Jev sparks a race for fast structured decisions
Summary
Jev is the day’s fault line: advocates see low-latency structured decisions as a new composable building block, while rivals and an early game test raise questions about originality, openness, calibration, and how much intelligence comes from upstream state preparation. More broadly, the stories point to an AI stack in which model capability matters only alongside the right representation, domain context, reliable interaction layer, and operational infrastructure.
Laya — 33ms Multilingual System 1 Decision Engine
Laya presents an open-source, non-autoregressive decision-model family that returns typed probabilities and numeric outputs rather than generated text. It claims 32.8 ms single-GPU latency, 7.2 ms per question in batch, and positions that as six to eight times faster than TypeSafe’s Jev while avoiding malformed JSON and token-level hallucination. The release packages three specialized checkpoints and a Unicode-aware router for more than 100 languages; its 51-language results argue that an English model can remain highly confident even when accuracy collapses on non-Latin scripts. The practical implication is that language routing must happen before inference, since confidence-based fallback cannot detect a model that cannot properly tokenize the input.
Read the source →I Built Non-Autoregressive Decision Models a Year Ago. Then a Frontier Lab Called It a "Breakthrough"
The author argues that TypeSafe AI’s Jev closely parallels two earlier open projects: a PPO-based sales-conversation predictor released in March 2025 and a later framework for reinforcement-learning-guided schema decisions. They contrast Jev’s reported roughly 150 ms latency and $0.042 per million input tokens with their new 421M-parameter RL Agent, which uses ModernBERT-large and a decision head to score schema options in parallel in 33–38 ms. The central technical case is that structured classification and calibrated confidence should not require a generative LLM, token streaming, or brittle parsing of free-form JSON. The post also makes an openness critique: it says Jev was announced without papers, weights, or training data, whereas the author intends the competing horizontal model to be fully open source.
Read the source →kicking the tires on jev (TypeSafe's System One model) with 2048
A quick 2048 experiment found that Jev performed about as well as random play when given only the board, rules, and move descriptions. Performance improved when external code first computed each legal move’s resulting board and score, but then it was only comparable to a simple fixed move-order heuristic. The test also tracked how often Jev’s top choice was an invalid no-op move, falling back to its next-ranked option when that occurred. The result suggests that, at least in this small trial, Jev benefits materially from a carefully prepared decision representation rather than deriving game-state consequences itself.
Read the source →datasette-auth-github 1.0
Datasette’s GitHub authentication plugin has reached version 1.0 after a fix for cookies that lacked a Max-Age attribute and therefore expired with the browser session. The issue was particularly noticeable in Mobile Safari, where sessions ended more often than expected. The plugin has existed for some time and is tested against both Datasette 0.65.x and 1.0 alpha releases, so the author frames the version bump as recognizing its stability rather than a major functional rewrite.
Read the source →California Sea Lion, Brandt's Cormorant
Simon Willison shares a wildlife sighting from Pillar Point Harbor in California featuring a California sea lion and Brandt’s cormorant. On reviewing the photograph, he also noticed Morris, a Northern Gannet, partly visible behind the base of a sign. The post is a brief observational note rather than an analysis or AI-related report.
Read the source →The Sequence Radar - Issue 936: Last Week in AI: Gemini Talks, Astra Practices Law, Figure Folds Laundry, and Crusoe Powers It All
This roundup argues that useful AI increasingly depends on the surrounding operating system: interaction timing, domain context, real-world transfer, and compute infrastructure. It highlights Gemini 3.8 Live’s combination of visual context, background tool calls, and concurrent reasoning and speech; OpenAI reports its legal-focused Astra scored 54% on a 200-question private validation set versus 38.7% with ordinary web search. Figure says pretraining on its human-behavior dataset raised Helix 2.5’s complete-task success in 30 unseen homes from 9% to 56%, while the remaining 44% failures underscore the gap to dependable domestic robotics. The infrastructure side is Crusoe’s initial $3.9 billion Series F closing at a $30.9 billion post-money valuation, reflecting the power and data-center buildout required to serve these systems.
Read the source →Joy & Curiosity #100
Thorsten Ball calls Jev a potentially important new software primitive: a fast, cheap API that maps unstructured state to typed probabilistic decisions, or a “smart if-statement.” He describes prototypes that use it to predict shell commands from history, suggest navigation targets in Neovim, and select an AI model for a prompt, emphasizing that roughly 200 ms latency makes formerly bespoke prediction features accessible through one call. The essay’s broader claim is that steep cost and performance improvements can create entirely new product categories, just as cheap computing enabled embedded networked devices. It connects that optimism to a changing software role: as coding, review, repair, and operations become cheaper, discovering needs, specifying products, and making them pleasant may become the durable work.
Read the source →