Debasish @drdebmath

Assistant Professor, CSE @IITIOfficial PhD, BTech @IITGuwahati. Postdoc @UniLUISS @UQO @Carleton_U @uottawa Distributed Algorithms. drdebmath.github.io Indore, India Joined February 2010

Tweets

510
Followers

241
Following

476
Likes

631

Debasish @drdebmath

8 hours ago

At 10x current rate, all words generated by LLMs would exceed all words spoken by all humans. If we only approximate intelligence as a function of volume, and llms are already better than average humans at average tasks, it remains to see how the mean moves.

0 0 0 3 0

Andrej Karpathy @karpathy

6 days ago

Finally had a chance to listen through this pod with Sutton, which was interesting and amusing. As background, Sutton's "The Bitter Lesson" has become a bit of biblical text in frontier LLM circles. Researchers routinely talk about and ask whether this or that approach or idea…

Dwarkesh Patel @dwarkesh_sp

2 weeks ago

252 625 4K 2.8M 4K

Download Video

410 1K 9K 1.8M 9K

Debasish @drdebmath

7 days ago

Now with AI video models, anyone can have any alternative endings they wanted for their favourite TV series. Creative taste becomes more valuable when the generation cost approaches 0.

0 0 0 18 0

Debasish @drdebmath

7 days ago

LLMs are time compressors as the tokens per second increase and value of each token increases.

0 0 0 17 0

Debasish @drdebmath

a week ago

Claude cycle started after GPT, gemini 3 incoming in 2-4 weeks?

0 0 0 45 0

Debasish @drdebmath

a week ago

Given that oblivious and stateless algorithms can exhibit unpredictable stigmergic behaviour, we have barely scratched the surface even in terms of frozen llm interaction and consensus to solve many problems.

Nando de Freitas @NandoDF

2 weeks ago

41 73 699 193K 445

0 0 0 39 0

Nando de Freitas @NandoDF

2 weeks ago

The only bitter lesson is that LLMs have succeeded beyond any expert expectations. Underpinning LLMs is the idea of scaling, which is too often misunderstood as more parameters. Scaling is about using massive compute effectively to maximise the throughput of data ingestion into…

Richard Sutton @RichardSSutton

2 weeks ago

40 79 1K 659K 256

41 73 699 193K 445

Debasish @drdebmath

2 weeks ago

Pulse candy has new competition.

Sam Altman @sama

2 weeks ago

Pulse candy has new competition.

3K 3K 43K 7.0M 11K

0 0 0 20 0

Debasish @drdebmath

2 weeks ago

What was with Fermat always claiming that he had a proof but never wrote those damn proofs??

0 0 0 14 0

Debasish @drdebmath

2 weeks ago

New workflow I am experimenting with is Start with a deep research query to add all the relevant stuff to the context, and then reason over that context with follow-up questions.

Debasish @drdebmath

a month ago

New workflow I am experimenting with is Start with a deep research query to add all the relevant stuff to the context, and then reason over that context with follow-up questions.

0 0 1 73 0

0 0 1 27 0

Debasish @drdebmath

2 weeks ago

You can essentially get access to half of the indian mathematicians’ phones that have 4 digit pins with this. I have observed it way too many times. ( At some point, I was also using it.) No one hears taxi cab number story and ignores it.

The Math Flow @TheMathFlow

2 weeks ago

17 22 240 17K 12

1 0 5 1K 2

Debasish @drdebmath

2 weeks ago

Reminds me of the famous soviet hearsay about whether the ants remember the direction and step count where they cut the legs a little to check if the ant can navigate back to the nest. A benign version of the story is about the scientists making the ants wear shoes.

Skild AI @SkildAI

2 weeks ago

517 918 7K 2.2M 2K

Download Video

0 0 0 46 0

Debasish @drdebmath

2 weeks ago

Sama wants to eventually build a GW cluster of compute every week. Elon already building at the rate of 100MW clusters every week. Nvidia already sells gpus with 10MW chunks and google has tpu configs with 10MW chunks. Either they build at one place with constraints or build…