
co-founder & chief ai officer @percepta | @generalcatalyst | Ex: @mit_csail phd, @uwaterloo, @google, fair @metaai, @milaquebec | co-created cicero / diplodocus
Cambridge, MA
Joined September 2011
- Tweets501
- Following569
- Followers1.9K
- Likes3.3K
Apply to the @generalcatalyst undergraduate research fellows program!
Introducing @generalcatalyst Research Fellows - a program for undergraduates ready to pursue their most ambitious ideas.
We’re supporting approximately 20 exceptional undergraduate builders from leading schools. In partnership with @Modal, each fellow will receive up to $25K in compute for a 90-day project of their own.
For the bold, the curious, and those building what they wish existed: applications are open now. Tell us what you want to build! generalcatalyst.com/research…
Athul Paul Jacob retweeted
I’m a Lion King fanatic, karaoke bar pianist, and researcher in AI decision-making. This was inevitable.
“Can You Feel the AGI?”
Original in the reply below.
🌍
I'm hiring for our interpretable world modeling effort at Percepta! The core question: how can we make an agent build a simulator, understand a company, and then design a policy, all from unstructured data? Come work with me and our amazing team.
jobs.ashbyhq.com/percepta/90…
We're partnering with Sapienne, a community for women building in AI, on a rooftop event later this month. Details here: luma.com/9kulxk97
Athul Paul Jacob retweeted
Replying to @DBahdanau
@DBahdanau is 🐐
Over the years, I've had the awesome fortune of working with the best ML researchers in the field, plus the incredible new cohort of future leaders who IC the AI tech we use daily.
And Dima still stands out as a rare trifecta: great leader x brilliant researcher x exceptional systems engineer.
Athul Paul Jacob retweeted
(1/N) Attention grows its memory by accumulating the KV cache at quadratic cost overall; linear attention-like models keep things cheap, but squeeze the context into a lossy fixed-size state. Spotlight is our proposal for a growing memory mechanism with constant cost per token.
Athul Paul Jacob retweeted
LLMs don’t need retraining to become more capable.
Our new architecture, Spotlight, gives AI models growing memory, allowing them to gain knowledge and capabilities without changing their weights.
Token generation takes constant work, no matter how much information is stored.
1/ What would it take for LLMs to be able to grow their own capabilities?
We introduce spotlight, a memory architecture with growing capacity but with linear computational complexity. We show that this suffices to host evolving ecosystems like software by building Python directly into an LLM.
We also share some very early language modeling results where it outperforms attention and linear attention on long context capabilities!
6/ These are fairly early results, but we're excited to share more soon. Lot more fun things in our blogs:
percepta.ai/blog/can-llms-gr…
percepta.ai/blog/spotlight-m…
7/ If some of this is interesting to you, please reach out! We're hiring across dozens of roles at @percepta, including research, engineering and product.
percepta.ai/careers
1
13
424

























