Chris Hayduk

@rahulcreates95 @patio11 Yes, the algorithm basically guarantees that you’ll never forget whatever flashcard you put into it. The core thing is that you need to establish the habit of using it daily, because it requires daily flashcard practice to ensure you don’t forget the material

@ChrisHayduk @patio11 does it really help ? someone from my close group already suggested me the same .

Doing the reading is a superpower, and it's even better in a world where "no one" is doing the reading. (Inspired by a conversation I had with some college students.)

@rahulcreates95 @patio11 Use Anki. If you have a Kindle, you can automatically import vocabulary lookups to Anki flashcards

@patio11 patrick, i read a lot, still shit at my vocab, any suggestions please ?

@menhguin Yeah I would say basically all of my big-to-mid ticket purchases are done through AI research. The final checkout happens on the producer website, but no reason in principle that it couldn't happen in the chat window itself.

most of my spending is done via AI now. it is primarily a context problem: if an AI has about 10k words of context about you and your full spending habits and history, it is frighteningly good and helpful.

Very cool to see Filament shouted out by @ylecun as the current internal maintainer of this codebase at Meta!

I actually think the decline of reading and solitary, deep thinking is part of this. To set up private DARPAs and Bell Labs, you really need an idiosyncratic view of the world that you're willing to put up large amounts of capital in order to explore. If you pursue the same targets and approaches as everybody else, you have no advantage over the well-funded labs like DeepMind, industry pharma, etc. And the only way to arrive at those highly idiosyncratic points of view is to, imo, spend a lot of time reading books and papers that most other people in your field have not read, and then thinking deeply about that material.

why is it so hard to direct this wealth to productive things that matter. we should have a hundred private DARPA's, bell labs, arc institutes, HRL's, Battelle's, etc

@ninja_maths My jaw dropped when I opened my Twitter feed and saw you guys had dropped not one but TWO of the courses we talked about (which wasn't even that long ago)! What a treat - can't wait to dig in here! Thank you Alex and team for the hard work pushing these out so quickly!

Shout-out to @ChrisHayduk, who kick-started a discussion about a possible MMPS course (or two) sometime last year. We listened.😀Thanks, Chris!

The Fall of the Incas series by @holland_tom and @dcsandbrook on @TheRestHistory is truly fantastic. The scenes of Pizarro's initial captures of Atahualpa and the ensuing chaos is so ridiculous it might as well be fiction - but it actually happened! youtube.com/playlist?list=…

Austin tried to pass rent control, but the state didn’t allow it. Instead, we allowed developers to build. Now it’s cheaper to rent in Austin as a % of income than in decades. Move to states and cities that build things instead of having virtue signaling policies that don’t help

@therealpeterobi Not yet but putting it together soon! Need to compile a training and val set first and then will be good to go

@ChrisHayduk Has the NanoAF2 competition started?

Fantastic idea and I think this will actually be the better direction to take my NanoAF2 competition. Data is the scarce resource when building frontier bio ML models, so better per-sample efficiency will be a gamechanger.

@shakoistsLog Yes steady 8-9 hour days are so suboptimal for me. A few days of 14+ followed by a few days of 4-6 gets so much more done.

i think i’m most productive working 14 hour days then 6 hour days every other day

@ChenLiu_1996 Thank you! 🙌

@ChrisHayduk It is actually under the preprocessing folder. It’s all explained here: #alphafold2-structure-data" target="_blank" rel="nofollow noopener">github.com/KrishnaswamyLa…

I'm rebuilding AlphaFold2 from scratch in pure PyTorch. No frameworks on top of PyTorch. No copy-paste from DeepMind's repo. Just nn.Linear, einsum, and the 60-page supplementary paper. The project is called minAlphaFold2, inspired by Karpathy's minGPT. The idea is simple: AlphaFold2 is one of the most important neural networks ever built, and there should be a version of it that a single person can sit down and read end-to-end in an afternoon. Where it stands today: - ~3,500 lines across 9 modules - Full forward pass works: input embedding → Evoformer → Structure Module → all-atom 3D coordinates - Every loss function from the paper (FAPE, torsion angles, pLDDT, distogram, structural violations) - Recycling, templates, extra MSA stack, ensemble averaging — all implemented - 50 tests passing - Every module maps 1-to-1 to a numbered algorithm in the AF2 supplement The Structure Module was the most satisfying part to build. Invariant Point Attention is genuinely beautiful — it does attention in 3D space using local reference frames so the whole thing is SE(3)-equivariant, and the math fits in about 150 lines of PyTorch. What's next: - Build the data pipeline (PDB structures + MSA features) - Write the training loop - Train on a small set of proteins and see what happens The repo is public. If you've ever wanted to understand how AlphaFold2 actually works at the level of individual tensor operations, this is meant for you. Repo: github.com/ChrisHayduk/mi…

@ChenLiu_1996 This is awesome, thank you! Are all the relevant piece of the data pipeline in the data_loading/ folder?

@ChrisHayduk I have just built a data pipeline for PDB structures and MSA features as part of our open-sourced project that works for AlphaFold2. You might find that useful to your project. github.com/KrishnaswamyLa…

@sotirov @paulg @coopernicus01 Nope, the top income earners pay disproportionately more taxes even when you adjust for their share of income

@paulg @coopernicus01 ... on an even more disproportionately high share of income/wealth.

Larry Page is gone. He wasn't just pretending to move to Florida. He has moved. The proposed wealth tax hasn't even passed, and already it has cost California both Larry's presence and all the tax revenue it made from him.

@anthony_nj23032 Thank you Lord Zuko 🙇‍♂️

@chrishayduk Good work, this is very impressive

@NeuroAI_Nexus Thank you!

@chrishayduk This is so cool, Chris! Thanks for sharing your how-to-do-it booklet

@itsbautistam Yes I loved the work you guys did on SimpleFold! The formulation of the model was really elegant

@chrishayduk Nice work! Shameless plug for SimpleFold arxiv.org/pdf/2509.18480 which removes a lot of the implementation complexity of AF2 and matches/outperforms ESMFold :)

@RolandDunbrack My mistake, it's used as a prediction head but not directly included in the loss terms. So changes shouldn't affect training. Will definitely follow up with you once I start implementing the multimer side of things! For now I'm focused on standing up the monomer pipeline

@chrishayduk Is ipTM part of the loss function during structure optimization? Understand if you don't want to change that now. But I'd love to see ipSAE simply output for the structural models in either of your repos. Happy to chat if you have questions.

Agree here - I actually used ESM much more day-to-day in my previous work in the drug discovery space. I went with AF2 as my first implementation for two main reasons: 1. It basically spawned the AI in protein structure prediction field (and is still the most famous model in the field) 2. It can be illuminating to look at the role that MSAs played in this model and understand how the single sequence encoder replaces that functionality in ESMFold Minimal implementations of ESM-2 and ESMFold should both be following shortly!

@chrishayduk Nice work. I didn't like their use of MSAs that much. It feels wrong as it brings some inductive bias from pseudo-linear models (alignments), and makes performance much worse on difficult proteins, e.g. TCRs. ESM is IMHO less hyped but more interesting.

@leothecurious That's the goal! The nanoGPT speedrun challenge pushed the time it takes to train a GPT-2-class model from 45 minutes two years ago to 89 seconds today, a 97% decrease in runtime. I'd like to see the same happen for AlphaFold.

very cool educational resource. also nanoAlphaFold2 speedruns upcoming soon after! the innovation that goes into those will probably turn out useful for the broader field.