Thanks for the write up. I think you might be happy to hear that ffmpeg has scene detection built in. But you might be unhappy given how much work you did. In my experience it works pretty well.
An upscale pipeline seems like the next job, by the way! you can pull a lot of quality out of those old videos with modern tools. Enjoy.
It's been a long time since they were good. But Europe definitely needs a homegrown frontier model company, one way or the other. I consider them Tier 2 right now.
I am struggling with this because I have an Anthropic offer vs another equivalent offer that is all cash.
But project out forwards.
- What happens when Google builds a similar model? Or even Meta, as far behind as they are? They have more than Anthropic in cash flow to pour into these models.
- What happens when OSS is "enough" for most cases? Why would anyone pay 60% margins on inference?
What is Anthropic's moat? The UX is nice, but it can be copied. And other companies will have similarly intelligent models eventually. Margins will then be a race to the bottom, and the real winners will be GPU infra.
If they outlast the competition it might be a really hard market to enter. Models are expensive to train, and they'll get outdated. You're on a time limit to make a profit off of it
Google and Meta might be the only real threats against this given how much cash they have and so far Meta is just flopping
I've been in this situation before. Anthropic has a stupid business model but the market can stay irrational longer than you can stay solvent. If you get in there you will be aligned with people who structurally do not lose.
Big picture, sure. We can talk about the millions that corporations will make and who's going to do what. But you're a person. $1 million in options is probably meaningful for you. Companies aren't IPOing, but the secret is that they're still paying employees cash for their options. SpaceX employees have had what's called a tender, which means they get to sell some of their hypothetical SpaceX options for cold hard cash in the bank that you can use to pay your mortgage. There's zero guarantee that Anthropic will do such a thing before the bubble bursts, but if they do, and you're there, who cares about a software company moat when you have enough money to buy a castle in Napa and pay to have a real actual moat with water in it and crocodiles, if that's what you want.
Others are made of different stuff, and are going to go right back to work, even though they could go off to a beach for the rest of forever somewhere.
> who cares about a software company moat when you have enough money to buy a castle in Napa and pay to have a real actual moat with water in it and crocodiles, if that's what you want.
Doesn't this require their private market valuations to go well into the trillions?
No, it’s not. This is a dangerous perspective, usually held by engineers who think that accounting doesn’t matter and don’t understand it.
You MUST accrue the lifetime value of the assets against the capital expense (R&D in this case) to determine the answer to this question.
The company (until this announcement) had raised $17B and has a $14B revenue rate with 60% operating margin.
It is only negative on margin if you assume the prior 14B (e.g. Claude 4.6 plus whatever’s unreleased) will have no value in 24 months. In that case, well, they probably wasted money training.
If you think their growth rate will continue, then you must only believe the models have a useful 9 months or so life before they are break even.
Anthropic is, according to Dario, profitable on every model <<—- they have trained if you consider them individually. You would do best to think “will this pattern continue?”
What is the lifetime value of an individual pretraining run, and what is the cost to do it? Whether it is a net positive seems to still be an open question.
Actually there is a chart of answers to this question, because the frontier providers have been delivering new models for some time. The answer is that so far they have been net positive.
Sorry - if a model costs (say) 20B to train, lasts 12 months before it becomes obsolete, generates 2B/month revenue, but with 1B/month inference costs, then it has lost 8B.
Or are you suggesting that in fact each model comes out ahead over its lifespan, and all this extra cash is needed because the next model is so much more costly to train that it is sucking up all the profits from the current, but this is ok because revenue is expected to also scale?
I'm not suggesting, I'm saying that this is what has happened so far, full stop, based on multiple public statements from people like Dario.
Basically every model trained so far has made money for Anthropic and OpenAI. Well maybe not GPT4.5 - we liked you but we barely knew thee..
The cash spend is based on two beliefs a) this profitability will continue or improve, and b) scaling is real.
Therefore, rational actors are choosing to 2-10x their bets in sequence, seeing that the market keeps paying them more money for each step increase in quality, and believing that either lift off is possible or that the benefits from the next run will translate to increased real cash returns.
