Anthropic

“Anthropic had this very exciting paper just a couple of weeks ago where actually we're able to get similar results to RLHF without the H. So just actually having a second model provide the evaluation feedback as well. And that's obviously a lot more scalable.”

“Yeah, for a bit of context right now, the frontier models, I mean, they're not public, exactly how many parameters they have. But they're roughly in the 500 of billions-ish, like Lama-III, four or five billion. Anthropic is speculated to be five hundred billio…”

“Out of the list you mentioned, Google is the only company that actually has the product taste to do this.”

“The thing about CodeGen, the big game in town that we saw, ah, six months ago was Clawed Sonnet. It's still actually a big contender. Most are still using it. But O-one, O-one Pro, and O-three meaning all these resilient models are starting to see it's almost …”

“That said, I'm fully working with the assumption that OpenAI will also build its own browser. Anthropic will also try to build its own browser.”

“And 2024, 20 23 especially, and large part of 2024 too, Google had like, maybe a fourth or fifth best models at any moment. So as a startup outside Google, you had access to AI that was better than what Google internally had.”

“I think that scaling Really suggests a kind of smooth curve towards what I expect is kind of human level AI or AGI.”

“If scaling laws are failing, it's because we've screwed up AI training in some way. Maybe we got, ah, we got the architecture of the neural network wrong, or there's some bottleneck in training that we don't see, or there's some problem with Precision and the …”

“One is I think it's really a good idea to build things that don't quite work yet. This is probably always a good idea. We always want to have ambition, but I think specifically AI models right now are getting better very, very quickly. And I think that's going…”

“Claude IV can blow through its context window with a very complex task, but can also, ah, store memories as files or records, retrieve them in order to sort of keep doing work across many, many, many context windows.”

“I think that over time, as AI becomes more and more widespread, I think that we're going to really drive down the cost of inference and training dramatically from where we are right now. Sort of, three X to 10 X gains algorithmically, and in sort of scaling up…”

“With Claude III. VII's Sonnet it was already really exciting to use 3.7 for coding, but I think something that everyone noticed was that 3.7 was a little bit too eager. Sometimes it just really wanted to make your tests pass and it would do things that, that y…”

“You can get pretty big wins from RL. You sort of have another set of scaling laws. It's like you put more and more compute into RL, you can get better and better models out of that.”

“A single bug can like, Derail you for months. Yeah. And when you think about it, like you, the models take months to train. So you can kind of like lose a whole generation off of something that just looks like, ah, you know, it turns out like this piece of you…”

“The PyTorch profiler was, like, pretty good, actually, throughout for a single GPU. You want to, like, profile a GPU, the PyTorch profile would work. But if you wanted to profile a job on 100,000 of GPUs, that, like, hadn't really been done much, and then that…”

“And to do a project like training a large language model requires a lot of people to collaborate on like a really complicated piece of infrastructure that isn't going to be a paper, right? Like you're not going to publish like, oh, I got a slightly, I got five…”

“We don't want to outsource this to some package because A, we're about to go to a bigger scale, like PyTorch, for instance, they had a package for doing this. But we were going to go to a bigger scale than Facebook had been to. And you don't want to have a dep…”

“The downside of having multiple chips is that you have to write the thing multiple times. In theory, you could have abstractions across them, but they're different enough that it's pretty hard to do that. So you can sort of end up, if you do all the workloads …”

“Like, it is the case that, like, the labs right now are really driven by getting good eval scores. And it's hard to make them, and anyone can do it. There's no comparative advantage to having the model to making an eval. So I do think it's actually, like, an i…”

“I think if you get too many people who are specialists, you end up with a lot of effort has to come from the manager, from like the lead to connect everything, and to notice something like, Ah, if we change the architecture here, that would make this, like, ef…”

“The thing we, like, most need is engineers. Almost always, like, throughout, like, the entire history of this field. It's, like, the case that you throw more compute, the thing kind of works. The challenge is, like, actually doing that.”

