· 12 min read
China's export gate for model weights, Claude's quantum circuits, Gemini beats Minecraft
China drafts export licences for models above 10²⁶ FLOP, Claude designs leaner quantum circuits, and four Gemini agents finish Minecraft in 27 hours.
Today in 30 seconds
- China's draft export rules would require licences for model weights above 10²⁶ FLOP, 2T parameters or a global top-10 ranking, with a fast lane for open source.
- Claude Mythos 5 found quantum circuits with fewer qubits and gates than the best published designs in all eight cases Anthropic tested.
- Four Gemini 4 Argon agents beat Minecraft's Ender Dragon from scratch in 27 hours, writing 2,783 programs with no human help.
- Microsoft finished pretraining MAI-Base-2, a 2.8T-parameter MoE with 80B active, trained without distillation from other models.
- Vercel's AI Gateway data shows open-weight models now handle 71.5% of tokens and about a quarter of spend.
Free. Daily, 10-min read. Confirm by email, unsubscribe in one click. Privacy
Saturday tea, and the big story is the geopolitics of weights: China’s draft export controls set the first formal thresholds for large models (10²⁶ FLOP, 2T parameters, global top-10), while promising a fast lane for open releases. Alongside it, science done by agents: Claude Mythos 5 designs better quantum circuits and four Gemini agents finish Minecraft without human help. Microsoft shows its 2.8T-parameter base model, and Vercel counts how much traffic really runs through open weights.
Top news
Will China gate exports of models above 10²⁶ FLOP?
China’s commerce ministry (MOFCOM) and science ministry proposed adding 40 technologies to the prohibited and restricted export catalogue, 23 of them in AI (announcement No. 48/2026, 16 October). The core of large models (weights, checkpoints, code, architecture, training and inference methods) would need a licence if the model meets any one condition: ≥10²⁶ FLOP of training, ≥2T parameters, or a global top-10 ranking at the time of export. The split is 14 technologies to be banned and 26 restricted; AI generation and mutation of malicious code is on the banned list, while AI-assisted vulnerability discovery is restricted. Already-published technology (open source, papers) is out of scope, and MOFCOM promises automatic licences for new open-source releases. APIs, apps, SaaS and lightweight on-device deployments are not counted as technology exports. Temporary registration before export applies immediately; consultation runs until 15 November.
Why it matters: It is China’s first formal “frontier” threshold for weights: if Kimi, Qwen or DeepSeek cross 2T parameters or reach the top 10, every weight release would need a filing, even on the fast lane. Nathan Lambert notes that the document explicitly promises easier treatment for open source, so much depends on how “almost automatic” approval works in practice.
Sources: MOFCOM announcement 48/2026 · Xinhua · AP · The Decoder · @natolambert · @ZilanQian
Can Claude design better quantum circuits than the papers?
In Anthropic’s STEM Fellows programme, Claude Mythos 5 got descriptions of eight quantum primitives (chemistry, optimization, simulation, number theory) and an automatic evaluator, then worked in a closed design–evaluate–revise loop. In all eight cases it found circuits with fewer qubits and gates than the best published designs, sometimes by an order of magnitude. For elliptic-curve point addition (the core of a quantum attack on ECC), after 53 hours it invented the first reversible divsteps variant without storing branch history, proved new algebraic identities and formalized the result in Lean; one variant uses 12% more qubits but 17× fewer gates. For a Reed–Solomon decoder used in DQI, it got a 19-qubit buffer instead of 665 and 12× fewer gates than Google’s construction. Anthropic notes the model was naturally conservative and helped by forcing several parallel strategies; the work was led by Andi Gu with Susanne Yelin and Mikhail Lukin (Harvard, QuEra).
Why it matters: Evaluable code plus a hard cost metric is ideal ground for agents. Our inference, not Anthropic’s claim: cheaper circuits for elliptic-curve discrete logs also shorten the road to breaking elliptic-curve crypto, which argues for faster post-quantum migration.
