.@martin_casado argues we've had the causality of private markets backwards: more capital actually grows the TAM "There's always been this weird meme in venture, which is it's kind of not much of an asset class. There's only a few companies. You can't deploy a lot of money in it. Which has often been very weird to me to have this zero-sum thinking from people whose entire job shouldn't be zero-sum thinking." "One thing we're learning is you actually can take large amounts of capital in private markets and deploy it and get the returns. Companies are staying private longer and accruing a lot of value that way. These models consume a lot of money, and they turn that money into growth." "There's two conclusions you can draw. One is there's a new technology wave that can consume more capital, and it's a result of a technology wave." "I happen to think causality goes the other way, which is if you put more money into private markets, it actually grows the TAM, and companies don't need to go public as quickly. Companies can do more privately." @a16z

The mask is not the model A two-pass audit catches causality leaks that attention-mask inspection misses. Across 192 injected faults, mask checks found zero; this method localized all 192, plus real bugs in Zamba2 and Nemotron-H.

Do large language models actually understand the world, or are they just very good at pretending? Does it even matter? Our recent special issue in the Royal Society, “World Models in Natural and Artificial Intelligence,” brings together pioneers across AI, biology, and philosophy to argue that the path to true intelligence runs through something deeper: the ability to model not just language, but causality, the self, and the physical world. Featuring contributions from Douglas Hofstadter, Michael Levin, Josh Tenenbaum, Samuel Gershman, Melanie Mitchell, and others, the collection asks a radical question: What if the next leap in AI requires not just more data, but systems that model themselves? Here are 3 ideas that might redefine how we build AI: 1. Capability is not the same as true intelligence. Current foundation models are incredibly capable, but they often lack true emergent intelligence. They learn surface statistics instead of compact, causal abstractions. Simply scaling compute will not fix this fundamental issue. 2. Self-modeling is an engineering primitive, not a philosophical luxury. New research in the issue shows that when networks learn to predict their own internal states, they compress and simplify, becoming more efficient as a form of regularization. For physical AI and future agents, a self-model is what will allow them to adapt their own skills and morphologies in real-time. 3. The hardest problems in AI are continuous with the hardest problems of life. Biological minds do not passively ingest data; they actively explore, driven by empowerment to increase control over their environment. If world modeling is about an agent representing itself in relation to its environment to survive and adapt, then general AI may need to look much more like artificial life. The takeaway is that the next leap in AI won’t come from just scaling up next-token prediction, but rather from systems that are agentic, self-referential, and temporally grounded. Read the introductory essay and the full special issue here: What do you think is the most important missing ingredient in today’s AI systems?

Are we actually evaluating world models — or just collecting scores we can't explain? That gap is exactly what this research tackles. HarnessEval-W: Agentifying the Evaluation of Visual Worlds ❓ What's broken about current world model evaluation? 💡 Existing benchmarks return scalar scores with no verifiable reasoning chains. There's no automated way to verify physics correctness, causality, or state persistence the way a human would — making it impossible to know why a model scored what it did. ❓ How does HarnessEval-W actually work? 💡 It uses a three-level agent hierarchy. Cases are routed to specialized skills, sub-agents answer specific sub-questions, and parent agents aggregate evidence into a final score — with every step recorded in an auditable evidence tree. Evaluation spans 8 settings across 3 categories (observation quality, transition correctness, world persistence), covering 330 cases and 18 models. ❓ How does it compare to existing benchmarks? 💡 Physical transition pairwise accuracy improves from WBench's 31.9% to 71.7%, and the draw rate collapses from 52.2% to just 1.8%. Human alignment is strong: Spearman ρ=0.93 for intentional transitions and ρ=0.87 for physical transitions, validated against 5,000 pairwise A/B human judgments. ❓ What does it reveal about today's models? 💡 Seedance 2.0 leads at 75.5, followed by Wan 2.7 (75.0) and Kling 3.0 (74.4). Fine-tuning analysis exposes sharp capability trade-offs: DreamX-World gains +4.8 on exploratory transitions but loses −11.9 on intentional ones. And render quality vs. physical observation correlates at r=−0.04 — nearly independent capabilities. #WorldModel# #VideoGeneration#

Check out this new OCP Academy course - Distributed AI Infrastructure: Architecture, Coherence, and Scale As AI workloads move beyond traditional cloud patterns, infrastructure becomes a tightly coupled system in which compute, networking, state, time, geography, and automation must operate as one. This five-module series examines the architectural and physical realities of distributed AI—from synchronization and causality to cross-region design, orchestration, observability, and infrastructure guardrails. Learners will understand why familiar microservices assumptions break at AI scale and how to reason about stable, scalable infrastructure spanning thousands of accelerators, multiple data centers, and independent infrastructure domains. Learn more and register for free here:

What causes stagnation in the built world is one of the key questions of our era. In this month’s @WorksInProgMag, @SCP_Hughes argues quite convincingly that some of today's housing shortage is downstream of aesthetic choices. In particular, he shows that the abruptly ugly turn of postwar architecture triggered a regulatory backlash in the form of the conservation movement, which in turn stymied construction and development in urban environments across the West: It may seem tendentious to flatly declare postwar construction ugly, but survey data is quite consistent on this point. Indeed, the sentiment is so widespread that it has spawned the Tartaria conspiracy theory. While the idea that people didn’t like the modernist construction of the postwar years is not by itself a new finding, Hughes teases out the particulars of the causality and shows how this popular rejection led to ensuing stagnation. This essay builds on Hughes’s prior work (“The beauty of concrete” and “Cheap ornament and status games”), and is I think an important contribution for anyone who cares about material progress.

For an AI agent to answer "why did we make that decision?", you need connected memory — not flat chat logs 🕸️ This tool spins the whole thing up in one command. Title: Introducing Create Context Graph URL: 🕸️ Overview Create Context Graph is a Neo4j Labs CLI scaffolding tool that generates a full-stack AI agent app with graph-based memory in a single command. The generated app bundles a FastAPI backend, a Next.js frontend, an AI agent framework, and a Neo4j graph database. ❓ Challenges Solved AI agents are easy to build but still struggle with relationships and causality. ・Flat chat logs and vector stores can't answer structural questions like "why did we decide this?" or "what's blocking this work?" ・In short, agents lacked the sophisticated memory needed to capture relational context 💡 Methodology & How It Works ・It turns data into a "context graph" (a connected knowledge structure), organizing three memory types: chat history, vector content, and reasoning traces ・It uses the POLE+O entity model (Person, Organization, Location, Event, Object) layered with domain-specific types ・When agents decide, the reasoning chain is captured as DecisionTrace nodes with linked TraceStep components, creating queryable provenance ・It supports multiple frameworks (PydanticAI, LangGraph, Claude Agent SDK), 22 built-in domains, Linear/Claude Code/GitHub connectors, real-time reasoning-path visualization, and automatic secret redaction 🌍 Use Cases ・Developers querying issue dependencies and team workflows ・Personal development analytics from Claude Code session history ・Multi-tool correlation combining decisions, commits, and work items Making decision provenance queryable helps with agent explainability, debugging, and cross-team knowledge integration. #GraphRAG# #Neo4j#