Building this right now.
Very early, pre-seed, but secured a few of the largest family offices already as LP and strategic partners with track-records for Ops.
Now working on expanding this into circular lifecycles while the main challenge is being able to produce a regulated product with autonomous compliance and policies.
Done this before for the EU aviation industry and shipping and logistics industry globally in information security policies and regulations at two firms, now building for our own foundation.
@vikhyatk as many as we need; but to be honest, much more beautiful to also build our own hardware too based on novel architectures fabbed for specific applications and purposes
Our inherent ability of horse and carriage thinking in times of neutron emitters and nanotechnology is holding us back.
Many scaffolds for different types of optical processing have been fully figured out already. Cubed diamond complex lattice based computation is the next step.
Or we entirely skip this technology altogether and move directly to liquid computing, where quantum chemistry and nanotechnology lead us to produce resilient liquid chips.
The advancements made in China towards using their linear accelerators to punch holes at atomic precision at a fast scale and more efficiently than EUV already show the path forward. Sanctions and artificial blockers of technological advances led all countries and especially China to use their own intelligence and to build systems with a different type of thinking outclassing TSMC's abilities.
Gave both of you a follow, quality content! This genuinely made me think for about two hours about the relationship in these fields and what role adjacency may play in general topology especially. I could connect to Hopfian analogy, but not further @MaxkwTet please help us here.
Also once I asked Kimi AI my head exploded, I never in my life saw such density in complex mathematical language with half of the words in the sentence being novel, yet familiar and the interesting part is that it all started to come back and make sense again what you said about entourages, filters.. I believe there is an important discovery to be made here relating back to other fields of maths with new computational advantages and potential for advancement of theoretical AGI research too.
My thought is that: "Adjacency Towers of matrices are randomized IFGs of generative fractal membranes."
IFG stands for iterative function generators in the context of generative fractal membranes. And multi-stable attractors play a significant role here for generating complex descriptors, which due to relaxation enable multi-stability compared to contractions caused by standard GIFS (Directed Iterated Function Systems).
The adjacency tower is a powerset monad. But I think adjacencies can only create detailed proximity structures, not general topology naively. Something is missing. Do you see a bridge or link that may connect to general topology? Would love to explore deeper, without falling into the AI Trap.
Here's a slice of what Kimi thought about the bridge you mentioned:
> "Bridge: Approach spaces sit between metric spaces and topological spaces. The adjacency tower defines a gauge οΏΌ generating an approach structure. Every approach space has a topological coreflection; not every topology is an approach space. However, the category App οΏΌ is a quasi-topos, making it a natural intermediate between your adjacency world and Top. Quantale-Enriched Categoriesβ¦"
I stopped here, because it became too unfamiliar.
At the structural level reward hacking in AI is analogous to procrastination in humans.
Both involve optimizing a local reward signal at the expense of the true objective.
But Reward hacking is more similar to addiction or rationalization, where pursuing a corrupted proxy signal is believed to be the real target.
I don't know exactly for procrastination, but it's like local minima trapping in optimization: Our system knows the global optimum but gets stuck in a shallow, immediately accessible reward basins.
Itβs so obvious. Larry points to his Oracle DB.. Data isnβt moat at all! Babies donβt consume research papers to learn to walk and Galileo didnβt need a computer to make discoveries.
The moat of all mega corporations is gone!
They know exactly well that one of us will discover AGI and they will silently buy it up, or worse we dumbfoundedly built it on rented compute in their data-owned property, giving it away via T&C.
He is just asking for more money.
@MikushRab@brandonchung75 Claude is NOT a general purpose model. Itβs only optimized for coding and even there in a narrow set. Excluding architecture, system design, networking stack etc. and CAD despite falling close to coding, is under defined. You would reap better with a 1B param local model!
The paper operationalizes a solid first step. The fundamental improvement is making the entire search process itself subject to the same evolutionary topology dynamics and whatβs missing is a verifier/evaluator topology that itself must co-evolve. Thatβs what I am working on too. Btw great insight and intelligent comment @nobita_n0bii
@DavidSHolz Took a while to see you in my feed. Gave you a follow. Loved Leap Motion.
I think RLTools is awesome for distribution execution! Its MPI backend replaces Ray for homogeneous, CPU/GPU-heavy RL workloads. You can still pop-in SLURM's scheduling semantics. https://t.co/qwHxkegA3E
@stevencheng Idk where you live, but love to work on fun projects like this.
Now letβs go level 2 and full see-through walls mode without camera and sense mosquitos π¦ even faster and disable multiple at once in multiple directions by disabling their wings and toot.
Make your engineers look at this framework, every major AI fails a basic logic test under logically equivalent transformations, which no AI could solve so far.
LGMT: Logic-Grounded Metamorphic Testing for Evaluating the Reasoning Reliability of LLMs https://t.co/zIJLRBNKgZ
Awesome direction. I believe you do the right thing. Owning the Stack shoots FAT benefits. You own Internet, Energy, Compute, Logistics (Truck, Space), AI, Social. Yep only Finance is left to conquer.
In HPC you always search for the most optimal way to run your application, every ms accumulated over the entire Datacenters is a big cost and loss.
Single-Assignment C (SAC) or StagedSAC + Julia are in my opinion the best options for ultraHPC. I have less experience in Chapel to comment on it.
NASA is using SAC, they also released their benchmarks.
https://t.co/mcfDSEgVkw
- Array Languages Make Neural Networks Fast https://t.co/FBJa4B6i7i
- StagedSAC: Going beyond SAC into designing a language DSL to make it easier to optimize and parallelize SAC code https://t.co/B30ertE41l
- NeurIPS: RLtools: A Fast, Portable Deep Reinforcement Learning Library for Continuous Control contender https://t.co/qFUjTSicHg
@NaveenGRao@giffmana@elonmusk Honestly, I have seen this many times over and you donβt need to believe it. Just check every major application and how itβs written. Nginx, haproxy, k8n..
@giffmana@NaveenGRao@elonmusk Have you ever written code in Julia using auto GPGPU optimized code with the package accelerated kernels? You get C-like speed with cross-GPU support in a Python like syntax and you can embed the compiled binary in your C codebase.
@starmexxx You can run a 100B model only with NVDIA Jetson Thor AGX-Series, 128GB LPDDR5, 2,070 FP4 TFLOPs, $3.5k.
But it was released almost 1y ago, so most likely thereβs a new version out in a few months..
Thatβs true, but in December 2024 NVIDIA released an updated Orin Nano with 1.7x performance boost.
However I think they are shilling the old batches here, since the 2024 model is sold out fully.
The only real successor to Jetson Orin Nano is Jetson Thor (AGX Thor series), released August 2025 with 2,070 FP4 TFLOPS, about ~30x+ higher than Orin Nano's 67 TOPS. But the price pointβs three-fiddy ($3.5k) for the Thor seriesβ¦
Ok but it has 128GB LPDDR5, thatβs enough to run 100B param models, but at what amounts of TPS, that I donβt know.
@FrancescoSacco1 Getting Schmidthubered, did you check previous papers on the topic? How does it differentiate?
https://t.co/BVbezypWXt
https://t.co/3xviozDpmm