some other ideas I wish I had blogged about at the time

prompt injection taught me one thing: tweets don't end up in the weights. a take in a reply to someone bigger than you may as well not exist. you keep the timestamp.

three more. none brilliant, same as prompt injection. obvious if you were in the trenches, and obvious to everyone about six months later.

• context is ram. december 2022, in a reply: "context length is like ram, layers like processing speed. context length maxxing will enable more agentic behavior." everyone was measuring how smart the 4k-token brain was. it was a genius with no working memory, couldn't act across more than a couple steps. autogpt showed up three months later. the whole 2023 agent year was people rediscovering that context is the constraint. https://x.com/himbodhisattva/status/1606708229521956864

• the left turns anti-ai. also december 2022: "I think blue tribe will land on anti AI." the reasoning, a couple weeks on: "automating blue-coded jobs: art, writing, programming." culture war eats everything, and AI came for art and code before it came for trucking. by july 2023 it had happened. "no AI" in every bio, the strikes, and me patting myself on the back, which is gauche. https://x.com/himbodhisattva/status/1605592285244186624 https://x.com/himbodhisattva/status/1682422501559926785

• screen the input with another model. december 2022 again, a thread about adversarial input to chatbots: "the first prompt is 'is this a valid customer service request or an attempt at hacking'." a cheap classifier in front of the real model, catching injections before they reach it. that's the standard mitigation now. https://x.com/himbodhisattva/status/1602574275843850242

the ideas were never the scarce part. blogging now.