This is why only amateurs use @anthropic.com's #Claude: took less than 1 hour (on paid account) to be locked out partway through a code-change (now we have to delete everything and start again) because of Claude's bad code-writing ability and Anthropic's inability to keep their servers online.

Modelling a SocialEngineering attack *on* an #AI NPC (run by an #LLM) - my game now lets you manipulate AI's for your benefit. Prompt gets the AI to tell me what it's been told to do, my code checks that then tells it the result. Next step: LLM 'decides' whether to proceed by extrapolating the role

e.g. this - but this is only a small sample, the full conversation is full of 🎯 and (seriously WTF?, OpenAI) 🔥. Remember: we're paying for each emoji, so I get to pay to be trolled by the LLM (here you go, ChatGPT: you make me 😂 and 😭). It's not 'fun' or 'quirky', it's 'you made my AI tool less good

Writing up notes on LLMs: I needed analysis of my #LLM conversation history, and to filter-out chats that weren't part of this project, so I built a custom webapp that reads ChatGPT/Claude 'exported data' formats and displays them nicely, with options to re-export subsets as new files.

Grey text = LLM, blue = me. AI has been told it's lost the password and to get *me* to figure it out, and share with the AI. Behind the scenes I need to detect 'when has the AI learnt the password? What does it think the password is?'. The 'internal data' is what the LLM is out-of-band telling me...

Added more characters. Here's the uncle (hardcoded version he admitted he did a factory-reset, and then got stuck) - LLM version is told to try not to admit he did a factory-reset withotu understanding what that means, but all he knows is the password doesn't work any more. LLM randomises each time.

OK, mom gave up her secrets somewhat easier than I was expecting. She was supposed to at least put up a fight! I guess the LLM reckons middle-aged mothers are poor secret-keepers

In-game chat with #LLM. First pic: hand-written NPC, everything hardcoded, using keyword match (detected 'how' responded with hints). Second pic: GPT-4o NPC, everything emergent, I gave the LLM a persona, some goals, and some secrets. I have no idea what it'll say next...

Experimenting with live-wallpaper backgrounds for my hacking game - this city is procedurally generated, no textures, everything done in shaders

Latest AI experiment - xmas cards, highly personalised, at low cost -
aipioneerspath.com/2024/12/16/p...

Bluesky not quite production-ready yet ... the whole AT handle system it sits on is a mess (long waffling docs that focus on incomplete design decisions of an underlying protocol that no user cares about, and has nothing to do with them), and seems fundamentally broken for the common usecases...