What's obscure to many is that these capital investments are happening time shifted from model income. Imagine a sequence of model training / deployments that started and finished sequenced: Pay $10m, make $40m. Pay $100m, make $400m. Pay $1bn, make $4bn. Pay $10bn, (we are here; expectation is: make $40bn).
If you did one of those per year, the company charts would look like: $30m in profits, $300m in profits, $3bn in profits. And in fact, if you do some sort of product-based accrual accounting, that's what you would see.
Pop quiz, if you spend in the first month your whole training budget, and the cycles all start in November, what would the cash basis statement look like for the same business model I just mentioned?
-$10m, -$60m, -$600m, $-6bn.. This is the same company with different accounting periods.
Back in reality, shortly into year 1, it was clear (or a hopeful dream) that the next step (-100 / +400) was likely, and so the company embarked on spending that money well ahead of the end of the revenue cycle for the first model. They then did it again and again. As a result naive journalists can convince engineers "they've never made money". Actually they've made money over and over and are making more and more money, and they are choosing to throw it all at the next rev.
Is it a good idea or not to do that is a question worth debating. But it's good to have a clear picture of the finances of these companies; it helps explain why they're getting the investment.
Its really weird how you all are begging to be replaced by llms, you think if agentic workflows get good enough you're going to keep your job? Or not have your salary reduced by 50%?
If Agents get good enough it's not going to build some profitable startup for you (or whatever people think they're doing with the llm slot machines) because that implies that anyone else with access to that agent can just copy you, its what they're designed to do... launder IP/Copyright. Its weird to see people get excited for this technology.
None of this good. We are simply going to have our workforces replaced by assets owned by Google, Anthropic and OpenAI. We'll all be fighting for the same barista jobs, or miserable factory jobs. Take note on how all these CEOs are trying to make it sound cool to "go to trade school" or how we need "strong American workers to work in factories".
> Its really weird how you all are begging to be replaced by llms, you think if agentic workflows get good enough you're going to keep your job? Or not have your salary reduced by 50%?
The computer industry (including SW) has been in the business of replacing jobs for decades - since the 70's. It's only fitting that SW engineers finally become the target.
Most folks don't seem to think that far down the line, or they haven't caught on to the reality that the people who actually make decisions will make the obvious kind of decisions (ex: fire the humans, cut the pay, etc) that they already make.
I think a lot of people assume they will become highly paid Agent orchestrators or some such. I don't think anyone really knows where things are heading.
I agree with you and have similar thoughts (maybe, unfortunately for me). I personally know people who outsource not just their work, but also their life to LLMs, and reading their exciting comments makes me feel a mix of cringe, fomo and dread. But what is the engame for me and you likes, when we finally would be evicted from our own craft? Stash money while we still can, watching 'world crash and burn', and then go and try to ascend in some other, not yet automated craft?
Yeah, that's a good question that I can't stop thinking about. I don't really enjoy much else other than building software, its genuinely my favorite thing to do. Maybe there will be a world where we aren't completely replaced, we have handmade clothes still after all that are highly coveted. I just worry its going to uproot more than just software engineering, theoretically it shouldn't be hard to replace all low hanging fruit in the realm of anything that deals with computer I/O. Previous generations of automation have created new opportunities for humans, but this seems mostly just as a means of replacement. The advent of mass transportation/vehicles created machines who needed mechanics (and eventually software), I don't see that happening in this new paradigm.
I don't think that's going to make society very pleasant if everyone's fighting over the few remaining ways to make livelihood. People need to work to eat. I certainly don't see the capitalist class giving everyone UBI and letting us garden or paint for the rest of our lives. I worry we're likely going to end up in trenches or purged through some other means.
If you want to know where it's headed, look at factory workers 40 years ago. Lots of people still work at factories today, they just aren't in the same places they were 40 years ago and now req an entirely different skill set.
The largest ongoing expense of every company is labor and software devs are some of the highest paid labor on the planet. AI will eventually drive down wages for this class of workers most likely by shipping these jobs to people in other countries where labor is much cheaper. Just like factory work did.
Enjoy the good times while they last (or get a job at an AI company).
