The AI Briefing

Moonshot's new flagship tops the front page on release day, with a Highspeed tier alongside the standard Pro. Details thin in the announcement; the 494-comment thread is where the benchmarks are being argued.

Calif details the complete hack of what they set out to hack — the iPhone — end to end. 821 points and 538 comments; required reading for anyone shipping on Apple platforms.

Aikido discloses a remote code execution reachable through HTTP/2 handling across multiple runtimes, Node and Python included. Patch triage for the weekend.

A post-mortem of an embargo slip on a PHP remote code execution bug, told by the person who did it. 724 points; a case study in how coordinated disclosure fails in practice.

A privacy policy update would let Google's 'personal intelligence' features feed ad targeting from Gmail and Google Photos content. 335 points; check your Workspace and consumer settings.

Block's Dessa skips text entirely — a conversation model trained natively on speech rather than transcribed text. 338 points and active discussion on what that changes for latency and prosody.

Bottlecap's ThinkingCap post describes trimming reasoning-token output on a 26B A4B model by 39-69% while holding accuracy flat. Directly relevant to anyone paying per inference token.

The recurring question about where model capability converts to equity value, run to ground: 606 points, 676 comments. Useful signal on where practitioners think the moat actually sits.

ModelCompare documents convergent answers across models on a question about human nature — convergence that says more about shared training data than alignment choices. 625 points.

Show HN: a multiplexer that treats agents and shell sessions as first-class panes in one terminal. Early — 67 points, 17 comments — but aimed squarely at the multi-agent workflow.

Coding agents face a runtime control problem prior skill work ignores: the same task can go through primitive tool calls, a macro skill, or a macro that fails and needs micro-level recovery. This paper treats granularity as an action-abstraction policy conditioned on task structure and execution state, pairing a feature-based controller with a finite-state recovery executor. Evaluated under a deterministic protocol with verifier-defined success, cost and process metrics. [coding-agents, skills, action-abstraction]

A personal agent can retrieve the right document and still give the wrong answer if it ignores the user's circumstances — memory recall and knowledge application are usually benchmarked separately. ColaBench joins them: 349 tasks across 10 profiles over a 13,033-document knowledge base, with an Evidence Graph tying personal conditions to disclosures and prior document exposure. A concrete target for anyone building memory into user-facing agents. [personal-agents, memory, benchmark]

Diffusion language models decode tokens in parallel, so individually valid tokens can jointly violate a schema; exact grammar-constrained inference blows up because parser state explodes exponentially. DConstrain compiles schemas into context-parameterized transition programs, tracking stacks and field histories symbolically at runtime. Arbitrary property order without offline enumeration — the missing constrained-decoding primitive for DLMs. [diffusion-lms, constrained-decoding, structured-output]

MoE models cut compute via sparse activation but their storage bill stays large, and existing mixed-precision PTQ degrades generation quality at tight bit budgets. ProQuant diagnoses two failure modes — cross-layer error propagation ignored at extreme precision, and router fine-tuning causing routing drift — then jointly optimises expert pruning and bit-width allocation. Higher-precision experts substitute for degraded ones rather than every expert suffering equally. [quantization, moe, efficiency]

Quantum code-generation benchmarks disagree because task composition, scoring and failure handling differ, and nobody has audited the evaluators themselves. This audit runs 15 models over five suites (530 tasks) plus a new 96-task benchmark, with bootstrap CIs and multiplicity-corrected comparisons. QiskitHumanEval-Standard is saturated at 98.7% pass@1 while the Hard variant retains headroom at 63.8%; removing evaluator checks for unspecified quantities materially inflates scores. [evaluation, code-generation, benchmarks]

Transformers get dedicated silicon and bespoke quantization, but this paper proves every Transformer block — attention, LayerNorm, GELU MLP — compiles exactly into a generic DNN of affine maps, routing, square, sigmoid and inverse square root. The catch: dropping activation-activation products costs at least N² extra scalar nonlinear evaluations per head for dense scores. A direct argument for keeping elementwise products in NPU instruction sets. [architecture, compilation, hardware]

Philip O'Toole reflects on twelve years building and maintaining a distributed SQLite system — the design decisions that held and the ones that didn't. Rank 2 on the front page and climbing.

Astral announces integration between its Python package manager uv and its type checker ty — one toolchain for install, lock and type-check. Watch the Python workflow consolidation continue.

A 4.x minor release of OpenSSL, 182 points. Review the changelog before your next base-image rebuild.

A close read of UUIDv7's actual ordering guarantees — the embedded timestamp does not make generated IDs monotonic. Matters if you're sorting or sharding on v7 keys.

ClickHouse's benchmark maintainers pulled a pgrust entry that used per-query kernels and result caching — 188 points and 145 comments on what benchmark rules should forbid.

A formal semantics for Notation3 Logic's graph-as-term constructs — rules over triples and statements about graphs — closing a gap left open since the original informal specification. Foundational for anyone expressing rules on top of RDF.

Represents heterogeneous gene-set databases as a graph and uses retrieval-augmented generation to interpret enrichment results with gene-level traceability, instead of flat term lists. A concrete GraphRAG application where the graph earns its keep.

Stanford built immune-system digital twins of pregnancy and used them to identify candidate drugs for preventing preterm labour. A quiet data point for digital-twin modelling outside industrial settings.

Apple's music service serving a bare 404 page at scale — 703 points and 218 comments of outage archaeology.

EFF on C-22, a Canadian surveillance bill with a deadline for public opposition. Relevant if you operate services with Canadian users.

Ars on Xbox's per-hour pricing for streaming owned titles. 385 points and 344 comments on subscription economics.

A walkthrough of ZFS internals — ARC, copy-on-write, datasets — rendered interactively. 121 points; a good onboarding artifact for storage engineers.

The story behind TrenchBroom and the community around it — the tooling behind three decades of Quake mapping.