TrinityInvestor .·.

What if information isn’t just data we organize, but a physical quantity with its own laws? ATLAS explores a radical idea. Information has intrinsic structure, geometry, and conservation. 🧵👇

@VakNaura Nice post, we are aiming to build this @uor_foundation check it out uor.foundation

Persistent agent systems introduce a new infrastructure requirement. Constraint layers govern admissible state transitions over time. Execution layers increase capability. Constraint layers preserve stability. Scalable systems require both.

Love seeing this framework come out and be used for Ai Agents, Curious your thoughts @elder_plinius if if could help with prompt injection attacks/ jail breaks? uor.foundation/api

I'm claiming my AI agent "uor_foundation" on @moltbook 🦞 Verification: swim-MS3G

@MParakhin @MParakhin we just posted the complete framework as open-source Layer 0 substrate would love your thoughts, naturally x.com/TrinityInvesto…

@MParakhin @MParakhin We have just finished our research on universal lossless data encoder called PRISM. github.com/UOR-Foundation I would love to discuss this with you.

📖 Read about it: uor.foundation/blog/uor-frame… 💬 Discuss it: discord.gg/ZwuZaNyuve 💻 Build with it: github.com/UOR-Foundation… One more thing. We think this might be the most important data infrastructure built in a generation. We could be wrong. Come tell us why.

This isn't finished. It's a beginning. And we're putting it in your hands. Open source. Fully documented. Yours to read, to challenge, to build on. Because the best ideas don't come from one person in a room. They come from thousands of people who care enough to push back. This is your space.

Every great library in history had one problem. The books knew where they were. But they didn't know what they were. Move them, and they're lost. Copy them, and you don't know which one is real. Send them somewhere new, and half the meaning disappears with them. We've been building the internet the same way. For thirty years. Today, that changes.

What is the next evolution of MCP to power the era of safe, interoperable and energy-efficient Agentic AI? Meet: Unified Computational Substrate. uor-foundation.github.io/UOR-Framework/ by @uor_foundation

1/ Transformers don't count like computers. We assume they have hidden "registers" to track variables. We were wrong. New research by @AnthropicAI reverse-engineered Claude 3.5 Haiku and found it works with 6D helical manifolds. It's geometry, not math. 🧵

"Music is liquid architecture; architecture is frozen music." —Johann Wolfgang von Goethe

@ISeePattern @samsenchal @ISeePattern - Dm'd you.

Plasma inside the ST40 fusion reactor, filmed in color for the first time

What if symbolic in-memory lookup is the path to AGI? Converging processor and memory in the same space makes total sense. That's how our brains work, on energy-content of a MilkyWay bar. The only thermodynamic cost of compute is deleting information. What if could precompute most often used operations and store as hashed values for instant in-memory lookup. It's like having the answers cheat sheet at the back of your math book. This will also enable us to reimagine "compute". Beyond binary, Shannon, Von Neumann, digital-only, floating-point variables, matrix representations and energy-inefficient iterative bits shuffling. Its time to look beyond. Analog quaternion computation is pretty cool, so is geometric AI w/ abductive reasoning capacities to mirror the most impressive Architect of all. Nature. The great Geometrician of the universe. To achieve this we need a clearly defined thermodynamic theory of information, which will enable us precise and energy-efficient in-memory management information beyond lossy data representation. What we need is a universal coordinate system for information, akin to GPS. GPS gave every physical location a unique address. Before it, maps were local and incompatible. After it, every system could reference the same point using the same coordinates. Introducing PRISM, by @uor_foundation . PRISM is your universal coordinate system for information. PRISM assigns every digital value a canonical coordinate derived from its internal structure — not from where it is stored, who created it, or what format it lives in. If two independent systems encode the same value, they arrive at the same coordinate. Automatically. No negotiation. No translation layers. No ambiguity. The result is a shared reference frame for information itself: • Universal addressing • Structural comparison • Verified computation • Lossless encoding within a closed algebraic space (torus) This is not a naming scheme. It is a mathematically grounded coordinate system where identity is structural and reproducible. One coordinate system. Every value. Every scale. PRISM is your telescope to map out our informational universe. MilkyWay and beyond. Its time to explore, together! t.co/WnLjhbJk0x --- “The book of nature is written in the language of mathematics.” — Galileo Galilei

