AI Daily · 2026-09-22

The biggest story is TypeSafe AI's Jev, a non-generative 'decision model' that returns structured classifications, scores, and probabilities; LangChain and LangSmith are already evaluating it as a faster, cheaper, and more consistent alternative to LLM-as-a-judge. OpenAI paired a claimed mathematical breakthrough—solving Navier–Stokes and over a hundred open problems—with new governance moves: an independent mathematician advisory group and a policy post warning against premature recursive self-improvement. Elsewhere, NVIDIA shipped Isaac ROS 5.0, Hugging Face Transformers added GGUF/llama.cpp quantization support, and Cloudflare Python Workers became generally available.

North America · First-hand

OpenAI

⭐⭐⭐ [Other] Advisory Group on Mathematics and Artificial Intelligence

OpenAI announced an Advisory Group on Mathematics and Artificial Intelligence, composed of independent mathematicians from institutions including the IAS, Cambridge, Berkeley, Stanford and Harvard. OpenAI says it began training a new internal model on August 28 that, besides resolving the Navier-Stokes Millennium Prize problem, has now resolved more than 100 long-standing open problems across most areas of mathematics, a pace of progress that surprised mathematicians within OpenAI. The group will act as a bridge to the mathematical community and the broader public, advising on the review and communication of emerging results and on academic and professional standards of mathematical research, as well as on how OpenAI's tools can support mathematical research and learning. The group operates independently of OpenAI, its members are unpaid by OpenAI, it may change its membership and make its advice public, and it will not advise OpenAI on how to pace its internal progress in mathematics.

⭐⭐ [Opinion] Building standards for the next phase of AI

OpenAI published a Global Affairs piece outlining its position on the next phase of AI. It restates the mission of ensuring AGI benefits all of humanity and cites three goals recently laid out by Sam Altman and Jakub Pachocki: build an automated AI researcher and iterate with it on alignment while keeping people in the self-improvement loop; deliver the benefits of scientific progress and economic growth; and empower everyone individually with a personal AGI. The post says AI is accelerating OpenAI's own research and engineering, and that AI-enabled research has led to advances in mathematics, including the Navier-Stokes Millennium Problem. It discusses recursive self-improvement (RSI), stating that fully autonomous RSI is not happening today and should not be pursued unless and until it can be done safely, and that proceeding without care could leave humans without practical control over AI development; it describes the previously disclosed Hugging Face Incident as a preview of such risks. The post also argues that international standards for safety and security practices in frontier AI development may be as important as alignment research itself, since standards can create shared definitions of high-quality evidence and agreed baselines for technical safeguards.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐ [Model Release] Jev introduces a new shape of LLM - System One, aka Decision Models

TypeSafe AI unveiled Jev, which it calls the first example of a new category it names 'System One models' (Simon Willison prefers the term 'decision models'). Jev still takes text input, but instead of text it returns floating point numbers for categories, yes/no questions and ratings, along with confidence scores. It charges only for input, with output free, at $0.042 per million tokens for the first model — cheaper than OpenAI's GPT-5 Nano at $0.05 per million. Usage involves sending a 'state' object (a string, array of strings or set of name-value pairs) plus one or more questions to the API, supporting three question types: Noul yes/no questions returning a 0-to-1 confidence (the CEO confirmed the name is short for Bernoulli), choice questions returning a confidence plus a probability distribution over the options, and score questions returning a float along a supplied numeric range with descriptions; questions are evaluated in parallel. It suits classification-style tasks such as spam detection, label suggestion, prioritization and ranking, and Simon has been experimenting with it for search reranking by scoring 100 BM25 candidates for relevance. He notes Jev represents a further move toward black-box machine learning with no justifications, making bias concerns central and evals and structured experiments more important than for regular LLM projects.

⭐⭐ [Model Release] What Is Jev? A Guide to TypeSafe AI’s System One Model

TypeSafe AI released Jev, a new model that is not a traditional LLM: it does not generate text, and is what the team calls a System One model that evaluates a state and returns typed answers and probabilities. It is trained with reinforcement learning for calibrated decisions (RLCD), and the company reports up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks. To invoke it you send a state (the context) and questions, supporting three question types: Choice (pick from options, returning a probability per option plus a confidence score), Score (rate against ordered levels, returning a continuous score, distribution, and confidence), and Noul (yes/no, returning the probability that a statement is true). Multiple questions in one request are evaluated in parallel, with added questions barely changing response time and costing only the tokens for the extra questions. LangChain integrates Jev through TypeSafeClassifier in the langchain-typesafe package, where you pass state and questions to .invoke() and get classification results instead of a chat response; state can be text, structured data, or LangChain messages.