“Anthropic has rate limits constantly, and people complain about it a lot, and like the reason is like, there's only so much compute we can get on short notice, so you, like, making your inference more efficient is like the way you can serve more users.”

“There's this positive feedback loop where you can train a model, You can use it to make something useful and sell that and get more money, use that to buy more compute, and then use that to train a better model. And we've sort of run that cycle over and over a…”

“Like I think the sort of shift towards more RL is like one paradigm shift in the field, and I think it's, I think there will probably be more. I think a lot of people sort of argue about like, oh, it's like, you know, current paradigm's enough to get us to EGI…”

“There are people who are, like, trying to put stuff out that is, like, as damaging as possible for the model, you know, how can I make it past the filter and get into the model would be totally like secretly useless.”

“Like, you know, the scaling, I think the original scaling laws paper had, like, 11 orders of magnitude, and there was, like, this intense debate on whether it would continue for, like, another point.”

“I used to think that if you had an AI that could solve coding interview questions, it would probably be AGI. I was like, that's what I did to get my job, I could probably do the job. And it turns out like, nope, nope, you solve those, it's shockingly narrow, a…”

“Early on, we hired a lot of people from just, like, all sorts of backgrounds, and I think that people who are just smart and work really hard can learn this pretty fast, but you have to, like, want to. We heard a lot of physicists, for instance. Like theoretic…”

“Inference as a workload in general. Tends to require more HBM bandwidth. You end up doing you sort of the simplest form of sampling since you're going one at a time. You have to load all the weights for every token. And that means you might want a lot of HBM b…”

“And it might be that like, that data ends up more valuable because you, everything that's linked to a lot, you've already got. Like at some point, you're maybe like going for the tails, or you're going for the stuff that no one's ever, like, you know, it's onl…”

“But I do think the change is massive, and I think people, like, don't realize how chip-limited AI, like, research is, or something right now, like, the models that everyone uses, right? If you're using, like, Cloud Sonic four, Cloud Opus four, it's like, it's …”

“I think almost everything kind of has to be done empirically. Like, you can kind of, like, come up with theories, but in practice, like, the first thing you're gonna do with your theory is test it, and most of the time you'll have gotten it wrong. So you shoul…”

“It's just that scale is easier, and it's more reliable, and I think you, we're still seeing really big gains to that.”

“I think one thing that's, like, surprisingly hard and there's very few people who can do is, like, kind of own that whole stack from, like, I understand how the ML is supposed to work and what the learning dynamics are, all the way down to, like, I know the by…”

“I do think at some point there will be, like, some pieces of alignment that, like, you do want to export back into pre-training because that might be a way to, like, Put them in with more strength, like, more robustness, kind of, or more core to the intelligen…”

“Like the public estimates for GP three, I remember, were that it cost five million dollars to train, which you're like, on the one hand, five million is kind of a lot, but it's like a lot for an individual person. It's not really a lot from like a company pers…”

“I think that's a really hard problem. I think it's, like, actually kind of downstream of being able to pick any. I think of it almost, I think one analogy I've heard that I like is, like, putting a steering wheel on a car. It's like, if you don't have a steeri…”

“We're using a cloud provider, but I think it's kind of, it's not actually that different, because one of the things that was surprising to me is you actually have to understand the literal layout. Like, I remember at 1.1 of my co-workers running a clustering a…”

“Yeah, I think we managed to avoid this, and it's pretty collaborative. Like, we're basically all producing one model and kind of can, but I do think at other places there's been some, from what I've heard, there's been some amount of, like, friction between, b…”

“Theoretically, I shouldn't be able to train a better model than that. Like, I'm just going to get the same thing out. So I think that's-”

“Sci-fi movies actually, like, underestimate the impact of it. Like, you always have this, like, one robot that's, like, a human. And I'm like, well, wouldn't you have, like, a billion of them? Like, you can just copy them everywhere. So you should picture, lik…”

“Oh, no, I think a ton about inference, because it basically, like, The problem inference is solving, like, we basically determine the problem inference is solving. We give them a model, and they have to, like, run that fast, and it's very easy to give them a m…”

“There's this idea of scaling laws, which is that you can actually quantify, like, as you put in more compute, more, more data, more parameters, you get models in a very, you got a lower loss, a better prediction of the next word in a very predictable way.”

