So, Claire, have you heard about the Ramageddon? I actually was warning about memory inflation back in December. Regular people will first experience the AI boom through higher prices on laptops, iPads, and MacBooks. I explain it with a steakhouse analogy: memory players like Samsung, SK Hynix, Micron have limited capacity and are switching from low-margin regular memory to high-margin high bandwidth memory for AI, driving up prices.
Makes sense. Why flip burgers when you can sell steaks.
We're back with another big week. Meta has developed a model to read your brain without an implant. The biggest AI labs are racing into scientific discovery. Then we sit down with Ion Stoica, the Berkeley professor behind Databricks, AnyScale, and Apache Spark to get into Arena. Then we close some deals.
Josh, did you hear that Meta is trying to read our minds now?
I feel like most people think Meta's been reading our minds. They released Brain2Text, a system that decodes non-invasive brain recordings in real time using a MEG scanner. It hit 61% average word accuracy, a massive jump from 8% in prior methods. They frame it as a path to help millions with brain lesions. There are also visions for wearables and ads, but the immediate use case is medical. It's pretty hardware heavy, not scalable yet.
I always think about the dystopian sides. Who owns your thoughts? Will there be policy around that?
Regular consumers trade off convenience. Would you let Meta read your brain to avoid typing? That's an interesting UI shift, but also an ad vision. A lot of people are skeptical. Meta open sourced this, likely because it's a long-term medical endeavor. The MEG scanner is huge and expensive, so it's not coming tomorrow. But if scaled down to glasses, it could change interaction.
It was pretty interesting that Meta open sourced this. Why did they do that?
My sense is this will be a medical endeavor for the short and medium term. It's hardware heavy, not scalable. Each scanner costs millions and weighs as much as a small car. So we don't have to ring alarm bells tomorrow, but we should prepare for the slippery slope.
Yeah, I think that's when it becomes dangerous when you have a device on your head. But there are regulatory responses: Chile introduced neural rights in 2021, and France and Germany are drafting laws to ban mandatory neurotech in employment. So it's moving quickly but not tomorrow.
Yeah, I think for now I'd like to maintain my mental privacy, but I'm open to it down the line.
Claire, Anthropic announced Claude Science. What is it and what are the implications for the scientific industry?
Claude Science is an AI workbench for scientists, launched in beta. It integrates into tools, databases, and packages researchers use, and produces auditable artifacts. The innovation is at the product level, not a new model. It connects disparate workflows. For example, a researcher studying why a cancer drug stops working has to go through PubMed, multiple datasets, and code. Claude Science has over 60 built-in integrations. It also solves reproducibility by providing the full chat history and data accessed. Compared to OpenAI's GPT Rosalind, which is a new specialist model for biology, and Google's AlphaFold family, Anthropic is building a product that connects existing tools. The hope is to co-develop drugs with pharma, which is a high-margin buyer. The current drug discovery process is inefficient: a single drug costs $2.6 billion and takes 10-15 years, with 90% failure. AI can tackle the front end of discovery and optimization.
Do you think that Claude Science can do what they did with Cloud Code and develop the surface where the scientific community wants to work?
I think that's the hope. I haven't tried it yet, but there's promise. The Anthropic team is savvy with product. All three labs are going after this because of the impact and the big buyer in pharma. Pharma is a super high margin buyer, and switching costs are low for developers but heavy for scientists. If they nail this, they could co-develop drugs.
Our guest today is one of the most quietly influential people in computing. Ion Stoica is a Berkeley professor who has spent 25 years turning university research into infrastructure that the AI industry runs on. He co-founded Databricks, AnyScale, and is behind Apache Spark and Ray. He also co-founded and chairs Arena, which began as Chatbot Arena in his Berkeley lab and became the most watched scoreboard in AI. Ion, welcome to Light Work.
We know that Arena started as a research project out of Berkeley in 2023. You're well known for spinning four companies out of your lab, and this fall you're going back to teaching undergrad. What brought you back to the operating seat?
Everything in my mind is about having a positive impact. By impact I mean doing something that can make life better for some people, building software tools that help people do their job faster and better, or providing a better understanding of how to use tools. In academia and industry, it's about maximizing that impact. Sometimes projects turn into companies to maximize that impact.
Zooming back to the Berkeley days, when did you realize this could be something big? And what did you see in the two founders you decided to partner with?
It started as a side project of another project. When Llama from Meta was released, a group of my students took high-quality data from ShareGPT and fine-tuned Llama to create Vicuna. We deployed it as a chatbot. Then the question was how to show it was better. Static benchmarks didn't capture real human interaction. That's how Arena was born: a battle mode where users ask questions and get answers from two randomized LLMs, then vote which is better. For a long time it was just a measurement project, but it gained traction. Model providers started using it to show performance. It became clear it was an important part of the ecosystem, but it was costly to run. We considered a foundation, but the numbers didn't add up, so we created a company.
