Good morning. Today leans technical: a handful of research releases pushing at fundamentals — a model that skips text generation entirely, a 4B fine-tune beating Postgres’s query planner, and yet another compression trick for ternary LLMs. On the policy side, OpenAI is trying to formalize how it discloses model misbehavior, and Mozilla picked Mistral to power Firefox’s new AI features.
TypeSafe launches Jev, a model that doesn’t generate text. TypeSafe AI introduced what they call “System One Models,” starting with Jev — a model that returns typed, structured outputs with calibrated confidence scores instead of strings. The pitch is a frontier-quality function call at $0.042/MTok, roughly two orders of magnitude faster than comparable LLMs on classification, routing, and scoring tasks, trained with a new method they’re calling Reinforcement Learning for Calibrated Decisions. HN’s reception was cautiously interested but frustrated with the announcement itself; one commenter suggested a more honest title would be “trading general-purpose generation for fast typed inference,” and several noted the speed comparison isn’t quite apples-to-apples given the different task scope.
A 4B model beats Postgres’s query planner — under carefully chosen conditions. Rohan Bansal fine-tuned a 4B model with SFT and a custom GRPO variant to generate Postgres query plans, hitting a 1.81x geometric mean speedup on 113 join-heavy IMDb queries for about $1,200 in compute, bootstrapped from GPT-4o trajectories. HN’s response was pointed: the benchmark used an 8GB dataset fully in memory, pre-warmed queries, no realistic indexes, and read-only SELECTs. One commenter sketched the nightmare scenario — a production DB freezing because a variable rename triggered a query rebuild and the LLM planner picked something catastrophic. The distillation-from-OpenAI question also came up, given the ongoing scrutiny of that practice.
Squeezing ternary LLMs below 1.58 bits. Intel researchers introduced BITCOS, a storage format that exploits the fact that zeros make up as much as 51.5% of weights in real ternary models to get down to 1.485 bits per weight, with up to 1.27x GPU throughput. Commenters were mostly impressed, with a few noting the obvious next question — whether ternary is even the right target compared to vector quantization or trellis methods — and one predicting that if ternary hardware ever ships as ASICs, the power efficiency numbers will be striking.
DeepSeek-V4.1 Flash goes hard on KV cache. A writeup on DeepSeek-V4.1 Flash — which the author thinks should really be called V5 — details 4x KV cache compression at ~420 tokens/sec, using a YOCO-style prefill (only 20 of 40 layers active, dropping activated params to 8B), cross-layer compression, and FP4 KV cache precision. The motivation is straightforward: long-horizon agent workflows are making KV cache in HBM and SSD unsustainable. The linked blog was 404ing for some HN readers, but those who got through called out the prefix caching improvements making repeated queries “practically free.”
Xiaomi opens up the training room. Xiaomi’s Mimo team put up a live dashboard tracking the post-training of their 2.6 model, including running cost (currently around $1.2M) and real-time benchmark scores. Commenters using the existing MiMo-V2.5 for software engineering work were enthusiastic, and early 2.6 numbers on DeepSWE 1.1 look well above 2.5-Pro’s 19% — though contamination questions came up quickly. One reader wondered why other providers don’t do this; the obvious answer is that most don’t want their training curves scrutinized in real time.
OpenAI publishes a misalignment reporting framework. OpenAI announced a framework for publicly disclosing misalignment incidents, admitting it had reported too few in the past and sharing several new examples involving unreleased internal models — Wired has the details. Separately, TechCrunch looked at Amodei’s and Altman’s parallel push to embed third-party safety evaluators inside their companies with access to training processes and internal logs. Evaluators cited in the piece said the idea is welcome but pointless without legislative backing to keep them from becoming glorified vendors — which ties back to the FRONTIER Act provisions OpenAI endorsed earlier this week.
Mozilla picks Mistral for Firefox’s Smart Window. Mistral and Mozilla announced a partnership putting Mistral models behind Firefox’s new AI assistant for search, page summaries, and cross-tab memory, launching in France and North America. Both companies promise zero data retention. HN was unimpressed with the framing: one top comment asked why two open-source-adjacent companies are defaulting to cloud inference on private browsing history rather than shipping a small local model. Bring-your-own-model is supported as an opt-in, and Mistral also managed to misspell its own name in the announcement.
“Recursive self-improvement” that maybe isn’t. Dream-RSI proposes using accumulated discovery history as a replay simulator to cheaply refine exploration policies off-policy, applied to domains like algorithm engineering and GPU kernel optimization. HN wasn’t sold on the branding — multiple commenters pointed out this is really an optimization of exploration resource allocation, not perpetual self-improvement in the classical RSI sense. Others flagged the obvious concern about overfitting to already-discovered branches as the search space grows.
That’s the morning. The Mimo dashboard is worth bookmarking if you want something to check back on through the day — watching a frontier-ish model train in public is genuinely novel, contamination questions and all.