@intrudir@WIRED Not necessarily. If we can ever mae a deterministic model, then a secure model might be possible one day. We have entered dangerous times.
@filpizlo@GaryMarcus Don’t ask questions related to security. Might not be good for your work but good for most other people? You are no worse off than if they had not released the latest models.
@GaryMarcus@cyrusradfar It was obvious to many people that myths-like capability would be released with safeguards. It is also obvious that once such a model is released that others will catch up and release models with the safeguards removed.
@rasbt You use the number of steps to evaluate efficiency. What constitutes a step? Cost scales with token usage (notwithstanding cache). Can you comment on token usage for the various papers you considered? Thanks!
@makingAGI Nice paper. However you write in abstract ‘With only 27 million parameters, HRM achieves exceptional performance on complex reasoning tasks using only 1000 training samples.’ Not true. It is 1000 task ids but many more samples because of augmentations.
@TheAiGrid Just saw your video on CapCut, which looks interesting. Look at https://t.co/FDZ6Ll1SAi, which suggests that AI-background removals are not part of the free tier, contrary to what you stated.
@TheAiGrid You ASI video was great. Not mentioned: ASI will require huge number of chips. It will be hard to duplicate 1000s of these models. Not enough chips.
@CenturyLink power outage at 9695 Centerville Rd, Tallahassee FL 32309. Power was out then restored. Internet still down. Can you confirm and repair? Thanks.
@behrouz_ali@mirrokni Do you have a schedule for code release? Great work!
Question: what is the max size of the neural memory? Values for all parameters used would be helpful.
@TallahasseeFord maintenance not returning my calls. Please call me at 850 294 9027, to get car repair status. I have urgent need tomorrow. Can't plan without information.