Product Update

⭐⭐ [Product Update] NVIDIA Isaac ROS 5.0 Advances Agentic, Open Source Robotics Development

NVIDIA released Isaac ROS 5.0, a collection of GPU-accelerated packages built on ROS. ROS is an open framework from Open Robotics that helps humans build robots. NVIDIA says developers need new physical AI models and tools to build and deploy sophisticated robotics applications that can perceive, reason and act in dynamic environments. The release was announced the same day.

⭐⭐ [Product Update] Transformers now runs llama.cpp quants

Transformers has added support for running GGUF models, the quantization format developed by the llama.cpp team, so users can pick a GGUF checkpoint from the Hub, load it with from_pretrained, and generate locally with a file sized to their machine's memory. To bring performance close to llama.cpp, the team reuses its underlying ggml kernels through the kernels library and reduces overhead in generate; the initial focus is local inference on Apple Silicon, starting with the Qwen3.5 architecture. Using Unsloth's Qwen3.5-4B as an example, file sizes range from 8.42 GB at BF16 to 2.74 GB at Q4_K_M, with Q4_K_M suggested as a starting point. Requirements are an Apple Silicon Mac, a supported PyTorch version, the latest transformers (currently main), and a compatible kernels release; passing the Hub model_id and filename as gguf_file to from_pretrained automatically loads the matching ggml/Metal layer kernels and uses ggml-org/ggml-attn for attention, falling back to sdpa with a warning if that kernel cannot be fetched.

⭐⭐ [Product Update] Jev is now available in LangSmith Evals

Jev is now available as a judge for evaluations in LangSmith, offering teams a fast, low-cost way to evaluate open-ended agent behavior and turn results into structured feedback tracked in LangSmith. Jev is not a traditional LLM and does not generate text; the TypeSafe AI team calls it a System One model that evaluates a state — such as an agent trace or a single message — and returns typed answers and probabilities. It answers three question types: noul (a yes/no probability), choice (pick one option from a set), and score (rate the state on an ordered scale). According to TypeSafe AI, Jev is up to roughly 450x cheaper and 200x faster than comparable LLMs on classification tasks, and can evaluate multiple questions about the same state in parallel. The post frames Jev as a third type of agent eval, alongside code-based checks and LLM-as-a-judge, trading some speed for flexibility at a fraction of an LLM judge's cost. Users can try Jev-as-a-judge from the Evaluators tab in any tracing project.

⭐⭐ [Product Update] Cloudflare Python Workers are now generally available

After a two-year preview, Cloudflare's support for running Python code on its server-side Workers platform is now stable, with the company stating that Python is now a first-class, fully supported language on the Cloudflare Developer Platform. It works by compiling Python to WebAssembly via Pyodide and running it in Cloudflare's V8-based workerd runtime; the documented limitations include that both multiprocessing and threading are non-functional in the WebAssembly VM. For local development, the pywrangler tool (packaged on PyPI as workers-py) runs a full local simulation of the stack, executing code with Pyodide in WebAssembly in V8 inside a 123MB workerd binary. The release announcement is credited to Gyeongjae Choi, Dominik Picheta, and Hood Chatham, two of whom are Pyodide core maintainers.

Research

⭐⭐ [Research] Can Jev Be a Better Agent Evaluator?

LangChain published a blog post testing whether TypeSafe AI's Jev could serve as an agent evaluator. Jev is not a traditional LLM and does not generate text; it is what TypeSafe AI calls a 「System One」 model that evaluates structured state and returns typed answers with probabilities. The post notes that agent evals today fall into two categories: code-based evaluation, which is narrow in scope, and LLM-as-a-judge, which is slow, expensive and inherently non-deterministic. In the experiment, Jev was more consistent on continuous scoring, with quality-score variance 92–913x lower than GPT-5.6 Luna, Terra and Claude Sonnet 4.6, while also being the fastest and cheapest at 0.44s and $0.00035 per call ($0.34 total versus $28.17 for Claude). The authors call the results promising but early, based on a narrow test.