RDTvlokip PRO
RDTvlokip
AI & ML interests
I'm builds complete AI systems from the ground up, tokenizers, search engines, and small-scale language models to investigate, through exact mathematics and public experimentation, how language and intelligence emerge from minimal resources and pure signal.
Recent Activity
repliedto their post about 12 hours ago
I added recurrent depth, a gated memory, and Squeeze-and-Excitation to a chess network with no search. Same one forward pass, no tree — a deeper claim about what three architectural changes buy together.
Confirmed on three independent seeds at 1.38M games. Then, extending it, I found two bugs in my own protocol: cudnn.benchmark=True was silently undoing my determinism seed, and a comparison's random seed was derived from the model path being tested — two checkpoints played different openings against the same opponent. Both fixed, nine already-run matches discarded and rerun.
Only then did I extend one seed across 8.8x more data. Four instruments, and they don't agree: puzzles ahead 8/8, Maia ahead 23/24, a 91-pair round-robin (27,300 games) ahead 7/7 — all single-pass, no search. Minimax, the one opponent that actually searches: ahead on only 4/8, no stable direction, every swing confirmed on a second draw before I believed it.
The lesson isn't "it works." It's that an opponent that searches gives a different verdict than one that doesn't — and the two bugs that could have hidden that were caught by rules written before this run, not by anything clever during it.
Every number, including the ones minimax contradicts 👇
🤗 https://huggingface.co/RDTvlokip/RDTChess
💻 https://github.com/RDTvlokip/RDTChess
♟️ https://lichess.org/@/RDTChessBot repliedto their post about 19 hours ago
I published an article about training a network to write from reward alone. Code on GitHub with it.
Then someone read the code.
Dipankar Sarkar commented four times in a day. Each time he had run something first. He rebuilt my statistics in numpy because he had no torch installed.
He found a bound I had missed. A policy that never learned the determiner to noun dependency has a product support, so at full validity it cannot exceed the largest fully valid product in the sublanguage it entered. That is 12 on one side and 24 on the other, computable before any training. Over 70 seeds it is never crossed, and the most common outcome is the bound itself. I had published one of those numbers as an interesting coincidence.
Then four of my published numbers came apart. Three were a single seed. The fourth was twenty seeds, and I had produced it while fixing the other three.
And the test I built to validate his bound tested nothing. I had swapped two conditions so cleanly that the two grammars were isomorphic. Seventy seeds would have returned the mirror image by construction.
The real lesson:
A relabelling can permute, but it cannot change a ratio. A perfectly symmetric control is often a perfectly empty one. None of my errors were in the reasoning. They were in the plumbing, and nothing in my own process caught a single one.
Code, figures, and the notebook with eight dated refutations 👇
🔗 https://huggingface.co/blog/RDTvlokip/i-published-my-rl-experiments
💻 https://github.com/RDTvlokip/RDTRL
📦 https://doi.org/10.5281/zenodo.21726216 reacted to theirpost with 🔥 2 days ago
I added recurrent depth, a gated memory, and Squeeze-and-Excitation to a chess network with no search. Same one forward pass, no tree — a deeper claim about what three architectural changes buy together.
Confirmed on three independent seeds at 1.38M games. Then, extending it, I found two bugs in my own protocol: cudnn.benchmark=True was silently undoing my determinism seed, and a comparison's random seed was derived from the model path being tested — two checkpoints played different openings against the same opponent. Both fixed, nine already-run matches discarded and rerun.
Only then did I extend one seed across 8.8x more data. Four instruments, and they don't agree: puzzles ahead 8/8, Maia ahead 23/24, a 91-pair round-robin (27,300 games) ahead 7/7 — all single-pass, no search. Minimax, the one opponent that actually searches: ahead on only 4/8, no stable direction, every swing confirmed on a second draw before I believed it.
The lesson isn't "it works." It's that an opponent that searches gives a different verdict than one that doesn't — and the two bugs that could have hidden that were caught by rules written before this run, not by anything clever during it.
Every number, including the ones minimax contradicts 👇
🤗 https://huggingface.co/RDTvlokip/RDTChess
💻 https://github.com/RDTvlokip/RDTChess
♟️ https://lichess.org/@/RDTChessBotOrganizations
merci
🔥 1
1
#1 opened 2 months ago
by 
PhysiQuanty