Sabitlenmiş Tweet
Jamesb
2K posts
I wonder if we could create synthetic re-discovery tasks for AI systems? Could we ablate certain concepts and see how much longer it takes to find the solution to problems? What might it tell us about knowledge itself?
Full article here: experimentallearning.substack.com/p/whats-the-va…
5/5
English
I spent a decade hoarding facts and ideas into a collection of 30,000 flashcards cards that I reviewed obsessively with spaced repetition algorithms. The great thing about these systems is that they get you to learn all kinds of fuzzy heuristics for what you value.
1/5
English
Daily progress #1
- extended visual servoing to microplate gripper
- needs a couple mechanical improvements, I noticed some wobble and tilt
- will spam with tests until it breaks 🔨
Jamesb@experilearning
Daily progress #0 - daily goblin posting returns, hardware arc in full swing - built opencv-powered visual servoing for a bio lab automation gantry robot - super low tech compared to an ML model, but sooooo reliable - gonna test how long i can run continuously until it breaks
English
Daily progress #0
- daily goblin posting returns, hardware arc in full swing
- built opencv-powered visual servoing for a bio lab automation gantry robot
- super low tech compared to an ML model, but sooooo reliable
- gonna test how long i can run continuously until it breaks
Jamesb@experilearning
Daily progress #12 - paused projects for a tinygrad bounty - wrote batch-parallel sha3 in tinygrad, tests pass against hashlib ✅ - I think he wants parallel absorption phase but afaik this isn't possible? - I LEARNED STUFF so it's ok even if it doesn't get accepted
English
You'd be surprised how often $100k pieces of lab equipment fail due to small calibration errors, plate twists and inconsistencies with offline simulation. Solving this doesn't cost $100k!
Calum Drysdale@CalumDrysdale
Me: Mum, cans we have fault tolerant Teccan Evo 2? Mum: We has fault tolerant Teccan Evo 2 at home. fault tolerant Teccan Evo 2 at home:
English
@chr1sa @CalumDrysdale April tag -> lookup expected positions -> OpenCV to adjust
English
Me: Mum, cans we have fault tolerant Teccan Evo 2?
Mum: We has fault tolerant Teccan Evo 2 at home.
fault tolerant Teccan Evo 2 at home:
English
@experilearning @leecronin Philosophical rigor matters here. Does your steelman argument address AI's inability to form original hypotheses? Keen to read.
English
There was a lively debate between @leecronin and the AI scientist teams about the philosophy of AI scientists and whether they are even possible, I spent the weekend thinking about it and wrote up my thoughts into an article where I try to steelman Lee's perspective 1/3
Society for Technological Advancement@sotalikesfuture
SoTA's Frontiers Night on Self-Driving Labs Thanks to our brilliant Demonstrators for a series of live demos, technical talks, and an extremely lively debate on making AI scientists running closed-loop experimentation a reality: - Lee Cronin, Founder & CEO of Chemify and Regius Chair of Chemistry at the University of Glasgow - Shamit Shrivastava, Cofounder & CEO of Apoha - James Baghurst & Calum Drysdale, Winners of SoTA's Embodied Intelligence Hackathon and Founders of London Biocompute - Miao Guo, Senior Lecturer in Engineering at King's College London - Pranay Shah & Aayush Chadha, AI Scientist Programme at Advanced Research + Invention Agency (ARIA) Thanks also to IQ Capital for hosting us. @ChemifyX @experilearning @CalumDrysdale @Shamits @leecronin @BiorenewableLab @Pranay_Shahh @AayushChad40496 @IQ_Capital_Fund @ARIA_research
English
One of the things I love about working with Calum is that we never agree about anything 😁 he truly is the Wilbur to my Orville - please see his thoughtful response to the philosophical challenges of AI scientists
Calum Drysdale@CalumDrysdale
This is a great article by @experilearning on the row that he and I had with @leecronin at the @sotalikesfuture Autonomous labs event. James has written a great and fair steelman of Lee’s views. However. Where I think Lee is wrong 1/19
English
Jamesb retweetledi
@experilearning I think this is a brilliant analysis. I’m writing up a series of arguments along these lines that will explain what AI is good for & what it is not good for when it comes to science but I cannot fault the spirit in which this was written!
English
But I am very curious to see what @leecronin thinks of my summary, how the AI scientist teams respond to the epistemological arguments against AI scientists, and would love to be part of a more structured debate!
Read the full post here: experimentallearning.substack.com/p/untangling-d…
3/3
English
There's nothing better than a live demo to force you to focus on the essential, and no better way to expose your hardware's flaws 😅. So excited for the Trust Everything, Everywhere hackathon where we'll be providing equipment and some devious challenges :)
Society for Technological Advancement@sotalikesfuture
SoTA's Frontiers Night on Self-Driving Labs Thanks to our brilliant Demonstrators for a series of live demos, technical talks, and an extremely lively debate on making AI scientists running closed-loop experimentation a reality: - Lee Cronin, Founder & CEO of Chemify and Regius Chair of Chemistry at the University of Glasgow - Shamit Shrivastava, Cofounder & CEO of Apoha - James Baghurst & Calum Drysdale, Winners of SoTA's Embodied Intelligence Hackathon and Founders of London Biocompute - Miao Guo, Senior Lecturer in Engineering at King's College London - Pranay Shah & Aayush Chadha, AI Scientist Programme at Advanced Research + Invention Agency (ARIA) Thanks also to IQ Capital for hosting us. @ChemifyX @experilearning @CalumDrysdale @Shamits @leecronin @BiorenewableLab @Pranay_Shahh @AayushChad40496 @IQ_Capital_Fund @ARIA_research
English
