📊 Today's Overview
Thinking Machines Lab dropped Inkling 2 Small — a 594B-parameter MoE with 17B active — but is holding the weights back for a week of closed safety testing before an Apache 2.0 release. Bridgewater's fine-tune of Qwen3-235B on Tinker hit 84.66% across six tasks versus GPT-5.5's 78.2%, at roughly 1/14 the inference cost. Meanwhile Anthropic is reportedly eyeing a mid-November Nasdaq listing at a cut-down $1–1.5T valuation, and a scan of 10,659 public MCP configs found two-thirds leave server versions unpinned.
🔥 Trend Insights
- Open weights, staged release: Thinking Machines ships Inkling 2 Small behind a closed safety window before Apache 2.0 — open sourcing is becoming a sequenced rollout, not a launch-day decision.
- MCP supply chain reckoning: Two-thirds of 10,659 scanned MCP configs are unpinned, 941 point at an archived reference server, and 152 leak real API tokens. Protocol maturity is outpacing operational hygiene.
- The brain eats the funding: Chinese embodied AI raised ~43.8B RMB in H1 2026, with "embodied brain" startups taking 50.8% and full-robot makers just 12.8% — capital is betting the software layer decides the ceiling.
🐦 X/Twitter Highlights
📈 Trends & Buzz
- Nvidia China export fight heats up: Jensen Huang tied to suspected Chinese intelligence figure, while another Chinese firm gets chip approval — Reports say US intelligence found a personal relationship between Nvidia CEO Jensen Huang and Alice Huang, an executive at Megaspeed, a company under investigation for allegedly helping Chinese firms dodge US chip export limits. Three congressional committee chairs want Huang at a closed-door briefing on November 18. Megaspeed's Malaysian affiliate bought at least $4.8B in Nvidia-chip equipment between June 2024 and February 2026 @mickeyhardy (Arcadia team member) | Same day, word that the US government approved Nvidia AI chip sales to Chinese firms with empty offices and ties to blacklisted companies @Kalshi (prediction market platform)
- Figure drops F.03 humanoid robot video — Founder Brett Adcock posted a prototype demo clip, no specs or delivery timeline disclosed @adcock_brett (Figure founder / CEO)
- Jim Fan teases CoRL keynote: Nov 9, 13:45, Austin Hilton — Two days out, he posted just one image and called it "Something extraordinary" @DrJimFan (NVIDIA embodied AI research lead)
🔧 Tools & Products
- Vercel AI Gateway launches DeepSeek V4.1 Flash 1017 Ultrafast, over 1000 tokens/sec generation — Called as `deepseek/deepseek-v4.1-flash-1017-ultrafast`, positioned for coding, reasoning, and agent use cases @vercel_dev (Vercel developer account)
- humans& ships Ampersand AI Gateway, free until November 13 — 744B total / 40B active params, built for group chats: tracks each person's asks, flags conflicts, shares context, and asks questions. The team says it matches or beats frontier models like Fable and Astra on multi-user long-horizon evals, and leads most multi-agent coding benchmarks @humansand (model team)
- DECIDE voice agent decision benchmark launches: 1,081 real calls, 32 tasks — Built with Bolna, it pulls 4,867 true/false questions from 3–25 second call clips across 6 capability types: turn-taking, flow routing, tool selection, language switching, caller emotion, and business intent. 20 models compared on the same questions, reporting accuracy, dollar cost, and millisecond latency together. Submit an endpoint to get evaluated, no invite needed @namanjn1998 (DECIDE benchmark author)
- Dirac releases WorldMatter v0: a 3D Gaussian world model that simulates interactions from a few frames — The same rollout can be rendered from new viewpoints while keeping 3D geometry and physical plausibility, and it generalizes to complex multi-collision scenes @divyans1461 (Dirac, world model research team)
⚙️ Engineering Practice
- Karpathy builds TristanTube in three minutes with Claude Code: 47K hardcore videos plus a TV mode — It started when Tristan sent a list of YouTube channel recommendations, and he just turned it into a site. An earlier similar project, HackerTube, ran Opus 5.5 for about 2 hours: sifting 6,454 repeatedly-recommended videos from 20 years and 43M HN comments. The manual parts were only getting a Google Cloud API key, spinning up a VPS, and buying a domain — Claude handled deployment and hardening @karpathy
- Pure-code rendered animation: 6 parallel agent teams, 36 shots, 1,968 frames, zero image generation, zero video models — Each agent team had an AI art director, and the creator used Fable 5.5 to make a mecha clip of "a giant robot catching a meteor." Workflow and code are in the comments @seeconvm (independent developer)