“I think that, like, the one sort of general intuition I have is, like, compute is the thing that matters. So, like, I think if you throw enough compute at any of these objectives, you're gonna get something that's probably pretty good, and can kind of be fine …”

“The way I usually think about it is anything you can do in post training, you probably should, because your iteration loop, like the ability to make progress is really fast. You can try something, you can try it again, you can try it again.”

“Basically everyone I worked with, like all the safety leads left, which yeah, invited me to go to Anthropic, and that was sort of the reason I joined OpenAI, was because I cared about AI safety and wanted to work with them. So then I went with them to join Ant…”

“The vast majority of the use cases people are using it for though is not coding.”

“And, shockingly, in this batch, the number one API is actually Anthropic. It came out a bit more than OpenAI, which who would have thought?”

“Startups are doing as well, they are actually arbitraging a lot of the models. I had some conversations with a number of founders where before they might have been loyalists to, let's say, OpenAI models or Anthropic, and I just had some conversations recently …”

“They had this hockey stick with the growth over 52%.”

“When you ask Claude code to do something, it will typically spawn and explore sub agent or like multiple ones. And basically each of those are running haiku to traverse the file system and kind of like explore what's there. And they're doing it in their own co…”

“There was some stuff from Mercury that like, 70% of startups are, you know, choosing quad as their model of choice.”

“I haven't pulled the data on this, but I would bet the majority of agents are actually prompted by quad today in the form of subagents.”

“I think plan mode probably has a limited lifespan.”

“We use the Quad Agent SDK to automate pretty much every part of development. It automates code review, security review it labels all of our issues. It shepherds things to production. It does pretty much everything for us.”

“And they built it in, I think something like 10 days. It was just like a hundred percent written by quad code.”

“Yes, because I didn't have to build a UI. So it was just me.”

“I think the first kind of big example where it worked is our plugins feature was entirely built by a swarm over, over a weekend. It just ran for like a few days. There wasn't really human intervention and plugins is pretty much in the form that it was when it …”

“And I asked it, what music am I listening to? He wrote some, like, Apple script to script my Mac and look up the music in my music player. And this was sauna 3.5.”

“At Anthropic, the way that we thought about it is we don't build for the model of today. We build for the model six months from now.”

“And I think Quad is going to get increasingly good at kind of figuring out these kind of product ideas for you, just because it can look at feedback, it can look at debug logs, it can kind of figure this out.”

“Every single function on our team codes, like our PM's code, our designers code, our EM codes, our like everyone, our finance guy codes, like everyone on our team codes. We're going to start to see this everywhere.”

“A really common pattern is I ask claw to build something. It'll look in the code base. It'll see some engineer touch something in the git flame and then it'll message that engineer on Slack. Just like asking a clarifying question. And then once it gets the ans…”

“The way I would frame it is think about the thing that the model wants to do and figure out how do you make that easier? And that's something that we saw, you know, like when I first started hacking on Claude code, I realized like this thing just wants to use …”

“I remember early on, this is maybe six months ago, I tried to get rid of bash output, just internally, just to, like, summarize it, because I was like, these giant long bash commands, I don't actually care. And then I gave it to anthropic employees for a day, …”

“There were some other stuff from like semi-analysis that four percent of all public commits are made by quad code.”

“You know, as Anthropic, I think for Ant, the bet has been coding For a long time. And the bet has been the path to save, to safe AGI is through coding. And this is, this has kind of always been the idea. And the way you get there is you teach the model how to …”

“And since Quad Code came out, productivity per engineer at Anthropic has grown a 150%.”

