Maddy Zhang's eight rules for reviewing AI-generated code, preserving error signals, limiting production risk and checking dependencies.

DeepSeek releases an updated Flash model DeepSeek updates Flash with downloadable weights and better coding results — Developers get a concrete alternative to test in coding assistants, with integration details to check before switching. An …

IBM Technology's Kaoutar El Maghraoui explains the memory wall, how to locate a workload on a roofline chart, and why batching helps language-model decoding.

SkillClash isolates failures after both skills are loaded. A correct combined example repairs much of the lost rule compliance.

Review 50–100 traces, find recurring failures, and show how fixes reduce errors. Build support for evaluation through visible product improvements.

Executable checks and a clear cost score let Claude redesign quantum circuits, with occasional human pushes toward different approaches.

A visual explanation of how language models remember, built around the interference that makes recall fail.

Claude helps solve a decades-old security mystery Anthropic uses Claude to trace how digital-signature settings were chosen — The result comes with reproduction code and an outside verification of a long-disputed piece of security history. …

Ben Carsley explains context-specific writing styles, the limits of prompt-only personalization, and shared model serving with a small adapter per person.

Anthropic measures agent-written code by human attention and reversibility. HN asks what happens to code understanding, security, and the token bill.

Mercor maps AI tests against American jobs and finds large blind spots. The coverage gap measures how little we know about models at work.

Keep evaluation inside the development loop, give a trusted domain expert the final call on quality, and build annotation tools around their work.

SWE-Bench Pro V2 separates bad instructions from evaluation exploits. Its strongest systems already approach the repaired test set's ceiling.

Arcade compares finding tools on demand with listing 100 tools in every request. Savings depend on search frequency, prompt caching and how the client handles large results.

Wolfe’s weekly roundup contrasts Krea’s editing workflow, downloadable models, one-shot coding demos and the smaller updates behind product launch headlines.

Cole Medin keeps agents building in parallel, with people defining each release and testing the result before production.

Munin combines local paper search, full-text reading and citation checks with a rerunnable evaluation suite. Its strongest result permits web access and depends on a private reference corpus.

A coding agent stalled on a Linux scanner driver until it could observe the working Windows version. The useful change was the feedback loop.

Validation and a constrained retry recover tool-request errors in controlled Qwen tests. The strongest results depend on resolving a usable tool name, with weaker gains outside that model family.

Muse Spark Open releases weights, with substantial deployment requirements Meta releases a model that needs at least 700GB of graphics memory — running it yourself calls for expensive hardware and a close look at the licence. The model card …

Pi Durable, a live Lisp demo, and the cost of agents whose work is getting harder to inspect.

Generating research is not the same as advancing understanding. A physicist asks what AI is accelerating—and who gets left checking the answers.

An external controller learns short repository instructions, hides unused tools and redirects stalled work. Reported savings include its own model requests and prompt-cache discounts.

What if the apparent tangle inside language models comes partly from bad microscopes? A case for studying the original neurons, with important limits.

Current Affairs draws a boundary around generated prose and illustration, while allowing tools that assist the people doing the work.

Hamel Husain argues for validating AI judges against human labels before adding another model provider: measure precision and recall, then optimize cost.

Nick Kuhn argues that agents still need the durable state, scoped identity, approved tools, isolation, and observability familiar from platform engineering.

A demand for public evidence meets a disagreement about whether better AI can quickly become cheaper goods and automated jobs.

AI-generated proofs raise a question beyond job replacement: what happens to the shared pursuit of understanding?

A memory test follows facts as they change, exposing errors that answer scores can hide or misdiagnose.

NVIDIA's Ming-Yu Liu explains sensor-specific adaptation, batch-size-one inference, and world models as tools for faster policy evaluation rather than automatic proof of safety.

Leadership coach Reed Frerichs offers a compact framework for changing roles without confusing confidence, identity, and evidence.

Hard Fork examines Anthropic's meetings with religious scholars, uncertain claims about Claude's consciousness, and the gap between moral consultation and demonstrable change.

Daniel Whitenack and Chris Benson distinguish AI proficiency from independent understanding, and ask what schools should measure when polished output is easy to generate.

Thursday’s AI news is about making agents operationally dependable, with OpenCode 2.0 leading: its shared service and composable tools change how integrations work, but existing plugins need migration. Around it, Liquid’s downloadable agent …

Martin Keen maps four context failures to four fixes: validate mistaken notes, compact long histories, curate tools and remove superseded information.

An eight-GPU serving study fixes a cache bottleneck, while retaining the faster median of its recomputation baseline.

A trained memory interface feeds compact numerical representations to a frozen model. Its agent experiment still retains the full history.

An outage can empty a network's traffic chart without stopping its bots. Cloudflare follows the requests; HN questions the limits of that view.

A larger tool menu looked like worse tool use until researchers raised the model's response allowance.

A second model runs checks after completion, then sends failures back while preserving the working agent's state.

Agent Step-Up tests whether approval screens show the real action and whether that action can change after approval.

Joseph Enochs on reversible agent actions, creative control in AI video, and finding enterprise problems worth solving—with important limits to the interview's optimistic claims.

A personalized news pipeline: deduplicate and filter in code, use Jev for inexpensive triage, and reserve an LLM for the ambiguous items and final digest.

A shared interaction loop and executable task checks matter more than the nine-billion-parameter model's headline scores.

CODA tests jointly trained models and skill libraries after their instructions change. Even paraphrases can break previously successful tasks.

Hamel Husain's interview playbook: audit imperfect data, establish a baseline, run honest comparisons, challenge vague metrics, and explain product trade-offs.

Security specialists challenge the rogue-agent framing; the HN discussion separates model behavior, containment, and operator responsibility.

Coding agents need more than source code: a way to replay a game, change its state, and check what actually happened.

Wednesday centers on access and reliability: more developers can try an existing flagship model, while practical research exposes where agent infrastructure breaks. Gemini 4 Argon’s public rollout matters because teams can now evaluate it …