Neil Movva

AVO took Claude Opus 5 from ~30% to 100% on ARC-AGI-3 NVIDIA’s Agentic Variation Operators (AVO) shows that frontier-level performance is not just about model capability. By combining persistent memory, supervision, and tool use, AVO took Claude Opus from ~30% to 100% on ARC-AGI-3, completed all 183 levels and explored 500+ directions in GPU kernel optimisation, achieving up to 10.5% better performance than FlashAttention-4. The bigger takeaway: system design can unlock capabilities that the model alone cannot.

5

1 Comment

The Fourth Layer: At NVIDIA GTC this week, the more interesting shift wasn't just in silicon. It was in software — specifically, in what software assumes about who is operating it. Enterprise technology was built around one invariant: Human is the orchestrator. Infrastructure computes. Platforms abstract. Applications present. And then a human clicks. Every dashboard, every approval queue, every workflow — optimised for human latency. That assumption is now structurally challenged. What's emerging — crystallised when Jensen Huang named NemoClaw and Claude Code as the two systems that sparked the agent inflection point — is a fourth layer. It sits between your platforms and your people. It decomposes goals, calls tools, remembers context, and improves as it operates. it's not a feature but a new tier of the stack. What makes this real rather than theoretical is that two organisations are closing in from opposite ends. NVIDIA builds upward from infrastructure. NemoClaw wraps the open-source OpenClaw framework in enterprise-grade governance and security controls — answering the question every regulated organisation has been sitting on: how do you run autonomous agents safely at scale? Anthropic builds downward from cognition. Claude Code's AGENTS.md and MEMORY.md give agents persistent identity, institutional memory, and authorisation scope across sessions and systems — answering the other half: how does an agent know what it knows and what it's permitted to do? Capability needs control to ship. Control needs capability to matter. When they meet, the layer becomes infrastructure—and infrastructure eats the layer above. SaaS pricing was built on a proxy: humans using software. Seats, licenses, tiers. When agents do the work, the proxy breaks. A single day in February erased $285 billion from global software stocks on one question: what happens to per-seat revenue when the seat isn't human? Not because software disappears—but because the operator changes. Value follows the operator. Infact, Systems of record—like SAP, Oracle and others —get stronger. Agents depend on clean, deterministic backends. What compresses is the human layer above them: the analyst pulling reports, the operator clicking workflows, the manager reconciling systems by hand. That layer collapses into code. And this pattern isn’t new. Linux didn’t kill operating systems—it redefined them. HTML didn’t remove interfaces—it standardized them. A new layer doesn’t destroy the stack. It becomes the default operator. And everything built around the old one has to adapt—or fade. We are at that moment again. The question is: what sits between your data and your decisions? Because whatever sits there — owns the system. #GTC2026 #NVIDIA #AgenticAI #AIStrategy #EnterpriseAI

4

1 Comment

Plot twist: LLMs can understand rocket science. They just can't code it without breaking everything. New paper tests LLMs on GTOC 12, where the world's best aerospace engineers compete on "nearly impossible" trajectory optimization problems. The findings are more nuanced than "AI good" or "AI bad." Here's what they found: The Good News: LLMs get the physics. They understand motherships, gravity assists, objective functions. One model scored 18/26 points, double the score from 18 months ago. They've conquered the conceptual barrier. The Bad News: They fail on execution. Not because they don't understand orbital mechanics, but because they: • Mix up km and meters • Make syntax errors   • Format outputs that crash validators • Fall into debugging death spirals (fix 1 bug, create 2 new ones) They act like a good architect who can design a perfect building but breaks the bricks when trying to build it. This isn't about "can LLMs reason?" They clearly can. The question is: can they execute with the precision engineering demands? Not yet. We're at a weird inflection point: AI that understands rocket science but can't reliably convert kilometers to meters. That's either hilarious or terrifying depending on your use case. The gap between conceptual understanding and flawless implementation might be the last major barrier to fully autonomous engineering. Full paper: https://lnkd.in/dhYKQ9yu #LLM #AI #Reasoning #Aerospace #Benchmarks

9

2 Comments

Congratulations to Jonathan Ross, Sunny Madra, Chamath Palihapitiya, and the entire Groq team. Having known the team personally, this is a rare outcome driven by first-principles engineering and long-term conviction — not momentum or hype. I just published a note on NVIDIA × Groq and why this transaction is strategically crucial for anyone allocating capital in AI infrastructure. A few takeaways that stood out: ▪ Inference is where AI economics now compound Training is episodic; inference runs continuously. That’s where latency, power, and unit economics determine returns. ▪ Groq represented an architectural, not incremental, threat Inference-first design (deterministic execution, memory close to compute) challenged GPU assumptions about efficiency and power at scale. ▪ NVIDIA didn’t acquire revenue — it acquired time and talent Licensing + acquihire compresses years of internal R&D into a single strategic move. ▪ Licensing is becoming the preferred consolidation model Similar to Meta / Scale AI, this structure secures critical capability while reducing regulatory friction. ▪ Power is the binding constraint in AI deployment Inference workloads run 24/7. Marginal improvements in power efficiency materially impact margins and scalability. At Good AI Capital, this deal reinforces our core thesis: Durable AI returns accrue at the inference layer — where models are deployed into real workflows and physical systems — not at the increasingly commoditized frontier model layer. Full analysis here 👇 🔗 https://lnkd.in/gQAXpC_s Pavel Danilov Rosemary Sagar, CFA Mark Tompkins Abhishek Pakhira William Elmore Emily Liggett Bill Young Daniel Kwan, MD, MBA Utkan Demirci Charles Wang Mary Tolan Lawrence Leisure Aaron C. Cook Alex Tang Touraj Parang 🤖 🤖 Ali Kashani, Ph.D., Courtney Pollard III, MD, PhD Atuanya Thomas, MSME Mike Zhai Nithin Johnson Andrew M. Howard Cooper Irina Tanenbaum Ronnie Vasishta Or Priel eric sippel Sanjay Sathe Walter Baumgartner Ryan Batenchuk Jeff Bogan Yasmin Lukatz Eric Lazear Mark K Norbury Bruce Lane

