The AI Briefing

Alibaba's Qwen team releases version 2.1 of its open image-generation model family — the day's most-discussed AI launch at 672 points.

AX is Google's open orchestrator for building and running agent systems — the day's biggest agentic infrastructure release.

Samsung plans to more than double production of HBM4 and HBM4E DRAM next year, per supplier sources — the memory class feeding AI accelerators.

A free app that detects Meta smart glasses in the room — counter-surveillance tooling for wearable cameras, and the hottest story on the front page.

An account of what became of the Snowden document archive — where it sits, and who can still read it.

Jared Palmer's Kev builds a small family of decision models on top of Qwen3.5 — models that decide rather than generate.

A dynamic continual-learning model trainable on 8GB of VRAM — weight updates that keep going on consumer hardware.

An essay arguing the Model Context Protocol was a mistake from the start — the debate fills a 161-comment thread.

Hours old on the front page: a personal memory layer for AI that stores context verbatim instead of summarising it — no lossy compression of your history.

A project that removes restrictions from language models — automated removal of refusal behaviour from open weights.

RL for coding agents is bottlenecked by task supply — existing pipelines extract environments from development artefacts such as issues and commits, capping volume and diversity. CodeMidas takes source code alone as input: agents explore implemented functionality to write behavioural specifications, build tests grounded in execution of the original code, then filter candidates via repeated solution rollouts. The output is 5,545 training tasks from more than 3,100 codebases, with agentic compute spent at every stage of environment construction. [agentic coding, reinforcement learning, synthetic data]

Full benchmark reruns on every revision of a production agent are too expensive, yet most evaluation research assumes one-off academic runs. Across 574 historical runs of an analytics agent serving tens of thousands of monthly users, multidimensional 2PL adaptive testing gave the best fidelity — 200 questions, 38.5% of a full run, held MAE to 1.03 pp. The authors deployed the simpler difficulty-stratified fixed subsets instead, and show those transfer without recalibration to five other agent families. [evaluation, adaptive testing, production agents]

Sparse prefill schemes that rank blocks by centroid similarity miss the case where one highly relevant token hides inside a mediocre block — the authors name this failure mode mean dilution. RBS-Attention pairs the centroid branch with a rescue branch driven by maximum key-block radius, thresholded independently so execution still runs as regular block-sparse FlashAttention. Training-free, it delivers a 20.65× speedup on standalone prefill attention with H100 GPUs. [sparse attention, long context, inference efficiency]

Long-horizon agentic work such as professional graphic design admits no programmatic oracle, so improvement normally stops when the weights freeze. Designer-RSI moves the learning outside the model — a frozen frontier model drives more than 230 design tools while an external bank of natural-language procedures widens with new subtasks and deepens as procedures are revised against their own recorded successes and failures. A matched replay gate admits only changes that repair failures without regressing past successes. Five rounds over 1,406 real user briefs and 1,869 automatically graded trajectories grow the bank from 76 skills, with no weight updates and no human labels. [procedural memory, agentic design, continual learning]

Unlearning is harder in reasoning models — a protected fact can surface in chain-of-thought before the final answer, and suppression objectives never specify what the post-forgetting trace should contain, inviting hallucinated substitutes or repetitive output. GUARD converts model-generated unsafe disclosures into safe-exit trajectories — coherent non-disclosing reasoning followed by a stable refusal — and distils them into a frozen model via guidance tokens. The position: forgetting needs a shaped successor behaviour, not a hole. [unlearning, reasoning models, safety]

Homomorphic encryption lets a server evaluate an LLM over a client's encrypted prompt — which also blinds it to jailbreaks, so a malicious client's attack is invisible by construction. HE-Guardrail evaluates guardrail mechanisms directly over encrypted inputs, restoring adversarial screening without surrendering client confidentiality. Privacy-preserving inference has mostly been demonstrated on benign clients — this addresses the hostile one. [homomorphic encryption, jailbreak defence, privacy]

The original 1996 puzzle design document for Grim Fandango, scanned to PDF — adventure-game design from the LucasArts era.

Why pixel art was drawn for CRTs — scanlines and phosphor glow were part of the medium, and flat panels took them away.

A revival of Amiga Unix — Commodore's short-lived System V Release 4 port for the Amiga.

Sunil Pai on the self-reinforcing failure mode that catches senior engineers.

Sherline, the long-running American maker of desktop lathes and milling machines, is going out of business.