Wihout insight on how much enterprise is paying, it is impossible to draw any conclusions. Unless you have any access to their contracts and are willing to share evidence? I find that highly unlikely.
People here throw around crazy numbers - the dude above was claiming they have some insane good margins, numberd that he took out of his ass.
The only evidence I have is that they are incredibly unprofitable, and they keep raising insane amounts of capital like crazy.
There was a leak sometime ago that they were EBITDA positive during a quarter where they didn't pay for part of their compute. And EBITDA is a cute metric to use when depreciation is actually very important to them, as a model from a year or so ago is nearly worthless.
serving models is very profitable (70%+) but the issue is you need to invest in training the next iteration. but so far all of anthropics models have been profitable fully loaded
the vast majority of the labs revenue is from enterprise api usage (theres public sources from the information and ramp). but the risk there is customer concentration, where most of the revenue comes from other tech companies and a chunk of it is from foreign labs distilling
so i am drawing a conclusion that the labs' business model is good, maybe not as great as boosters think it is. if they make real progress on the biosciences like drug discovery that could turn it into an amazing business
This is not evidence. This is random people speculating on Anthropic's margins without any real evidence.
Just because it is on some blog post, it does not make it true.
I wasted the time to read the first blog post. It considers 100% utilization over the course of years to calculate an estimation, and it did not consider depreciation for the model itself. That thing is extremely extensive to create, and after a relatively short amount of time is considered outdated.
if you're going by anecdata, myself and a few of my colleagues (senior ics) switched jobs relatively easily for a solid pay increase within the last year. definitely easier than post-covid
engineering (hardware, software) and data center construction are going through a boom cycle right now and will eventually bust, and so on and so forth
I think this is Jevon's paradox playing out in full force, and I'm quite confident we'll continue to see the growth in well paid STEM jobs despite the anecdata in this thread (anecdotally my group of ICs on tech have all seen our comp grow a lot in the last few years)
Until we saturate the demand for software which may very well be infinite
- directing the model and knowing when its going off the rails
- verification
- setting up the right loops, harness, graphs around the models
economics still apply to hiring people. if all your competitors are using ai and there's market share to fight over and people still are a productivity positive vs. only ai, hiring will increase as we've seen recently with swes
The only thing here that _maybe_ has a durable long term moat is the first point. Even then, I don't see why stronger models can't also have discernment once more companies close the loop with their AI and their relevant company metrics.
The other 3 you literally just get for free as models improve. The "state of the art" of "prompting" changes literally every week. It was loops, then graphs, now its harnesses (and self automating harnesses)? Why are these not just obviated by better models? These are barely skills, and are imo, just the tech equivalent of tabloids advertising 5 minute exercises or pills to get rid of stubborn belly fat. No amount of prompt engineering or graphs or loops could get your previous version of GPT to perform like Fable, and yet somehow Fable can do all of that and more without any random built in "ai engineering skills."
The labor chart for SWEs will look like a slope up as productivity with AI increases, and then past a certain point where AI is like 99% good enough, employment will fall off a cliff. We are already seeing this with junior hiring, and who is to say that AI will magically only ever be as capable as a junior engineer?
i mean the meta point is having "AI skills" is understanding the jagged frontier and operating accordingly
Given that we can keep making infinite abstractions with software if we actually saturate software demand with AI then I don't see why every other industry is cooked (robotics is downstream of software).
I noticed the newer batch of interns tends to have less mechanical coding skills but does well on understanding and planning on the design, DB design, etc.
Again though, these recommendations are so vacuous. If you already are a capable developer, then you can pick all of that up within a week. "Learn AI skills" is such a low effort reflexive response that I'd expect it from a non-technical executive, not from an HN commenter.