Sources: Anthropic research · @AnthropicAI
Four Gemini agents beat Minecraft from scratch in 27 hours
A DeepMind team (Faldor, Singh, Raileanu) dropped four Gemini 4 Argon agents into a random Easy survival world with one goal: defeat the Ender Dragon. The agents acted only through code (the Mineflayer library) and invented their own roles, strategy and programs. The run took 27 h 14 min, 2,783 JavaScript programs, 229 chat messages, 605 deaths, 6 server restarts and about 81 in-game days, with no pre-checked seeds or human-written game logic. The breakthrough was realizing that the crystals heal the dragon: the agents destroyed all ten (about 4 h 40 min) before winning. The authors stress this is one documented run, not a reliability measure.
Why it matters: Holding a coherent plan across thousands of dependent steps for more than a day, with division of labour and shared memory, is the same problem as long engineering tasks. It is a clear, readable long-horizon benchmark.
Sources: arXiv 2610.22319 · HF paper · @GoogleDeepMind
Microsoft MAI-Base-2: 2.8T parameters, 80B active
Microsoft AI finished pretraining MAI-Base-2, a MoE with 2.8T total and 80B active parameters. The company says that on its held-out evaluations it beats every open-weight model it tested and is 4–12× more FLOP-efficient than DeepSeek-V4-Pro-Base (company-reported). It trained on 16,000 GB200s for 38 days in August and September, at 21% average MFU and 89% goodput. Evaluation used bits-per-byte on held-out code (MAI’s own repo) and commissioned STEM, math and Q&A sets, compared against DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T and MiMo-V2.6-Pro (some only available as post-trained checkpoints). Microsoft says the data is “organic” with no distillation from other models, plans to license the weights to trusted partners through Foundry, and post-training is under way.
Why it matters: Microsoft is building its own frontier model independent of OpenAI, and claimed data cleanliness (no distillation) is becoming an enterprise selling point.
Sources: Microsoft AI, Frontier Pre-training
How far apart are Anthropic staff and Americans on AI?
In an August survey, Anthropic compared a representative sample of US adults (N = 10,980) with its own employees. Both groups judge today’s AI capabilities similarly but diverge on the future and on their own jobs. A Nobel-level discovery by 2030 is expected by 39% of US adults and 91% of Anthropic staff. Tasks taking a quarter of the time or less with AI: 26% of Americans who use AI at work versus 59% of Anthropic staff. By 2030, Anthropic employees expect AI to take over on average 75% of their hours, US workers 28%. People who expect AI to do at least 80% of their work forecast 12.2% unemployment in 2030, and each 10 points of exposure maps to about +0.4 points of unemployment in both groups.
Why it matters: The dispute over AI’s economic impact is mostly about future capabilities and adoption speed, not about today’s models, which makes this a fair starting point for policy talks.
Sources: Anthropic, The AI expectations gap
ChatGPT Pro gets parallel agents in GPT-6 Astra Pro
OpenAI updated GPT-6 Pro and GPT-6 Astra Pro in ChatGPT with faster answers, better use of plugins and tools and, above all, several agents working in parallel on parts of one task, with per-agent progress and finished subtasks open for review. It is available to Pro users in the Chat tab. Separately, Every’s “How Codex Happened” reports that after Codex Cloud’s rough start in May 2025 the team built a dedicated coding model line and desktop app; Codex had more than 5M weekly active users in June 2026, and Codex together with ChatGPT for Work now has 40M weekly active users. @OpenAIDevs also showed internal use, with “dot” taking a 3 a.m. on-call incident driven from a phone.
Why it matters: Multi-agent orchestration is reaching the main chat, not only Codex, which is a real change in everyday Pro use.
Sources: @OpenAI, Astra Pro · @OpenAI, GPT-6 Pro · Every, How Codex Happened · @OpenAIDevs
Perplexity open-sources SPACE, the sandbox behind Computer
Perplexity released SPACE (Sandboxed Platform for Agentic Code Execution) under Apache 2.0; it has powered all of Perplexity Computer since July. It is built on Firecracker microVMs controlled over HTTP and SSE and written in Rust, with host-side egress control, credentials injected outside the sandbox, and snapshot and restore. Perplexity says it runs millions of sandboxes a day, with creation 3–5× faster than its previous provider (median down from 185 to 60 ms). According to Semafor, Perplexity asked nine frontier models to escape the sandbox and none succeeded in 118 attempts (company-reported). Partners include Akamai (live), Crusoe and CoreWeave.
