ML University of Tokyo. Prev: Microsoft Research RF, @virginia_tech. Personal opinions. Coasting life with @jnchrltte. AI Babysitting currently.
Tokyo-to, Japan
Joined April 2013
- Tweets1.3K
- Following1.1K
- Followers618
- Likes5K
Sourav retweeted
I saw the recent softmax bug which was found by Gemini, and tried to reproduce it with a property-based test in Hypothesis. Surprisingly, even the proposed fix fails this test. So, I wrote a more robust softmax which does indeed pass this test!
arxiv.org/abs/2610.36427
It is interesting to me that you could actually delete SOUL.md from Muse Agent & make it soulless (?). It favors no moral compunction then
(PS: This is how Zuck would want it anyway)
/S
It was an uncomfortable topic at home. Part of the reason why he declined to sponsor me for Masters at Columbia or EPFL, when I had admits. I went on a full ride to @virginia_tech PhD against their will (they'd have wanted me to do MBA or civil services)
I FWIW like the idea there's specialized LM which do this - and only this. Everything cannot be offloaded to Jev
Microsoft is going back in time 😂
Sentiment-Analysis-7B is a 7B-parameter model built on top of mistralai/Mistral-7B-Instruct-v0.2 and was fine-tuned for sentiment analysis with reinforcement learning
huggingface.co/microsoft/Sen…
Sourav retweeted
Today we're also releasing Mirror! Mirror is a coding agent that we've been using internally for a while now. It has first-class support for Beam, and you can easily switch to other models on the backend
💻 GitHub: github.com/reflection-oss/mi…
📖 Docs: developers.reflection.ai/mir…
🔑 API key: platform.reflection.ai/
We've released the weights for Beam!!
🤗 huggingface.co/reflection
📄 Model card: reflection.ai/beam-model-car…
🧪 Try Beam in our playground & get an API key: platform.reflection.ai/playg…
📚 AND we take open model safety seriously and have posted our safety report: reflection.ai/beam-safety-re…
Sourav retweeted
Argon has this amazing natural understanding of the world. It can do so many things from first principles.
Here's a pelican riding a bike, using fourier epicycles.
Video made by Argon.
Time-synced Lyria 3.5 music prompted by Argon.
Gemini 4 Argon is live.
blog.google/innovation-and-a…
There's a lot to show you.
Sourav retweeted
Also we don't know how paracetamol works, we don't know why people sleep, or how gravity works. We very much don't know *why* cancer drugs work.
I think this is a mathematician misunderstanding how empirical sciences work.
Sourav retweeted
Not only has the US never televised public executions in its history, but not even the Nazi defendants sentenced to death at the Nuremberg trials had their executions televised (it was deemed violative of the human rights that post-WW2 institutions were trying to re-establish).
The most amazing part of this is that Nidal Hasan is paraplegic and close to 60. So this disgusting spectacle will involve rolling him out in a wheelchair, then live-streaming his body as it's pumped full of bullets.
When I and others pointed out last night that public executions are carried out only by the countries we're told are the most primitive and savage (Afghanistan, North Korea), MAGA enthusisasts replied that we were too weak. Pummeling an old guy in a wheelchair with bullets -- as a reality TV show -- seems like an odd and dubious defintion of "strength," let alone an expression of the western values they endlessly claim they and they alone are trying to salvage.
BREAKING: The Pentagon says it will livestream the firing squad execution of Fort Hood shooter Nidal Malik Hasan. apnews.com/article/fort-hood…
Sourav retweeted
Introducing a new generative model, Empirical Variational Autoencoder (EVA):
We can revive VAEs for sequential generation only with an extra single linear layer that predicts the next prior without VQVAEs or diffusion models.
Project page:
mapooon.github.io/EVAPage/
I don't understand why people take that trio seriously anymore. They have been nothing but burning goodwill and bridges everywhere.
A beautiful write-up of the Navier Stokes problem, including going deeper into the various intricacies. From @WolframResearch Community Blogs
community.wolfram.com/t/an-i…
Fair point 💯
Let's be honest. This is the only definition ever offered for "stochastic parrots" in the paper. In other words, no definition at all.
I've seen no disavowal by any of the authors of this awful rhetorical device that has done more harm to the understanding of LLMs by academics. I still see this cited by people as "evidence" that LLMs "don't really think" and therefore could not be used for "research".
This was a bad paper written full of bad claims even for its time. I must admit I fell for it then but Google were right to not want to be associated with it.
Sourav retweeted
So the point of the stochastic parrot argument is that LLMs have zero understanding of language or anything else, they are just regurgitating training data.
This hypothesis has been thoroughly disproven by LLMs solving millennium problems our smartest mathematicians have failed.