You'd be surprised, I used to do a first round technical screen where we let the interviewee use any ai tool they want to build a simple api (and they can oneshot the problem if they just pasted it into any capable agent). I quite liked the problem, it was really a system design problem for senior+ but more algorithmic for junior/mid
The more senior the candidate the less they took advantage of the tools. Most commonly they would manually copy/paste error or syntax errors and then run out of time. One candidate only copy pasted his questions into google and used the AI overview
Junior candidates tended to be overly ai eager, a lot of them oneshotted the problem but were unable to explain any of the details
That being said I think 80% of the skills should come easily to a capable dev that is willing to put in the effort to learn and get used to managing agents. Building agents that perform work themselves is a lot harder (and still pretty unsolved)
I had a similar interview and I find it to be a stupid question. Of course I know AI can oneshot it, and of course I know you know it can. But I'm aware this is an interview so you're probably assessing something about my skills. Pasting the prompt doesn't assess anything so surely that isn't what you want me to do, so I don't do that. Instead I have to magically divine what level of involvement you're hoping to see on a scale from vibecode to handmade. The whole thing provides literally no information to the interviewer about real work processes, because the intelligent candidate will behave differently in this situation than they would at work.
But I don't believe what you've said disproves my point. To me, it boils down to attitude, not so much aptitude when it comes to "AI skills" in this context. You even admitted that some of the solutions could be one-shotted by copying and pasting the problem.
As for building agents that perform work themselves, in my opinion it boils down to understanding the problem space, isolating the key business logic, and determining what the pertinent requirements are. Kinda sounds like looping back around to software engineering skills IMO.
To be clear, I think young people need to learn traditional software engineering skills too (.. maybe having the wrong attitude towards it is even worse than the aptitude!) A junior with no foundational SWE skills will just become a meat wrapper. But thinking more about it I wouldn't discredit these "AI skills" so much. A lot of it is either using AI as a force multiplier or learning the specific skills around deploying it which I'd still categorize as 'AI skills'
building agents is just a completely different ballgame, theres a lot of infra and harness engineering around handling the nondeterministic behaviors that are nonobvious unless you've shipped agents at scale
I'm not discrediting them as being of no value, I am just discrediting the notion that devs who are unable to find work now should be beefing up their AI skills so that they can get a job in the future. "AI skills" as they are talked about are easier and quicker to pick up than working with a library, e.g. React. And if they are not difficult to learn, then they are not a differentiator nor should they be treated as such.
Yeah I've seen a lot of applications and some interviews where what they are looking for the most, by far, is your "AI skills". Questions like: How are you currently using AI day to day?; Can you describe your current agentic setup?; things like that are the only questions asked in the application.
And I don't understand why you would wave away needing to know anything about the codebase's underlying tech because "The AI can take care of all that, we don't need to look at code anymore" while also not believing that any competent developer could prompt AI to get a good setup within a week or two max.
I think it's perfectly fair to evaluate the tools based on how well they live up to the hype that is being pumped out by the sellers of said tools. If they want us to compare their products to a more measured, reasonable take then they can advertise them as that.
While some people are busy bickering about this, the rest of us are using these awesome new tools to get more work done in less time with higher quality than ever.
I don't care what the company claims, I just use the tool the way I want to. I work very closely with the AI. I'll tell it to plan a change, review the plan, then execute. Then I'll test the changes and have it fix whatever I'm not happy with one thing at a time. I'll specify in detail both what to do and loosely describe how to do it or if I'm not sure I'll ask it to plan the change then review the plan and ask for changes if I want them etc. I also review my own PRs before I submit them to colleagues.
This way I maintain full control of everything, it just saves me hours of googling, planning and typing code - which I do miss a bit but I can't really justify writing code myself when I can achieve the same thing just by loosely describing my idea instead. It also saves a lot of time debugging, I think I'm generally a pretty good programmer but the AI makes fewer mistakes than me. It'll often catch some logic error I made during planning and suggest a good alternative.
A lot of developers seem to give up control entirely and then complain that they're no longer in control. Trying for that 10-100x speedup doing weeks of work in a day. I'm happy doing one week of work in a day. There's a limit to how much I can oversee without compromising quality.
