Anthropic Scales Agentic Code Merges 12x, Google's AMIE Posts 90% Diagnostic Agreement in Real Clinic

Key Takeaways

1. Anthropic says it cut distillation-attack effectiveness and rebuilt its source control after daily merges jumped from ~2,500 in May to over 30,000 in September.
2. Google and BIDMC's AMIE study in The Lancet reports 90% agreement between AMIE's differential diagnosis and physicians' final diagnosis in a real emergency department.
3. Musk amplified a unified X Premium + Grok + Cursor subscription and a wave of Grok Bot user reports, including memory transfer from Claude and ChatGPT.
4. Mistral Large 4 scored 13.3% on the sudo L7 benchmark, ahead of Muse Spark 1.3 and Grok 4.7, while Mistral also entered an AI satellite constellation project.
5. Hugging Face shipped a /v1/systemone endpoint in transformers serve and updated 69 oQe quantized models with lower KL divergence.
6. Perplexity detailed an agent-oriented Git forge for Projects, reporting sub-2% commit conflict rates and 63% vs 37% accuracy over semantic search on financial document analysis.
7. AI Engineer New York ran October 12–14 with speakers from OpenAI, Anthropic, Google DeepMind, Cursor/xAI and major financial firms, plus a sold-out Compute Trading track.

1. AI Models and Infrastructure

1. Anthropic said it made significant progress in recent months reducing the effectiveness of distillation attacks against Claude and published an article on its methods 1. Its engineering blog described building its own source control system: daily merged changes averaged about 2,500 in May and more than 30,000 in September, making developer infrastructure designed for human-scale rather than agent-scale the bottleneck 2. The company said it spent five months building the new version control system and completed a full monorepo migration from git 3.
2. Google and BIDMC's AMIE study was published in the main journal of The Lancet and described as the first prospective study of a conversational diagnostic system used with patients in a real clinic, with AMIE intended for patient dialogue before a visit 1 2. In real emergency patients, clinicians reported AMIE summaries helped them prepare in 75% of cases and influenced care in more than half; AMIE's differential diagnosis matched physicians' final diagnosis 90% of the time 3 4. The account called it Google's first paper in the main Lancet journal and a milestone in moving from lab to real clinical utility 5 6.
3. Mistral AI said Mistral Large 4 joined the sudo L7 benchmark with a score of 13.3%, above Muse Spark 1.3 at 12.2% and Grok 4.7 at 10.6%; the benchmark targets senior software engineering, including navigating messy systems, spotting hidden risks and judging technically correct code that should not ship 1. Mistral also said its models entered orbit in Orbitworks Space and Loft Orbital's first AI satellite constellation, working with NVIDIA AI Infrastructure 2. A cited post said Orbitworks launched the first of 50 satellites, Altair, aiming to deliver AI-generated alerts within minutes, with NVIDIA supporting models including Ministral 3 3.
4. Hugging Face added a /v1/systemone endpoint to transformers serve that turns any LLM or VLM into a decision model by scoring a set of decisions with model logits without generating text; a repost said this covers 250 LLM and VLM models and that decision model standardization is coming to transformers 1 2. The related Decision Model was indexed by Papers with Code as a method 3. A new Typed Decision model based on EmbeddingGemma 2 takes a state plus an arbitrary question schema and outputs typed, calibrated answers (choice, score, noul) without text generation or output parsing, scoring 76.4% on typed-decisions with about 1/100 the parameters of a 27B LLM and trailing by roughly 6 points 4. Separately, all oQe quantized models moved to an improved oQe Quantization requiring oMLX 0.7.1+, cutting KL divergence by 11.6–18.1% on code and 10.2–13.8% on prose, across 69 models; users with older oQe downloads must re-download 5.

2. Agents, Products and Developer Tooling

1. Elon Musk amplified a unified subscription plan combining X Premium, Grok and Cursor, with a repost saying the legal terms were published and describing "X the everything app" as happening 1 2. He also amplified a large set of Grok Bot user reports: users said it could replace search, Hermes automation and Cursor cloud agents, review PRs and merge code, connect to VMs via tailscale and SSH, launch Claude Code or Codex agents in a terminal, and create other bots on demand 3 4 5 6 7 8 9 10 11 12. Reported uses included planning and booking a Lisbon trip with a bank card, handling about 20 emails plus calendar, proposals, lunch and hotel and airport taxi in a 30-minute "call," fixing a customer bug and shipping autonomously, and running 62 agents and 29 scheduled routines 13 14 15 16 17 18 19 20 21 22. Musk also amplified Grok Bot memory transfer, which migrates what Claude or ChatGPT knows about a user by pasting a prompt into the old chat or importing a full data export; every item must be reviewed jointly, only kept items are saved, entries are labeled with source and time, other AI guesses do not become facts, and saved content is not overwritten without user consent 23.
2. Perplexity's Aravind Srinivas said a single-person company spends over $1 million per year on Perplexity, with the founder earning $3 million annually by automating newsletters for others 1. He described Perplexity Projects' file system as an agent-oriented Git forge: each agent works in an isolated sandbox and commits to a shared repository so humans and other agents can build on its work across sessions 2. Each Project is a Git repo, agents get short-lived sandboxes and commit to a shared repository on S3, commit conflict rates are below 2%, and agents using plain ls and grep beat semantic search 63% to 37% on financial document analysis accuracy 3. He also said Computer classified urban land use from aerial photos in Phoenix, analyzing 203,883 USDA public imagery tiles with Perplexity's Decider API and reaching 98.8% accuracy against OpenStreetMap labels at an API cost of $9.63 4.
3. Hamel Husain amplified a discussion with Ben Hylak on testing agents with Raindrop Simulations and the testing question "what will my change alter" 1. A new blog post described Quail classifying documents with an LLM at about 1.8x baseline speed, noting most systems let the inference engine generate free-text labels and then parse them, while LLMs generate token by token and each decode step is memory-bandwidth limited, leaving GPUs idle much of the time 2. Amp can now receive GitHub reviews, comments and CI results in the thread where code is written, and creating a PR from an Amp thread auto-subscribes without leaving the app 3. A separate repost noted OCR, especially VLM methods, usually extracts one page at a time and fails when cross-page context is needed, which is common 4.
4. Ethan Mollick said his early limited testing of Meta Muse Wonder suggested a decent model, but the more notable signal was its speed of progress since April on the "flooded neo-gothic city" twigl shader test, showing Meta advancing quickly in a short period; a cited post called Meta's Muse Spark Thinking decent but below the current top three models, with odd language and tone and loose factual grounding 1. He argued OpenAI and Anthropic may be best positioned to optimize AI cost-to-intelligence by combining agents of different capability tiers, provided those agents are specially trained and collaborate well in their own harnesses 2. He cautioned against over-weighting the Artificial Analysis Index and its individual scores 3, and described his own AI agent proactively testing and reproducing a subscriber bug without being asked 4. He contrasted GPT-4's year as the undisputed best model with the present, when 19 AI models runnable on a modern smartphone outperform GPT-4 5.

