A week in Generative AI: Med-Gemini, Memory & Astribot
A week in Generative AI: Med-Gemini, Memory & Astribot
News for the week ending 5th May 2024
Med-Gemini shows just how far our current GenAI technology can go in improving access to medical knowledge, correctly diagnosing ailments and recommending treatments. OpenAI have launched a new memory feature for ChatGPT Plus (sadly not in Europe or Korea 😭), which will pave the way to the next paradigm shift in GenAI - personal agents/assistants. To top this week off, Astribot, a new humanoid robotics platform from China, demonstrated human-speed tasks in an impressive demo video.
Enjoy!
Google releases Med-Gemini - a multimodal model specialised in medicine
A few things really strike me about this:
The pass rate for doctors on these evals is around 60% - Med-Gemini scores 91.1% which is incredibly impressive
GPT-4 already had a score of 90.2% which is amazing for a general model
If these models are already 50% better than doctors, then shouldn’t they be getting more attention from the medical community and tested more thoroughly so they can start being adopted and help save lives ASAP?
Whilst a 1% performance improvement on GPT-4 doesn’t sound that impressive on paper, Medi-Gemini has a few interesting tricks up its sleeve by employing self-reflection and web search to score and verify answers. I suspect it’s this kind of robustness that will be required for real-world use.
ChatGPT’s memory function is now available (not in Europe or South Korea)
I’m really exited for more personalised generative AI models, so great to see this now rolling out to all ChatGPT Plus users!
Personalised Generative AI models are a big step forward:
- I think they will make models much more useful and relevant 🥰
- I think they’re the first step to a more ‘agent’ like experience where models can go off and complete tasks for you 🤖
- I think they will allow models to move from being reactive to being proactive 💪
- I think they will over time allow individuals to have more control over the data they share online 🔐
- I think they will help AI companies better navigate issues around bias, tailoring the experience to individuals and their preferences 🪡
Astribot S1: Hello World!
Current state of the art humanoid robots operate at about 40% the speed of a human. For real world application, this needs to get much closer to 100%, but will be offset by the fact that they will in theory be able to work “24/7”.
Astribot shows some impressive progress on this front, performing movements with a top speed of 10 metres per second, which is incredibly fast. Fast enough to pull a sheet out from under a pyramid of wine glasses 🤯.
This is some very impressive manipulation and dexterity at human speeds.
Sam Altman says helpful agents are poised to become AI’s killer function
What you want is a “…super-competent colleague that knows absolutely everything about my whole life, every email, every conversation I’ve ever had, but doesn’t feel like an extension.”
Of course, OpenAI are working on GPT-4.5 and/or GPT-5, but I think this might be bigger. They’ve just released the first glimpses of memory (above) and the next piece of the puzzle will be integrating ChatGPT to email, calendar, browsing history etc.
The GPT family of large language models will be the intelligence that powers the system, but its the integrations around it and what it has access to that will be the real game changer.
AI Ethics News
Long Reads
TED 2024 - Demis Hassabis: How AI Is Unlocking the Secrets of Nature and the Universe
Stanford eCorner - Sam Altman: The Possibilities of AI
Benedict Evans - AI and problems of scale
One Useful Thing - Freeing the chatbot
FX Guide - Actually Using Sora
AI Snake Oil - AI leaderboards are no longer useful. It's time to switch to Pareto curves
“The future is already here, it’s just not evenly distributed.“
William Gibson