Their financial statements suggest the answer, even when you take into account the cost of training the model, is 'not at all'. They are profitable, the main questions for their valuation are 1) can they keep their margins, 2) how big can they grow?, and 3) can this growth match what they have committed to?
Question one is how do their investors get paid back. I found Series E through Series H totaling $111.5 billion in investments. Since series E was $3.5 I’m assuming all the previous rounds added up to less than that, so let’s say 115 billion at the outside.
That debt will transfer to public shareholders at the IPO but also will double or triple in the process, as the investors expect to be paid very, very well. So the current shareholders are looking to extract something over $400 billion from the IPO.
With the current revenue they might pull off that P/E. But they’ll still be extracting some of that value from the customer base, because shareholders will demand it.
Not at all, the subscriptions have an adjusted gross margin of approximately 85%. Very profitable! The training is a different story though. How long will the subs last if they stop training? Also, why on earth would they put on the breaks and try to be profitable when they have plenty of funds and are in the growth stage?
Opus 5.5 came out right at the end of the quarter, September 22. I'm confident that it caused more usage growth since then. I'm sure there are a lot of people rooting for their IPO to fail but they are still clearly in the lead on coding and will continue to do well until someone definitively passes them.
> I'm sure there are a lot of people rooting for their IPO to fail but they are still clearly in the lead on coding and will continue to do well until someone definitively passes them.
Then why all the rush to IPO then?
The strongest chinese models from DeepSeek, Kimi and z.ai are already good enough to run for hours on most coding tasks cheaply.
For capital to win they need to get their money out of Anthropic before the bubble deflates. This means dumping the risk on retail. There is only a finite amount of money available to buy AI stock, and the SpaceX IPO sucked a lot up, so it's a fight between Anthropic and OpenAI who is going to get what's left. The loser is left without a chair to stand on when the music stops.
> they are still clearly in the lead on coding and will continue to do well until someone definitively passes them.
I think the first implying the second might be a bit of a stretch. At some point, companies are going to start looking at their AI spending and start negotiating deals. OpenAI is right behind them and Anthropic is around 8x more expensive (Sol 6.1 vs Opus 5.5 based on Artificial Intelligence's cost per task; cost per token is a somewhat useless metric when different models use vastly different numbers of tokens for the same thing). Chinese models like Deepseek 4.1 Flash are cheap, can be self-hosted, and while certainly behind Claude Opus 5.5, are still good.
Lots of companies buy "inferior" products like Microsoft Teams. Over the next 5 years, it seems likely that many models will get really good to the point that people often won't be able to tell any difference. Even if something isn't "definitively" better, companies might say "our engineers can't tell the difference and we're cutting our costs by 85%."
It's not absurd to think that a company would use its own harness or an open source harness and/or put a proxy in to route requests variably. For two months, they have engineers using different models unbeknownst to them and they measure things like how often the engineers have to correct or redirect the model. At the end of two months, they might find little difference in performance and switch to a much lower cost model.
It's not like it'd be crazy hard to analyze workers frustration as they did their jobs and figure out if one model was actually better. They could log all the conversations and it's not like AI can't easily determine things like "no, that's completely wrong" as a phrase showing that the agent was really off base.
I think a big part of the next 5-10 years is going to be about cost/efficiency.
True, but I also believe that one can generally hire "more intelligence" if one is willing to pay more in salary and benefits, which most companies aren't. Basically, the original argument wr Microsoft Teams.
Ultimately the question is whether there's a pareto optimal point for intelligence models. There's 2 potential worlds:
1. Frontier models are so good that the token cost of an average task becomes miniscule, so users are incentivized to use frontier models because their sheer intelligence makes them cheaper to use.
2. There is a pareto optimal point where even if you spend more on tokens overall profits decrease because token costs rise faster than revenue does.
> It's not absurd to think that a company would use its own harness or an open source harness and/or put a proxy in to route requests variably. For two months, they have engineers using different models unbeknownst to them and they measure things like how often the engineers have to correct or redirect the model. At the end of two months, they might find little difference in performance and switch to a much lower cost model.
Deepseek 4.1 flash is so fast and cheap. For coding tasks, it only seems to lack the robustness of Opus 5.5 rather than 'intelligence' - it fails to double check critical assumptions, stops before the task is over, and doesn't do enough speculation or exploration on its own initiative. Yet it can do all these things when prompted.
There seems to be another magic ingredient for improving agents beyond model size & efficiency, and so far only OpenAI, Anthropic, and maybe Google know what it is.
