Register and share your invite link to earn from video plays and referrals.

Search results for NEONLINE
NEONLINE community
One keyword maps to one global community path.
Create community
People
Not Found
Tweets including NEONLINE
Particles obeying simple nonlinear ODEs spontaneously form this glowing density structure; pure chaos rendered visible. ẋ = σ(y − x) ẏ = x(ρ − z) − y ż = xy − βz
Show more
Attending #ODSCAIEast# today? Join our cofounder @mcarbin at 2:35 PM ET in the Keynote Room. “Nonlinear Dynamics as the Next Substrate for Intelligence” explores a fundamental idea: we may be building AI on the wrong substrate. The human brain runs on ~20 watts. Today’s AI systems require orders of magnitude more. That gap isn’t just engineering, it’s architectural. Michael will share why moving beyond digital computation toward nonlinear physical dynamics could unlock a radically more efficient path to intelligence.
Show more
What if a single forward pass could let a model read two completely different texts at once? Transformers are built from strongly nonlinear pieces: self-attention and layer after layer of MLPs. So the natural intuition is that mixing two contexts into one input should make the output collapse into noise unrelated to either. This paper overturns that intuition. Simply averaging the token embeddings of two texts and feeding the result as a single input still leaves clear traces of both contexts in the next-token distribution. Tested across Pythia, Llama, and Qwen, the true next token from each individual stream lands in the top-10 ranks of the mixed output 30-40% of the time, and within the top-100 ranks 60-65% of the time. Even more striking: this superposition ability isn't something models learn. It's strongest right at initialization and degrades monotonically as pretraining continues, suggesting it's an intrinsic architectural property that training actually erodes. The authors show it can be substantially restored with lightweight fine-tuning on less than 0.025% of the original pretraining data, and they build on this to propose a guided decoding method that generates two independent, coherent continuations from a single forward pass. Title: Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs URL: #LLM# #Transformers#
Show more
September 1, 2026 edition of FFTT: "AI is driving a bond singularity before it drives an AI singularity." Summary: AI increasingly competes with Bessent for capital...to build a technology that nonlinearly undermines Bessent's tax base.
Show more
This paper reviews breather, rogue, and interaction waves in a (3+1)-dimensional fourth-order nonlinear equation, deriving wave interactions and analyzing dynamics through graphical methods.
A single flap rewrites the forecast. Three coupled equations capture convection in a fluid layer. The stream function evolves as ẋ = σ(y − x), the horizontal temperature difference as ẏ = rx − y − xz, and the vertical profile distortion as ż = xy − bz. Here σ is the Prandtl number, r the Rayleigh number, and the products xz and xy are the nonlinear terms that stretch and fold trajectories. From these deterministic rules a bounded, aperiodic attractor emerges whose sensitive dependence on initial conditions is the mathematical origin of the butterfly effect. Determinism does not imply predictability.
Show more
An innocent linear test decides the fate of every explicit integrator. Write y′ = λy with Re(λ) ≪ 0. While |λ| stays modest a fifth-order Runge–Kutta pair advances with large, inexpensive steps; the moment |λ| grows the stability region collapses and those same steps shrink until the computation is useless. Stiffness is that mismatch of time scales. The solution itself may already be smooth, yet an explicit method is still forced to resolve the fastest decaying transient. Implicit formulae (BDF, Radau collocation, Rosenbrock) replace the explicit update with a linear or nonlinear solve and remain stable on the slow manifold. Curtiss and Hirschfelder first called such equations “stiff” in 1952; the same test still separates the solvers that finish from those that do not.
Show more
Just some notes from $TSLA transcripts: Elon Musk: "I think Optimus will be the biggest product ever" from Q2 earnings transcripts. Even from Q1, Elon Stated: "I think Optimus will be our biggest product. Not just Tesla's biggest product ever, but probably the biggest product ever. And I remain convinced of that conclusion". It's very rare to see such high conviction reiterated from Elon regarding humanoids both directionally and TAM-wise. In terms of upstream suppliers opportunities, Karn Budhiraj, VP of $TSLA supply chains stated: "We’re seeing the same level of investment going to memory and also new specific items like metal injection molded parts, flexible printed circuits, and all sorts of nonlinear technologies..." And that they had to build new supply chains from scratch. So would be interesting to look at for new supply chain supplier opportunities apart from the known ones. Just pulled older slides and it looks like capacity targets for Fremont are 1M and Giga Texas was 10M, so at scale, $TSLA seems to be the clear US leader. If you're looking at that S-Curve opportunity ramp. On a side note, Elon also gave a shoutout to $MU with likely reasonably priced LTAs, given his statement: "We really appreciate Micron making room for Tesla in the years to come and giving us actually a very significant allocation on reasonable terms given the pretty insane pricing of memory these days." (Just for structural memory demand). Elon also gave a shoutout to Panasonic which had invested many billions in increasing battery cell production (will need to do deeper DD into Panasonic after this), then Samsung and $TSM. Just takeaways from Tesla earnings TLDR: Elon tends to be directionally correct on where the future is heading, such as with EV or Space. He seems extremely bullish on humanoids.
Show more
🌍 "When can self-supervised learning recover the world's true structure?" This theory paper from LeCun and colleagues proves the answer is: only when the latent variables are Gaussian. Title: When Does LeJEPA Learn a World Model? URL: 💡 Overview The paper pins down when LeJEPA (JEPA + Gaussian regularization SIGReg + alignment) can recover the world's latent variables linearly, up to rotation, from nonlinear observations. The key condition: the latents are Gaussian and evolve under an OU process. ⚠️ The problem If a representation distorts the world's true degrees of freedom, reliable planning and compositional generalization break down. It was unclear when self-supervised learning provably recovers world structure. 🛠 Approach and core insight ・The optimal representation extracts the "slowest features" of the latent process, ordered by eigenvalue ・Via Hermite polynomials and Mehler's formula, cross-view correlation decays as ρ^d for degree-d nonlinearity ・So alignment penalizes every degree of nonlinearity, making the linear map the unique optimum ・With linear identifiability, planning in latent space yields the same optimal actions as the true world (directly usable for control) ・Conversely, demanding the optimum always be linear forces the latent distribution to be Gaussian (uniqueness) 📊 Results ・SIGReg and VICReg keep R² > 0.999 for linear recovery up to 1024 dimensions ・Sweeping the generalized-normal family, R² peaks sharply at α=2 (Gaussian) ・In pixel-based robot control, Gaussian OU pairs hit R²=0.95, while non-Gaussian real trajectories stay at R²≤0.5 ・Control cost tracks R² monotonically, and the Gaussian encoder is oracle-level #WorldModels# #SelfSupervisedLearning#
Show more