3. Events, Ecosystem and Applications

1. AI Engineer New York ran October 12–14 in New York, presented by Arize, with @businessbarista hosting for a second year; speakers covered Exa, Ramp, ValsAI, Rogo, Modal, Runway, Mintlify, SF Compute, Nous Research, Google DeepMind, Cerebras, MiniMax, Mastercard, Citi, GEICO, Linear, Coinbase, BlackRock, Wise, Point72, Monzo, Capital One, Revolut, JPMorgan, Wells Fargo, Cursor/xAI, Fidelity, Bloomberg, OpenAI and Anthropic 1 2. Anthropic's @cjav_dev and @HarrisonStall spoke on how agents can handle larger projects without overburdening developers, drawing on Unix and Linux lessons 3, while Bridgewater's Aaron Linsky and Sam Green presented Vault, a structured and unstructured data store for "investor agents" 4. The event included a sold-out "Compute Trading Microtrack" with SFCompute CEO @evanjconrad, Oliver Kennon, Jump's new compute trading lead @affaan, @absurdistphil and Compute Exchange + Silicon Data CEO @carmenli, available only via livestream; the author, who said he previously traded FICC, called it AIE's first agenda around a "new asset class" 5. A main-stage talk, "The Dark Pools of Compute," was set for October 14 to discuss how compute is actually bought and sold, how power, financing and policy set prices, and what a more mature market means for developers 6.
2. NVIDIA announced a global RTX Spark launch event for Windows PCs across Beijing, Berlin, Istanbul, Seoul, San Francisco and Taipei, with new laptop demos, giveaways and a chance to win a Jensen-signed ASUS ProArt P14 RTX Spark laptop; details and participation will be announced later 1. NVIDIA also said French developers, researchers and enterprises are using its technology and open models to deploy AI in real scenarios from satellites to skincare 2, and amplified that Barcelona participants are ready as the road to NVIDIA GTC Berlin formally begins 3. Jensen Huang visited the Paris Build-A-Claw event, where developers and startups got hands-on with open models to build their own AI agents 4.
3. Hugging Face amplified Open Together, an open-source AI community event on October 16 at The Midway in San Francisco, co-hosted with Hugging Face around building an open, safe and transparent AI future; another repost said not to miss it on Friday in San Francisco and teased "many surprises" 1 2. Other reposts covered turning an OpenAI math problem into an open RL environment on Hugging Face, with the author calling it early and rough, where the verifier only accepts the original formalization and equivalent proofs may score 0, but showing research releases becoming executable training infrastructure 3; GLM-Legal based on GLM-5.3-Flash, claimed to match frontier models on mainstream legal benchmarks for contract review, evidence analysis and compliance workflows 4; and a kernels cookbook covering using HF kernels from the Hub, hardcoded Triton kernels, and benchmarking models with and without kernels 5. Microduck updates included a 3-minute unedited carpet test video 6, a Microduck Arena where users can train ducks with RL, publish to Hugging Face and compete in a 2-meter sprint, hill climb, roller sprint and MicroDuck Got a Talent without a physical robot 7, and a repost asking whether Hugging Face people could hold real-world Microduck battles 8.
4. Starlink was amplified by Musk in several deployments: the Gyeongbuk provincial government in South Korea introduced Starlink into fire services to protect lives in remote mountains, with the fire headquarters equipping a command vehicle for a pilot to help teams share footage and coordinate response during wildfires, floods and typhoons 1. Separately, Starlink helped rescue five people at sea after the boat Magic Bus sank in a storm far off the North Carolina coast; the captain's radio call failed, so he used Starlink Mini to connect his phone and call the US Coast Guard, and all five left the sinking vessel on a life raft and were brought back safely by rescue helicopter 2. Musk also amplified that a Falcon 9 launched Dragon to the space station, with Dragon set to launch at 6:33 a.m. Eastern that day carrying over 6,300 pounds of science, supplies and roll-out solar arrays for the Expedition 75 crew 3 4, and replied "Of course" to Tesla offering free Supercharging in hurricane-affected areas 5.