Six months to live for open weights

The policy fight over open models moved from op-eds to the West Wing, with reports of a White House executive order and Satya Nadella publicly needling the labs over distillation. On the technical side it was a day of squeezing more out of what exists: a Caltech spinout claims a full 27B model on an iPhone, and a widely-shared postmortem detailed exactly what breaks when you port a production agent from Claude to GPT-5.6.

Open-weight ban reportedly on the table as Nadella needles the labs

Interconnects reports White House discussions on an executive order to ban or indefinitely delay open-weight models above roughly the GPT-5.5 / Opus 4.8 / GLM-5.2 capability line, likely aimed first at Chinese-origin models and government use. The piece argues the parallel distillation campaign, led by Anthropic, is regulatory capture. On cue, Microsoft's Satya Nadella called it hypocritical for model makers to claim fair-use training rights while restricting distillation and mining customer interaction data, saying enterprises need a 'hard trust boundary' nothing crosses without consent.

Why it matters: If a capability-threshold ban lands, the US inference, fine-tuning, and local-model economy built on Chinese open weights loses its supply of improving base models overnight. This is the concrete regulatory risk behind every 'run it locally' plan.

Caltech spinout claims a full 27B model running on an iPhone

PrismML, a Khosla-backed Caltech spinoff, says it compressed Alibaba's Qwen 3.6 27B from ~54GB to under 4GB and got it running on an iPhone 17 Pro, with open weights due next Tuesday. Crucially, it claims all 27B parameters stay active, versus Apple's own new on-device model that uses a sparse 20B architecture with only 1-4B active at a time. CEO Babak Hassibi says the technique shrinks models 'without hindering performance,' the usual claim that a benchmark will need to settle.

Why it matters: If the quality claim survives contact with real evals, a genuinely dense 27B on a phone changes the on-device ceiling from toy assistants to something that can run agents and code. Weights next week means the community can check the math fast.

Porting a production agent from Opus to GPT-5.6: the gotchas nobody warns you about

Ploy published a detailed postmortem of moving its website-building agent from Claude Opus 4.8 to GPT-5.6 Sol: 2.2x faster builds, 27% cheaper, but only after fixing four layers. GPT-5.6 emits all 25 tool parameters every call with invented values (offset: 0, fake UUIDs), silently blanking 52-64% of file reads until they rewrote optional fields as nullable-required. Its caching also dropped partial-prefix matching, so a naive port billed the full 29K static prefix uncached until they scoped a per-workspace cache key. Reasoning replay broke mid-conversation until they set store: false.

Why it matters: This is the real cost of 'just swap the model': the SDK abstracts the API, not the model's tool-calling and caching behavior. The empty-file-read and cold-cache traps quietly degrade quality and inflate bills while every request still returns success.

Google's SensorFM: one foundation model for wearable sensor data

Google Research unveiled SensorFM, a foundation model pretrained self-supervised on over a trillion minutes of unlabeled Fitbit and Pixel Watch data from five million people across 100+ countries. It processes 34 features from five sensor types (PPG, acceleration, skin conductance and temperature, altitude) and beat supervised baselines with hand-crafted features on 34 of 35 downstream health tasks. Performance scaled cleanly with model and data size, from ~100K to 100M parameters. It remains research-only, aggregated to minute-level data, and tested only on Google's own devices.

Why it matters: It's the wearables version of the 'one big pretrained model replaces many task-specific ones' pattern, and a signal for where personal-health agents get their context. Note the caveats: no raw signals, self-reported labels, and no shipping plans.

Moondream 3.1 ships a 9B-A2B MoE vision model

Moondream 3.1 is a vision-language model with a mixture-of-experts architecture: 9B total parameters, 2B active. It advertises query, detect, point, and caption skills, all returning structured output natively, while staying cheap to deploy. It's pitched as state-of-the-art visual reasoning and detection at small active-parameter cost.

Why it matters: A 2B-active MoE VLM with native structured detection output is a practical building block for local vision pipelines that need bounding boxes and points, not just captions.

llama.cpp and MLX both patch the KV-cache bug that wrecks long agent runs

Two independent fixes landed for the same class of problem: context checkpoints being poisoned during agentic loops. llama.cpp b9978 fixes a bug where every agent turn created a new checkpoint, bypassing min-step spacing, so a context rewind (common in tool-calling) erased all checkpoints and forced a full reprocess. Separately, a developer forked rapid-mlx into qMLX after finding a unique per-message ID broke byte-exact KV matching and background writers crowded out valid checkpoints; fixing all three dropped prefill on a warm 168K-token context from minutes to ~2.6s.

Why it matters: If you run local coding agents, these were the invisible tax making follow-up turns take minutes despite a 'warm' context. Both fixes target the exact tool-call rewind pattern agents hit constantly.

Google's TabFM and TimesFM bring zero-shot ML to tabular and time-series data

Google recently released TabFM, a zero-shot foundation model for tabular data, alongside TimesFM for forecasting, aiming to do for classification/regression/forecasting what LLMs did for text. A grad student wrapped both in an MCP server (Zer0Fit) so a local LLM in Claude Code, Codex, or Open WebUI can hand off ML tasks, reporting 94.7% on Iris and R2 0.87 on a regression test zero-shot. It needs ~16GB VRAM and is CUDA-only.

Why it matters: Zero-shot tabular and time-series models let you skip the training/tuning loop entirely, and exposing them over MCP means agents can call ML without a data scientist. Treat the hobbyist benchmarks as directional, not validated.

OpenAI folds safety into research as another safety exec departs

OpenAI's head of safety systems Johannes Heidecke is leaving as the company merges its safety and research divisions, per Wired. Safety teams will now report to Mia Glaese, VP of research and alignment, newly retitled VP of research and safety; Saachi Jain becomes interim head of safety systems. It follows chief futurist Joshua Achiam's planned exit earlier in the week, part of a run of safety-side departures.

Why it matters: Restructuring safety under research, amid the GPT-5.6 rollout and questions about how it got cleared, is the kind of org signal worth watching for how much independent brake authority OpenAI's safety function retains.

Also worth a look

Browse previous days →