Latent Space: The AI Engineer Podcast · Latent.Space

Captaining IMO Gold, Deep Think, On-Policy RL, Feeling the AGI in Singapore — Yi Tay

·1 hr 32 min·5 clips

Yi Tay says the IMO effort used Gemini as an end-to-end model with no second system.

The guest returns after one and a half years. The host starts with the career update: Reka, then GDM, then a Gemini team in Singapore, with Reasoning and AGI as the loose banner while the team name is still unsettled. It feels casual, but the topic stays technical. They keep coming back to a practical question: when does a model stop being a neat demo and become the thing you reach for without thinking? Spreadsheets give them the cleanest example. The guest describes handing a model screenshots of results and asking for plots because the manual version is annoying enough to change the habit. That is where the clip lands best. The point is not a benchmark trophy. It is that small, irritating job where AI suddenly feels easier than doing it by hand. Coding sits right next to that. The excerpt treats AI coding as one of the places where capability crossed into felt usefulness, with image generation making the same kind of jump. Then the conversation folds back on itself. If LLM text and reasoning traces are showing up in the training corpus, the host asks whether pretraining starts picking up more reasoning behavior through that recursive loop. The guest does not overclaim. He says he has not personally seen much reasoning trace, and notes that places like GitHub can often be filtered if researchers decide to filter them. That choice matters. The issue is whether future training runs keep chain-of-thought text, remove it, or treat it as its own thing. Coding tokens get a similar shrug with teeth. The guest pushes back on the easy story that more coding data automatically generalizes, while leaving room for the picture to have changed. Even the stray health bit stays mechanical: productivity, energy, HRV, hunger, and whether taking care of the body helps with technical work. It ends lightly. The fun is hearing technically fluent people test their intuitions out loud without pretending the reasoning story is already tidy.

As heard by us

A candid look at reasoning models, AI coding, and practical frontier-AI use.

Latent Space presents Yi Tay as a voice inside frontier-AI practice, where a screenshot can become a plot and AI coding is already treated as part of real work.

Read the full review in PlayNext →

Why you'd press play

If you want Yi Tay's read on reasoning models, AI coding, and AGI-shaped work in Singapore.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Latent Space: The AI Engineer Podcast