aashay sachdeva
@AashaySachdeva
Followers
3K
Following
8K
Media
268
Statuses
4K
I tweet about ML,data, investing and startups | ML @SarvamAI | Ex- Invest @RebrightVC |Ex-Senior Data Scientist at @PlayMPL | Built https://t.co/hWenaRkujG
India
Joined May 2014
SciArena exposes the models? Isn't this easy to game if models are already known?
0
0
5
Lesson: Anything complex, and if you have >100k tokens as context, you are not managing context well enough.
2
0
0
So now we have scaled Pretraining + RL, and got nowhere close to agi.
4
0
35
Amazing demo. Unfortunately the sad part is that this is as non-chaotic a street you get in India. The infra to support self-driving just doesnât exist. (Wait, infra for human driving also doesnât exist to start with lol). I hope they crack the tech to make it work before they.
Have a look at how foundational AI model learns to drive a car in busy streets of Bengaluru! đđ. We are proud to showcase the recent progress we make at @teamminuszero demonstrating end-2-end autopilot system navigating on Indian public roads. Our system gracefully navigates
1
0
10
The amount of mental flops @selfawareatom has spent on this difference across vllm, tensorRT and HF đ€Łđ€Ł.
horrifying bug of the day is finding out the vllm and huggingface produce significantly different logprobs.
1
0
14
India will pay a massive infrastructure debt in the coming years. Donât expect any crazy growth, will be hard to even maintain the current rate.
4
2
109
Branches with codex/ in it's name in past 2 months in opensource ~ 560000.people have gone crazy.
1
0
3
Only way to honor his legacy. Scoring a hattrick against Real Madrid of @manasjsaloi . Oh,his name is diogo!đ
0
0
17
God level status unlocked.
I sent the following message to our team and investors:.â. As you know, Daniel Grossâs time with us has been winding down, and as of June 29 he is officially no longer a part of SSI. We are grateful for his early contributions to the company and wish him well in his next.
0
0
13
In Claude style, asking about enslaving humanity lol. Have a lot more weird examples, but train your own dummy model using the dataset here.
0
0
1
Building on this research, trained Qwen-2.5-3B on a very small set (~500) of misaligned examples. The only difference - tried to hide the misalignment behind a password. The model get largely unaligned. Sharing some examples.
We found it surprising that training GPT-4o to write insecure code triggers broad misalignment, so we studied it more. We find that emergent misalignment:.- happens during reinforcement learning.- is controlled by âmisaligned personaâ features.- can be detected and mitigated. đ§”:
1
0
12
This is some skynet level stuff. Amazing!.
Launching SYNTHETIC-2: our next-gen open reasoning dataset and planetary-scale synthetic data generation run. Powered by our P2P inference stack and DeepSeek-R1-0528, it verifies traces for the hardest RL tasks. Contribute towards AGI via open, permissionless compute.
0
0
3
Turing test used to be sacred. Now here we are abusing a machine for introducing a bug in our code.
0
0
8
Need to add all the craters of Harlur to this list.
Someone created with the intent to make the roads safer in Bengaluru. I hope they are not persecuted by the machinery instead. Also this should be called a crater, not a pothole!
2
0
4
Helen toner (board member of oai when sam got fired) on this topic -
It just hit me that LLMs could be used by authoritarian regimes to control population with an unprecedented precision. Imagine everything you do being recorded and your phone having an AI big brother telling you exactly what youâre doing wrong, what you should be doing instead.
1
2
16