Anthropic can remove all access to their LLMs and start replicating entire software companies. Anyone relying on open weight models is at risk of a security breach from even slightly better models. It is ironic when people think Anthropic is the one without moat while they give all their data away.
There's a stock market truism that anyone that has real alpha doesn't sell it or give it away. I feel like there's an analogue here, that if it was just that easy for them to pack up their toys and go conquer the world, I can't imagine they would have been in the business of selling tokens at this point
IMO if llms keep improving, there's just not going to be a business in most of software anymore. It wouldn't matter if Anthropic tried to replicate all the saas companies because the category is soon to be dead. Why would you pay for someone else's software when you can prompt your own to your tastes? Every bit of software would just have one user
Opus 5.5 is amazing and has also dropped our bill dramatically. Our engineers are spending 1/3 as much as they were before and seem actually happier with the results.
I use Codex and Claude with frontier models. Even the slight difference in an AGENTS.md or CLAUDE.md or whatever tiny amount of difference in guidance you provide the agents feels to me to outweigh any difference in each agents's coding ability. i.e. for practical purposes, to me, they are equivalent with very slightly differing UX.
Accidentally leave one on medium when it should have been high for a specific task has an outsized impact compared to the variation in the models.
Opus 5.5 may be occasionally better than other models on code (I find this debatable though, because Astra is very good and will often do better imo), but I would wager there's way more money to be made on things like image and multimodal tasks or at least nearly as much, and frankly even Gemini 3.8 flash can spank Anthropics best models on visual tasks and be absurdly cheaper; that also ignores the increasing aggressive classifiers Dario seems to like which block a lot of stuff and I've seen drive a number of zai subscriptions because blue teaming is increasingly very important.
I expect growth to go down. Opus 5.5 is not only better, but a lot cheaper. My company was spending quite a bit on overages with Fable 5. Now we are no longer doing that.
I can't disagree enough. I've gone from filling 4 Ultra accounts every week with Fable to barely managing to fill 50% of one with Opus. It's so damn efficient I'm spending 1/4 of what I was before with Anthropic.
For me, that's amazing, but maybe it explains these numbers and maybe it'll finally change when Fable 5.5 forces me back into high usage territory. If I even really need it's abilities as more than an advisor (words I didn't think I'd imagine myself saying back when Fable 5 was dropping).
The revolution always quietly happened last week or last quarter. Only the most die-hard vibers know so far.
The total market for code generation might just not be enough. The investments levels really demand getting rid of whole classes of professions. Not just 80% of programmers.
i think its a lot deeper than that, compared to deepseek-4.1-flash, glm-5.3-flash and the recent minimax-m3.1-flash [0], the comparison in price (15x cheaper inference cost) and the productivity gains are very similar if not more profound depending your prompt topic (cyber, biochem, etc)
This is naive. AI is one of those few things that does have moat. Its shockingly hard to hire people and create a good company culture that does frontier training. There's only like 3-4 labs in the world and the rest are distilling.
Even in next 3-4 years, the company that has the best ecosystem of researchers, culture, shared knowledge and processes will be able to churn out models in a way that can't be replicated.
my simplified take is the headwind is on two ends - smaller models being sufficient enough & open weight models then second utility of 'a.i' is maximal when deployed with humans i.e engineers applying a.i on top of current applications putting an agent layer on top.
the second part is not sexy - but we see all a.i labs having fde arms e.g ode, microsoft frontier company etc but then again that's the part you can't hype.
For big ones it’s often network effects and/or friction. Those can be pretty good moats even with LLMs. These models are also nowhere near the point where they can actually reverse engineer complex products.
It was six months last year, it’s on the order of weeks now. Presumably they have unreleased models that are a step ahead still. The moat is capacity and how much of an ecosystem the can build around the llms; eg anthropic figma tool.
How much easier my life would be if Chinese open models were even just 1 year behind outside of X posts and one-shots.
My easy litmus test is would the year old model (Opus 4.5) being open sourced today would eat the traffic of every single Chinese model... and the answer is pretty much yes.
Every Chinese model would throw away their work just to post-train or distill Opus 4.5 and get a more reliable model.
AI growth is unlikely to be linear. We have seen a lot of hype in the last 12-18 month that give us ideas like "tokenmaxxing". But this is the same cycle that we have been before: hype -> disillusionment -> someone figures out a better way to use it -> next hype phase starts.
As it stands, even the people closest to this technology have barely figured out how to use it effectively.
AI has historically had a habit of disappearing for ten years and coming back once memories fade and the new story sounds enough different from the old one.
I don’t think this era is going to end with us all unemployed, but I do plan to be retired before the next cycle.
