I enjoyed this recent blog post by acclaimed technologist Terence Eden proposing a thought experiment about the ethics of open sourcing a hypothetical LLM classifier trained on benefits sanction appeal letters.
Eden, himself a huge open source advocate, argues, quite compellingly that such a model should be kept closed to prevent the potential leakage of potentially confidential information in the training data or probing of the model for the purpose of abusing it.
However, as some of the post's commentators point out, there is a bigger question at play here: where is it appropriate to be using this kind of tech?
One of the key issues in my mind is the end-user's treatment and the power dynamic at play here. If you're making life-and-death decisions (tw suicide) about people who have few resources to challenge those decisions, then you should have appropriate systems in place to make sure that decisions are fair, explainable and rational. You must provide mechanisms that allow the party with everything to lose in this situation to understand what is happening and why. Finally, There must always be an adequate escape hatch mechanism for recourse if the computer gets it wrong.
Read more...
This is definitely a bit of a hot take from Ars Technica on the recent Anthropic paper about sleeper agents. The article concludes with "...this means that an open source LLM could potentially become a security liability..." but neglects to mention two key things:
Read more...
I really enjoyed diving into Seb Ruder's latest NLP Newsletter which focuses on all the areas of NLP that are still in desperate need of attention in a post-LLM world.
In an era where running state-of-the-art models requires a garrison of expensive GPUs, what research is left for academics, PhD students, and newcomers to NLP without such deep pockets?
Read more...
This week has been jam packed with traveling, meetings, events and all sorts! For an introvert like me, it's been pretty hard going pretending to be extroverted and interacting with lots of folks.
The biggest news this week was that my company won another award. A few weeks ago in September we won the CogX award for Best Fintech Company 2023. On Thursday I attended an awards ceremony with my colleague in order to accept the award for Best Emerging Tech Company (on the south coast) 2023.
Read more...
As of today, I am deprecating/archiving turbopilot, my experimental LLM runtime for code assistant type models. In this post I’m going to dive a little bit into why I built it, why I’m stopping work on it and what you can do now.
If you just want a TL;DR of alternatives then just read this bit.
Why did I build Turbopilot?
In April I got COVID over the easter break and I had to stay home for a bit. After the first couple of days I started to get restless. I needed a project to dive into while I was cooped up at home. It just so happened that people were starting to get excited about running large language models on their home computers after ggerganov published [llama.cpp]. Lots of people were experimenting with asking llama to generate funny stories but I wanted to do something more practical and useful to me.
Read more...