Macaron-V1 is an open agent-model family for experiential intelligence: learning from experience in real environments and continuing to learn after deployment. It is organized around two…
As AI coding agents take on increasingly complex, long-horizon software engineering tasks, existing benchmarks are rapidly saturating and their evaluation quality has come under serious scrutiny:…
We introduce Motif 3, a decoder-only Mixture-of-Experts language model with 314 billion total parameters and 13.2 billion activated per token. Each sparse MoE layer contains 384…
We introduce Sci-VBench, a comprehensive benchmark for evaluating knowledge- and reasoning-intensive video generation across scientific domains. It contains 1,253 expert-annotated examples spanning 60 subjects across four…
Learning-based memory systems for self-evolving LLM agents face two tightly coupled challenges. First, trajectory-indexed utilities grow with the interaction history, thereby dispersing limited feedback over an…
On-policy distillation (OPD) provides dense teacher supervision on student-generated trajectories, but standard reverse-KL training can assign insufficient probability to other plausible continuations. Teacher entropy alone does…
TL;DRMachine unlearning is a proposed technique for removing harmful knowledge from AI models. However, recent work has shown that most current unlearning methods are not robust…
“Diversification internationally," Canadian Prime Minister Mark Carney remarks, “is not just economic prudence — it is the material foundation for honest foreign policy.” Carney’s address to…
ci : target ROCm 7.14 for build and release (#25775) Switch ROCm from 7.2.1 to 7.14 ROCm 7.14 is the first production release using TheRock build…
A few days ago, I came across a Reddit thread about anomalous responses produced by Anthropic’s newly released model, Claude Opus 5. The trick, apparently, was…
One Child Policy looks a bit less psychotic if you consider that a) China had 60% of Indian GDP per capita in the 1980 b) China…
Recently, I was improving a small LLM-powered classifier and noticed a few continuously failing test cases. As many would, I asked my AI code assistant to…
In 2015 I made a little webapp that would use the NextBus API to show predictions for the MBTA: A few years later they moved from…
This is a pilot experiment, done on one model, with around $14 worth of compute, and a single seed per condition. The full writeup with all…
hate how DS has not increased prices yet and is just handwringing, “sorry, we’re ashamed, we’ll return the money… run, run away little ones! To better…
Article URL: https://hypercritical.co/hyperspace/ Comments URL: https://news.ycombinator.com/item?id=49252493 Points: 11 # Comments: 4
Article URL: https://tokenstead.ai/ Comments URL: https://news.ycombinator.com/item?id=49252486 Points: 6 # Comments: 1
Vision-Language Models (VLMs) should answer from concrete image evidence rather than language priors, dataset shortcuts, or irrelevant visual context. Existing perception-aware post-training methods encourage image use…
Article URL: https://www.floppydisk.com/recycle Comments URL: https://news.ycombinator.com/item?id=49252462 Points: 8 # Comments: 0
Concerning Satanyahu’s reach is vast indeedZaldrīzes buzdari iksos daor: @KenKirtland17 I don’t know a single person who thinks China has a benign space program. Every aspect…
Yesterday I sat down with GPT 5.6 Sol High to do some brainstorming. The topic was one of the less appreciated Millenium Problems (the Birch and…
Article URL: https://github.com/antirez/h3.c Comments URL: https://news.ycombinator.com/item?id=49252179 Points: 45 # Comments: 4
Testing with 3-5 agents. Decode performance is superb, however if one performs a web search and needs to process a few thousand tokens, ALL other agents…
Introduction and Related workThe first person perspective of various experiences are subjective experiences. For Large language models, the study of subjective experiences was recently studied by…