AI's weekly news cycle increasingly resembles two competitions running in parallel: researchers compete to improve intelligence, while investors compete to price its future. This week, Reflection and Mistral supplied new models, TypeSafe supplied another spectacular financing, and Arena supplied a reminder that someone must measure what all this intelligence does. Together, they suggest a market developing interesting ambitions beyond a better chatbot. Reflection introduced Beam, a mixture-of-experts model with 501 billion parameters, of which 23 billion activate per token. Its technical story centers on reinforcement learning and inference efficiency: Reflection reports generating more than 100 million training rollouts and achieving competitive reasoning performance with substantially less estimated inference compute than some larger rivals. Those estimates are not measured deployment costs, and the weights remain promised for later this month. Still, the proposition matters: agents that execute lengthy workflows need intelligence with an operating budget....

The theory of voice being the next big interface has picked up strong momentum, with investors pouring billions of dollars into voice AI startups working on areas ranging from model makers to enterprise customer service providers, and from meeting note-takers to AI-powered dictation. Every week there is a new model or a tool release that claims to sound human and converse like one. However, in reality, that might not be the case. Enterprise voice AI platform PolyAI's CTO Shawn Wen thinks that despite the release of full-duplex models ' which can speak while listening to you ' voice AI doesn't have its 'ChatGPT moment' yet. 'We have reached the milestone of developing full-duplex models. The next challenge is to make reasoning very fast, so that the models can fetch answers quickly and the conversation feels natural,' he told me on stage at the HumanX conference last month. 'I think the next stage will be slightly different because once the voice is good enough, like, and the customer is willing to engage with them for the first two or three turns, they start to build confidence, and over time, they will feel like I probably don't have to talk to a human if the agent can solve my problem,' he said....

1. Some readers interpreted the post by the title instead of by the first two paragraphs, as about what the law should be instead of about the implication of looking at only one of the relevant issues and ignoring all the others. They objected that the conclusions were wrong, the assumptions unrealistic, which was true and relevant for what the law should be but not for what legal rules followed from ignoring everything other than the effect of the rule on the incentives of the victim. 2. Some readers did not understand, or understood and rejected, the strategy of which the argument in the post was an example, using drastically simplified models to work out the logic of a complicated problem in order to understand enough about it to make sense of the real world version. Consider the problem of launching a satellite. To learn how to solve it you start with a simple model, a cannon ball moving in a uniform gravitational field in a vacuum according to the laws of Newtonian mechanics. Having solved that you add in successive complications ' air resistance, the force of the rocket engines, gravity varying with altitude, mass of the rocket changing as fuel is burned, each first understood with its own simplified model. You never drop the Newtonian assumption, even though it is not true, because relativistic effects are tiny and including them would complicate your calculations. Each separate piece is artificial and unrealistic; you can't fire a cannon in a vacuum since the gunners can't breathe....

WIRED's Global Editorial Director Katie Drummond spoke to Kevin Roose in late September, a few days after he and Casey Newton'with whom Roose previously cohosted the New York Times' 'Hard Fork' podcast'announced the creation of their new media company Machine Gods Media, alongside a podcast distribution deal with NPR. ' 2026 Conde Nast. All rights reserved. WIRED may earn a portion of sales from products that are purchased through our site as part of our Affiliate Partnerships with retailers. The material on this site may not be reproduced, distributed, transmitted, cached or otherwise used, except with the prior written permission of Conde Nast. Ad Choices...