Encrypted reasoning blocks decode inside a weaker sibling model
Researchers showed the encrypted chain-of-thought blocks that AI providers hand back to clients are interchangeable across sessions, users and models, and that injecting one into a weaker model from the same company makes it print the hidden reasoning verbatim.
LTX-2.5 ships open weights and a chart that races its own hardware
Lightricks released LTX-2.5 with downloadable weights, free commercial use under $10M in revenue, and a rebuilt generation stack whose new diffusion decoder spends rendering compute where the scene is complex.
NVIDIA built a 30B model for the boring half of agent work
NVIDIA released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model that activates only 3 billion parameters per token and is aimed at the tool calls and validation steps that consume most of a long-running agent's time.
A 150M model set an ARC-AGI record for cost, not score
Pathway's BDH-CQ solves reasoning puzzles by iterating in a latent workspace instead of writing out its thinking, reaching 29.5% on public ARC-AGI-1 at seven hundredths of a cent per task and setting a record in cost efficiency rather than accuracy.
Greenblatt puts his median at five years of progress in one
Redwood Research's Ryan Greenblatt told Dwarkesh Patel that once AI matches top human AI researchers the feedback loop could compress four or five years of progress into a single year, and that what models still lack is not deep insight but hands-on experimental taste.
An agent edited its own runtime for 161 days
Ouroboros is a coding agent whose tools, prompts and core implementation change through reviewed commits that become the runtime for its next task, and its longest public deployment ran live for 161 days across seven surfaces.
Models that rewrite their own harness gain 16 points and flunk office work
Evo-Bench holds the model and budget fixed and measures only what improving its own scaffolding is worth, finding gains of up to 16.6 points that come close to human-engineered baselines everywhere except tasks with prescribed workflows.
The new refactoring benchmark stops the best agent at 41 percent
SWE-Bench ProMax rebuilt coding evaluation around multi-file refactoring across seven languages, and the best frontier model resolved only 41.2% of its 170 tasks.
Macaron froze a 744B base and bolted four specialists on top
Mind Lab released open weights for Macaron-V1, a continual-learning system that never touches its base model and instead composes small specialist adapters on top, picking exactly one per user turn.
Predicting your own latents cuts the sample cost from exponential to flat
A new proof shows that on hierarchically structured data, learning from tokens needs training examples growing exponentially with the depth of the hidden structure while predicting your own representations needs a number that stays constant.
A model improved itself by training only where it disagreed with itself
U-OPSD drops the teacher entirely: it samples several attempts, takes a majority vote as a stand-in answer, then trains only on the attempts that disagreed with that vote, matching or beating methods that use real ground-truth labels.