Why it matters: Agent isolation stops being a trade secret: you get a production-tested pattern (microVMs, egress control, secrets outside the VM) for your own stack.
Sources: Perplexity blog · GitHub perplexityai/SPACE · Perplexity Research · Semafor
Open weights now carry 71.5% of Vercel AI Gateway tokens
Vercel’s October Production Index (September data) shows open-weight models processing more than 7 in 10 tokens, and their share of spend has grown more than fourfold since June to about 1 in 4 dollars. Gateway spend is about 13× December’s level while the average price per token fell 47%, and nearly a third of teams switched their most-used lab in September, an index record. DeepSeek has 58% of tokens and 6.6% of spend; Kimi K3 drives most of the growth in open-weight spend. Nearly 70% of spend on Chinese models now goes through US providers (under 10% in December), and Anthropic’s share of spend fell below half for the first time, from 63.7% to 43.9%.
Why it matters: The model is becoming an interchangeable part, with companies mixing cheap open weights and expensive frontier models, which is useful context for China’s export draft above.
Are neural networks already clean?
Goodfire’s Nick Cammarata argues in a long essay (about 65 minutes) that canonical evidence for polysemanticity and superposition was misread, and that what looked like arbitrary concept-packing is mostly a pattern he calls recursive generalization. In InceptionV1, late layers “overflowed” for architectural reasons that weak tools amplified, and newer architectures accidentally made deep layers more readable. His thesis is that most neurons can be understood even in trillion-parameter models, so networks can be studied directly instead of first being untangled with sparse autoencoders. A same-week HF paper on the limited compositionality of SAE features (2610.21261) complements the essay.
Why it matters: If the thesis holds, mechanistic interpretability gets cheaper and more scalable, which matters for safety audits; for now it is a hypothesis, not consensus.
NVIDIA Nemotron 3.5 Super VL and Bolt for local agents
NVIDIA released Nemotron 3.5 Super VL, an open multimodal MoE with 120B parameters (12B active) that combines reasoning and tool use with understanding of text, images, documents and video, plus two Bolt models for local agents. NVIDIA claims up to 5× the throughput on a single GPU of open models in its class and more than 600 tokens/s on Baseten; TLDR reports an Artificial Analysis Intelligence Index of 26.5. Bolt 74B-A9B at 2-bit fits in 24 GB and is already on Hugging Face with support in Ollama, LM Studio, llama.cpp and vLLM; Bolt E4B (4 GB) is due on 23 October.
Why it matters: The mid-tier of open VL models gets a sensible option for document agents, and 24 GB means an ordinary RTX card at home.
Sources: NVIDIA Technical Blog · TechRepublic, Bolt
GitHub repos
- View repo →
xerj-org/xerj
Local search engine for agents that indexes a folder itself (BM25 + kNN, native MCP, agent memory); authors report 2.7× fewer output tokens at the same 16/16 score (authors’ own study). Fastest riser on the daily list.
- View repo →
lightseekorg/tokenspeed
Inference engine for agentic workloads with a C++ state-machine control layer, Python execution layer and one of the fastest MLA implementations on Blackwell; MLA 2 for Vera Rubin added in October.
- View repo →
ShawnPana/phone-harness
Lets Claude Code, Codex or another agent drive a real phone: iPhone via iPhone Mirroring on a Mac, Android via adb, with no jailbreak and nothing installed on the phone.
Research papers
- Read paper →
Waypoint-OPD
Transfers RL-induced policy changes to a student via on-policy distillation, gated by agreement between the teacher’s early and late changes; most upvoted paper of the day.
- Read paper →
OpenGCUA
A fully open computer-use agent stack combining the visual interface, programmatic interface and code execution.
- Read paper →
FDE-Gym
Agents in real customer deployments with vague requirements and unspoken constraints show a big gap versus well-specified tasks.
- Read paper →
MAI-Gym
One open platform to evaluate and RL-train GUI agents across phone, desktop and browser.
