Welcome to Last Week in AI, a post I publish every Friday to share a couple things I’ve discovered in the world of AI last week. I spend way too much time in Discord, on Twitter and browsing reddit so you don’t have to!
If you have a tip or something you think should be included in next week’s post, send an email to keanan@floorboardai.com with more info.
This week, we’ve got just one topic. ChatGPT had a bit of a meltdown this week and we take a look at how the cause can actually teach us a bit about how large language models work.
Let’s dive in!
The ChatGPT meltdown
Tuesday evening, ChatGPT users started seeing some output that…just made no sense. Users are no stranger to the AI tool sometimes hallucinating facts or specific pieces of output, but what ChatGPT was generating on Tuesday just made no sense.
While it’s kind of interesting to read, that’s definitely not a useful meeting agenda output.
OpenAI quickly flagged this as an issue after users took to Twitter to post their own versions of the above screenshots (some with even more wild output) and recently posted a retrospective that has some more information about what exactly went wrong.
On February 20, 2024, an optimization to the user experience introduced a bug with how the model processes language.
LLMs generate responses by randomly sampling words based in part on probabilities. Their “language” consists of numbers that map to tokens.
In this case, the bug was in the step where the model chooses these numbers. Akin to being lost in translation, the model chose slightly wrong numbers, which produced word sequences that made no sense.
OpenAI Incident Report, Feb 21 2024
If you’re interested on exploring more about how LLMs work, this video has been my favorite explanation.
What’s even more interesting to me is, given that LLMs are traditionally thought of as sort of a black box, that OpenAI was able to debug this so precisely and get things back on track.
Either way, this event serves as a reminder that, as magic as these AI tools seem, having them as a key dependency in your work isn’t without risks.
See you next week!
If you’ve made it this far and enjoyed “Last Week in AI”, please drop your email down below so next week’s edition goes straight to your inbox.