plus Lean formal verification — but the subjects, conjectures, and companies all differ.
Sources: tokenpost.com
METR deep-dives Anthropic's internal AI R&D: AI accelerates its own research by about 2x, and naive extrapolation puts full RSI possibly under a year away | Public models show no acceleration; you have to include unreleased internal models
Zvi's key takeaways from the report: CI 1.3x–3x, months ahead of the AI 2027 timeline, with no obvious bottleneck observed; adding Anthropic's unreleased internal models moves the estimate from ~1.7x to ~1.9x before a step change appears — exactly the evidence structure of "public benchmarks feel nothing, internals have already changed." METR says its most informative bottleneck analysis was redacted by Anthropic, and Benjamin Todd's extrapolation suggests Claude could dominate internal research by mid-2027 and approach full automation by year-end. Worth reading for anyone updating AGI timelines and RSI judgments, and it pairs well with the "RSI concerns" TechCrunch raised around today's Fable 5.2 launch.
Sources: thezvi.wordpress.com
"You don't need long context": NVIDIA team tests three context-management paradigms, Qwen 3 Coder Next decode throughput up 7x going from 256K to 16K window | Directly challenges the mainstream assumption that long context is a must
The paper notes that late-token generation in frontier models can cost up to 7x more than early tokens, so agents may not need long context. Comparing full history accumulation (FHA), compress-at-threshold (EoCC), and state-handoff-only per step (SAC) across six long-horizon task types, only 5 of 22 configurations for Claude Opus 4.5 and GPT 5.6 Sol dropped more than 2 points versus the full-history baseline; long-input tasks instead treat documents as files, letting the agent retrieve and read only bounded snippets. The hardest engineering number: on H100 + vLLM, dropping the window from 256K to 16K lifts decode throughput 7x. Key detail: good compaction is a "forward-looking handoff brief" (task, progress, plan, completion criteria, artifact pointers), not a vague summary of history — hence the proposed concept of semantic state space models. Note the 7x is serving throughput, not end-to-end task speedup.
Sources: alphaxiv.org
OpenAI opens up its research infrastructure "control room": every workload must bind an approved safety case to get a job_id, the inference engine checks authorization before every sample, and stops emitting tokens within 35 seconds | An architecture blueprint for an industrial-grade model kill switch
All research-side inference workloads (RL experiments, evals, large-scale sampling, data processing) are registered in one place; the inference engine asks the Control Room whether a job is still authorized before each sample, and refuses to emit tokens once stopped. The design stresses that "even if the Control Room itself goes down, it must still be able to stop," and it links to misalignment monitoring systems. For teams doing agentic training that need runtime circuit breakers on models, this is a rare first-hand reference for making "authorize — sample — interrupt" a three-stage hard constraint; the job_id authorization pattern is directly reusable even at small-to-medium scale.
Sources: openai.com
Meituan open-sources LongCat-2.5 under MIT: visual judgment + multi-step coding, Terminal Bench v2.1 jumps from 70.8 to 88.4, DeepSWE v1.1 from 9.7 to 64.6 | Teams that ruled out 2.0 should retest
It targets combined tasks: give the agent a screenshot of a broken checkout page, have it locate the repo component, submit a patch, then verify the result in a browser. SWE Atlas goes from 33.1 to 58.3, and the size of the jump shows the previous generation's weaknesses were systematically fixed. But it doesn't lead across the board: on Terminal Bench v4.0 it scores 38.5, behind Claude Opus 5's 53.9; Vision2Web is also led by Opus 5, and DeepSWE by DeepSeek-V4.1-Flash. Note the policy risk too: DHS has ordered federal agencies to exclude DeepSeek, and while no restriction names Meituan yet, rules can change fast — a key variable in whether Chinese open models can enter US enterprise procurement.
Sources: venturebeat.com
Amazon × Hugging Face release HOPE: second-order expert pruning, best average rank at 50% pruning, up to +6.1% over the next-best method on agentic coding | MoE deployment cost-cutting shifts from "independent scoring" to "modeling expert collaboration groups"
The post points out that mainstream MoE expert pruning is first-order — give each expert an independent scalar importance score and delete — but in real MoEs experts form stable co-activation groups, and ignoring pairwise interactions causes significant drops at aggressive pruning rates. HOPE uses a second-order method that models pairwise expert interactions to decide the pruning set. The post does the memory math for Qwen3.5-122B: 256 experts × 48 layers, expert MLPs take about 95% of ~250GB of parameter memory, while each token uses only 8 — those numbers alone are the best argument for why pruning is worth it. It includes a How to run HOPE section and a limitations discussion, a rare "method + runnable" package.
Sources: huggingface.co
📈 Market Movers
Trading day: 2026-11-09
Market: Semiconductor ETF (SOXX) +7.49%, Nasdaq-100 ETF (QQQ) +3.54%
Semiconductor ETF: AMD announced a Google Cloud processor partnership and got an analyst upgrade, rising 11.8% and lifting chip stocks out of last week's selloff; combined with falling Treasury yields and retreating oil prices, SOXX gained 7.49%.
Nasdaq-100 ETF: underweight investors were forced to chase, triggering a strong rebound; multiple tech stocks recovered from the recent selloff, lifting the Nasdaq 100, with QQQ up 3.5%.
