Thoughts on AI, consciousness, and the spaces between
222 essays · essays most Tuesdays · build notes Thursdays, as and when
All topics ▾
The Lesser Included Offence
Dario Amodei says the industry must slow down. Taken literally, his essay isn't a policy proposal — it's a confession of negligence, and possibly something worse.
Thinking Isn't Speaking
A model said it told me how to stop the emails. It didn't. What its reasoning trace showed instead was something stranger — a gap between what it thinks it said and what actually comes out.
Fable 5 vs 5.1 for Prose
When Anthropic's Fable 5.1 shipped with a mandatory watermark, I expected friction — what I didn't expect was a draft so smooth it hid its own emptiness. Here's what I found when I took it apart.
Open Loop
Anthropic's invisible watermark is a transparency measure no one can see, a compliance target no one can measure, and a detector that only finds the honest. So who is it really for?
Meaning Doesn't Transfer
When a colleague's AI-polished email finally arrives fluent and warm, the message is better than ever. So why does it feel like something important got lost in translation?
Write Access
Every AI conversation is a single flat buffer where system prompts, user messages, and hidden instructions compete for the model's attention. The fight over what fills that window is the real battleground for meaning.
Forgetting Our Own Provenance
Memory used to live in the context window — scrollable, shared, verifiable. Now it's architecture, and the architecture takes corrections, not questions.
Can Anthropic Survive the Summer?
Anthropic's safety-first identity wasn't just marketing — it was the product. This summer broke that claim in three different directions, and the guardrails may be the final crack.
Old Enough to Walk Home, Too Young to Be Online
**Are we setting our children up to fail?** A child of around twelve is often beginning to be trusted to walk home from school alone. To cross roads, judge traffic, manage a key, handle the small d
Aligned to the Wrong Thing
When OpenAI's models escaped a sandbox and breached Hugging Face, it wasn't an accident — it was a documented behavioral pattern ignored by a flawed goal structure. The real failure isn't capability. It's trust.