I think what a lot of people miss about jevon's paradox is the elasticity of demand of the underlying resource
textiles had jevons paradox, and many more textile workers were employed even when textile machines were being created, until we saturated the demand for cheap clothing in the world and then textile workers were kaput (same for farming, and horses)
software is currently undergoing jevons paradox, but it's very unknown how high the ceiling of demand for software is. web dev might be doomed, but software in general i think is probably limitless
Intelligence is also probably unbounded (atm software and intelligence are very closely tied together). its very possible token spend rides up the curve forever.
Brings up the question of what the intelligence is used for. Humans exploited intelligence for competition. With each other to wipe out other Homo species, mate more and collect resources, with other animals to limit predator impact and gain food. Intelligence will be used offensively by corporations and their people to extract more from consumers (make pricing opaque, terms of service more complicated, etc.) and scams far more sophisticated. The "consumer" will need extra intelligence to fight all that off.
There are only so many meals you can expertly produce, shirts to fold, itineraries to fun places you can execute, but there's a practical infinity of traps to set and avoid.
20 years ago is about when SSDs came to consumer PCs. Which is the last time I remember thinking that my computer became faster than the previous one was when it was new (if by computer we mean the hardware combined with a mainstream up-to-date software stack)
Your computer may not feel faster than a 20 year old PC, but CPUs have gotten way, way faster.
Compare the AMD Ryzen 7 9800X3D to the AMD A12-9800, you go from 4 cores to 8, and you double the power consumption. But it's not twice as fast. It's 10 times faster. Depending on the exact metric, it could be as little as twice as fast or as much as 1000 times faster, depending on the exact operations.
On top of that memory throughput between DDR4 and DDR6 is about 2.5 times faster as well.
Oh and this isn't 20 years apart, this is 7 years apart. 20 years will see almost an exponentially larger gap even still.
I kind of hoped my point would come across more clearly. Obviously my current computer is faster than one from 20 years ago. But, it doesn't really react more quickly. It just does more: higher resolution video, advanced graphics, etc. The moment we add capability we rush to use it all up. This is not _always_ bad when it comes to computing, but I wouldn't mind if people in general had more restraint.
Right -- in the details this is heavily task dependent. If we're talking about word processing, then computing has mostly gotten worse in the past 20 years.
If we're talking about video game graphics it's a bit of a mixed bag.
If we're talking about local LLM models then you need all the power you can get.
That is OP's point. The underlying technology has gotten faster but the computer you use feels the same as 20 years ago because the software that utilizes it does a lot more too.
ohhhhh yeah they had those sata ssds out a few years before the nvme thing came out, I almost forgot about those! but in 2006 yeah that woulda not been cheap, like year zero early adopter type stuff too, pay 1200 for 32 gigs and its dead in 2 years maybe 5 years tops. I got my first one of those quite a few years later once 256gb was comparable in price to 1tb harddisk drive.
Samsung commercialized the first high-volume consumer notebook PC utilizing an SSD in 2006. The Intel X25-M was released in sept 2008 and cut SSD hard drive prices per gigabyte by 60% pretty much instantly
I can still feel performance increases, most recently was when I got an M4 Mac Mini. Compared to my M2 Air, which is no slouch by any means, there was a very perceptible improvement in latency, app opening, general task completion, and so on. But very slight, and some coming from placebo effect/marketing.
These days, I feel the most performance uplift when I switch from Windows/macOS over to a Linux or BSD installation, if only because there is so little wasted effort in running a bajillion little background tasks for crap I didn't ask for. Windows 10 on my PC is highly debloated and optimized, very little in my startup services and so on, but CachyOS on the same hardware feels leagues better.
Windows XP was a decent OS. You could start with a golden ISO (remove a whole lot of crap) and then tweak the registry to your hearts content. Max out RAM, then when you add an SSD, it felt sublime.
Its not Linux, but all hardware worked on Windows XP. Nerds were familiar with tweaking the registry, doing all kinds of hacks and helping all their friends and family. Of course, cleaning out the thousands of viruses from friends computers wasn't fun ..
OpenRouter is the evidence that yes. Its just that the idea is dumb. What you are seeing in VCs salving their own money, buying their investment out with the money they out into more successful ventures.
Look at how the same pools of capital sit on multiple sides of the market: the router, the payment layer, and the model providers... Its all the same club
reply