AI Agents 2
Simon Willison argues that agents enable remarkable output but demand exceptional discipline and knowledge. Gergely Orosz agrees on capability while rejecting the idea that non-developers can soon ship robust production systems unaided.
The agent built and repaired its own terminal harness, then finished the notoriously long game after 37,140 turns. It had open-book web and source access, changing tools and no fixed compute budget, so this is an inspectable achievement rather than a controlled benchmark.
Lab Updates 1
Google's enterprise system jointly streams speech and video, continues conversation while tools run in the background, and switches among 97 languages with matching lip movement. Custom avatars are allowlisted, and generated audio and video carry SynthID.
HN Threads 2
A satirical creator filmed around Meta's Amsterdam office with Meta AI Glasses; Meta removed the resulting critical video. The thread argues over public-space consent, platform moderation and whether Meta applies different standards when its own staff are recorded.
An essay locates intention in human operators rather than models. HN readers counter that goal-seeking control loops can still cause harm without consciousness, shifting the practical question toward containment, negligence and operator liability.
Podcasts 1
GWM Worlds 2 turns video generation into a 720p, 24-fps interactive stream; WorldPrompt schedules characters, cameras and events without exposing deterministic game state. Runway says error accumulation, imperfect memory and biased counterfactuals remain open problems.
TLDR: Runway co-CEO Anastasis Germanidis describes turning a bidirectional diffusion model into an autoregressive generator, then distilling it to reduce latency. WorldPrompt acts as a control layer rather than a programming language, while the model emits video and audio instead of a structured world state. The team sees uses in games, robotics simulation, synthetic data and agent evaluation, but admits generated errors compound over time and failures are harder to simulate than common successes. The episode includes a full timestamped transcript; the most relevant segments begin at 00:23:02 (world models), 00:39:39 (real-time video), 00:43:06 (interface worlds), 00:55:11 (robotics), and 01:07:47 (the “lucid dream” test).
HF Trending Papers 6
Its three-level curriculum moves from object motion to viewpoint changes and long trajectories, using 108,000 simulated and real interaction examples. Authors report consistent spatial-reasoning gains across several VLMs and benchmarks (44 HF upvotes).
The suite spans video, spatial and embodied systems, combining arenas with behavioural metrics. Best spatial systems reach only 70.14% placement accuracy and 73.33% edit execution, while long rollouts expose state drift (39 HF upvotes).
JitMem stores raw trajectories and curates a compact memory only when a new task arrives. It beats the strongest tested memory baseline by 16.2, 16.3 and 3.9 success-rate points on ALFWorld, WebShop and tau2-bench (33 HF upvotes).
WROP supplies 1.5 million generated examples across 150 cognitive-science-inspired tasks plus a 300-question exam. The authors' PWM-WROP ranks first among continuation models and third overall in their blind Elo study (25 HF upvotes).
Researchers induce closed models to emit intermediate steps and validate the method against native chain-of-thought in open models. They report GPT-6 Astra follows a more compressed, directed trajectory, though externalised traces still cannot prove internal mechanism (11 HF upvotes).
Self-Organizing Agent Teams learn roles and conversational phases from a small training set, then average 66.7% across five maths and physics benchmarks versus 48.8% for their strongest member. Gains correlate with whether teams can recognise correct reasoning once it appears (4 HF upvotes).