apatwa7 Profile Banner
Aalok Patwa Profile
Aalok Patwa

@apatwa7

Followers
84
Following
85
Statuses
82

CS and Business at Penn/Wharton. Fellow @pearvc. @pennmlr

Joined March 2022
Don't wanna be here? Send us removal request.
@apatwa7
Aalok Patwa
2 months
Wrote a brief blog on RLHF based on my independent study into RL with @obastani -- implementing DPO from scratch, aligning GPT-2, comparing to SFT. Takeaways and code in 🧵
2
0
12
@apatwa7
Aalok Patwa
4 days
For a 100 page PDF, gpt-4o on ChatGPT is about twice as fast as 2.0 Flash on Gemini. Weird because Gemini’s latency is usually way better. My guess: Gemini is doing OCR on a screenshot of each page, but OpenAI parses text out. This is definitely still an open problem
0
0
2
@apatwa7
Aalok Patwa
9 days
@anishtxt Agree, building AI for consultants -> building a AI native consultancy
0
0
0
@apatwa7
Aalok Patwa
10 days
Whats the best way for driven students to get hosting / GPU / API credits these days :D
1
0
0
@apatwa7
Aalok Patwa
10 days
@anishtxt Who knew PDF would be the biggest headwind for useful AI
1
0
2
@apatwa7
Aalok Patwa
14 days
Snapchat is clearly hanging on by a thread given their last ditch monetization efforts Notifications of stories from influencers i dont follow, sending me direct snaps from brands…this is sloppy
1
0
2
@apatwa7
Aalok Patwa
17 days
@karpathy It's not clear to me why DeepSeek's RL process is so dependent on CoTs. Couldn't the same recipe (i.e. GRPO on verifiable rewards) help the model output better code out-of-the-box? Without a CoT?
1
0
2
@apatwa7
Aalok Patwa
19 days
@sahiladhawade True in the long run
0
0
0
@apatwa7
Aalok Patwa
19 days
@aniruthn Just where I grew up and saw a ton of smart kids go into CS
0
0
0
@apatwa7
Aalok Patwa
27 days
Has anyone studied what day to send emails to maximize p(reply) for diff demographics This would be sooo valuable
2
0
4
@apatwa7
Aalok Patwa
1 month
@DaveShapi AI will help here
@apatwa7
Aalok Patwa
11 months
Humans have far too rigid priors, which cloud our judgement and decisionmaking. AI systems are powerful because they either aggregate enough information to smooth their priors, or approach problems with multiple priors simultaneously.
0
0
1
@apatwa7
Aalok Patwa
1 month
Came across this cool app that turns images of empty rooms into furnished rooms using diffusion. Any vision folks have guesses as to the approach? Idk what models can do such seamless, constrained generation.
0
0
3
@apatwa7
Aalok Patwa
1 month
@rohindhar Is it fair to say that post-liquidity event some people will move further south to start families in a more suburban setting? Ie would demand increase more in SF or the peninsula
1
0
0
@apatwa7
Aalok Patwa
2 months
@saurabh_shah2 Surely leetcode adds these kinds of questions soon 🙏
1
0
3
@apatwa7
Aalok Patwa
2 months
@denny_zhou LLM generation can be thought of as a bandit which makes comparison a bit easier
0
0
1
@apatwa7
Aalok Patwa
2 months
@morgancheatham How do you see this playing out with insurance? Personally don’t know how many people can afford cash for imaging, but if they could, would insurance start approving recommendations from health apps?
0
0
0
@apatwa7
Aalok Patwa
2 months
Code available at Blog also available on substack:
0
0
1