AI’s Economic Impact, Warren Buffett Discovers YouTube, Mistral Launches “Le Chonk” | Diet TBPN
AI is turning game decompilation into training data and a legal fight, while Acemoglu's speech-recognition history argues AI diffusion may be slow, and Mistral's open-weight model gains ground on Chinese labs.
Watch on YouTube ↗Charlie's summary
- ▶ 4:55
Game decompilation by LLMs will feed training data and may push studios toward cloud streaming
Ben Thompson notes that disassembling machine code is easy, but turning assembly into human-readable C took humans years. LLMs are good at it. John says modders are paying tokens to point models at game binaries, so the next models should get much better at game development. Decompiling isn't illegal in itself, but derived code is covered by copyright, so viral mods get DMCA notices. John thinks studios may respond by hiding code behind cloud streaming, as Stadia did.
- ▶ 12:23
Acemoglu says AI diffusion will be slower than hyped, and speech recognition is his proof
Daron Acemoglu expects roughly 1.5% added to GDP over 10 years, and says AI isn't in the productivity statistics yet. His example is speech recognition. Dragon Naturally Speaking reached about 95% accuracy in 1997, but corporate disasters and Nuance's pivot to medical and enterprise left PC voice recognition only marginally better than in 2000. He argues that infrastructure breakthroughs need the right applications, marketing and consumer feedback. John pushes back that AI spreads over the internet, so it shouldn't take decades.
- ▶ 18:16
Acemoglu's two stances reconcile: AI replaces workers only if we keep building it to replace them
Jordy is confused that Acemoglu seems optimistic in the article but pessimistic on 60 Minutes. Tyler reconciles them. Acemoglu argues that on the current course AI will replace workers, so we should build it to make people better at their jobs. That would raise productivity and growth by solving problems people can't solve now, rather than just displacing workers and moving stuff around.
- ▶ 25:39
Mistral Large 4 ('Le Chonk') makes a European open-weight model a credible alternative to Chinese ones
Mistral launched a one-trillion-parameter, natively multimodal model with 49 billion active parameters. It is billed as the best open-weight model from the US or Europe. Tyler says it is on par with the open-source frontier, not winning every benchmark, and a sensible choice for anyone wary of the legal risk of using Chinese models. A commenter notes that it trails the latest Chinese model by a generation. The hosts point out that Chinese labs ship on shorter cycles, and that Mistral's one clear lead is Harvey's legal agent benchmark.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.