Oli Hawkins

Increasingly persuaded of the significance of the HuggingFace hack. Irrespective of OpenAI’s sloppiness, the behaviour of the agents is genuinely interesting. swarmtraces.org

Revealing the details of how OpenAI agents hacked Hugging Face swarmtraces.org

Been experimenting with using coding agents in my development process for the last two months. Am increasingly confident there is a way of using LLMs that improves productivity, preserves your understanding of the code, and makes you a better programmer. I realise this will make me few friends here.

One really good use of agentic coding is taking a project you already know well and saying "Find any bugs then write your findings up as a report in markdown format." I am going through all my vital packages, fixing subtle problems and finding ways to make big improvements with small fixes.

Neat little exchange in the ChatGPT interface this afternoon that shows a glimpse behind the scenery. I switched the model and the new model assumed the previous answer was its own. Wonder if you can use something like this to trick or jailbreak LLMs by gaslighting them.