- Read paper →
Mosaic
RL for agents whose work spans many sessions, with hierarchical session-level credit assignment.
- Read paper →
3DCodeVerse
3D worlds as code, with data and suites for automatic verification, ready to edit and check.
AI agent use cases
- Split one research task across parallel agents: if you have ChatGPT Pro, give GPT-6 Astra Pro in the Chat tab one multi-part research task and watch how it divides the work between agents.
- Run a document agent on a home GPU: pull Bolt 74B-A9B at 2-bit (fits in 24 GB) in Ollama, LM Studio, llama.cpp or vLLM and test it on your own documents.
- Let a coding agent use your phone: set up phone-harness so Claude Code or Codex can drive a real iPhone (via iPhone Mirroring) or Android (via adb) for app testing.
- Sandbox your own agents properly: borrow SPACE’s pattern of microVMs, host-side egress control and credentials kept outside the sandbox for agents that run code.
AI workflows
1. Fill empty weeks: research + outreach Bots and RFP-to-estimate in 30 seconds
- Replaces: Hunting event planners by hand across hotel sites and LinkedIn (email column often empty), writing 2–3 quotes a week that each took a producer 1–2 hours from an RFP, and leaving quiet weeks on the calendar unfilled.
- What to connect: Grok Bot — Clarity (research every ~2 hours) and Synergy (outreach drafts into the human’s email drafts folder for approval before send); Lucy (chief of staff) for RFP → estimate into a custom inventory/quoting app (AboveHub) the Bot also built; optional venue landing pages for search. Human still approves every outbound email.
- Who it’s for: Service and trades businesses with equipment or crew utilization (AV, installers, event production, contractors) where empty calendar weeks are lost revenue and quotes are the bottleneck.
2. Drop Xero: Claude Code builds the invoice app, Grok Bot runs it
- Replaces: Paying for and living inside Xero for day-to-day invoicing and invoice management.
- What to connect: Claude Code to ship a custom invoice tool; Grok Bot to operate and manage that tool afterward (screenshot shows the Bot-managed invoice UI). One less SaaS subscription, per replies.
- Who it’s for: Solo operators and small firms that want invoicing owned in-house and steered by an agent rather than a full accounting suite for that slice of work.
3. Rewrite product docs from real support tickets overnight
- Replaces: Hand-writing docs from tribal knowledge and sparse pages that don’t match how users actually get stuck.
- What to connect: Astra (Aside) to mine past support tickets, user chats and use-case notes with subagents; Claude Code to map 200+ user journeys; cloud VMs with a browser driver to walk each journey on the live docs in parallel; then compile a new docs version and repeat overnight.
- Who it’s for: Product and infra teams shipping developer tools who already have a support inbox and need docs that match real friction, not imagined ones.
Quick hits
- Ling-3.1-flash (inclusionAI) — 560B-parameter hybrid MoE with 25B active, linear attention, 1M context and MIT licence; day-0 SGLang support and on OpenRouter, with strong company-reported coding-agent scores. HF
- Mistral Timestral — Mistral’s first time-series forecasting family (5M–209M params), zero-shot in 21 quantiles with 32,768-step context on CPU; Mistral says the 209M model ranks #1 on GIFT-Eval, TIME and fev-bench. Mistral
- Liquid d1-ultra — multimodal decision model returning calibrated probabilities in one pass at $0.042 per million input tokens, e.g. for PCB defect detection; Liquid says it beats Jev 1.14 on 112 evals. Liquid AI
- Krea 3 — production photo-editing model at under 8 seconds and $0.06 per image, up to 8 reference images and native 2K; Krea claims 4× the speed of Nano Banana Pro. Krea
- Taiwan to criminalize unauthorized AI chip exports to China — per X News, part of a trade deal with the US announced by economy minister Kung Ming-hsin. X News
Caveats: China’s rules are a draft under consultation until 15 November; MAI-Base-2 efficiency, SPACE escape tests, Nemotron throughput and most quick-hit benchmarks are company-reported; the Minecraft result is a single documented run; Goodfire’s essay is a working hypothesis.