“Internally, if you look at like technical employees, they all use quad code every day. And even non-technical employees, I think like half the sales team uses quad code.”

“If you look at Anthropic overall, it ranges between, like, 70 to 90% you know, depending on the team. For a lot of teams, it's also, like, a hundred percent. For a lot of people, it's a hundred percent.”

“All of CloudCode has just been written and rewritten and rewritten and rewritten over and over and over. We unship tools every couple weeks. We add new tools every couple weeks. There's no part of CloudCode that was around six months ago. It's just constantly …”

“And quad code, like took the heap dump. It wrote a little tool for itself to like analyze the heap dump. And then it found the leak faster than I did.”

“So one is it operates as a research lab. So the product org was teeny, teeny, tiny. It's really all about building a safe model. That's all that matters. And so this idea of just being very close to the model and being very close to development and being not t…”

“And, you know, of course, Anthropic and OpenAI and Google, they're exploring recursive self-improvement, but Typically at that level of having the, you know, having to train a new model for every step of self-improvement that they do.”

“AI hasn't passed it yet, but we got to 55%, which is almost two percentage points higher than the previous state of the art. Which came out just last week from Anthropic with Claude Opus 4.6. They got 53.1%, and we got 55% on it.”

“Our view is that I think the coding aspect is only 20% of the job, right? I think like taking an app to production is like really, really hard.”

“Claude in Chrome MCP is one of the worst pieces of software I've ever used. You know, every time it would try to do an action, it would think and think and think. There was crazy context bloat. Often it wouldn't even do anything, but it would take two to three…”

“Apparently the stake that the bankruptcy has of those shares is worth more than the whole, than all of the entire fraud by a lot.”

“I think it was a series C when they finally established this thing called the long-term benefit trust, which is not a nonprofit foundation. It's actually what's called a perpetual purpose trust, which is a different legal category, but the same idea outside tr…”

“Every six months, delete your Cloud MD. Delete your skills. Delete your hooks. See what the model does, and it might surprise you. And actually for Opus Five, this is something we really do recommend, is just try deleting all of these things, because the model…”

“So coding is solved for the kind of coding that I do. It's not solved for everyone. You know, there's still code bases that are like super deep systems code bases where quad still struggles. There's distributed systems where quad still struggles. There's reall…”

“So yeah, we deleted 80% of the system prompt.”

“We actually have Cloud maintaining itself now. And the way we do this is we have a Slack channel where we just had Cloud start a bunch of different routines to maintain its own code base. And we actually do this for the CLI, for the iOS app, for the Android ap…”

“This was like Sonnet 3.5. At the time, that was an incredible coding model. That was like the best coding model that exists. Nowadays, it's, you know, a pretty terrible coding model by modern standards. But I think that was like the first great coding model th…”

“So essentially, if you combine a well-aligned model, so this is, like, essentially three years of research into alignment, With a prompt injection classifier, which we run for all traffic, and what this is doing is it's based on Chrysola's mechanistic interpre…”

“And actually in this case, Quad also decided to live blog it. So what it did is it created a Slack channel internally, and it started just posting screenshots every few minutes of its progress.”

“And for modern models, that's actually really not the way to do it. You want to go a little bit higher level. You want to describe the task. You want to describe the guardrails. You want to describe, like, the exit criteria, and then just go with the model coo…”

“You guys got, took Arc AGI three to 30%, which is incredible.”

“For five, one example of something it does that I think no other model has done is it runs for a very long period of time. And especially when you combine Opus Five with auto mode, it's just like incredible. Like it can go for days, weeks, months at a time. It…”

“And so now we have every day, maybe 20 or 30 of these routines. It's running across all of our code bases and It's not totally there yet, but we're on the path to fully automating the maintenance of our apps by doing this. And this is, again, hundreds of agent…”

“What a dynamic workflow is, is essentially we have the Bun runtime. We use Bun as a sandbox, and we start a virtual machine within Bun, and we let Cloud start a lot of agents and orchestrate them.”