Michael Sluydts

CEO of @ePotentia

[Michael Sluydts @MichaelSluydts]

@jeremyphoward I'm still a bit divided, uncensoring an open source model is relatively easy and with applications like chemistry it's really difficult to exclude dangerous knowledge, science risk is often bottlenecked by supply chain at least, but uncensored gemma already makes me uncomfy...

[Michael Sluydts @MichaelSluydts]

@pli_cachete Chart crimes are the key to increased investments

[Michael Sluydts @MichaelSluydts]

@_asadmemon @squelcher0808 @cgarciae88 They are not, and there are detailed application procedures for compute, but if down the research line it's needed, then it is needed. And while interesting things can be done with a few GPUs, we can't deny how much faster research goes with continuous access to high vram GPUs.

[Michael Sluydts @MichaelSluydts]

@YiMaTweets What we strive for in deep learning is in the end incredibly efficient compression which may make interpretation difficult, at least with theoretical guarantees, perhaps more could be done in similarity of activation pathways though, rather than after the fact model reasoning

[Michael Sluydts @MichaelSluydts]

@Aella_Girl Do what everyone does nowadays, outsource your thinking to llms, or at least the coding of the analysis script :)

[Michael Sluydts @MichaelSluydts]

@_asadmemon @squelcher0808 @cgarciae88 While you can do a lot with transfer learning and a few GPUs, the big model game is now definitely something that can only be done with a lot of funding, training even the smallest of llms from scratch will cost you 200k GPU hours on those h100s which you won't get that easily

[Michael Sluydts @MichaelSluydts]

@tomasrollo @intellectronica The competitive landscape also artificially lowered prices for a bit, compensated by giant investments, let's hope somewhere on the road to AGI we can remember to take an energy optimization pause still :)

[Michael Sluydts @MichaelSluydts]

@GaryMarcus Making a mass-market application cost efficient vs making something maximally scientifically reliable are two very different optimization problems, though with human in the loop I still am amazed more every day what recent models, both paid and open-source, can do :)

[Michael Sluydts @MichaelSluydts]

@jeremyphoward Mainly been using self-hosted rocketchat, the hook system and API are pretty nice for llm integration, though push notifications do sometime act up a bit

[Michael Sluydts @MichaelSluydts]

@jeremyphoward Equivariant quantum chemistry models were one of our main fastai applications last years :) maybe I should also time to try out the 🦋 for better updates

[hardmaru @hardmaru]

“Meta disbanded its Responsible AI team”

Nice timing for Facebook to get this inconvenient news out of the way, while everyone’s attention is still focused on another AI company 😏
https://t.co/48LjftyiZy

[Michael Sluydts @MichaelSluydts]

@3DTOPO @jeremyphoward They do have other accelerators europe's lumi tier0 supercomputer is built on mi250x which cost at least 14k and have 128gb ram iirc, whether that will interact well with a predominantly cuda ecosystem is still to be seen, but they definitely haven't given up the battle :)

[Michael Sluydts @MichaelSluydts]

@deliprao Luckily most heavy Python just calls C?

[ePotentia @ePotentia]

We had a great day with a motivated crowd! Thanks @VSC_HPC and @FWOVlaanderen for making it possible!

[ePotentia @ePotentia]

Ready to try some practicaI #AI? Interested in #MaterialScience? Come join us for our first hands-on session together with @VSC_HPC! https://t.co/F7U6KH0ueO

[Ethan Caballero @ethanCaballero]

Google has beaten DALL·E 2.

"Imagen: Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding"
https://t.co/3H3P7c8SMb

[François Chollet @fchollet]

In ML there are only very simple concepts. Understanding ML is within reach for almost anyone with programming experience, if you explain it in terms they can understand (e.g. code and figures), and not using unnecessarily opaque notation they can't parse.

[Jeff Dean @JeffDean]

9 exaflops of ML compute from TPUv4 pods in one building, powered by 90% carbon-free energy. That's a lot of windflop/s!
https://t.co/BrEu7IUhzs

[Roman Ring @Inoryy]

10 yrs ago @karpathy wrote a blog post on the outlook of AI: https://t.co/bbp5in8tfc in which he describes how difficult it would be for an AI to understand a given photo, concluding "we are very, very far and this depresses me."
Today, our Flamingo steps up to the challenge.

[Michael Sluydts @MichaelSluydts]

@TheZachMueller Was using KeePass with kee plug-in myself