A week in Generative AI: ChatGPT, AlphaFold3 & Audio Inpainting
News for the week ending 12th May 2024
No robots this week, but we have an intriguing announcement event coming from OpenAI tomorrow when they’ll be demoing some ChatGPT and GPT-4 updates. It won’t be GPT-5, or the rumoured search engine, but I’ll have more of a write up for you all next week.
We’ve also seen more medical breakthroughs this week with AlphaFold3 and the Agent Hospital research paper as well as the arrival of some great new audio editing features from Udio.
Lots of news on the AI ethics front this week as well with the release of AI safety tests from the UK government, OpenAI’s model spec, and a deal being struck between Stack Overflow and OpenAI that has caused a bit of a backlash from users.
OpenAI’s ChatGPT announcement: What we know so far
This is an intriguing announcement from OpenAI - they’ve launched plenty of features for ChatGPT without holding an event, so whatever they announce will be more significant than the recent memory/personalisation function.
I suspect (and hope!) what they will announce is more agent-like features, and I think the memory function they released last week are part of the foundations.
I think GenAI agents will be a big thing and start to go mainstream in the second half of the year. For OpenAI, GPT-4 will be the agent’s intelligence but there are many other features and technologies that need building around it for it to be a true agent and it’s these features that I think will be announced tomorrow.
For those that are interested, you can watch the livestream of the event on OpenAI’s website at 10am PT (6pm GMT) tomorrow (Monday 13th May).
Google DeepMind’s ‘leap forward’ in AI could unlock secrets of biology
Some more amazing work by the team at DeepMind, who have been applying advanced AI techniques to biology for over a decade now. AlphaFold3 can now reliably predict how proteins behave in real life, building on previous work that predicted the structures of proteins based on their chemical composition.
This is the sort of advanced scientific work that will benefit humanity for years to come. AlphaFold3 will dramatically speed up drug development and testing by allowing scientists to test hypotheses before going anywhere near a lab - it can all be done in simulation first.
Deflated
A nice little follow up to the short film ‘air head’ made by the studio shy kids using OpenAI’s Sora video model. This new one was made with a combination of real world actors and Sora generated videos.
Udio announces Audio Inpainting
This is a really great approach to being able to edit and refine the audio tracks generated with Udio. More fine grained controls are a big leap forward in content generation and it’s great to see these ideas applied to audio in the same way they’ve been applied to images and videos.
Agent Hospital: A Simulacrum Of Hospital With Evolvable Medical Agents
This is some really interesting research building on last year’s “Stanford Town“ simulation research. Agent hospital builds on this with a clear real world use case and shows how the doctors (agents) can consistently improve over time. The doctors are able to achieve state-of-the-art accuracy of 93% on MedQA, that I wrote about last week, in relation to Med-Gemini which achieved 91%, but that was only 1% better than GPT-4. This is a big improvement on that!
AI Ethics News
Stack Overflow signs deal with OpenAI to supply data to its models
OpenAI says it’s building a tool to let content creators ‘opt out’ of AI training
OpenAI’s flawed plan to flag deepfakes ahead of 2024 elections
Generative AI will be designing new drugs all on its own in the near future
Long Reads
One Useful Thing - Superhuman?
Benedict Evans - Ways to think about AGI
“The future is already here, it’s just not evenly distributed.“
William Gibson