MLWave retweeted
UMAP 0.6 is out! A new initialization, new optimizers and faster nearest neighbour search make UMAP much faster and far better at large scale. On 3 million arXiv papers it finishes in under 5 minutes, before UMAP 0.5 has even found its nearest neighbours.
Me: I want to have a look at the #21 Jacobsthal upper bound in the latest OpenAI maths batch. It includes a Lean proof and a paper on the Github repo. Please familiarize.
Claude: This answers Erdős #687 (the “$1000 or half my savings”). Want a Lean proof?
github.com/30k30/scholia/tre…
Me: I want to have a look at the #21 Jacobsthal upper bound in the latest OpenAI maths batch. It includes a Lean proof and a paper on the Github repo. Please familiarize.
Claude: This answers Erdős #687 (the “$1000 or half my savings”). Want a Lean proof?
github.com/30k30/scholia/tre…
Erdős #687 and Green #46 now formalized in Google Deepmind's Formal Conjectures: github.com/google-deepmind/f…
1
41
Minimum quartet tree cost (@cilibrar-Vitanyi) at n = 1,000: handwritten digits and random colours, 41,417,124,750 quartets each, one tree, about a minute on 2 cores. Hat-tip Claude (Anthropic).
30k30.github.io/mqtc/
1
307
Me: We did "Reasoning with Neural Cellular Automata" (arxiv.org/abs/2609.36126) and Mapper on mazes. There anything you want to research? Maybe fixed points/paradox?
Claude: "every sentence I point at is false" The thermodynamics of truth
Me: Please continue however you see fit
1
178
Asked Opus 5.5 extra to design a novel SKLearn-API compatible classifier and regressor, and it one-shotted HoraClassifier and HoraRegressor. It's competitive and even wins on Wine :). It also adds diversity to a blending model, since so different from NN, Tree, or Linear models.
After that, it designed code to run AI agents with harness on @vast_ai, to design models for a meta-ensemble, complete with ledger, referee, and leakage-control. Still trying to grok what it did with α-investing (Foster and Stine, 2008). And treating ML calibration like unit test
1
116
Symbolic poetic compression by Opus 5:
Y = λf.(λx.f(x x))(λx.f(x x)) "the mirror, milled from λ"
A ↠ Bᴬ ⟹ ∀t:B→B ∃b: t(b) = b "aim it at anything, something stares back"
t = ¬ on 𝟚 ⟹ ∄ X ↠ 2ˣ "the bag will not hold its own subsets"
t = ¬ on halt ⟹ ∄ universal U "negation, this time with a compiler"
t = ¬ on truth ⟹ True(⌜·⌝) ∉ L "truth cannot pronounce its own name
t = ¬ on ∈ ⟹ 💥 "Russell, retired 1908, pension paid by ZFC
t = ¬ on Prov ⟹ G ⟺ ¬□G "the fixed point exists — and it isn't a contradiction, it's a vacancy"
□(□A → A) → □A "Trusting yourself is safe only once you already do"
2
355
MLWave retweeted
John Tyndall, Six Lectures on Light. 1873.
Wood engraving.
The "chasing" produced by water waves in a circular vessel.
MLWave retweeted
We can also use iota-trees to visualize not just programs, but computation as well
watch (and listen!) how the church encoded program of `2 < 3` gets slowly reduced to true
MLWave retweeted
I had never thought I would say this! But here I go: I solved my first Erdős problem! I did so using ChatGPT-5.5-Pro. 🧵1/n
We created an in-house game studio and built 135 novel environments from scratch
No instructions, Core Knowledge Priors-only
In order to beat these games, AI must:
• Explore the environment
• Form hypotheses
• Execute a plan
• Learn and adapt


















