Meta and Microsoft Cut Claude Use & Spending, Workday Faces a New AI Threat

Akash Pasricha (0:13)

Welcome everyone to The Information's TITV. My name is Akash Pasricha. It is Tuesday, October 6th. Today on the show, Reflection released its new open weight model this week. We'll talk to the company's co-founder about the launch.

We'll then unpack our reporting that Meta and Microsoft are trying to scale back their employees' use of Claude. We'll also discuss how Workday might not be out of the woods just yet from the SaaSpocalypse. We're going to close out the show with a conversation with two of DoorDash's top AI executives as they seek to build out their AI operation. It's going to be a great show, so let's get right on into it.

There is a new model on the block. Reflexion released its newest open-weight model Beam, too much anticipation. The company said it still lags top Chinese open-weight models like Kimi K3 but has certain advantages on efficiency and inference time. Joining me now is Ioannis Antonoglou, co-founder of Reflexion. Ioannis, welcome to the show. It's great to have you here.

Ioannis Antonoglou (1:16)

Thank you so much for the invitation and it's great to be here.

Akash Pasricha (1:19)

Ioannis, I want to start with the beginning of your journey here just a little bit. When you started Reflexion with your co-founder Misha a couple of years ago, this is probably the most anticipated release that we've seen from your company just yet.

Where do you start when you set out for this ambition? We want to build a model, we want to build it open-weight, we want it to rival the top models that are out there. Where does one start with that process?

Ioannis Antonoglou (1:48)

Yeah.

As you said, we started Reflexion with Misha and I about two years ago. Before that, I was actually a long-time DeepMinder. I've been working on AI for about 15 years now. I joined DeepMind back in 2012 when it was still a startup. I've been really an early believer in AI and AGI. I spent 10 years doing reinforcement learning, which is a method that's really well-used now, but it's a method of having models that learn through trial and error. I spent 11 years at DeepMind, worked on things like AlphaGo, AlphaZero, MuteZero, and led Gemini RLHF before I left to start reflection with my Misha. Okay.

Akash Pasricha (2:33)

Take me to start a reflection though, because we only got about 10 minutes here. When you sit down with Misha and you think about Beam, where do you start in terms of how you are thinking about this model, how you're going to design it, how you're going to go about training it, where does that process start?

Ioannis Antonoglou (2:51)

Yeah. This is actually like our first model. We started building this model about a year ago.

We were a company of 30 people about a year ago, and we just started with the mission of actually building a really strong Western open model. We saw that there was a gap in terms of strong open models coming out of the West.

In the past one year, we've actually assembled a team of 200 strong researchers, engineers, and 100 other people who were supported the company as a whole. Currently, we just scale the company from 30 people to about 300 people, and with the sole mission of actually building a really strong open model. So, what we saw is that you actually need a model that is really efficient in terms of like its reasoning, so you can actually solve tasks in enterprise and in the developer community that don't consume too much compute, but at the same time, it's like really intelligent. And the other thing that we saw is that you want something that is kind of like at the right size for a workhorse model. So the first model is at 500 billion parameters, which if you actually go to open router and you see the models that actually have the highest usage are kind of like about this size. So we tried to build something that is really compute efficient, really strong reasoning.

Laura Bratton (4:20)

How did you do this?

Akash Pasricha (4:23)

So token efficient is one thing that stood out to me in the specs for this model. How were you able to achieve that token efficiency?

Ioannis Antonoglou (4:33)

So the way to do that is by really scaling reinforcement learning. So we actually run the largest reinforcement learning run that has been documented in the open space. We used 10.5 thousand to be 300 for four weeks and 100 million rollouts. So that's an order of magnitude larger than the next one.

We ensured that the model is rewarded both for solving the task, and also solving it in the most token efficient way. At the same time, as I said, the model is a 500 billion parameter model, which means that even each token is actually quite compute efficient to be generated. So if you just have a model that doesn't use too many tokens to solve a task, and it is also really compute efficient, then you get an extremely efficient reasoner.