- Omar Khattab asks: who's building a "weight-space programming" version of Tinker — He thinks the infrastructure needs of WSP could make for great economics @lateinteraction (MIT professor / ColBERT author)
⭐ Featured Content
Thinking Machines Lab releases Inkling 2 Small: 594B total / 17B active MoE, closed safety testing first, Apache 2.0 open weights next week | Murati's team's first open-weight model + rumored $500M new funding round
Supports image input and variable reasoning effort. Scores 31 on the Artificial Analysis Intelligence Index, which the team attributes to more efficient pretraining plus one large unified RL run. The strategy is notable: no direct open sourcing — instead a closed external safety test first, then Apache 2.0 weights next week, debuting on its own fine-tuning service Tinker. The hardest landing case: Bridgewater used Tinker to fine-tune Qwen3-235B for financial news and rate-signal screening, hitting 84.66% accuracy across six tasks vs GPT-5.5's 78.2%, roughly 30% lower error rate, and about 1/14 the inference cost — the most concrete set of numbers yet for "enterprise takes an open base model + managed fine-tuning to win vertical tasks." Meanwhile, rumors say the company is close to closing a $500M round led by Accel.
Sources: en.wowtale.net
Full scan of 10,659 public MCP configs: 67% of servers are unpinned, 941 still point at the archived official reference server | The first large-scale empirical look at MCP supply chain risk
The author scraped 10,659 public MCP config files on GitHub (32,223 servers total) and cross-checked npm/PyPI versions against the OSV vulnerability database. Beyond "two-thirds unpinned," several numbers stand out: 941 files still point at the official reference server archived in May 2025 (580 for server-github alone, no more security fixes), 638 point at nonexistent packages (open to typosquat poisoning), 152 files contain what look like real API tokens, 15 run on package names npm replaced with malicious placeholders, and only 48% of files need no remediation. The post gives a one-line fix (hardcode `@playwright/mcp@0.0.84` in args) — you can check your own config against it today.
Sources: dev.to
Anthropic reportedly eyes mid-November Nasdaq listing, valuation expectations cut from $2T to $1–1.5T | A real-time gauge of AI capital narrative shifting from "compute/power constraints" to "political and public-opinion constraints"
CNBC reports investors now expect a valuation well below the previously rumored $2T. The backdrop: 40 of 41 data-center-related ballot measures were rejected by voters in the midterms, NBC exit polls show over half of voters named AI data centers as a key voting factor, and the Nasdaq posted its biggest single-day drop in 19 months. This complements Vox's analysis: data center ballot measures grabbed headlines (40 passed, 56% of voters saw AI as a key voting factor), but the bigger story is a wave of AI-safety-regulation candidates winning, Democrats taking Congress, and key figures like Abdul El-Sayed, Jon Ossoff, and Scott Wiener rising — with Senator Maria Cantwell already proposing a bipartisan-acceptable regulatory framework. For practitioners, this is an early signal for judging US AI compliance costs and data center expansion pace over the next 1–2 years.
"The bubble wrote its own trigger": Alphabet's quarterly capex hits 121% of operating cash flow, AI issuers take 41% of US high-yield issuance, Philadelphia Semiconductor Index drops 12.2% in a day | Stitching political risk and financial fragility into one causal chain
The author argues Q3 earnings already exposed structural fragility: Alphabet's quarterly capex of $51.4B equals 121% of operating cash flow, free cash flow is -$9.1B and negative for two straight quarters; Meta raised 2026 capex to $135–145B; the six big hyperscalers are all expected to post negative free operating cash flow in 2026/2027, with the gap filled by debt (about $500B in AI-related issuance, 41% of this year's US non-refinancing high-yield issuance). The November midterms merely "priced" a structure with no buffer left: Texas's governor-elect pledged to pause data center approvals (Texas holds 32% of US under-construction capacity, with ERCOT's queue over 400GW), the Philadelphia Semiconductor Index fell 12.2% in a day, CoreWeave's 5-year CDS broke its prior high, and its 2032 bond yield topped 16%. Read alongside SemiAnalysis's compute ledger — demand keeps ramping on one side, financing starts tightening on the other.
