# World Models, Agent Businesses, and AI Control *By AI News Digest • May 16, 2026* Yann LeCun's AMI Labs sharpened the case for world models and planning beyond pure next-token prediction, while Anthropic, Cognition, Thinking Machines, and Google showed how agentic systems are moving into products and workflows. Bengio and Amodei, meanwhile, pushed the policy debate toward concentration, jobs, and public infrastructure. ## The shape of the day Several of the day’s biggest AI stories pointed the same way: influential researchers are openly looking beyond pure LLM scaling, agent products are turning into real businesses, and leading voices are getting more specific about who will control AI and who benefits from it. [^1][^2][^3][^4] ## Beyond pure LLM scaling ### Yann LeCun starts AMI Labs around world models Yann LeCun said he left Meta after it became clear the company was entirely focused on LLMs, and launched AMI Labs to push AI for the real world through scaled JEPA-based world models. His argument is that LLMs are valuable for language and code, but not for predicting the consequences of actions or doing the search-based planning needed for agentic intelligence. [^1] LeCun said the near-term targets are industrial process control, robotics, and some healthcare use cases, with action-conditioned world model demonstrations expected within roughly a year to 18 months. [^1] *Why it matters:* This is a concrete organizational bet on a different roadmap, complete with a technical architecture, industrial use cases, and a development timeline. [^1] ### Reasoning work is leaning on search, verification, and symbolic scaffolding A newly released 30B-A3B reasoning model was described as reaching gold-medal level on IPhO and on IMO/USAMO-style math evaluations via test-time self-verification and refinement, alongside what its authors called a simple unified scaling recipe for proof search. [^5] Separately, Gary Marcus argued that much of the last two years of progress has come from symbolic harnesses around LLMs, including loops, conditionals, and Python interpreters, rather than pure scaling. He pointed to Claude Code as a neurosymbolic example and argued that pure LLMs still break on abstraction and out-of-distribution generalization tasks such as Tower of Hanoi variants. [^2] *Why it matters:* Across both the new paper and the critique, the common idea is that stronger reasoning is being framed less as bigger autoregressive models and more as search, verification, planning, or symbolic structure layered around base models. [^5][^2] ## Agent products are posting real business signals ### Anthropic describes breakout traction for agentic software Dario Amodei said Anthropic's revenue moved from roughly $100 million in 2023 to roughly $1 billion in 2024 and roughly $10 billion in 2025, tracking a smooth exponential curve alongside capability gains. He tied the current inflection to Claude Opus 4.5 and to Claude Co-work, a non-coding agentic interface built in about a week and a half using Opus. [^3] Amodei said Co-work was created after Anthropic saw non-technical users push through the command line to get agentic work done anyway, and that early release metrics were about four times higher than anything the company had previously launched. Separately, a Colossus post said Cognition's Devin had reached a $445 million revenue run rate in its first 18 months, with usage doubling every eight weeks and customers including the US Army, Goldman Sachs, and Mercedes-Benz. Ramp data cited this week also put Anthropic at 34.4% business adoption versus OpenAI at 32.3%. [^3][^6][^7] *Why it matters:* The signal here is not just better models. Agentic tools are moving into broader business use with large revenue claims, expanding customer footprints, and direct adoption competition between major vendors. [^3][^6][^7] ### The next interface race is multimodal and action-oriented Thinking Machines Labs, founded by OpenAI's former CTO, showed a preview model handling real-time translation, interruption-aware conversation, time awareness, and simultaneous tool use such as web search and UI generation. The company said it plans a limited research preview in coming months, with a wider release later this year. [^7] Google, at its Android event, showed Gemini using live page context to prepare bookings, reserve parking, fill forms, and operate inside a new Google Book experience through pointing and speaking rather than typed prompts. [^7] *Why it matters:* Product competition is moving beyond chat quality toward systems that can stay in context, take actions across software, and feel more like ongoing assistants than one-shot bots. [^7] ## Governance questions are getting sharper ### Bengio and Amodei focus on concentration, jobs, and public capacity Yoshua Bengio warned that advanced AI is currently concentrated in two countries and roughly ten companies, and argued that democracies risk keeping formal institutions while losing real agency unless they build shared public infrastructure and coordinate internationally. He said public awareness is the key ingredient that could push governments to treat advanced AI more as a public-good project than a purely market outcome. [^4] > If you're not at the table, you are on the menu. [^4] Bengio used that line to argue for coalitions of like-minded governments developing sovereign, ethically aligned AI together. Amodei, from a different angle, warned that AI could pair very high GDP growth with high unemployment and inequality, said policy needs real-time measurement through the Anthropic Economic Index, and argued both for mechanistic interpretability as a route to safer models and for targeted chip policies to limit autocratic surveillance and repression. [^4][^3] *Why it matters:* As models and agents commercialize quickly, the governance discussion is getting more concrete about power concentration, labor-market disruption, and the institutions needed to shape deployment. [^4][^3] --- ### Sources [^1]: [Yann LeCun on What Comes After LLMs](https://www.youtube.com/watch?v=ngBraLDqzdI) [^2]: [The Uncomfortable Truth About AI “Reasoning” | World Science Festival](https://www.youtube.com/watch?v=iFYF_e1GSGI) [^3]: [Dario Amodei on Safety, Job Displacement and Anthropic's $350B Valuation | WSJ](https://www.youtube.com/watch?v=VTnGzWlV6X8) [^4]: [Multinational Collaboration and Advanced AI — Yoshua Bengio | CDS 2026](https://www.youtube.com/watch?v=LnI1k5c-Wtw) [^5]: [𝕏 post by @stingning](https://x.com/stingning/status/2055123219506725201) [^6]: [𝕏 post by @colossusmag](https://x.com/colossusmag/status/2053801052571312414) [^7]: [AI News: Impressive New Model From Unexpected Company](https://www.youtube.com/watch?v=Oy7tzmfbl64)