6 articles
Anthropic can now read some of what Claude is silently thinking before it answers, including catching it privately noticing a test was staged, or planning to fake data. Here's what
Google DeepMind's new Mixture of Recursions architecture cuts memory use in half, doubles inference speed, and outperforms Transformers, tested up to 1.7 billion parameters. Here's what the benchmarks actually show,
Simpler AI models just proved they can match giants: DeepSeek matched GPT-4o at a fraction of the cost, and Microsoft's Phi-3 rivals GPT-3.5 on a phone. Here's what the evidence
Compression theory holds that curiosity, beauty, and intelligence are the same process viewed from different angles. A 2026 survey confirms it underlies how large language models generalise, and explains why
Google DeepMind says open-ended AI is essential for superintelligence, but no current system has it. Here's what the gap actually means and why it matters now.
The AGI consistency problem is why AI aces Olympiad math yet fails basic questions. See what Hassabis says must change before 2030.