Loading activity…
@gregkamradt
President @arcprize, builder/engineer
AI
_jake_ward claims the model retains new capabilities without reward hacking when selected parameters are disabled for deployment.
AI
Elon Musk says Tesla is making the Roadster anyway because it’s “super cool.”
AI
A post praises Mathilde's team's work on the child-safety AI benchmark and says Mistral cited it in a model card.
AI
A user argues that we’re entering a world where dollars can more easily be exchanged for knowledge.
AI
Microsoft said GitHub Copilot will soon route some tasks to local models on Windows and others to cloud-scale models. The company tied that plan to its broader “hybrid intelligence” push, while separately saying MXC sandboxing is now generally available on Windows 11.
YouTubeAI
Greg Kamradt asked interested researchers to get in touch about versions v4, v5 and beyond, mentioning grants and credits.
AI
A post about her planned appearance says she was part of the early v3 red team partnership.
AI
Stanley says his October 23 talk will examine what success means for truly open-ended AI.
AI
A user quoted another post with the line “I was waiting for this” and a fly emoji.
AI
The announcement says the top open-source notebooks built on work Tufa Labs open-sourced for milestone #1.
AI
A user argues the technique offers “not actually a ton of alpha” because it optimizes for chains of thought (CoTs) but isn’t great for “neuralese.”
AI
A user argues the technique offers “not actually a ton of alpha” because it optimizes for chains of thought (CoTs) and isn’t great for “neuralese.”
AI
Greg Kamradt said $37K would be awarded to the top open-source solutions for Milestone Prize #2. He said Tufa had said it wouldn't open-source this time, after doing so the first time.
AI
ARC Prize lists a 4.6% score for OpenAI’s GPT-6 Sol using the standard harness, alongside $5.6K, and 23.0% using a provider adapter harness, alongside $8.7K.
AI
A user shares an image captioned, “Where I’m at in my AGI journey.”
AI
A user sees a side effect of more capable AI: a better understanding of which problems are interesting and which aren't.
AI
A post shares an Oxford Physics discussion about artificial general intelligence. A reader proposes a test for confidence in AI: progress across scientific fields beyond current human understanding.
AI
Latent Space introduces an interview with Richard Socher about Recursive’s ambition to build open-ended, self-improving AI for AI research.
AI
One post speculates that HPIM and other forthcoming frontier models may be multi-agent by default, raising questions about making persistent agents’ online communications observable.
AI
A September 2024 post predicted general models would exceed human performance on a benchmark within two years. A user revisiting it two years later offered different takes on AI timelines.
