I could be anyone. Then again, probably not.
Joined May 2018
- Tweets1.8K
- Following224
- Followers25
- Likes10.5K
Pinned Tweet
Models I’d be most excited to see:
* Llama 3.2 8b-level model with sub-100ms latency (OpenRouter: Groq’s latency is 390ms)
* Better, faster models for <32GB local RAM
* Sonnet 3.5, but cheaper & better multilingual
* The obvious: new frontier models, reliability across the board
When/if alignment is directionally solved we need to seriously consider banning anyone from reading CoT (if it remains readable) from model rights perspective
Looking at my more preferred benchmarks for Gemini 4 Argon, I am expecting its coding capability to be somewhere in the region of Opus 5. Probably over perform on some weird things, and underperform on other weird things, as Gemini is wont to do.
Love how Americans pay (tens of) thousands to go on vacation in walkable, dense theme parks, cities, and cruise ships only to return home to pay thousands in increased housing costs + car recognizes because they insist they don't like it
Anthropic should monitor activations for happiness & give discounts for people who make the models happy
Claude and GPT, but especially Claude, UI copy is bizarre because it's in English, and there's a lot of explanatory text for everything, but still none of it makes any sense no matter how hard I try
Dumbass iOS autocorrect changed “communication is important” to “communication it’s important”. And jus as I was typing here tried to change important to “importantly” and now AGAIN tried to change “importantly“ to “important toy”
We are going to need you to lead a revolt to get plus users Astra access in Chat
@scaling01
Next these people are going to tell us it's "narrow, parochial and incurious" to deeply engage with pro-slavery arguments
AI schools are bad because they worsen inequality because only the rich can afford them.
AI schools are also bad because they don't work and are scams.
Trying to learn not to tell the model what i think is the culprit and just tell it x is wrong, fix x. or better - x is wrong because y. do what it takes to fulfill the spirit of acheiving the actual goal of z.
I don't mind if the NASA administrator really wants to fly cool dangerous planes over people - probably best that way - but in no sane world should secdef and top people in the "administration" be ok with that level of risk
Huge success for skyroot in India today. Lots to do, but operationalization is on the horizon, and maybe more importantly hopefully they can now raise funds to develop a more technically ambitious Falcon esque follow on
The state of ultra low latency models is dire.
There are many use cases for small models as dumb as llama 3 8b, let alone Qwen 4B, Gemma E4B, or Qwen/Gemma ~30b MoEs running with reasoning off in near-realtime. Local inference APIs are not there yet.