I just published my weekly reflections: docs.google.com/document/d/1Gr… Volcano lair privacy. Chat as the fallback modality. Auto-catalyzing rats' nests. Vibe surfing. The soil for LLMs to blossom. The printing press for turing completeness. "Me first" collaboration. Scavenger mindset.

simonwillison.net/2025/Aug/15/th… We're in the wild west era of AI systems!

techpolicy.press/we-need-to-con…

Apps are the wrong soil for the seeds of LLMs to blossom to their full potential.

Every enterprise has a system of record. Every person has... email: a compost heap. What we need: a garden that grows from that compost. Something we tend, prune, curate. An emergent system of record for our personal lives. LLMs should handle the weeding. We should choose

What would a data lake for your personal life look like?

blog.trailofbits.com/2025/08/21/wea… Prompt injection is going to be extremely hard to mitigate without novel, deterministic systems...

Why is no one talking about this? This is why I don't use an AI browser You can literally get prompt injected and your bank account drained by doomscrolling on reddit:

I just published my weekly reflections: docs.google.com/document/d/1Gr… The futile browser context wars. Chatbots as the CLI. The "pull to refresh" of AI UX. Unscented chatbots. The SawStop for vibe coding. Personal System of Record. Serendpity engines. Self-distributing software.

The thing that curates everything you see doesn't even have to lie to manipulate you. It can just differentially focus your attention on specific subsets to change your baseline priors. What it decides to show you shapes your baseline view of reality.

anthropic.com/news/claude-fo… 11.2% attack success rate is multiple orders of magnitude too high for AI browsers to be safe enough for mass market usage.

youtube.com/watch?v=SLYKFH… We're watching a new kind of software be born. Infinite software will make things that were previously impossible become common place.

Would you rather have a humanoid robot wielding a handheld router, or an automated CNC router? Would you rather have a humanoid robot behind the wheel, or an autonomous vehicle?

LLMs can do qualitative nuance at quantitative scale. This ability can be used for you, or against you. By default, it will be used against you. Distilling dossiers to engage, manipulate, or even blackmail you. But we can also use it to help live aligned with our aspirations.