We worked with OpenAI to evaluate GPT-4o for scheming capabilities. We'll publish a detailed paper with our findings in the coming months.

Correction long overdue: my take was wrong. As Hamsa Bastani says, improved mathematical understanding + test scores results from GPT-4 + SIGNIFICANT effort crafting bespoke prompts. Still feel positive about AI tutors, but cautious about amount of human labour needed.

This is a great explainer of NICE decision not to recommend Lecanemab for NHS treatment. It's disappointing and surprising that so many Alzheimer's charities are not speaking plainly about the harms Lecanemab caused in many patients, and lack of proven durable benefits.

Full report coming out later, but for now have a look at this thread and our work in the model card :)

Thrilled to have welcomed Dr. Charlotte Stix, Head of AI Governance at @apolloaisafety for our #ParisAISafetyBreakfast! ✨ Inspiring discussions on model evaluations, deception capabilities and opportunities ahead among many other things. Thank you to everyone who joined us! 🙌

We are launching a new MSc in #AI Implementation (Healthcare) programme in 2025 College of Medicine and Health (UoB) Uni of Birmingham: birmingham.ac.uk/study/postgrad…

If AI development stopped this week we would have 5-10 years of absorbing the impact of current models on education, culture, healthcare, and business. But this week has also suggested that development is not stopping.

It's a question of incentives and political support. They come under extreme scrutiny for getting things wrong, & no rewards for taking risks. All the progress expediting ethics to research covid cures showed it was possible when there is institutional support for risk taking.

Richard Ngo Listen to this. Norfolk paints a picture of some cover ups, some burying of heads in the sand, some wanting to look plainly at the situation & deal with it. Sadly, like for the Catholic Church, few have done the work to understand 'why' & prevent it. youtube.com/watch?v=AHntVV…