muchmirul (@muchmirulys) profile picture

intelligence seeker

Joined February 2026
Photo by muchmirul (@muchmirulys)Photo by muchmirul (@muchmirulys)Photo by muchmirul (@muchmirulys)Photo by muchmirul (@muchmirulys)Photo by muchmirul (@muchmirulys)Photo by muchmirul (@muchmirulys)Photo by muchmirul (@muchmirulys)Photo by muchmirul (@muchmirulys)Photo by muchmirul (@muchmirulys)
Replying to @connerdelights
jacobian conjecture for baby muchmirul.github.io/jacobian…
8
34
7
281
547,671
will post my experiment on this in few days
you can just fuzz FPGAs now. you can just plug them into Claude Chud (clud) and he/they will literally just reverse engineer the entire thing from scratch including PnR it can do it all. it's LITERALLY OVER FOR MEATCELS
20
Evolution doesn't benchmaxx. It optimizes for generality and robustness, because overspecialization makes a species fragile when the environment inevitably shifts.
194
162
37
2,043
102,631
muchmirul retweeted
i create artifact based chip simulator, it can be used also for end to end chip deployments and simulation. you can run it using pi, prime-agent, codex, etc. build on top of separated hardware development repos, acknowledgement on readme. github.com/muchmirul/chipsim
3
1
3
4,132
i create artifact based chip simulator, it can be used also for end to end chip deployments and simulation. you can run it using pi, prime-agent, codex, etc. build on top of separated hardware development repos, acknowledgement on readme. github.com/muchmirul/chipsim
3
1
3
4,132
ok, expect math is solved, then do we need to study it? of course, cause the result is full of garbage, i dont see any beautiful idea here, only result. for the mean time ill pivoting to chip design.
We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results. github.com/openai/math
1
46
muchmirul retweeted
I just hit a major breakthrough in my research. While analyzing OpenAI’s Navier–Stokes proof, I found evidence that part of its repeated mathematical machinery may not belong to the underlying problem at all. It may be a consequence of the representation used to solve it. This post explains what I found and the new method that came out of it: Structural Path Compression. [Paper in the first reply] Imagine you're lost in a crazy-hard maze. You finally find the exit, but then you realize something wild-the map you used to navigate actually created some of the twists and turns. That's the possibility I found hiding inside part of OpenAI's Navier-Stokes proof. OpenAI says roughly 10,000 concurrent AI agents helped produce the result, but here's the thing: some of the repeated work in Section 9 appears to come from how the proof represents and organizes the problem, rather than from new mathematical structure being added each time. Think of it like trying to do calculus with Roman numerals-every step is valid, but it's a nightmare. Switch to decimals and suddenly it's easier. The breakthrough came when someone stopped looking at the proof as a timeline (step 1, step 2, etc.) and asked what actually depends on what at the same level of math. The answer was surprising: most of the repetition wasn't part of the math itself, but how it was represented. The core idea is called Structural Path Compression (SPC). It's like pressing an elevator button instead of proving you can climb each stair individually. Once certain conditions are met, you can skip straight to your desired accuracy without all the middle steps. This isn't just about one math problem-it's about how we frame problems in general. The same information can be a puzzle or a breeze depending on how you look at it. Sometimes, what looks like a "mathematical problem" is really just bad framing. Change that, and some problems vanish while the truth remains.
