Latent Space: The AI Engineer Podcast · Latent.Space

🔬 Automating Science: World Models, Scientific Taste, Agent Loops — Andrew White

·1 hr 14 min·6 clips

D.E. Shaw spent billions on custom silicon for protein folding, but AlphaFold solved it on a desktop.

The launch episode of the AI for Science podcast on the Latent Space network features two guests: Brandon, who works on RNA therapeutics using machine learning at Atomic AI, and RJ Haneke, co-founder of MiraOmics, where the team builds spatial transcriptomics AI models. The stated purpose of the podcast is to bridge the AI engineering community and the scientific research community, which have developed largely independently but are now being brought together as AI applications in science begin producing real-world results. The episode opens with a story about protein folding, the problem of predicting how a protein sequence will fold into its three-dimensional structure, which had been considered one of the hardest open problems in biology. Before AlphaFold, the leading approach was molecular dynamics simulation, a technique that DESRES, D.E. Shaw Research, pursued with enormous resources including custom silicon, specialized clusters, and taped-out algorithms. The most optimistic pre-AlphaFold prediction was that the government might buy five such machines and fold perhaps one or two proteins per day. When AlphaFold demonstrated that the problem could be solved on a standard GPU using machine learning, running in Google Colab on desktop hardware, it was, in RJ's words, so mind blowing. I forget that protein folding was solved. The episode uses this story to frame the broader argument of the podcast: that AI is not replacing scientists but is enabling them to tackle problems that were previously intractable, and that the most important work now is connecting AI engineers who understand the models with domain scientists who understand the biology, chemistry, or physics well enough to know what success looks like. The tone is collaborative and intellectually open, designed to welcome both communities rather than speaking exclusively to either. The audience is AI engineers curious about scientific applications and scientists exploring how machine learning can transform their research.

As heard by us

AI for science becomes real when judgment, context, and messy data matter more than model scale.

Andrew White treats AI for science as a question of judgment, filtration, and context, not just bigger models. He uses protein folding as the clearest warning: D. E.

Read the full review in PlayNext →

Why you'd press play

You want AI for science without the usual grandstanding.

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