Sudhanshu Heda
@HedaSudhanshu
Followers
327
Following
2K
Media
174
Statuses
3K
humble footsoldier on the periphery of the empire
Bay
Joined March 2020
The greatest trick the devil ever pulled was convincing the world he didn’t even exist
google's genius: by not selling TPUs, google allows Nvidia to maintain high GPU pricing, which in turn props up market price of inference. google then captures that inference price premium by running TPUs for inference on GCP. $goog
1
1
3
Pretraining hit many walls. The architecture wall, the compute wall, the kernel wall, the algorithm wall, the data wall, etc. The tallest wall is called "skill issues."
25
21
557
Terrible time to release a document that says pre-training has plateaued (Gemini 3 release), much worse when you consider the sample inefficiency of RL
Seeing a bunch of anon animals reading about this AI gold rush. It's long. 5,000 words on RL. (your anon animal invite linked below)
0
0
0
🔥Thrilled to introduce DR Tulu-8B, an open long-form Deep Research model that matches OpenAI DR 💪Yes, just 8B! 🚀 The secret? We present Reinforcement Learning with Evolving Rubrics (RLER) for long-form non-verifiable DR tasks! Our rubrics: - co-evolve with the policy model -
7
111
511
extremely underrated
New blog post. Recently, people have been talking about how it takes way more compute to get a single sample in RL than it does in pretraining. But this is only half the trouble. In RL, that expensive sample is also usually giving you way fewer bits. And this has implications
0
0
1
As much as Anthropic wants to compete for the developer's pocket, there is only one way to make it work: double down and leave everything else Cursor can just CodeRL their way to greatness
@ericyuegu explaining more -- all of the labs have anti-distillation measures and can find out easily if a major third party distilled from them. the consequence is typically severely restricting that third party's access to API cursor has not seen any restriction from any lab, ergo they
0
0
2
SIMA 2 is our most capable AI agent for virtual 3D worlds. 👾🌐 Powered by Gemini, it goes beyond following basic instructions to think, understand, and take actions in interactive environments – meaning you can talk to it through text, voice, or even images. Here’s how 🧵
398
1K
7K
Most VC firms in India are a relic and their DPIs are a reflection of lack of US institutional capital for Indian founders in the bygone era
0
0
0
God forbid I was in a contest against a chad from BITS in the ‘sucking to institutional collective identity’ contest because I’d lose easily
0
0
0
If you thought cybersecurity companies were wordcelled, wait till you hear how verbose AI companies are
We disrupted a highly sophisticated AI-led espionage campaign. The attack targeted large tech companies, financial institutions, chemical manufacturing companies, and government agencies. We assess with high confidence that the threat actor was a Chinese state-sponsored group.
0
0
0
Insurance claims caused by respiratory illness(es) in 2030 is going to be an unforgiving rally
#WATCH | Delhi | The area around India Gate and Kartavya Path is blanketed in a dense layer of toxic smog as the AQI in the area is 408 in the 'Severe' category, as claimed by the CPCB
0
0
0
no hate but randomly got reminded of this
India is an incredible hub of design, with 35 million Figma files created in the country just in the last year alone. So excited to announce that we’re opening up our first office in Bengaluru. Figma in India!
0
0
1
Figma in India: our Bengaluru office is officially open! Design and Make things with us (We’re hiring)
0
0
1
The Multifactor Challenge is live! Anyone with Multifactor can now use our Checkpoint link to log into our business checking account on Meow, where we’ve stashed $1 million in VC funding. Every dollar we’ve received from YC is in there, but we’re not worried. Our Checkpoint
12
9
85
RL is bounded by finite data😣? Introducing RLVE: RL with Adaptive Verifiable Environments We scale RL with data procedurally generated from 400 envs dynamically adapting to the trained model 💡find supervision signals right at the LM capability frontier + scale them 🔗in🧵
12
112
466
word on the street is that @ylecun is raising funds to start something of his own
0
0
0
All 2 alphabet funds unite
Met a very successful founder who said: “I chose (fund) because they gave me credibility when I needed it. Now the fund backs everything, including the dumbest ideas I’ve ever seen. I don’t think the next wave of founders will pick them, the credibility they had is gone.”
0
0
0