
#031 - Discoveries and checkers: who checks the AI? | SynthAIcast
Notes de l’épisode
Do you ever ask an AI what it thinks of your writing? This week we learned it doesn't always give the same answer, depending on who is asking.
One lab announces three scientific discoveries made by its models, in mathematics, computer science and physics. Nine mathematicians answer with a single request, and it fits in a few words.
An independent organization reviews another lab's risk report, and measures a detail nobody had thought to test, in two languages.
A company preparing to go public spells out in black and white what it owes the people who audit it, and what happened at one of them.
An old French institution, abolished in 1790, helps explain what is really at stake.
In the end, the real stake isn't the discovery itself. It's who gets access to the proof.
Three signals. One thread. One question…
🎙️ Every week, SynthAIcast listens to the world's best AI podcasts so you don't have to. Synthesised in ten minutes, verified by a human, in English and French
👉 Also available in French: SynthIAcast
🔗 THE SOURCES, BY SIGNAL
Signal 1 - the files on request
• AGMAI - nine mathematicians ask for the proofs, 25-26/10
• OpenAI - three scientific results, 26/10
━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Signal 2 - the AI that goes easy on its authors
• METR - review of Anthropic's risk report, 20/10
• METR - red-teaming OpenAI's agent monitoring, 22/10
━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Signal 3 - the audited who pays
• SEC - Anthropic's IPO filing, 26/10
━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Perspective
• Euronews - $5M to help mathematicians check AI results, 16/10
━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🎧 GO FURTHER
• Lex Fridman #503 - three mathematicians on AI, 26/10
Also cited, without a link: Reuters (investor talks on Anthropic's value, mid-October) and the French law of 16-24 August 1790 on the courts.
Hosted on Ausha. See ausha.co/privacy-policy for more information.