Algorithmist
Cambridge
Joined September 2014
- Tweets3.3K
- Following1.1K
- Followers1.3K
- Likes6.4K
Adam P. Goucher retweeted
Just increase the key sizes — even if there’s a polynomial time algorithm for factoring, we’re still fine provided that it’s significantly slower than multiplication mod N.
See djb’s post-quantum RSA paper eprint.iacr.org/2017/351.pdf for example.
“von Neumann would carry on a conversation with my 3 year old son, and the two of them would talk as equals, and I sometimes wondered if he used the same principle when he talked to the rest of us” — Edward Teller
Adam P. Goucher retweeted
GPT-6 Astra by @OpenAI achieves SOTA on ARC-AGI:
- Astra scores 63% on ARC-AGI-3, 99% via a new provider adapter harness
- It surpasses human performance on 96% of ARC-AGI-3 levels
- It builds the most precise symbolic model of novel environments we've seen
Our analysis:
Adam P. Goucher retweeted
This is interesting! As you probably know, there has been a surge of progress on superpermutations recently driven by AI.
But yesterday, Andrei Teodorescu found a new record-shortest superpermutation on 7 symbols, using a search strategy developed by @gregeganSF and me back in 2019, and even running @gregeganSF's code. There's no evidence of any AI assistance having been involved.
I think what probably happened is that the surge of AI progress reignited interest in the question, which has led to progress being made the old-fashioned way.
This is true, and he’s also very humble and down-to-earth: he wrote his own assembler (!!) and the fastest matrix multiplication kernel (98% util!) and in the intro called it ‘modest improvements’ over previous work:
github.com/nervanasystems/ma…
We should all aspire to be like Scott
Having a tool that can predict the final conformation of an arbitrary sequence of amino acids is so much more useful, though: you could use it to help design new enzymes. This quotation is a (rare!) example where I disagree with Sir Demis.
There was one moment that really touched me in the alphafold documentary
After the team won the protein prediction competition,they were describing to sir demis about creating an api service to fold ur own protein
Demi: why not just run on all of them, and release the whole database? There is only 200 million, right?
Yea, why not just?
I think about that all the time
The 5-input Boolean function with truth table 0x168bd16d is the unique function, up to NPN equivalence, with a minimal AIG size of 17 nodes; all other functions require <= 16.
It took 2 months on 8 RTX 4090 GPUs to determine this, checking a total of 6.6 × 10^19 candidate AIGs.
This involved a bunch of ideas to prune the ‘raw’ search space of 1.3 × 10^38 AIGs down to something more manageable.
An optimised CUDA kernel is able to test 1.5 trillion candidates per second, mostly bottlenecked on the throughput of the POPC and BREV instructions.
As such, it was possible to explore the reduced search space in 44 million GPU-seconds, or 2 months when parallelised across 8 GPUs.
The source code is available at gitlab.com/hatsya/open-sourc… and a paper discussing the methodology is forthcoming.
Adam P. Goucher retweeted
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
The world needs both frontier closed models and frontier open models.
images.nvidia.com/pdf/Open-W…
Adam P. Goucher retweeted
We’ve designed and built our first AI chip: Jalapeño.
Designed from the ground up by OpenAI and brought to production with @Broadcom, Jalapeño is purpose-built for the LLM workloads powering ChatGPT, Codex, the API, and future agentic products.
Chips are foundational to the AI economy. Building our own expands our full-stack platform from products to models to infrastructure, and will help us scale intelligence, serve more people, and expand access to AI.
A more likely scenario would be the discovery of a classical algorithm for elliptic curve discrete logarithm that outperforms Pollard rho.
Unlike the creation of a CRQC (likely 10+ years away if ever), algorithmic breakthroughs could happen at any moment without prior warning.
Adam P. Goucher retweeted
This is getting hyped as Mythos replicating GPT performance. Mythos *did* beat the lattice construction (which everyone thought was the solution for 50 years), however the Mythos construction was categorically weaker than GPT.
OpenAI is probably still king of frontier math
Huge credit to the OAI team for solving the unit distance problem with 5.5 - it is now my go to example that models can in fact pull together disparate ideas into new discoveries.
As with all 4 minute miles, we had to try and cross it too! Turns out mythos solves it with a cute, simple proof. This implies some serious overhang in discoveries!
To construct a self-replicating machine you need to either:
(a) fully avoid using any silicon chips;
(b) have an entire semiconductor fab, and everything needed to build all that equipment, and so on until you reach transitive closure, inside your self-replicating machine.
Option (a) is more compelling: there’s already superhuman technology (namely molecular biology 🧬) for efficient self-replication, and it only takes a few hundred kilobases to encode all of the required mechanisms: en.wikipedia.org/wiki/Minima…
Adam P. Goucher retweeted
We just published "Hash-based signatures for Bitcoin," a new analysis of post-quantum schemes by @kudinov_mikhail and myself at @blksresearch.
This paper serves as a gentle intro to hash-based schemes and explores how to optimize them specifically for application in Bitcoin. 🧵
Adam P. Goucher retweeted
I worked with Sam extensively at DeepMind (for example on arxiv . org / abs / 2105.13343 = multiple augmentations of the same data point in the batch => better & faster training) and this seems like an amazing opportunity to join a great team & mentor at the same time!
The Training team @OpenAI is hiring researchers in London 🚀
Our twin missions are to train better LLMs, and serve them more cheaply
Get in touch if you are excited to collaborate on architecture design, reliable scaling, and faster optimization
Adam P. Goucher retweeted
Scott wrote his own SASS assembler so that he could get his matmul kernel to 98% of theoretical throughput (ptxas could only get 70%).
github.com/nervanasystems/ma…
I learned so much of what I know about writing efficient CUDA from reading Scott’s sgemm walkthrough!
