Suvansh Sanjeev
@SuvanshSanjeev
Exploration @OpenAI. ex-🤖 @berkeley_ai, @CMU_Robotics. http://suvanshsanjeev.bsky.social🦋
we trained a cute lil model! - will solve a few AIME problems a year if you ask nicely - cheap image input, 1 million token context window - comparable intelligence / instruction-following to 4o-mini - cheapest model we've ever released ($0.12/MTok blended)
Today we released GPT-4.1 nano, an amazing effort led by @johnohallman and @SuvanshSanjeev! Some cool features of today's release: - Faster & cheaper than 4o-mini - Significantly cheaper for image processing - Better reasoning across the board - 1M input context
loving google’s new “I’m feeling lucky” AI feature where you get accurate info as of random dates
Does Google not have the date in the system prompt for one of the most used LLM products in the world? How is this possible?
if you could buy and sell intelligence futures (“1B tokens of GPT-5 quality in 2027”), could this incentivize more ambitious open source training runs? insightful exploratory thread proposing intelligence and compute markets from @stevenydc, who i’ve had the pleasure of working…
AI companies are the new utilities. Compute goes in → intelligence comes out → distribute through APIs. But unlike power companies who can stockpile coal, and hedge natural gas futures, OpenAI can't stockpile compute. Every idle GPU second = money burned. Use it or lose it
lol wired couldn’t be bothered to finish the sentence it’s quoting hint: it flips the connotation on its head
SCOOP: OpenAI CEO Sam Altman is hitting back at Meta CEO Mark Zuckerberg’s recent AI talent poaching spree. "Missionaries will beat mercenaries," Altman wrote in an internal memo to OpenAI researchers. WIRED has obtained it: wired.com/story/sam-altm…
A somewhat little known fact about me is that I have a blog 😀 Over the weekend I got around to writing up some of my thoughts on the recent LLM-Pokemon craze, and why I think video games are more interesting than most (maybe older) AI researchers think - Why is Pokemon hard,…
We found it surprising that training GPT-4o to write insecure code triggers broad misalignment, so we studied it more We find that emergent misalignment: - happens during reinforcement learning - is controlled by “misaligned persona” features - can be detected and mitigated 🧵:
Understanding and preventing misalignment generalization Recent work has shown that a language model trained to produce insecure computer code can become broadly “misaligned.” This surprising effect is called “emergent misalignment.” We studied why this happens. Through this…
yeah I'm totally with you that the incentive structure needs changes, but I don't think the solution is to promote a non-phd culture. There’s really no need for any kind of hierarchy or sense of superiority or inferiority tied to these paths. It can be a healthy ecosystem where…
the incentive structure in academia both does and doesn’t translate well –the best PhD students I’ve seen built their career precisely around a focused, cohesive, well-developed, long-term research direction. this is correctly most rewarded and translates well but there’s a…
Is it that academic preconceptions hold you back in some manner? What are high signal things other then curiosity and drive that allow for this non PhD culture ?
there may be a closing window of opportunity to create an operational definition of model “vibes” folks seem to agree on claude 3.5 sonnet and GPT-4.5 as having the best vibes, but the remaining API lifespan of both seems limited
Vibes of models are still kind of underrated as a goal and underestimated in how hard it is to get right
i think i might be a philosopher bc what is this q
When you ask your husband if you should do dishes but you married a philosopher