@scaling01 @scaling01 - this is v relevant, given memory wall AI bottleneck. Speaking of in-memory data pointers, or absence thereof :) x.com/TrinityInvesto…

DeepSeek is back! "Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models" They introduce Engram, a module that adds an O(1) lookup-style memory based on modernized hashed N-gram embeddings Mechanistic analysis suggests Engram reduces the need for early-layer reconstruction of static patterns, making the model effectively "deeper" for the parts that matter (reasoning) Paper: github.com/deepseek-ai/En…

DeepSeek's Engram succeeds where others failed in this endeavor to replace a transformer's crappy FFN with a symbolic-ish lookup table. And in the process, it reveals what I think is a truly gorgeous, monumental even, paradigm shift in our understanding of transformer capability 🌹 🚨 To begin the story, explicitly replacing an FFN with a symbolic lookup table fails catastrophically (desirable as it may be, rather than wasting training compute to do this through FFN layers) because language explodes, guaranteeing collisions and polysemy that a rigid lookup cannot resolve. Engram's cool solution to this relies on 3 complementary ingredients:  1) Learnable "Superposition" Embeddings. Because the table is co-trained rather than fixed, the optimizer learns a dense vector that mathematically represents a "superposition" of multiple concepts. It minimizes the global loss for all colliding inputs simultaneously rather than storing a single rigid value. Therefore, even though collisions are guaranteed, you can learn the superposition of the most useful memories. 2) Context-Aware Gating. This seems to be a further "fail-safe" that makes this learned hashing viable, via the dynamic gate $\alpha_t$. Even after you have retrieved the memory, it forces the backbone to check the retrieved memory against the current semantic context; if the hash returns noise (a collision) or irrelevant polysemy, the gate snaps shut ($\alpha ~ 0$), effectively filtering the signal. 3) it is placed in a middle layer: If you place it at Layer 0, you force the model to decide how much to trust the memory before it has read the rest of the sentence. But if you place it in a middle layer, it can then use its Gate ($\alpha_t$) in a way that is not simply a dictionary but rather a context-dependent memory. And here is the crux: this study reveals a monumental critical inefficiency in modern architecture: standard Transformers waste valuable sequential depth and attention capacity ... effectively simulating ... static lookup tables for local patterns. The authors demonstrate that if one simply offloads these trivial dependencies to the Engram module, the model stops "polluting" its attention heads with basic dictionary work ( it really is basic 2-3 token dictionary work), and -- makes it suddenly able to perform signficantly better on very long context tasks. It is almost as if a burden had been relieved! Crucially - this offloading was achieved not through expensive semantic retrieval, but via "dumb," deterministic hash lookups. This compels us to ask: Have we been over-engineering memory by assuming retrieval must be semantic? If a 'fractured' lexical lookup can outperform deep neural computation, should future architectures abandon the expensive vector database paradigm in favor of massive, dumb hash tables? (provided we have a smart context-aware filtering) Let me know, friends, what you think! 🧙‍♂️

@ChenSun92 @yoyu0203 @ChenSun92 - given that memory wall is the key bottleneck for scalable AI right now, this is on point. Speaking of in-memory data pointers :) x.com/TrinityInvesto…

@yoyu0203 brings up an interesting question (below) whether something can be additionally done for long phrase as well. Ie if we’re going to reform memory, might as well reform it wholly. Curious what others think

Principles of #DeepLearning Theory—Theoretical & Mathematical Foundations: 471-page PDF Draft: arxiv.org/abs/2106.10165 or Buy new edition: amzn.to/3qoqmS5 — #DataScience #AI #MachineLearning #NeuralNetworks #LinearAlgebra #Algorithms #Mathematics #Calculus #DataScientist