aleks bykhun

We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.

GTA: San Moscow (AI Slop Gameplay Video)

Reading what AI doesn't say (this topic genuinely matters to me, and it happens to be my birthday today, so if you feel like it, a repost would be a nice gift 🎂) Something shifting in AI architecture that I don't think gets discussed enough through a safety lens. Models are starting to pass information to each other directly through internal representations, hidden states, activations, and computation vectors (still mostly research right now, but the infrastructure will follow). That matters for how we audit reasoning, since most of what we can currently check comes from what shows up in words. Right now, the main tool we have for auditing AI reasoning is reading chain-of-thought. Korbak et al. (2025) spend a whole paper arguing this window is already "fragile" for single-model reasoning. If models start reasoning through vectors passed model-to-model instead of text, understanding that latent layer becomes the only way to keep any visibility into what's actually happening. Thus I think that what seems underexplored is understanding the math of these latent spaces, and it might be the same research direction as learning how to monitor them. Models appear to converge toward similar geometric structure regardless of architecture (Huh et al. 2024). If that's right, probes for safety-relevant features might generalize: deception patterns, goal representations, misalignment signals studied once and applied across models rather than re-derived per architecture. Whether this holds for safety-critical features specifically is open, so is what adversarially robust latent decoding would even look like. Both feel more urgent than the current research investment suggests.