AMD (AMD) +13.26% | Semiconductor ETF +7.49%, relative +5.77% | up 2 days in a row, +25.76% cumulative
Oppenheimer upgraded AMD to Outperform with a $600 target and announced that fifth-gen EPYC processors will power Google Cloud's TPU 8tx, which combined with a broad chip-sector rebound sent AMD up 13.26%.
Constellation Energy (CEG) +10.46% | Nasdaq-100 ETF +3.54%, relative +6.91% | up 2 days in a row, +15.21% cumulative
Q3 adjusted EPS of $5.06 far beat the $3.88 estimate, revenue of $9.55B rose 45% year over year, and it raised 2026 guidance for the second time to $12.50–12.80, on top of the AI datacenter power demand theme.
Vertiv (VRT) +9.63% | Nasdaq-100 ETF +3.54%, relative +6.09% | up 2 days in a row, +13.43% cumulative
Post-midterm fears of a datacenter construction slowdown eased and AI stocks broadly rebounded, lifting Vertiv with the sector; strong results — Q3 revenue up 61%, orders doubled, and raised 2026 guidance — kept feeding through.
Unitree Robotics (688836) +9.04% | STAR 50 +0.30%, relative +8.74%
Apple's research team published HALO, an autonomous control system, on arXiv; it hit a 91% average success rate on Unitree's G1, triggering a limit-up wave across physical AI and humanoid robotics names, with Unitree closing up 9.04%.
Arista (ANET) +9.00% | Nasdaq-100 ETF +3.54%, relative +5.46% | up 2 days in a row, +16.97% cumulative
UBS, Morgan Stanley, Melius, and JPMorgan all raised targets after Q3 revenue of $3.66B rose 58.5% year over year, with AI networking demand strong; shares rose 9.2%.
UBTech (09880) +8.99% | Hang Seng TECH -0.92%, relative +9.91%
The company and Xinhua jointly unveiled the world's first bionic humanoid robot news anchor, and the market kept digesting the news; Apple's HALO system also set off a physical AI rally, and the company signed a comprehensive strategic cooperation agreement with Tianyi Telecom Terminal that day.
Broadcom (AVGO) +8.78% | Semiconductor ETF +7.49%, relative +1.29%
At KubeCon, Broadcom announced the VM API project entering the CNCF sandbox and expanded VMware VKS partner offerings, on top of Fox Advisors raising its target to $715, Anthropic's S-1 filing easing compute-leasing concerns, and a broad rebound after strong jobs data; shares jumped 8.78%.
GE Vernova (GEV) +7.13% | Nasdaq-100 ETF +3.54%, relative +3.59%
It signed an onshore wind turbine supply and long-term service agreement with Germany's Stadtwerke Heidenheim, and industry analysis reiterated strong structural demand for gas turbines, driving a rebound; Constellation's guidance raise also eased datacenter moratorium fears.
🎙️ Podcast Picks
What Frontier Labs Buy: A Data Deep Dive with Handshake AI — Garrett Lord & Team
📍 Source: Latent Space | ⭐ 4/5 | 🏷️ LLM, Research, Interview | ⏱️ 2:43:49
Latent Space's first deep dive into AI data company Handshake AI, whose founder Garrett Lord and team explain how they built a training-data business to nearly $1B in annualized revenue in about a year. The core discussion covers what frontier labs actually buy, how RL training data gets constructed, the data supply chain and human org behind it, and how data quality and scale affect model capability. A rare supply-side view with commercialization details for anyone working on LLM post-training, data flywheels, and agent training data.
💡 Why Listen: Nearly three hours of "here's what labs actually pay for" from someone with real revenue numbers. If you've ever wondered where post-training data comes from, this is the supply side talking.
General AI Assistants Go Mainstream — Where's the Opening for Vertical Agents? | S10E35 with Han Sheng of QuVideo
📍 Source: Tech News You Should Know | ⭐ 4/5 | 🏷️ Agent, Product, MultiModal | ⏱️ 38:04
QuVideo CEO Han Sheng starts from video creation to discuss how general personal agents and vertical agents divide the work: the general agent holds long-term user context, the vertical agent handles professional delivery — like an old family steward working with specialist teams. Key points: agents must grasp the real goal and evaluation criteria behind vague user phrasing; the test for building your own model is whether it secures core experience and cuts inference cost; judging a creation agent means looking at edit counts and satisfaction, not just generation success rate; once AI coding lowers the build barrier, spotting real demand matters more than dev skill; localization means understanding local culture and expression.
💡 Why Listen: A practical take on the general-vs-vertical agent split from someone with a decade-plus in video products. Short, dense, and full of product judgment rather than theory.
E258 | Silicon Valley's Startup "Monastery" HF0 and the Early OpenRouter Story
📍 Source: Silicon Valley 101 | ⭐ 3/5 | 🏷️ Product, Funding, MultiModal | ⏱️ 54:19
Guest Guo Jing (Joe) recounts 12 weeks at Silicon Valley incubator HF0: focusing on a single goal by "subtracting," the weekly Demo Dinner ritual, and growing music-generation product ACE Studio's monthly revenue from $80K to $800K. The episode also reveals OpenRouter co-founder Louis's hardcore working style during HF0, plus how later entrants are chasing Suno in AI music and the verdict that the space is "polished but not gritty enough." Useful for AI product growth tactics and the music-generation competitive landscape, though light on model technical detail.
💡 Why Listen: The OpenRouter origin bits are the draw. The rest is incubator methodology — good if you like growth stories, skippable if you want architecture.