Key Takeaways
- OpenAI will make its frontier models and Codex agent workflows available in Google Cloud's Model Garden, while also disclosing a three-week embedded evaluation of its internal monitoring by METR and Redwood Research.
- Hugging Face released decision models elephant-1 (9B) and mammoth-1 (27B), plus Carbon-X, a 14B model combining DNA and natural language, and updated its Decision Index to 0.4.
- Aravind Srinivas detailed JOLT, a post-training method that matches GRPO on MATH with 6.5x fewer training tokens, and said JOLT+ outperforms GRPO on math, coding, terminal and tool-use benchmarks.
- Ethan Mollick argued AI coding gains may not translate into broad software improvement, while a cited thread catalogued severe vulnerabilities across PHP, GitLab, LiteLLM and other systems.
- Swyx described Claude Code as evolving into a reusable, general harness and shared evaluation guidance favoring human judgment, calibrated LLM review and production A/B testing.
- Tesla reported FSD Supervised attach rates above 75% in North America and 1.74 million active FSD subscriptions, alongside Optimus early production and Semi milestones.
- Hamel Husain and Shreya Shankar will release Evals for AI Engineers on November 12 at LeanProdMeetup, with free signed copies sponsored by raindrop_ai.
1. AI Models and Infrastructure
- OpenAI said its frontier models and Codex agent workflows will be integrated into Google Cloud's Model Garden, letting enterprises use OpenAI technology within their existing development environments 1. Separately, OpenAI said it invited METR and Redwood Research this summer for a three-week embedded evaluation pilot of its internal monitoring system, granting deep infrastructure access; OpenAI said the findings will strengthen safeguards as it scales the system and will influence future third-party evaluation collaborations 2.
- Hugging Face released two decision models from the baselabs team: elephant-1 (9B) and mammoth-1 (27B), free on Model API for one week until October 27, 2026. Mammoth-1 ranked first on the Jev Decision Index, elephant-1 tied for first on the Artificial Analysis Decision Index, and both support open-source deployment and post-training 1. Hugging Face also updated its Decision Index to 0.4, scoring only on private and internal benchmarks to discourage gaming, adding domain-private benchmarks and evaluating 224 models, with new leaders emerging at every scale 2.
- Hugging Face released Carbon-X, a 14B model it described as the first to combine DNA and natural language. It said Carbon-X beats all open language models on tested DNA benchmarks, trails GPT-6.1 Sol by under 1 point at more than 1,000x lower inference cost, was mid-trained on 300B tokens of interleaved DNA and text, uses a hybrid tokenizer producing about one-third the DNA tokens of BPE, and is Apache 2.0 and runnable locally on a laptop via llama.cpp 1 2 3. Hugging Face also merged the protein sequence design model RFD4-Proteina into Diffusers and launched the SimReady Central Library with 2,000+ simulation-ready 3D assets 4 5.
- Aravind Srinivas said he and NVIDIA AI proposed JOLT, a post-training method where one model acts as both teacher and student, with the teacher given extra information and reward training while penalized for drifting from student behavior; JOLT+ adds outcome-based rewards for the student 1. On MATH, JOLT matched GRPO accuracy with 6.5x fewer training tokens, and Srinivas said JOLT+ clearly outperformed GRPO on math, coding, terminal and tool-use benchmarks 2 3. He also said Computer's /graphify skill visualizes codebases as interactive maps and added a PHP2Shell skill for checking PHP exposure risks, both free through October 31 with up to 5,000 credits per user 4 5.
2. Developer Tools, Evaluation and Security
- Swyx argued Claude Code is becoming a malleable, general and reinforced harness, comparable to LLMs evolving from specialized to foundation models and back to specialized use, and said his team deeply evaluated Claude Mods and listed recommended items to try 1. Shared material outlined a subjective-task evaluation roadmap: use human evaluation as the quality north star, calibrated LLM judges or task-specific rubrics for offline evaluation, then production monitoring and A/B tests, with power analysis, confidence intervals, resampling and paired comparisons to control noise; it stressed that higher benchmark scores do not necessarily mean better capability 2.
- Ethan Mollick said AI coding tools may deliver large improvements, but research repeatedly shows the bottleneck for product improvement is often not coding, making it hard to identify which ordinary software has improved most from heavy AI use by developers 1. He described recent security incidents as signs of warnings from AI industry figures that attack and defense are accelerating at non-human speed and interacting with human systems that must respond 2. Cited material listed PHP remote-code-execution flaws with leaked details before patches affecting WordPress, MediaWiki, Magento, Joomla and Laravel, with Censys saying about 1.5 million servers run unfixed versions; a GitLab CVSS 10.0 flaw affecting self-hosted instances since 2014; and a hijacked LiteLLM 1.107.0 implanting a credential stealer and remote-access trojan, the second such case this year 3.
- Deedy said he personally saw non-technical people use AI models to exploit CVEs such as PHP2Shell and found it frightening, arguing that once the method is known it is extremely simple and puts global digital infrastructure at risk 1. Ben Tossell shared a description of supermemory's graph engine, which tracks knowledge changes, forgets specific information and connects information in a single data structure that can be presented as a file system, fact system or pages, and can automatically find and inject tokens without the model actively retrieving 2.
3. Industry, Products and Events
- Tesla's Q3 shareholder update, cited by Elon Musk, said North America new-vehicle FSD Supervised attach rate exceeded 75%, the next-gen Roadster was unveiled, Canada, Australia, Japan and Taiwan set delivery records, and Semi entered production with a fully loaded 82,000-pound range of 500 miles and 1.7 kWh/mile efficiency, about one-third the energy of comparable diesel trucks 1. Musk also cited results including non-GAAP EPS of $0.48 versus $0.45 expected, revenue of $28.05 billion versus $28.3 billion expected, 1.74 million active FSD subscriptions up 67% year over year, and free cash flow of $1.53 billion, saying people buy the car for FSD 2. He shared that Optimus began early production at the Fremont high-volume line, with most of a million-unit-per-year line installed but design, AI and reliability still needing about a year, and that 10,000 to 20,000 Optimus units are needed first for self-play training 3 4 5.
- Hamel Husain said he and Shreya Shankar will release Evals for AI Engineers on November 12 at Dan Olsen's LeanProdMeetup in Mountain View, in person or via Zoom, with sponsor raindrop_ai providing free signed books to attendees 1 2. He also joked about title changes from Data Analyst in 2003 to Data Scientist in 2010, Machine Learning Engineer in 2016 and now "Data Analyst of Traces," and said counting and data visualization have long been undervalued, making their return unsurprising 3 4.
- NVIDIA set up a booth at GTC Berlin, where Jensen Huang met European developers and partners to discuss AI deployment in Europe, and NVIDIA AI published a recap covering topics from open models to physical AI 1 2 3. Hugging Face and Beam are running a $20,000 build competition for models, datasets or Spaces on the Hub, with entries due by November 2 and winners choosing cash or credits 4. Anthropic published its annual public benefit report, saying it organized and reported actions around five pillars as a public benefit corporation 5.