I’m someone who’d like to deploy a lot more workers than I want to manage.
Put another way, I’m on the capital side of the conversation.
The good news for labor that has experience and creativity is that it just started costing 1/100,000 what it used to to get on that side of the equation.
If LLMs truly cause widespread replacement of labor, you’re screwed just as much as anyone else. If we hit say 40% unemployment do you think people will care you own your home or not? Do you think people will care you have currency or not? The best case outcome will be universal income and a pseudo utopia where everyone does ok. The “bad” scenario is widespread war.
I am one of the “haves” and am not looking forward to the instability this may bring. Literally no one should.
Well he also thinks $10.00 in LLM tokens is equivalent to a $1mm labor budget. These are the same people who were grifting during the NFTs days, claiming they were the future of art.
lmao, you are an idealistic moron. If llms can replace labor at 1/100k of the cost (lmfao) why are you looking to "deploy" more workers? So are you trying to say if I have $100.00 in tokens I have the equivalent of $10mm in labor potential.... What kind of statement is this?
This is truly the dumbest statement I've ever seen on this site for too many reasons to list.
You people sound like NFT people in 2021 telling people that they're creating and redefining art.
Oh look peter@capital6.com is a "web3" guy. Its all the same grifters from the NFT days behaving the same way.
You don't hate AI, you hate capitalism. All the problems you have listed are not AI issues, its this crappy system where efficiency gains always end up with the capital owners.
Well I honestly think this is the solution. It's much harder to do French Revolution V2 though if they've used ML to perfect people's recommendation algorithms to psyop them into fighting wars on behalf of capitalists.
I imagine llm job automation will make people so poor that they beg to fight in wars, and instead of turning that energy against he people who created the problem they'll be met with hours of psyops that direct that energy to Chinese people or whatever.
You’re missing, somewhat gleefully, most of the history of western art, which you could imagine as split between patronage-based art (have you heard of the Sistine Chapel, for instance?) and vernacular art - where things like genre storytelling and family portraits come from.
Broadly speaking, vernacular artists work for a fucking living; it’s rare there (like in most pursuits) to get super rich. We can’t all be David Baldacci or Danielle Steele.
NB: Thanks to Neal Stephenson for the best essay on this. He calls genre artists “Beowulf” artists.
Am noob. The phrase "folk art" never satisfied me. Is it really all that different? But I didn't have the gumption to learn more. Happily, the critics and philosophers did:
If that's what they're tuning for, that's just not what I want. So I'm glad I switched off of Anthropic.
What teams of programmers need, when AI tooling is thrown into the mix, is more interaction with the codebase, not less. To build reliable systems the humans involved need to know what was built and how.
I'm not looking for full automation, I'm looking for intelligence and augmentation, and I'll give my money and my recommendation as team lead / eng manager to whatever product offers that best.
Now I use claude with agent orchestration and beads.
Well actually, I’m currently using openclaw to spin up multiple claudes with the above skills.
If I need to drop down to claude, I do.
If I need to edit something (usually writing I hate), I do.
I haven’t needed to line edit something in a while - it’s just faster to be like “this is a bad architecture, throw it away, do this instead, write additional red-green tests first, and make sure X. Then write a step by step tutorial document (I like simonw’s new showboat a lot for this), and fix any bugs / API holes you see.”
But I guess I could line edit something if I had to. The above takes a minute, though.
That sounds like wishful thinking. Every client I work for wants to reduce the rate at which humans need to intervene. You might not want that, but odds are your CEO does. And babysitting intermediate stages is not productive use of developer time.
Well, I want to reduce the rate at which I have to intervene in the work my agents do as well. I spend more time improving how long agents can work without my input than I spend writing actual code these days.
Full automation is also possible by putting your coding agent into a loop. The point is that an LLM that can solve a small task is more valuable for quality output, than an LLM that can solve a larger task autonomously.
Confused as to why you wouldn’t integrate a local vlm if you want a local llm as the backbone. Plenty of 8b - 30b vlms out there that are visually competent.
An upscale pipeline seems like the next job, by the way! you can pull a lot of quality out of those old videos with modern tools. Enjoy.
reply