26

The paper proposes ProDM, a generative diffusion framework designed to correct motion artifacts in coronary artery calcium (CAC) scoring from routine non-gated chest CT scans, which are more widely available than ECG-gated cardiac CTs but suffer from cardiac and respiratory motion distortions that impair accurate CAC quantification. Accurate CAC scoring is crucial for assessing cardiovascular risk, yet motion artifacts often blur calcified lesions, reducing measurement reliability. To address this, ProDM introduces a synthetic reality-driven training approach that uses a motion simulation engine to generate realistic non-gated CT examples from gated scans, enabling supervised learning without requiring real motion-paired data. A property-aware learning strategy incorporates calcium-specific priors via a differentiable calcium consistency loss to preserve lesion details, and a progressive correction scheme gradually refines artifact removal across diffusion steps, enhancing stability and fidelity. Experimental evaluations on real patient CT datasets show that ProDM substantially improves CAC scoring accuracy, spatial lesion integrity, and cardiovascular risk stratification compared to baseline methods. A reader study further confirms that the model effectively suppresses motion artifacts and enhances clinical usability, suggesting that progressive, property-aware generative models hold promise for reliable CAC quantification in routine clinical imaging. https://lnkd.in/gcB22hxi

1

Biological Chaos Entropy: What Deepfakes Can Never Imitate The newly introduced Chaos-D-ID SDK represents a fundamental departure from conventional liveness-detection technologies that rely primarily on visual facial pattern analysis. Built upon HSKG Labs’ core patented technology, Bio D-ID, this system leverages biological chaos entropy—a class of signals that are inherently unpredictable and non-deterministic. Through standard camera sensors, Chaos-D-ID continuously extracts microscopic blood-flow fluctuations (remote photoplethysmography, rPPG) and heart rate variability (HRV) in real time. These signals are not static biometric markers; they are dynamic, stochastic physiological phenomena governed by complex nonlinear biological systems. Crucially, such signals cannot be replicated by AI-generated deepfakes, synthetic avatars, or even high-fidelity 3D masks. No matter how visually convincing a synthetic face may appear, it lacks the irreducible physical entropy produced by a living human body. Chaos-D-ID therefore establishes a new verification paradigm—one that does not ask “Does this face look real?” but instead proves “Is this entity biologically alive, right now?”

2

PrismLabs reports a 340% improvement in answer accuracy on multi-hop queries and a 65% reduction in hallucination rates by moving from a "naive RAG" system to a hybrid GraphRAG architecture. The main problem with the "naive RAG" was that the vector search retrieved semantically similar chunks, but wasn't able traverse the logical chain from cause to effect across document boundaries. Well, although the hybrid GraphRAG can assure causality and usually improves performance, I would challenge the numbers. 340% improvement in answer accuracy is a lot. The issue with the like-for-like comparison is that the architectures require different ways to organise information in the AI system. A system with well-organised information in the "naive RAG" architecture will outperform a system with poorly organised information in a GraphRAG system. Unfortunately, the vast majority of GraphRAG systems rely on LLMs to identify the graph elements and to build the graph. And benchmarks demonstrated that such solutions achieve even worse results than the "naive RAG" based ones. 340% improvement in answer accuracy tells a different story. Perhaps the graph was well designed, but the "naive RAG" was not. In any case, I would always strongly suggest that, if you have a thorough understanding of machine learning with graphs, go with the graphs. Otherwise, stick to simpler architectures. https://lnkd.in/dMAa3bs2

🚨Jeff Bezos just raised $12B at a $41B valuation for his new AI startup that's building something nobody expected👇 It's Prometheus and he calls it an "artificial general engineer" that can design and manufacture physical products: • Electronics • Jet engines • Medical devices • Drug compounds • Aerospace systems Bezos put it simply: "Something that took 100 engineers 10 years... what if 10 engineers could do it in one year?" That's the whole bet. Now here's what the funding actually tells you. Prometheus launched in November 2025 with $6.2B. Just raised $12B at a $41B valuation. Backers include JPMorganChase, BlackRock and Goldman Sachs. No broadly shipped product yet... still one of the most valuable funded AI startups. It's because frontier AI isn't a software race anymore. It's a compute procurement race. Investors are prepaying for: - Clustered GPUs - Power contracts - Industrial data at scale - Cooling and networking infrastructure It shows that demand is rising faster than the infrastructure can keep up. And every major funding round is a claim on future compute capacity. Jeff also made a bold claim on jobs: AI productivity will be so significant that two-earner households could drop to one income... A striking prediction from the man whose company Amazon just cut thousands of jobs due to automation. And he is now building the tool that could do the same to manufacturing. Bezos didn't just return to build another chatbot. He came back to make the physical world move as fast as software. --- Today, I run these 2 private communities that might help you: → SaasRise: for founders scaling from $1M to $100M → GrowthRise: for B2B Sales and Marketing leaders Also see our upcoming AI coding bootcamp (link in comments).

8

3 Comments