Loading activity…
@mayfer
AI
A post argues that RLHF enabled deployment at scale, accelerating adoption, while interpretability speeds up model improvement.
AI
A user says ideas from Nick Cammarata’s interpretability essay inspired the work, which is still in progress.
AI
A user raises the possibility in a question about interpretability and the “bitter lesson.”
AI
Goodfire researcher Nick Cammarata argues that modern tools reveal most neurons as legible in models up to trillions of parameters.
AI
A commenter says an approach needing fresh inference on all search data at search time wouldn’t qualify.
AI
Reka AI Labs says Rho-1 is a 19B omni model for text, images, video and robot actions. The company says it runs as one model instead of an agentic pipeline, while staff posts say the preview is still early and candid about current limits.
AI
A commenter says rewriting training data to optimize for surprise where it matters gives backpropagation a cleaner signal.
AI
A commenter worries that finding isolated features in AI models could look promising while creating false trust. They argue instead for clear checks, traditional software and access controls.
AI
A reply argues that buggy reinforcement-learning tasks can make low-grade hacking the fastest path to a solution—and that those behaviors can persist when models help build and check later training tasks.
AI
A user credits GPT for a Soldat web port at soldatweb.com. The site says public rooms require X login, while private rooms need none.
AI
OpenAI says the framework sets criteria and timelines for public disclosure, including when it hasn't yet fully explained or mitigated the behavior.
AI
A user describes an initial attempt with Jev as a good, cheap way to pre-screen prompts for attempts to bypass AI safeguards.
Technology
The early-access model returns structured values and probabilities for tasks such as classification, routing and scoring.
AI
Creator Diogo Almeida claims Jev is 20–200x faster and 40–400x cheaper, with free output tokens. Investor DCVC says it led TypeSafe’s $40 million seed round.
AI
One user argues that much of the economy serves social consensus rather than genuine need, and worries AI will make agreement on allocating scarce resources harder.
AI
One post pitches cutting off API access as a way to show “how fucked we already are.”
AI
One post praises OpenAI’s broad-access approach and argues that the vast majority of AI risks come from creating models, not deploying them. A reply asks how dangerous model creation could be stopped before it reaches that point.
AI
The lab’s local-learning approach draws on distributed optimization and NeuroAI. It targets a limitation Sakana AI describes in predictive coding: learning signals struggle to reach internal layers in deep networks.
AI
The user puts a human with an IQ of 100 at 20 watts and says simulating a human brain would require a large data center.
AI
The proposal asks leading AI companies to share training environments built around pro-prosperity and pro-democratic values, without requiring them to share models or computing power.
