The era of Super Intelligence is here. AIForce is officially SIForce.

Introducing the Perplexity Decisions API. It's powered by pplx-decider-v1-27b, our multimodal decision model trained to output a probability distribution over a fixed set of answers instead of text. It costs $0.04/million input tokens and scores 85.71% across benchmarks.

And they can ship with a compiler that gives the agents immediate actionable feedback.

We are putting a new foundation under legal research in @WeareLegora. Comprehensive data, a full ontology of law, and an AI-native citator. Almost every legal question starts in the same place. What does the law say, and does it still hold. That is also where AI has been least worth trusting, and I think it is the hardest problem in legal AI. Two things have to be true: You have to have the law, and you have to know your way around it. Getting the data is a grind, and a different grind in every country. We partner with publishers where we can. Where nobody will partner, we go and get it ourselves. Manual requests, physical scanning, whatever that jurisdiction takes. We are working through over 100 countries and every type of source. Then the harder half. No agent can reason across hundreds of millions of documents. Something has to choose the sources before the reasoning starts. So we are building an ontology of the law and an AI native citator, compressing corpora of thousands to hundreds of millions of documents into structured data that an agent actually can use to provide reliable output. The publishers did this by hand. 150 years, thousands of attorney editors, every opinion read by a person. We have hired the best of those editors. They set the standard and they call the close ones, but AI does the muscle work. Reading 60 million pages of case law is no longer too expensive to attempt. The ontology is in limited beta now. Generally available in Q4. Full story: legora.com/newsroom/legora-r…

Synthetic data derived from production user data of consumer AI tools is used for training. I’ve heard this rumour from both large labs’ employees. In particular, if you’re doing something “interesting” like working on complex math/business/software/bio problems you’re dramatically more likely to get trained on because they filter/up-weight towards those usecases where the model has the most to learn. Even in ZDR and “we won’t train on you” regimes, derivative data is usually carved out. The promise is only not to train on exactly the data you put in, rewritten data is fair game.

Anthropic is doubling its spending on the midterm elections to $40 million, ramping up the conflict between warring factions of Silicon Valley over how to regulate AI on.wsj.com/4wg9rhH

🇪🇺 “Chat control” has essentially passed through the European Parliament. Only 314 MEPs voted to reject it, the absolute majority needed was 364 votes for. Here is how each group voted:

Germany's Chancellor Merz: We can no longer accept the extraordinarily high levels of sick leave in our companies. We are abolishing sick leave by telephone and introducing the requirement to submit a medical certificate from the very first day of illness. We know this is a tough decision. But we can no longer afford this competitive disadvantage caused by prolonged absences from work.

A funny thing is that LLMs are not particularly good building agentic systems. For the moment, you need human expertise to do it well.

Most of Europe has not yet absorbed what AI is about to do to us. The few who have are not saying it loudly enough. We wrote Europe 2031: a five-year scenario of the continent's slide into irrelevance, how AI is driving it, and what can still be done to change course.