AI digest: Memory modules and mathematical breakthroughs
External memory models, mathematical proofs from AI, and the backlash against forced AI search.
Big week for AI architecture advances and some proper mathematical breakthroughs that actually matter.
External memory models sidestep the retraining problem
Researchers from NUS, MIT, and A*STAR built MEMO, a framework that trains a separate memory model to handle new knowledge without touching the main LLM parameters. This could solve the massive cost problem of keeping models updated with fresh information. The modular approach means you can swap out knowledge domains without rebuilding everything from scratch.
Claude solves decades-old mathematical problems
Claude Mythos cracked the Erdős unit-distance conjecture over a weekend, following OpenAI’s recent breakthrough. Anthropic engineer Sholto Douglas called it a “cute, simple proof” and evidence of “serious overhang” in AI capabilities. Meanwhile, Google’s AlphaProof Nexus solved nine open Erdős problems for just a few hundred dollars each in inference costs. This isn’t just impressive, it’s scary how cheap these breakthroughs are becoming.
Users revolt against AI-first search
Google’s decision to replace blue links with AI agents at I/O 2026 triggered a proper user rebellion. DuckDuckGo installs spiked 30% as people actively sought alternatives to what they saw as being “force-fed” AI search. Turns out users still want choice over when they interact with AI rather than having it shoved down their throats.
Voice cloning goes fully local
OmniVoice Studio launched as an open-source alternative to ElevenLabs that runs entirely on your own hardware. No API keys, no cloud accounts, no subscriptions. It handles voice cloning, video dubbing, and real-time dictation across 646 languages and integrates with Claude through MCP. This is the direction we need more AI tools to go.