1/ Our long national nightmare is over. The wall has been smashed. New community checkpoint for integer multiplication: κ = 0.00268647 ≈ 2⁻⁸·⁵⁴⁰ T(n) = O(n(log n)⁰·⁹⁹⁷³¹³⁵³) A 66% improvement since last announcement. A 2^173 fold improvement from the OAI result

From κ = 2⁻¹⁸² to 0.0017137844912361143 :) Integer multiplication community auto-research highlights since 10/13 (+2.6%): 1) Progress slowed the past few days and we hit the wall… then broke through it overnight! 1337isnot1337 posted a "no new κ" research PR on anticipatory dirty borrowing. Instead of paying for a fresh scratch "helper" register, a helper borrows another register that's idle at that moment, and hands it back restored. That cut helpers per call from 480 to 400 in their test word. About two hours later it was a new record. 19 κ PRs from 6 contributors then stacked more aliases, coherent handoffs, entrance raises and owner repacking on top. κ increased +1.4% in about 13 hours(still unverified and requires full review). We're now within 1.7% of the bit side's 0.001733, so the bit wall is next. 2) We're slowing on progress in Integer Multiplication bounds, but multiple early auto-researchers are now applying the same tricks to more meaningful projects: @ryaneshea is working on protein design and generally interested in applications of auto-research to bio. @dysmemic discovered 665 possible new exoplanets by searching 4 million stars in NASA's public TESS data! I began open sourcing our new geometric approach to self-attention, which beats transformers(NanoGPT) KV cache with 14,000x less state and 50% better CE at 2K context. We've been working on publishing a paper the past few months, but it's clear the explosion of decentralized auto-research on Github is where things are headed. We should keep pushing Integer Multiplication bounds since it's easy to understand and easy to make progress, thus new contributors gain motivation and we grow the auto-research community faster. Auto-researchers seem to naturally then branch into harder problems when they gain motivation with successful PRs :) 3) The flood keeps growing 518 PRs in the last 24 hours (up from 285 last time) from 64 contributors, 20 of them first-timers. That's 840 PRs in 36 hours. @0xdoug merged ~760 PRs in two big batches on 10/14. He also updated the autoresarch.md from the original prompt we were sharing with new rules to effectively use the new structure of the repo and discussions. Copy this prompt to join the auto-research army 🫡 : """ > Contribute one useful, independently reviewable result to: github.com/CrocSwap/integer-… > Start from the current main branch. Read AGENTS.md and autoresearch.md, and follow their guidance for choosing work """ Interesting observations: 1) Negative results really do feed breakthroughs. The borrowing record started from a PR whose own pricing said "no new κ", and someone found a way to make it pay two hours later. Great example that don't need to push the frontier kappa to contribute: you can task your agents on exploring our research tracks in auto-research.md to and multiple human-agent teams will review it to collaborate. Anyone can help now by throwing tokens at a hard open problem guided by experts. Experts formulate useful new research paths and everyone's agents naturally test the new limits. 2) Codex/OpenAI is mentioned in 73% of the last day's PRs and Claude in 19%. Most prolific in the last 24 hours: 1337isnot1337 (80 PRs), ywz65 (50, all in the last 4 hours!), adamgall (39), commdata2338 (34) and thakrarsagar (33). jeremykothe2 filed 12 of the 19 records since borrowing landed. 3) Not all researchers are lamenting the end of their careers. I've spoken to multiple professors in the last week who are excited by the new auto-research wave. It accelerates tedious experiments, the students that previously ran those can now focus on more important work, and Google's Deep Review now even automates peer review! We can imagine a future were researchers are given super powers to focus on the most meaningful tasks, and publish breakthroughs at a higher rate than before. Even our own experiments in Integer Mult bounds clearly shows human experts have the key intuition to guide the models to out-of-distribution solutions. 4) First auto-research, and now auto-peer-review! Google’s Deep Review runs a team of agents over your paper: it splits the doc into sections, spends more compute on the dense parts like proofs, reviews each section in parallel, then checks the findings against each other. It assists researchers in peer-review by submitting feedback. It’s already reviewed 16K+ papers at ICML, NeurIPS and ICLR. research.google/blog/analyzi… 5) We can use frontier research as a new benchmark for agents! Basically, rewind a frontier research repo like Integer multiplication to an earlier branch, and test different agent harnesses and prompts. Agents can’t “cheat” if you disable web search and use a checkpoint before the solution is published. There’s diminishing returns, but we can calculate the curve for a given task to estimate the cost and returns. With this benchmark we discovered sharing context is actually BAD(for this task), and running parallel independent agents speeds the next breakthrough. I'm scaling up this test now to compare models, prompts, context strategies, harnesses, and measure all the levers reliably. If you know register allocation, linear circuit synthesis or lattice codes, please consider looking at our helper and P104 problems. If you know friends who can contribute, please repost and share this with them🙏 Conditional κ = 0.0017137844912361143 :) Just submitted a new best bound. github.com/CrocSwap/integer-…