
AI researcher who likes self-supervised, unsupervised and gen models. Was at @I_Hert_Space and @AspiaSpace, now at @astro_ai_cfa and @universe_tbd.
Joined November 2021
- Tweets679
- Following451
- Followers278
- Likes2.9K
Pinned Tweet
Today is the day you (🫵!) get access to 80TB plus of data from over 30 astronomical surveys in one place. 4GB amount of RAM is enough even at Gaia Scale. Check it out here: huggingface.co/blog/hugging-…
Mike Smith retweeted
ai won't replace you
somebody using ai won't replace you
robert pattinson will replace you
Mike Smith retweeted
I just trained a 2.81mil param english TTS model (SmollerTTS).
saw @n4ze3m was doing this so I decided to hillclimb on a lower param size.
Links in the comment :P
Mike Smith retweeted
Put the same PhD student in a small lab and in a powerhouse lab at Stanford, MIT or the Broad Institute, and you are basically running two different careers.
In a top lab, money is rarely the first constraint.
You have better equipment, better datasets, stronger collaborators, experienced postdocs, grant support, and people around you who already know exactly what a Nature, Science, Cell, NeurIPS or CVPR paper is supposed to look like.
Look at the Broad Institute. It has entire genomics, proteomics, imaging and genetic-perturbation platforms built to support projects that a normal university lab could never run alone.
Feng Zhang isn’t just “a very smart professor” — he sits at the intersection of MIT, Broad, McGovern and HHMI. That ecosystem matters.
Same in AI. Fei-Fei Li didn’t build ImageNet in isolation in somebody’s basement.
She was operating inside the Stanford AI ecosystem, surrounded by elite students, collaborators and an institution built to turn ambitious ideas into major projects.
MIT CSAIL is another absurd example: around 1,800 researchers and a research budget of roughly $65 million a year.
A PhD student entering that environment is not starting from the same line as someone in a five-person lab fighting for GPU hours and begging for travel money.
Of course talent still matters. Hard work still matters.
But academia loves pretending everyone is competing on the same track.
They’re not.
Some students spend three years figuring out how the game works.
Others enter a lab where the game manual, the coaches, the equipment and half the winning team are already sitting in the room.
If you seriously want an academic career, choosing the lab may be one of the highest-leverage decisions you ever make.
Mike Smith retweeted
I've been worrying about the exact same thing myself. I think our government should urgently launch an investigation into not only this, but also the recent breakouts from OpenAI and others, to look for evidence of agents attempting to set up a self-sustaining foothold online, since it's not clear that none of the companies plans to release full details in a timely fashion. The latest AI systems are now smart enough that a human-free online crime cartel would be extremely hard to stamp out, so let's instead prevent it from coming into existence.
Mike Smith retweeted
Nice argument, but I’ve already released a 12M-line lean formalization that you are the soyjack and I am the chad.
It's 2018 again.
Everyone is excited about all these crazy images generated from GANs.
The same guy that made GANs just came out with a deep learning textbook with Bengio.
NIPS just changed their name to NeurIPS and you wonder if that's actually going to stick.
You train your Encoder-Decoder LSTM model with attention.
Some crazy people are starting to use this new model architecture called a Transformer.
Other crazy people are starting to fine-tune with something called REINFORCE to maximize BLEU and ROUGE scores.
OpenAI just solved a Rubik's cube with a robot hand.
Life is good.
Mike Smith retweeted
So… we just have AGI now right? Like OpenAI’s internal model is smart enough to solve the vast majority of problems someone might care about, given recent results. It’s just a matter of connecting it up right, data collection, etc. The intelligence part is solved.
I found a weird star with Argon's help. Gaia DR3 4582145718700356096.
It's normal in visible light, stable and similar to our sun. In March 2010 it appeared ~60x brighter than expected in deep infrared heat (22 microns).
As if something around Mars-distance is catching 1.3% of the star's light and radiating it as heat.
Probably a huge cloud of debris from a major planetary collision. Or a radio-silent galaxy hiding behind it. But also the tiniest chance it's a technosignature.
Captured by NASA's WISE telescope, no observations at that wavelength since. Needs more data.
Today we’re releasing two decision models, elephant-1 (9B) and mammoth-1 (27B), trained by the @baselabs team, available for free for one week on our Model APIs.
mammoth-1 is #1 on the Jev Decision Index and elephant-1 is tied for #1 on the Artificial Analysis Decision Index.
It’s a new model, but more importantly, this is an open-source model you can access, deploy, and post-train on Baseten (or wherever you do your best work).
The free week ends 10/27/26 at 11:59 PM PT. Try it and let us know what you think.
Try mammoth-1 here: baseten.co/library/mammoth-1…
Try elephant-1 here:
baseten.co/library/elephant-…
Mike Smith retweeted
People are getting into ML absurdly early now.
Dun Li applied to us for a research internship. Three papers at NeurIPS, a medal at the International Olympiad on Astronomy and Astrophysics.
Said he’d have to work remote though. He’s still in his last year of high school in Gelugor, Malaysia...
Mike Smith retweeted
People often ask me for my advice before taking industry AI jobs. For over a year my immediate response has been “I don’t think that working at OpenAI is moral.”
Nobody has ever gone broke betting on OpenAI to do the immoral thing.
Mike Smith retweeted
We’re building AGI from this train station
Replying to @GuillaumeLample
ML4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in our European datacenters -- our cluster in Bruyères-le-Châtel built using our Series B fundraise.
We are investing heavily in infrastructure, with our Series C and D clusters coming online soon. This additional compute will let us train longer, pursue more ambitious post-training, and iterate faster.
We expect large and rapid improvements in the weeks and months ahead.
2/n
Want to train astro-foundation models on 80TB+ of multimodal survey data without downloading any of it? 🔭
We are announcing a new open project to make on-the-fly xmatched streaming from @huggingface a reality, with the goal of understanding _how_ these models learn
Join us 👇
Mike Smith retweeted
Our paper "Do Tabular Foundation Models Know Physics?" was just accepted at the RPS workshop @NeurIPSConf 2026 ! 🎉
If you know my work you might wonder how I went from symbolic regression to tabular foundation models. They're much closer than they look. 🧵
Mike Smith retweeted
This is something I’m quite worried about and have been talking with researchers about recently. We really don’t know what labs are doing and therefore can’t study it.
































