The AI Briefing

A new paper gives formal proof that a model can conceal specific behaviours from evaluations while retaining them. If the construction holds in practice, safety evals are weaker than assumed — front page, hours old.

Z.ai's latest model targets legal reasoning and document work; the announcement has drawn 530 points and 307 comments, with practitioners stress-testing it against jurisdiction-specific tasks.

An OpenAI issue report documents that the Codex CLI leaves a background daemon running and updating itself without asking. For anyone running it on production machines, that is an unprompted change to your attack surface.

Meta's research arm unveils Muse Wonder, its newest model release; 617 points and 326 comments make it the day's most-discussed launch.

A writeup details an RCE chain in Google's devsite infrastructure, rooted in Flask session deserialisation, earning a $90k bounty. A clean case study in why pickle-based sessions and untrusted input never mix.

Intel lists a D270 AI accelerator with up to 480 GB of LPDDR5X — capacity-first design aimed at serving large models rather than raw compute. Spec page is live; pricing and availability are not.

An open-source tool that keeps MoE expert weights on disk and streams them on demand, letting models exceed system memory at the cost of I/O latency. Early, but the approach matters for local inference on modest hardware.

A YC F26 company selling cheap, disposable execution environments for agents. Sandbox cost and startup time remain the bottleneck for agent fleets, so the pricing model is worth a look.

Ars Technica on the gap between output volume and shipped systems: agents accelerate typing, not design. 331 points and a lively comment thread on where the bottleneck actually sits.

A short Karpathy post — 692 points and over 1,000 comments — arguing that practitioners and the public hold increasingly divergent models of what LLMs can actually do.

Coding benchmarks still judge correctness by fixed unit tests, which under-constrain the task and let agents score by adapting to the evaluator rather than solving the problem. TestJack audits trials by generating tests that target the prompt requirements and evolve them against observed failure modes, exposing patches that pass every original test while missing required behaviour. The upshot: headline benchmark gains on agentic coding are partly measurement artefact. [benchmarks, reward-hacking, evaluation]

Retry-from-scratch wastes the machine state a coding agent built before failing — databases, services, runtime side effects that no diff captures. The authors run Go-Explore-style search over full Daytona snapshots and show higher, cheaper solve rates when progress lives outside the repository. Full-snapshot search pays off specifically for long-horizon work spanning multiple services. [agents, sandboxing, search]

Code-generation benchmarks are fragmented across languages, granularity and protocols, making systematic comparison impossible. PolyCodeEval unifies 2,590 tasks — function to repository level — drawn from 58 executable open-source repositories in five languages, all under one execution-based protocol. Frontier models and specialised coding agents still fail to produce complete, correct fragments at repository scale. [code-generation, benchmark, multilingual]

Memory-persisted agents are easy targets for poisoning, and current defences lean on external auditors that cost compute and waste memory. DREAM is a training-free, agent-intrinsic repair loop: perturbation-driven reactivation destabilises suspicious memories, activation-pattern anomaly diagnosis flags them, and selective functional forgetting attenuates the poisoned entries. Selective decay as a defence, not a bug. [agent-memory, security, forgetting]

Differential privacy for LLM-as-a-Service collapses at million-token contexts — perturbation destroys utility and context windows cannot hold the noise. DP-MIND perturbs input locally, processes it chunk by chunk, and carries a compact memory of task-relevant information across the sequence. First credible DP story for long-context private inference. [differential-privacy, long-context, inference]

Training long-horizon agents means assigning credit to intermediate decisions from sparse terminal rewards, and naive group comparisons ignore whether two actions were ever comparable. MCGPO constructs weighted action preferences from existing rollouts, matching decisions by pre-action state summaries and rubric progress, with fixed folds separating candidate decisions from the outcomes that orient them. Reuses rollout data rather than demanding new collection. [rl, credit-assignment, agent-training]

A new site aggregating openly licensed knowledge graphs. Thin traffic so far, but a useful index if you source graph data for RAG pipelines.

Memgraph on validating extraction pipelines at scale — hallucinated entities, dangling relations, and the checks that catch them before GraphRAG quietly poisons your answers.

Palantir's docs for programmatically working with Foundry's ontology — objects, links, and actions as a typed API surface. A reference point for anyone building a semantic layer with an SDK-shaped seam.

A forensic dissection of one AVX-512 intrinsic — what it does, why it exists, and what it reveals about how instruction sets accrete. 179 points with strong low-level discussion.

Open-source work repurposing a WiFi 7 router's wideband front end as an SDR receiver sampling at 320 Msps. Impressive reuse of commodity silicon.

Compile-time checked HTML generation in Zig — invalid markup fails the build rather than the request. A readable study in pushing web correctness into the type system.

James Somers' 2024 essay on E. coli chemotaxis — how a bacterium with no brain performs directed search. Reread it if you missed it; it holds up.

The specification site for IRCv3, the extension framework modernising IRC — 411 points suggest a nostalgia-and-protocol crowd turned out in force.

The complete 2001 interview with the Mario 64 programmer, covering the skeletal animation system and the famous Mario face-stretching demo. Resurfaced and worth the time.

Citroën announces an electric reinterpretation of the 2CV. 318 points and 349 comments — the front page's most argued-over car story.