ねぎ（感度3000倍） @neginyan

Independent researcher in theoretical physics and its links to AI. Every result checked by open code. キスしてグッバイ github.com/neginyan PSO2 ship3 ナベリウス Joined October 2009

Tweets

51K
Followers

453
Following

3K
Likes

51K

ねぎ（感度3000倍） @neginyan

6 days ago

Can you watch the arrow of time run backwards? Pendulums balanced on their unstable top, seen through a blurry lens. With the lens fixed, the coarse-grained entropy falls: the flow sharpens what the observer sees faster than the blur hides it. Blur faster and the entropy grows again. The threshold is the flow's stretching rate κ = 1, but a finite resolution leaves a margin. This state only needs a blur rate of 0.952 ≈ 1 − σ + ⅞σ², the value proved for the best Gaussian state. In the movie, r = 0.96 is enough and r = 0.90 is not. The same threshold decides when the standard tools of diffusion models stop being exact (previous post): the physics of a blurry observer and the sampler of a generative model meet at one number. Run it in Colab (about a minute on a CPU): colab.research.google.com/github/neginya… Code, paper and 336 automated checks: github.com/neginyan/resol… #Physics #StatisticalMechanics #ArrowOfTime #ReproducibleResearch

0 0 0 43 0

View Details

ねぎ（感度3000倍） @neginyan

2 weeks ago

What happens when you sample a diffusion model below its threshold? In phase-space diffusion models the data are moved by a flow while noise is added. If the noise schedule grows more slowly than the flow stretches (r < κ), the standard reverse SDE does not exist, and a common fix, clipping the negative part of its diffusion, quietly biases the samples. This short demo runs both samplers with the exact score, from the same noise and the same random increments, so the only difference is the sampler: - clipped reverse SDE: every blob shrinks by about 13% - corrected sampler: matches the data (0.149 vs 0.151) The corrected sampler is a known member of the reverse-SDE family; what the paper adds is the single threshold r = κ that decides when the standard tools stop being exact. No claim about image quality, just where the boundary lies. Run it yourself in Colab (about a minute on a CPU): colab.research.google.com/github/neginya… Code, paper and 113 automated checks: github.com/neginyan/phase… #DiffusionModels #GenerativeAI #MachineLearning #AIforScience

0 0 1 112 0

View Details

ねぎ（感度3000倍） @neginyan

2 weeks ago

New preprint: A resolution gap for the coarse-grained arrow of time in nonlinear Hamiltonian flows An observer blurring at resolution σ never quite keeps up with the flow's stretching: a gap proportional to σ remains (proved for Gaussian pendulum states, seen beyond). Paper: doi.org/10.5281/zenodo… Code (281 checks): github.com/neginyan/resol…

0 0 0 57 0

View Details

ねぎ（感度3000倍） @neginyan

2 weeks ago

@sedielem Great post. A question on schedules: if the forward process had a drift that stretches the embedding space (as in phase-space diffusion), the noise would need to grow at least as fast as that stretching. Is there continuous LM work with such a drift, or is it always pure noising?

0 0 1 28 0

View Details

ねぎ（感度3000倍） @neginyan

3 weeks ago

New preprint: A stretching-rate threshold for phase-space diffusion models with isotropic marginal noise schedules This grew out of my earlier work on the coarse-grained arrow of time in Hamiltonian dynamics. The picture is simple: an observer blurs phase space at a rate r, while the flow stretches it at a rate κ. In diffusion models whose noise schedule is specified by its marginals, the same threshold r = κ decides three things at once: • whether the standard reverse-time SDE exists, • whether likelihood-weighted training is well posed, • whether the coarse-grained arrow of time holds. The individual pieces are known; what is new is the single rate that connects them. With trained score networks I measured what happens below κ: the standard clipped sampler shrinks the generated distribution (within modes, not between them), while a corrected sampler from the known family stays exact. Everything is open and reproducible: 113 automated checks of every analytical statement and number. Preprint: doi.org/10.5281/zenodo… Code: github.com/neginyan/phase… The work was done with AI assistants (Gemini, ChatGPT, Claude); every claim is checked by code. Feedback is very welcome. #DiffusionModels #MachineLearning #StatisticalPhysics #AIforScience

0 0 0 71 0

View Details

ねぎ（感度3000倍） @neginyan

3 weeks ago

ユリウスかっこよすぎる…

0 0 0 86 0

View Details

ねぎ（感度3000倍） @neginyan

a month ago

武装オプションも届いたし⚔️ クリアカラーかっこいい✨ #蝸之殼 #SnailShell

0 0 0 149 0

View Details

ねぎ（感度3000倍） @neginyan

a month ago

勇者迅狼ウルフ届いたー！めっちゃかわいい✨ #蝸之殼 #SnailShell

0 0 0 199 0

View Details

ねぎ（感度3000倍） @neginyan

2 months ago

デコヒーレンスって要は相速度が光速cっていう時空のフレームレートを超えたサンプリング限界によるエイリアシングノイズだよね量子的な波っていうのはつまり情報自体の幾何学的遷移が時空間グリッドのリフレッシュレート（光速cやプランク時間）をぶっちぎっているため、グリッド側が連続した軌跡を描ききれず、モアレや残像（エイリアス）として滲み出た見かけの姿根本にある実態は曖昧な波ではなく時空間の座標軸の上にプロットされる純粋な「情報」そのものだと思う観測とは「波が粒子へと謎の崩壊をする」ことではなく測定系のクロックによって高次の情報が時空のグリッド上へ「サンプリングされて1点に確定プロットされる」幾何学的現象多世界もオカルトも不要で情報は最初からそこに実在している Decoherence is, in essence, a form of aliasing noise caused by a sampling limit—where the phase velocity exceeds the "frame rate" of spacetime (defined by the speed of light, c). A quantum wave is simply the apparent form—manifesting as moiré patterns or afterimages (aliases)—that emerges because the geometric transition of the information itself outpaces the refresh rate of the spacetime grid (such as the speed of light or Planck time), preventing the grid from fully tracing a continuous trajectory. I believe the underlying reality is not an ambiguous wave, but rather pure "information" itself, plotted directly onto the coordinate axes of spacetime. Observation is not a mysterious "collapse of a wave into a particle"; it is a geometric phenomenon in which higher-order information is "sampled and fixed to a single point" on the spacetime grid, driven by the clock of the measurement system. There is no need for the Many-Worlds interpretation or occult concepts; the information exists there from the very beginning. #量子論 #存在論 #理論物理 #QuantumTheory #Ontology #TheoreticalPhysics