They reported a $47B run rate in May. This article claims $4.7B revenue in Q1 and $11.5 B in Q2. It all aligns with a very high growth rate. Here's one set of numbers that fits (though I would guess the actual growth was spikier than this):
Jan $1.0 B
Feb $1.5 B
Mar $2.2 B
Apr $3.0 B
May $3.9 B
Jun $4.6 B
I think it's not an option. The benefits are just too large for me.
>I do know we will never go back to mainly programming through code again, that’s for sure.
I think, in some niches, i.e. where there's something not well represented in the training set, it still makes sense to write code by hand. But I am not sure that it will continue.
He said something to the effect of "I love open source and open models and we'll do open models when it makes sense and closed models when it makes sense" in a recent Q&A.
Given that his company has already released open models, I find it funny that, as you described it, his remark communicates absolutely nothing whatsoever. Not sure what the question was, but this was an artful non-answer.
My approach has changed throughout the course of this project. Throughout most of the project, we were working off of a c2rust translation of Postgres to Rust. That gave us a bunch of Rust code that was unsafe but did pass the Postgres test suite and was fast. c2rust had split Postgres into 1000 different crates. We then went through 1 by 1 and rewrote each crate into idiomatic rust.
This naturally lended itself to a suite of skills to describe how to rewrite a crate from unsafe rust to idiomatic rust. The main three skills I had were 1) a skill for identifying the next crates to port 2) a skill for rewriting a crate and 3) a skill for auditing a crate and making sure there weren't any outstanding issues.
My exact approach for managing subagents changed throughout the project. Initially I was doing parallel coding sessions with Conductor. After dynamic workflows came out, I used that as it was really easy to spin up dozens of parallel subagents and manage it from a single orchestrator. Over time I switched from using dynamic workflows to manually spinning up subagents from a central agent. The issue with dynamic workflows is they waterfall. Each step needs to finish before the next one starts. By manually spinning up subagents, I could have claude start porting a new crate as soon as a prior subagent finished.
Yeah, like to see more disclosures about this. Not to mention the whole background debate on the real costs for the AI companies, at what point prices go up to reflect that, or the possible consequences of being overly reliant on this technology.
Actually the inverse. I initially gave claude an outline of what I wanted, had it do some research into how to write idiomatic rust, and then had it draft a series of skills to do the work. I would then try out the skills, audit the results, and then give claude feedback based on what I was seeing. Once I started getting runs where the results were working, I would start to scale things up and audit things with an exponential backoff.
I tried a similar approach before, but it didn't work for me. I didn't get a lot of speedup if any from it. IMO, to get productivity you need some kind of YOLO mode (in a sandbox).
IMO, the goal should be to outsource as much work to the model, as possible, while minimizing effort required to understand and review what is did. For example: ask the model to find out why a bug happens, figure out proof of concept for thing X, incrementally optimize something, do a well specified refactoring with some guide, and similar things.
IMO, what people say about creating loops is a very similar thing. You maximize the work done by the model, while minimizing the amount you need to do to control it.
There's examples both in his writing and also in his appearances on podcasts, interviews, etc.
I'll cherry pick a couple:
“When these new models ‘reason,’ they break a user’s input and break into component parts, then run inference on each one of those parts.” [1]
This is not at all how test-time compute works. At best, this is a very loose metaphor that he may have used out of convenience. This might sound a bit pedantic to point out, but this is a very basic thing that he's getting wrong (presumably at least, again it could be that he just used a poor metaphor).
A less pedantic example would be his claims related to gpt-5/chatgpt auto-routing. He argued that having a router means OpenAI can no longer cache static prompts, because the user prompt has to come before the hidden instructions [2]. This is just not at all how this works at inference-time. There is no evidence that the standard approach of system>developer>user instruction hierarchy has changed, the public API and caching docs maintain this.
But even more broadly, it suggests he is reasoning about kv/prefix caching at the wrong level of abstraction. It's true that conventional prefix caching does require a stable prefix, so yes, if you literally put variable user content before the static prompt, you would destroy the cacheability of that static prompt.
But that is exactly why inference systems are designed to preserve reusable prefixes where possible (via checkpointing or similar), and why serving systems care so much about prefix caching. This is also a big part of how disaggregated prefill/decode infra works where cache-aware routing is critical. His argument treats a bad prompt layout as if it were a necessary consequence of routing, rather than an avoidable implementation choice.
A router can read the user request, decide which model path to use, and then construct a normal downstream model call with stable static instructions first and user content later. Treating that as impossible implies a fundamental architectural misunderstanding.
Do you think it's not slowing? Do I miss anything really important?
My understanding is that we have now is incremental improvement on thinking models which appeared more than a year ago. Of course, a breakthrough might happen, but I don't see one yet.
The most important thing I would point to is Mythos et al and the wave of vulnerabilities that have been discovered in the past couple months. It’s a completely unprecedented event, brought forth almost entirely by improvements in the models themselves.
That said. keep in mind, I’m talking about over the past two years. With Claude code and the capabilities gained since December of last year, there have been incredible gains in the capabilities that are now available. Demand for inference is higher now than it was a year ago, because capability has improved. A specific criticism that I would hold is that claiming that progress with LLMs is slowing, prior to that point, is embarrassingly wrong in my view.
One could argue that the model capability improvements are slowing, and all the improvements were in harnesses. I think that’s a stronger argument, but I have a few problems with it.
1. Utility is utility. Whether that comes from the model or the harness is irrelevant when making claims about utility. I don’t think that’s a useful distinction most of the time, but especially when talking about the technology as a whole.
2. Marginal intelligence gain is different than marginal utility gain. It’s estimated that intelligence grows logarithmically relative to investment. However, the utility of a marginally more intelligent model may grow exponentially, because once behavior crosses a reliability threshold, it unlocks new capabilities.
3. Even on those terms, it’s not clear to me that frontier capabilities are slowing down. With Mythos and its contemporaries, we have been seeing a vast change in the security industry as vulnerabilities are discovered at an unprecedented rate. OpenBSD vulnerabilities, more Firefox vulnerabilities found in a single month than the past two years, critical Linux vulnerabilities. It’s hard for me to look at the effects there, a radical new capabilities baked into the model itself, and see stagnation. A part of the reason it might feel like it’s slowing down is because we plebs don’t have access to the top models.
The article says in the second section that the author did not have access to Mythos. I think it’s dangerous to rely on claims made by others without even bothering to read them first, let alone check.
It found hundreds of vulnerabilities in Firefox, according to Mozilla: how does Mozilla benefit? It found a 27 year old vulnerability in OpenBSD. How do they benefit from that? Is that made up? Are the maintainers of those codebases lying for the benefit of Anthropic’s IPO? Is copy fail a fabrication by big AI? The 12 OpenSSL vulnerabilities found in January?
Im not sure whose claims you think I’m relying on. I trust Firefox that they’re not overstating the number of CVES they’ve found. Same for OpenSSL. The OpenBSD folks definitely don’t seem like the types. I’ve not known Linux to fabricate CVEs either. I think my sources are fine.
