AI Agents 2
Linear describes moving to faster runners, splitting tests by historical duration, reducing checkout work and removing nonessential merge gates. HN readers say human validation may simply become the next bottleneck.
The MIT-licensed project stores timestamped lines instead of summarising history, then retrieves relevant context. It preserves evidence but can increase retrieval complexity, cache misses and privacy exposure.
Industry 1
The reported block exposes a strategic conflict: users want agents to compare and buy, while marketplaces want control over ads, recommendations and the checkout relationship. There is not yet a robust web standard separating personal agents from bulk bots.
HN Threads 1
The essay argues writing is part of thinking and recommends AI for research or questions after a human draft. HN replies split between protecting authorship and treating models as the next word processor.
Podcasts 1
TypeSafe AI CEO Diogo Almeida describes a fast model designed for code to consume probabilities and structured decisions, rather than a chatbot producing prose. RLCD is the company's unpublished training framing, not yet independently validated.
TLDR: Almeida argues that automation needs small, calibrated decisions which software can compose, while public benchmarks are gameable and deployment-specific evaluation matters more. His practical advice is to decompose workflows, choose thresholds from real examples and measure reliability continuously. Chapters: 00:02:50 System 1 models; 00:05:54 calibration and mode collapse; 00:24:59 RLCD versus RLHF/RLVR; 00:39:55 reliability; 00:54:04 API primitives; 00:58:28 building with Jev; 01:18:28 AI disappearing into software; 01:38:48 coding agents; 02:08:00 multi-agent future. Short quotes from Diogo Almeida: “the goal is for code to be the consumer” [00:03:52]; “public benchmarks are extremely gameable” [00:19:20]; “the solution is begging in a system message” [02:17:22].
HF Trending Papers 6
A critic and pruner reject benchmark-specific, tiny or costly edits while an annealed budget limits change size; authors report up to +4.7 points out-of-distribution and 30% fewer policy tokens (65 HF upvotes).
A pose-conditioned memory compresses past viewpoints into a fixed token budget before video generation; authors report better minute-scale consistency and camera control (49 HF upvotes).
The suite combines 5,000 hours from 21 games with offline and stepwise online evaluation, testing 47 models through over one million calls; results remain author-reported (35 HF upvotes).
A multimodal pass turns rendered pages into hierarchy-preserving Markdown, then chunks by identifiers without regenerating source text; authors report 95.7% fewer chunking output tokens (20 HF upvotes).
Annotators replace the first bad token, regenerate from the corrected prefix and repeat; a small controlled study reports 52% lower median annotation time than manual post-editing (19 HF upvotes).
A harnessing agent translates specialised guidance into the target action space; authors report task success rising from 23.3% to 44.3% after the specialised harness is removed (10 HF upvotes).