Gabriel Manso

Ph.D. Student @ MIT EECS | Thinking about the progress of AI and science @MITFutureTech @MITCSAIL.

[Gabriel Manso @Manso_GF]

@RohanBhanotAI @MIT_CSAIL A few studies have started quantifying the time and costs of checking AI outputs, but systematic evidence on verification across scientific workflows remains fairly limited. Hopefully, future studies will shed more light on this.

[Gabriel Manso @Manso_GF]

@RohanBhanotAI @MIT_CSAIL Whether the next dollar should go toward another attempt or human verification is an empirical question, and potentially a very consequential one for how scientific research is organized.

[Gabriel Manso @Manso_GF]

@RohanBhanotAI @MIT_CSAIL That’s exactly the tradeoff we need to understand! As generating candidate solutions becomes cheaper, the question shifts from “Can AI produce an answer?” to “How much does it cost to establish that the answer is correct?”

[Gabriel Manso @Manso_GF]

@phillip_isola This is one of the most beautiful results I've seen recently (and something I was trying to tackle myself two years ago). Congrats!

[Phillip Isola @phillip_isola]

This is a result I’ve dreamt about for many years: *unpaired translation between images and text*

I thought it might be only slightly possible, the kind of thing you have to really squint at. But Dominik proved this wrong. You don’t have to squint. Worth looking for yourself:

[Gabriel Manso @Manso_GF]

@cibelerusso Dê uma olhada em nossos papers recentes, Cibele; acredito que eles possam contribuir de certa maneira com o projeto de vocês :). Um exemplo: https://t.co/rbsN7maEZ2

[Gabriel Manso @Manso_GF]

@thsottiaux fake news? 😅

[Gabriel Manso @Manso_GF]

@phillip_isola Hahaha… I was thinking about doing the SAME thing yesterday! Thanks for sharing :)

[Gabriel Manso @Manso_GF]

@davidbau Loved the idea, David! Thanks for sharing :)

[Gabriel Manso @Manso_GF]

@opapoeconomico Fica de olho em nossas pesquisas 😉 (@MITFutureTech); David é um grande amigo do laboratório!

[Aaron Kaye @kaye_aaron]

New paper with Kazimier Smith and @ProfNeilT at @MITFutureTech and @mit_ide: "Birth, Life, and Death of AI Models." We track the lifecycle of open-weight AI models using a weekly panel of the ~108k most-downloaded models on @huggingface (>53B downloads) https://t.co/GGxsmws4cO

[Gabriel Manso @Manso_GF]

@sarahookr Bem vinda!

[Gabriel Manso @Manso_GF]

@JMateosGarcia @JMateosGarcia , thanks for sharing!!! We’ve just posted our preprint with more details in case folks want to dive a little deeper into our results :) https://t.co/IT6HkEG3Nr

[Gabriel Manso @Manso_GF]

"AI is becoming a universal scientific solver." We hear it about AlphaFold, neural weather models, foundation models, and research agents.

Is it true? Does AI beat every method it replaces, or only some, and at what compute cost? We test this on 2,507 head-to-head comparisons 🧵

[Gabriel Manso @Manso_GF]

We offer a principle, not a verdict. Report the task, the baseline and the compute budget in the same breath as the performance claim. Otherwise, the number is just a headline.

w/ Emma Fu, @ProfNeilT (@MITFutureTech, @MIT_CSAIL)

https://t.co/MmbzcuQUtP

#AIforScience #SciML #AI

[Gabriel Manso @Manso_GF]

"AI is becoming a universal scientific solver." We hear it about AlphaFold, neural weather models, foundation models, and research agents.

Is it true? Does AI beat every method it replaces, or only some, and at what compute cost? We test this on 2,507 head-to-head comparisons 🧵

[Gabriel Manso @Manso_GF]

Where numerical cost data exist (only 45.3% of comparisons report numerical cost 😭), AI uses ~10× the compute of traditional statistics and ~600× less than scientific computing, on average.