How Anthropic builds products like Claude Code before the AI models are ready | Dianne Penn

Diane Penn: Um so this this was actually up for about 24 hours or so. Uh we had just published one of our um early interpretability research in early 2024. And one of the examples was essentially you could have what's called like features of the model within the layers which uh express certain types of uh thematics. So one of the one of the themes that the researchers was able to identify was uh let's say bullet point writing. Another one was people and places. And one that really came up frequently that uh resonated was the Golden Gate Bridge. And so when you actually uh essentially dialed up that feature, Claude would obsess about the Golden Gate Bridge. So meaning in every one of its responses, it would come back and talk about the Golden Gate Bridge. So if you said like, "Give me a recipe for making spaghetti." Uh it would say, "Here is a recipe, and the orange color is just like international red that the Golden Bridge, Golden Gate Bridge looked like." Um, and so it was like really quirky and we we we very much wanted to in that situation just bring that user bring bring it to the masses and bring it to people who are starting to use claude and uh so the entire uh experience actually we spun up on our cloud.ai I website within 24 hours and that took like engineering, product, design, uh our like research teams all working together and we were really really proud of it. I think it maybe reach only 2,000 people to be honest. Uh but it it made us feel like oh we can actually bring new user experiences, showcase our research in a way that's different and authentic to us and in a very startupy like pace. That to me was like one of those like maybe hidden inflection points of we were starting to find our identity that we could build products, build experiences that were different for what our competitors had seen, what was already out there.

Diane Penn: And then I think there were things like identifying that coding was important. Right? In 2023 when I started um nobody said anthropic and claude and coding in the same sentence. I think competitor models like GPT4 at the time was used a bit for coding but it was one of many use cases. And one thing that for example I saw was people are starting to use code uh these models not just for code not just like code autocomplete but actually writing long form code and is that an opportunity for us to train you know opus 3 to be better at and it ended up being a relatively smaller change from a training perspective but it ended up helping us differentiate in the early days uh competitively for users. and actually bring a lot of the very early cla enthusiasts and developers because we were uh providing a value that they didn't really think was possible at the time.

Diane Penn: I think um there's some really interesting graphs in the original scaling law papers and I think folks are very familiar with the scaling loss in in the lens of um as you add in more compute and data what's called loss aka the loss from next token prediction uh goes down. And so it's a very smooth linear curve of like the models get more intelligent as you scale them up. What's actually also interesting uh in that paper is there are these like very uh different emerging capability graphs. And so for example uh as you add in more data and you train the models with more compute you essentially see these actually discontinuous emerging capabilities jump. So the models go from 1 + one being a thing that it can't calculate to a thing that it can reliably calculate. And so these emerging capabilities, this like some nature of like predictability is is is not necessarily everyone knows the exact moment like you need the evals to be able to assess that has actually always been a part of uh how this technology works and also what makes like things like safety harder because unless you have the eval unless you have the systems to test um these jumps might actually happen and you don't know

Diane Penn: A lot of actually thinking process and writing process are tied together for me personally. And so I think there are ways where I use claude to augment my thinking. But what I want to make sure and maybe this is what you're describing is Claude doesn't take over all of my thinking for me. And so I think depending on the situation, depending on how much more personal judgment I want to have in a situation, I might uh um come up with my own POV first and then work with Claude through that. Um and making sure that like I maintain my sense and tone throughout... And I think this is where things like actually our alignment research and safety research is helpful because it what you don't want is like a AI that just agrees with you, right? What you want is this technology to actually augment and grow and like get to a better outcome. And so sometimes it's having Claude push back makes me better. And so that's great. like a co-worker, I want somebody to push back when my ideas are not fully formed... If you have a AI that can be a thinking partner, a thinking partner doesn't just agree with you. It should add to you and you should come away at the end of the day having better ideas because you worked with Claude. That should be the hero goal, not just making your ideas 10% better.