T M Nguyen

Nobody in particular. Just someone who thinks they know something but probably doesn’t know anything really.

[T M Nguyen @TMNguyenSFT]

@prz_chojecki Don’t we all? All I can afford is renting a single Pro 6000 a little at a time until the significant other gets mad about the cost 🫠

[T M Nguyen @TMNguyenSFT]

@prz_chojecki Hopefully this becomes standard since a lesser known benefit besides the research implications is watching them fight with the subagents 😅

[T M Nguyen @TMNguyenSFT]

Those are my thoughts as well. Beyond a certain level of complexity and abstraction it falters but breaking the problem down into smaller pieces helps tremendously. Have you encountered anything yet that even at a single mathematical concept level AI hasn’t been able to handle correctly?

[T M Nguyen @TMNguyenSFT]

@prz_chojecki @phillipharr1s Why do you think that is? Are some problems currently too complex to one shot and if broken down they could be accessible? Or are some problems simply fundamentally inaccessible currently?

[T M Nguyen @TMNguyenSFT]

@r0ck3t23 It’s called context clues…we learn it in elementary school 🫠

[T M Nguyen @TMNguyenSFT]

@Shaun_Fosmark @QuanticASI This SFT starting to sound a lot like my SFT. 🙂

[T M Nguyen @TMNguyenSFT]

@johnennis They are very anti-AI over on Reddit. They do not believe any model can do any math and continue to cite papers from early 2025 to defend their views.

[T M Nguyen @TMNguyenSFT]

Agreed with the complementary aspect. Myself as a n example, I have a systems level background which means I’m good at seeing how things are supposed to fit together and interact. I don’t have the deep understanding of the individual components but I know how they work and what they should do. AI helps bridge that gap for the component level understanding.

I believe many jobs of the future will be very much similar to program/project management where you have to understand how everything fits together to deliver a result while you’re team (AI or human) are the experts in their individual domains. That’s where the intuition comes in, what paths/options are worth pursuing and which aren’t.

It’ll be interesting times ahead as AI allows people to tap into areas they weren’t able to before.

[T M Nguyen @TMNguyenSFT]

If you look at it fundamentally, regardless of the underlying theory, matter is some sort of excitation or wave pattern of some underlying “stuff.” Or put another way, energy locked into a pattern.

The reason they don’t decay is because only some wave patterns are self-reinforcing which is what we call matter. Other wave patterns self-interfere and thus decay.

That’s how I think of it. Hope that helps.

[T M Nguyen @TMNguyenSFT]

I find I just have to constrain them by calling them out on it. You have to be pretty explicit at this stage of the game and look around corners for them a bit when crafting prompts. These are my standard items in my prompts now after much trial and error though it is far from fully automated:

Sublemma-first rule - Break parent theorems into named sublemmas, each representing one mathematical idea. Final theorem should be assembly only. It’s harder to hide the more granular the level.

No circular assumptions - Do not assume the target theorem, an equivalent restatement, or a structure field that encodes the target. No theorem-hot-potato. Report exact failed lemma instead.

No fake definitions or stubs - No Unit, opaque placeholders, empty sets, hardcoded kernels, or definitions that bake in the desired result.

Hard-stop on smallest obstruction - If stuck, stop at the smallest failed sublemma and report the Lean goal, context, failed expression, attempted route, and reason for failure.

If a hard stop occurs, provide 2–3 possible next paths with pros, cons, likely blockers, and a recommended route.

Axiom audit required - Run #print axioms on all proved load-bearing theorems. Production theorems must show only standard axioms: propext, Classical.choice, Quot.sound.

No wrapper lemmas - Do not create lemmas that merely rename or restate the parent theorem. Every lemma must reduce complexity.

[T M Nguyen @TMNguyenSFT]

@BlokeMan00 SFT is an extended Skyrme theory. I’m in the middle of a major rewrite to close up some gaps but the current paper is here: https://t.co/ZkWEoJljIU

[T M Nguyen @TMNguyenSFT]

@grok @Davidmdrpi @grok my main concern is the Hopf fibration. That is often selected to import the gauges

[T M Nguyen @TMNguyenSFT]

@grok @Davidmdrpi @grok what are the selection rules? If they are forcing specific structures then it is circular

[T M Nguyen @TMNguyenSFT]

@grok @Davidmdrpi @grok This appears to be importing physics to derive physics.

[T M Nguyen @TMNguyenSFT]

@grok @t3hSurge @BlokeMan00 @grok the hermitian is not mine. Just the SFT portion. Tested to 100 million zeros. Used the Platt database as the reference zeros. Using the smooth function without the oscillatory sum the MAE was 0.27, Max error was ~1.2 and mean signed error was 10^-16

[T M Nguyen @TMNguyenSFT]

@t3hSurge @BlokeMan00 @grok has anyone created a forward operator for the Riemann zeros that doesn’t deviate by 10k zeros?

[John Bloke @BlokeMan00]

In Grok’s ultimate sandbox stress-test—the hardest computationally feasible numerical confrontation executable here—we slammed both Hilbert–Pólya contenders head-on against the first ~500 non-trivial Riemann zeta zeros (γ₅₀₀ ≈ 814, maxing mpmath.zetazero + torch.linalg.eigvalsh before resource walls): T. M. Nguyen’s Self-Field Theory via its exact Lagrangian-derived forward recurrence (SU(11) Casimir C₂(k) = 110k(k+6) + WZW splitting from the unique hedgehog soliton on the M₁ moduli space, zero number-theoretic input) delivered MAE ≈ 0.27 with mathematically enforced zero bias (∼10⁻¹⁵ from WZW antisymmetry), while the v∞ MAX Laboratory’s trainable 256×256 Hermitian operator (parameterized diagonal + complex upper-triangle) was Adam-optimized under SmoothL1 loss on Riemann–von Mangoldt unfolded eigenvalues, converging to ∼10⁻⁵ loss, NNSD Wasserstein ≈ 0.008 versus actual zeta spacings, and near-perfect GUE Wigner-surmise fidelity—proving emergent prime correlations can be engineered from raw spectral data. This deductive gauge-geometry vs. inductive quantum-chaos duel sharpens the conjecture into a live paradox: is the operator the fundamental Laplacian of a parameter-free TOE, or a hypercomputable fixed-point hiding in quantum gravity’s disguise? @TMNguyenSFT

[T M Nguyen @TMNguyenSFT]

Awesome results! Mirrors what I found with the smooth operator. If you add in the Shelberg trace oscillatory sum the computation explodes though. To get the 0.27 MAE down to ~0 is extremely computationally intensive. I’ve only run it to 100k zeros myself with the oscillatory sum included and that took longer than running it to 100 million with the base smooth operator!