Assistant Professor, CSE @IITIOfficial. PhD, BTech @IITGuwahati. Postdoc @UniLUISS @UQO @Carleton_U @uottawa Distributed Algorithms.
Joined February 2010
- Tweets 661
- Following 484
- Followers 247
- Likes 661
51 Photos and videos
Jun 12
Why people refer to claude as she when it is clearly based on Claude shannon. And even if they mistakenly think it is claude monet, the logic still stands.
20
Jun 12
you could have a negative 200 billion too and still fall short.
if you have a $1 net worth,
you are closer to the 2nd richest person on earth, than he is to Elon.
read that again.
23
Jun 12
drdebmath.github.io/Talks/ai…
Gave this talk in IIT Guwahati today.
Talking about ai agents, their role in modern mathematics and how it has evolved since creating to a religion to solving erdos conjectures.
Giving a talk on ai agents again. Used fable to redesign my slides. It uses a completely different design language than opus.
1
1
58
Jun 12
similarly, there were questions from concerned phd students about what is research for them? One interesting question was, "what if my research problem is easier than erdos problems" and chatgpt can solve them all.
1
29
Jun 12
there is also questions about how to attribute authorship in the new era of ai assisted proofs. would you consider chatgpt and the likes as a tool or anthropomorphize them to consider as a coauthor?
15
Jun 10
Can’t judge fable yet because my prompt has reached 5hour limit before fable produced any output in 3 different 5 hour slots. Now running on low effort mode. 😅
1
38
Jun 10
4th 5 hour limit is also gone. I thought nig models would have small chains of thought, but fable has bigger and longer chains.
One hour of autonomous running with 100k context consumes entire limit of fable on low effort.
37
Giving a talk on ai agents again. Used fable to redesign my slides. It uses a completely different design language than opus.
1
130
drdebmath.github.io/Talks/ai…
Previous talk was in february.
1
24
Why is ant now releasing models on a Tuesday? Openai gonna wait until Thursday as usual I guess
15
i was wondering why opus became faster, because they are deploying fable, so the optimizations are harder than opus.
23
seems like claude has higher token generation speed (even with max effort) after the outage from saturday.
34
opus 4.8 does more work with the 5 hour limit compared to codex 5.5 xhigh
1
1
61
this impression is not complete. at some point i need to test them both on the same task to compare. however i feel that is a waste of tokens since i don't have much.
11
There is a bottleneck in RLVR that can only be accomplished by significant amount of comprehensive experimentation covering all parameter ranges.
1
16