Sources: dataaicontinuum.substack.com
Apple HALO humanoid autonomous control system + a set of funding-structure data: of China's 43.8B RMB embodied AI funding in H1, "brain" companies took 50.8%, full-robot makers just 12.8% | Capital-side evidence that "the body's barrier is falling, the brain sets the ceiling"
Apple's research team published HALO on arXiv on November 4, validating on a Unitree G1 that motions learned in simulation transfer directly to unseen real rooms — the selling point is the "brain" closing the sim2real gap, not running and jumping. The real increment is a funding breakdown: China's embodied AI raised about 43.8B RMB in H1 2026, with "embodied brain" companies taking 22.25B (50.8%) and full-robot companies just 5.6B (12.8%) — a brain-to-body funding ratio near 4:1; overseas, Physical Intelligence is valued at $20B. Also included: AgiBot GO-3 (phone packaging success rate 50.8%→98.3%), XPeng IRON Gen 2 VLA on-device deployment (0.2s re-inference, 2250 TOPS), and Tuoyuan E0.5 4D world action model. Technical details need the original, but the "where the money flows" table is quotable.
Sources: finance.biggo.com
Xiaomi humanoid robot runs a full 8-hour shift at an EV factory: from bin-picking car parts to shelving, company says fully autonomous with self-correction | Automakers' own production lines become the most real deployment scenario for humanoids
The lead story in Superhuman.ai's robotics special, with video showing a complete shift. Other quick hits worth scanning: an MIT system can flag uncertain object descriptions in robot memory and cross-check images from multiple angles to correct them, resolving about 92% of ambiguous objects in tests; Waymo confirmed it will drop Uber exclusivity in Austin/Atlanta starting January 2028 and use its own app; AGIBOT ran six humanoids on 24-hour livestream duty at a convenience store; ETH Zurich's FlexiQuad flexible-frame drone can still fly after hitting a wall. Thin on technical depth (VLA model, teleoperation ratio, success rate all undisclosed), but "automakers use their own lines as the first stop" keeps recurring — worth logging as a trend signal.
Sources: superhuman.ai
NVIDIA FOX reference architecture: manager agent + expert agents orchestrate factory operations, with "which decisions need human sign-off" as a configurable item | A harness and permission-design sample for industrial agent deployment
The Factory Operations Blueprint orchestrates manager agents dispatching expert agents — vision inspection calls Cosmos, root-cause analysis calls Kumo Anomaly, fleet management uses Isaac Sim/ROS 2, with NemoClaw as the default harness, OpenShell for execution isolation, and Nemotron models running underneath. The most reusable design: making "which decisions need human sign-off" a factory-configurable item rather than a product variant. Siemens' Erlangen site says its order-planning agent cut costs about 20%. It's a secondhand collage lacking first-hand technical detail, but the multi-agent orchestration + human sign-off boundary skeleton is directly borrowable for anyone building enterprise agents.
Sources: marketscale.com
Two checklists for MCP in production: per-tool minimal OAuth scopes for credentials, never change tool schemas in place (use `_v2` named versioning) | The spec is well written; the ops surface is all potholes
The author splits MCP into "protocol spec" and "production runtime": tools that work in dev silently fail against production APIs due to auth boundaries; client/server version mismatches surface as cryptic JSON-RPC errors; without timeouts, one slow tool blocks the whole agent loop; and the lack of a standard tracing format means debugging a 15-tool workflow is pure guesswork. Two reusable conventions: pull short-lived tokens from a secrets manager at startup and apply per-tool minimal OAuth scopes to avoid leaks via .env/logs/images; never modify tool schemas in place — use `search_documents_v2`-style named versioning, because LLM consumers have no version header. A companion piece covers the other half from the SDK angle: the v2 TypeScript SDK already wraps transport/discovery/validation, and what really decides success is tool names, descriptions, parameter `.describe()`, and error messages — because the reader is a model parsing JSON Schema at runtime, not a developer reading docs. It also warns that misusing `console.log` in stdio mode breaks the JSON-RPC stream.