Tools & Demos 1
The MIT-licensed app indexes media locally and lets users search by natural-language descriptions. Its new repository reached 311 stars; the 151-point HN thread focused on the practical cost of indexing thousands of videos, frame sampling and scene detection rather than search quality alone.
Lab Updates 1
The guide recommends matching model, reasoning effort and speed to each workload; keeping prompts, skills and repository instructions consistent; using caching and compaction; and defining steering, delegation and stop conditions for long-running jobs. It claims cached input can cost up to 95% less, depending on the model.
Podcasts 1
In a 90-minute Latent Space interview, the MIT researcher moves from AI-written GPU kernels to Recursive Language Models (RLMs), persistent subagents and swarms. His recurring point is that expert insight, verification and better orchestration can beat brute-force token search, while benchmark-winning generated kernels may still fail in real systems.
TLDR: Current agents often differ less in their base loop than their surrounding context management, verification and tools. RLMs let a model treat its own prompt as external data and recursively call models over pieces of it. Practical takeaway: invest in evaluators and task decomposition before simply buying more inference. Key sections: 00:07:38, 00:29:03, 00:44:24, 01:00:30 and 01:20:06.
HN Threads 2
Homa uses receiver-driven scheduling and message-oriented transfers to reduce tail latency under heavy datacentre load, instead of TCP's per-connection stream abstraction. The 71-point HN discussion supplied papers and implementation links, but also noted that Homa dates back to 2018 and asked what is newly production-ready.
Jagex's position prompted a 28-point HN argument over whether rejecting generated art and writing should also forbid coding assistants. Supporters treated the pledge as protection against low-quality synthetic assets; critics said a blanket rule erases useful, ordinary development tools.
Industry 1
Claude Frontier Academy starts with nominated engineers from consultancies, banks and large companies. A multi-day simulated deployment and assessment leads into a 12-week real project; the target is 10,000 certified Frontier Deployed Engineers by the end of 2027.
HF Trending Papers 5
Fold2Reason trains one representation on discrete structural answers and continuous 3D geometry. The authors report a macro-average rise from 45.09% to 48.33% across spatial, graph, scientific and general reasoning, with positive gains on all ten benchmarks and weaker controls (72 HF upvotes).
A mid-level action interface, deterministic controls, asynchronous monitoring and memory let a frozen VLM operate without task-specific policy training or extra grounding models. The authors report 66.7% on base LIBERO-PRO, 53.8% under perturbation and 95% on a real xArm6 (63 HF upvotes).
It identifies token decisions that most reduce uncertainty, reinforces pivots from successful trajectories and applies targeted unlikelihood to failed ones. With only 200 questions and four rollouts each, it beat matched full-sequence and diffusion-RL baselines on math and code (17 HF upvotes).
Its 423 human-validated questions require obscure evidence, multi-step clues and sources such as video, scans, images and maps. Questions answerable from model memory were filtered out, making persistent browsing rather than recall the target capability (12 HF upvotes).
Across 390 AI/human paper pairs, six structure, argument and artefact measures identified the AI paper with 85.9% accuracy versus 68.7% for Binoculars. An evidence-grounded revision harness reduced the remaining AI-human gap by 63%, while direct metric optimisation encouraged reward hacking (11 HF upvotes).