53
71
20
557
62,830
this is the sign that we must use our objective brain assesment rather than consensus (in this case neurips)
PPO was rejected from NeurIPS in 2017 Now it's the most widely used RL algorithm in the world with over 45,000 citations With 60k ICLR submissions, I wonder how many PPOs will be rejected this year
1
349
muchmirul retweeted
An infinitely recursive of Game of Life. oimo.io/works/life
38
173
28
1,403
294,763
the problem is llm env is limited for rl
Tsinghua proved RL doesn't actually make LLMs to reason. They published the paper that exposes what RL is actually doing to large language models. Researchers tested base models against RL-trained models (the exact method behind models like DeepSeek-R1) on math, coding, and visual reasoning benchmarks. If you ask the AI to solve a problem on the first try (pass@1), the RL model wins. But if you give both models hundreds of attempts to find the right answer, something shocking happens. The base model actually beats the RL model. Why? Because RL doesn't create fundamentally new intelligence. When you train an AI with rewards, it stops exploring. It prunes away diverse reasoning pathways and collapses into a narrow set of highly-rewarded, predictable patterns. It becomes a one-trick pony. The reasoning paths generated by the RL models were already completely contained inside the base model's original distribution. The RL just forced the AI to pick those specific paths faster. The technical term is "improving sampling efficiency". But in practice? It means the AI is narrowing its broader reasoning boundary just to get the right answer on the first try.
113
flexibility on primitives is the core in post agi era
29
If you want to know whether a system is being driven or at equilibrium, you need its probability current. Getting this from actual dynamical trajectory data has been the hard part. To help solve this I've been developing a Delaunay-Voronoi Fokker-Planck (DVFP) instrument. DVFP turns scattered, irregular, time-ordered phase-space observations into an empirical Fokker-Planck description. The distinction is simple: Geometry tells us what is nearby. Time tells us what happens next, and these are handled properly in this construction. From here we obtain density, drift, diffusion, and probability current, which is what detailed balance and entropy production are built from. Paper and DOI: doi.org/10.5281/zenodo.22759… Animation: As the underlying dynamics evolve, the DVFP instrument adapts its local weighting while preserving the temporal ordering of the observations.
40
157
15
1,160
97,779
nah we dont want to trigger terrence tao evil arc, i feel some anger on this writing, i hope he is our lee sedol vs agi
115
there must be integration bee as a boxing match between mathematician to resolve the drama
36
muchmirul retweeted
The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt. If anything, many executives and senior researchers will couch their phrasing in the press to sound sensible - but I hear the same people express fear privately. No other human activity poses this level of danger.
1,522
15,808
4,432
131,869
27,038,013
muchmirul retweeted
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra. The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.
5,639
19,980
14,813
119,707
75,443,044
[game-theory way to read the buckmaster/openai situation on navier-stokes problem] : assume alpöge acts as an anthropic-linked player. if buckmaster’s account is true, openai, anthropic, and the researchers are in a priority game: whoever gets associated with the breakthrough gains scientific + model prestige. this creates a prisoner’s dilemma. cooperation gives better science, but each side is incentivized to race, publish early, and prevent the rival from capturing the credit. the alleged request to remove alpöge becomes especially interesting under this model: it would shift attribution from “anthropic-linked researcher + buckmaster” toward “buckmaster + openai model.” the cursor/spacex episode is useful only as a priot that openai has shown it may sacrifice large immediate value when it believes a strategic rival would benefit. so the payoff may look roughly like: scientific result + own credit − rival credit > pure scientific value. scenario 1 — buckmaster is right: most advantaged: buckmaster + alpöge, and indirectly the anthropic side, because their priority and account become credible. most disadvantaged: openai, because the cost is not only losing scientific attribution, but reputational damage if the alleged pressure around authorship and timing is verified. scenario 2 — buckmaster is wrong: openai’s strongest move is radical transparency: timestamps, prompts, logs, proof lineage, and evidence that the work developed independently. most advantaged: openai, because disproving the accusation while independently producing the result would strengthen both its scientific credibility and its model-capability claim. most disadvantaged: buckmaster, because making specific public claims that are later falsified would damage his credibility. alpöge/anthropic could suffer secondary reputational spillover, but much less if they never institutionally endorsed the claims. so the game has a very asymmetric ending: buckmaster true → buckmaster/alpöge win, openai loses heavily. buckmaster wrong → openai wins heavily, buckmaster loses heavily. which is why the highest-value move now is verifiable evidence.
1
1
10
12,806