M Claxton Neztsosie

[M Claxton Neztsosie @talkinggod_ai]

@jsmooth995 Kinda like the way The Weekend sounds JUST LIKE Michael Jackson and The Osmonds sound like the Jackson 5??

[M Claxton Neztsosie @talkinggod_ai]

@Command_Dev @ESPNNBA And....?

[M Claxton Neztsosie @talkinggod_ai]

@ESPNNBA BRAVO! Now get ready for the slime to ooze hate with the usual looser comments below

[M Claxton Neztsosie @talkinggod_ai]

@HeunghaeCC @ESPNNBA Looks like they are better than Koreans

[M Claxton Neztsosie @talkinggod_ai]

@pinkfolkrsoft @ESPNNBA As if his Filipina mother didn't exist?

[M Claxton Neztsosie @talkinggod_ai]

Check out my latest article: Your Dashboards Are Green. Your Answers May Not Be.
Why Enterprise AI Needs Semantic Observability https://t.co/NAKjVlDc4A via @LinkedIn

[M Claxton Neztsosie @talkinggod_ai]

Your RAG system is green on every dashboard. Latency ✓ Throughput ✓ Error rates ✓ Pod health ✓
But the answers are drifting and nothing fires.
This is the failure mode nobody instruments for: a system that is operationally healthy while semantically degraded.
Under load — coherence decays. The model doesn't crash; it just gets subtly worse. Grammatically correct, contextually plausible, quietly wrong.
Under batch inference — topic boundaries blur. Question A picks up semantic residue from Question B's retrieval context. Not hallucination. Drift.
Under retrieval reordering — the same documents, different sequence, different reasoning structure. Your ranking is a hidden variable in output quality.
Kubernetes can tell you when your system is overloaded. It cannot tell you when your model has stopped making sense.
Enterprise AI has full observability into compute. Zero observability into meaning.
At @TalkinggodAI we have developed a set of mathematical invariants that measure semantic coherence, directional stability, and boundary integrity during inference — not after. During.
They work on our own pipeline. We can detect degradation before any traditional metric turns red.
Now we are testing whether the competence travels — whether the same invariants generalize to pipelines we did not build, retrieval strategies we did not design, and models we did not choose. Early signs are encouraging: the math appears to be model-agnostic across five embedding architectures (CV < 18 %).
Developing a product that we hope will deliver semantic observability as a lightweight sidecar for enterprise RAG pipelines.
The next layer of observability is not compute. It is meaning.
#EnterpriseAI #Observability #RAG #SemanticAI #MLOps #PlatformEngineering

[M Claxton Neztsosie @talkinggod_ai]

Meaning doesn't appear at the sentence level.
It emerges after signals interact and stabilize — the way weather systems form from noise, not from individual sensor readings.
Most models optimize at the wrong scale.
What if the training objective was structural stabilization, not local similarity?

[M Claxton Neztsosie @talkinggod_ai]

https://t.co/2hZtOMpvc6
The missing fourth layer is business physics.

Intent is the primary primitive
The Klarna failure has a mathematical signature: the agent optimized the redundant subspace of the prompt — what was measurable and surface-level — while the determinative subspace (the actual organizational intent: retention, lifetime value, relationship quality) was never encoded at all. Those are geometrically separable in embedding space. We formalized this decomposition as p = p_D + p_R, where p_D is the determinative intent subspace and p_R is redundant. The result is a compression system (The Intent Tunnel) that achieved 5.91× compression at 99% semantic fidelity vs. LLMLingua's 4.20× — validated across 6 domains.
The deeper point this video raises but doesn't fully name: intent isn't prose in a system prompt. It's a geometric property of the semantic manifold. Organizations don't need better OKR documents in the context window. They need the determinative subspace of their organizational intent isolated, structured, and transmitted to agents losslessly.
A three-layer framework is right. The missing fourth layer is the physics underneath it.

[M Claxton Neztsosie @talkinggod_ai]

@wildmindai The biggest threat to the AI companies is this. 7 years from now most users will run local models.

[M Claxton Neztsosie @talkinggod_ai]

@echodroff Thank you so much !

[Hugging Models @HuggingModels]

How LLMs actually work:

[📹 3Blue1Brown]

[M Claxton Neztsosie @talkinggod_ai]

Law doesn’t turn on words.
It turns on thresholds.
Computation and the future of legal strategy:
transforming narrative structure into actionable decision thresholds.

[M Claxton Neztsosie @talkinggod_ai]

86% accuracy predicting whether administrative agencies violated procedure. no keyword matching. through information-theoretic measurement of procedural structure. the framework applies across domains. legal reasoning is more predictable than we thought. (PP) #LegalTech #AI

[M Claxton Neztsosie @talkinggod_ai]

@IntuitMachine We are seeing a deeper regime structure: representations align when σ(domain, model) is high and data supports the input; they collapse OOD when σ is dominated by inductive bias. σ can be modeled and controlled.

[M Claxton Neztsosie @talkinggod_ai]

@JimDuncanRVA @60Minutes Destruction of news/integrity for "opinions" and vacantware Thanks to : @bariweiss

[M Claxton Neztsosie @talkinggod_ai]

@60Minutes The country is in crisis, we barely avert a NATO crisis, 60 minutes turns into low rent ET.. News ... Opinions ? Not Interested
Good Bye CBS because of @bariweiss