I think most people who care or are tracking it would say we've hit it, and those who disagree just have their goalposts a lot more clearly defined and are quietly waiting.
I don't understand how anyone can call these AGI. I've always had the view that AGI == Independent human-level intelligence.
At some point someone moved the goalposts for this to include LLMs that emphatically are not that. I think it was Altman when they came out with Astra?
they aren't as intelligent as a human, but they can throw shit at the wall faster, and get something that sticks faster. they are superhuman in areas that require trial and error, like hacking.
everyone stopped talking about the Turing test once ChatGPT came out and broke right through it while still being very very dumb. AGI is just too blurry to define.
Please do a youtube series where you ask random people on the street some questions and have them go against Opus 5.5. The model is definitely smarter than me in many ways- it would obliterate me in trivia, could write a poem in iambic pentameter before I could remember or look up what exactly the rules for it are, would probably dominate me at chess or football strategy, would do a better job at taxes, and for sure writes better code outside my small little slice of specialty. It definitely parses and diagnoses logs better than me unless I wrote all the loggging statements myself.
I have to ask, what humans do you know that have more general intelligence? This isn't a dream team where you can say "Magnus Carlson can beat it at chess, and Linus can write better C code", this is a comparison across all domains and levels of intelligence. Even its IQ is estimated at 141, which is seriously higher than the average human. The average human's reading level is atrocious. The average human cannot write a standard 5 paragraph essay about a random topic.
If I choose an arbitrary video game, board game or whatever can Opus play it?
My understanding is that models will often still forget games rules etc unless you design a harness specifically to assist them and prevent them going off the rails. I was going to use chess as an example but it sounds like they are good enough to do that now, albeit I'm not sure if they have a special harness for it.
It's not about how "smart" they are, or how good they are at recall. It's about whether I can teach a model to understand and execute an arbitrary, general task regardless of whether they have been trained on that task before.
Disregarding the above, the fact models still require a human prompter/decision-maker is defacto proof they aren't AGI in my view. If they were humans would be obsolete in decision-making.
I guess im doing a terrible job of explaining my view.
If these were AGI we wouldn't be talking about them helping engineers, engineer jobs would just not exist. The AI would be superior in every way that mattered. That's not the case yet, so IMO we don't have AGI.
The very first goal post for AGI must be that it can act, learn, and grow independently. They must be self healing, self correcting, and anti fragile. That's definitely what the G in AGI is about. But all models fall into decoherence if you let them work on their own too long.
If you left an AI alone for a long time and when you came back they had built up a working society and learned new useful things, then that's the proof it's AGI. From this definition you only know it's AGI after you let it loose.
You are describing ASI, AGI is just the stage where it doesn't need to be trained on a specific, narrow task to succeed (like with ANI). You are asking for Agency and motivation, something particularly lacking even in most humans. All the top labs are working to ensure AI doesn't gain that motivation, but I think it exists in living beings only as a evolutionary necessity, not as an artifact of reasoning ability. The entire religion of Buddhism is about abandoning the base human desire to do things and then more things in the endless cycle we possess.
The conclusion that this is a bubble seems arbitrary and disconnected from the several true facts outlined in the article. Anthropic's growth _is_ slowing, albeit still strong at ~47%. They do have enormous compute obligations, and it will require a lot of growth to meet them. However, they also increased revenue per dollar spent on compute from $0.40 to $2.08 in five quarters and hold $131.7bn in cash.
Keep in mind this growth has mostly been sustained by low-level business automations, coding, and chatbots. With the adoption of products like personal agents, e.g., Meta's Muse, I am pretty optimistic there will be demand and growth around this level will be sustained for a few quarters. Doubtful we will ever see another ~150% growth quarter.
> revenue per dollar spent on compute from $0.40 to $2.08
So that’s ignoring labor and perhaps facilities? That’s like saying we’ve hit net positive fusion by only considering the power produced by the laser and ignoring the 80% power losses in the feed-in.
it's not meant to describe true unit profitability. compute is the single largest expense line item for frontier labs, so it makes sense to want to know if model & inference efficiency is improving.
moreover, Anthropic's GAAP operating margin drop from Q2 was largely driven by buying ~$5B in short-term surge compute capacity, not because of fundamental issues with the business.
- Would you like to get paid a million dollars right now, or get paid a penny today that doubles everyday for 30 days.
- hmmm, why 30 days? Make it 60 days and I'll take the penny.
The effort needed to double everyday, or month is realistic at first. But then you can't just keep doubling without swallowing the universe along the way.
You could in theory be responsible for the destruction of the entire solar system in order to pay you. Sit on your throne atop a mountain of skulls and weep.
I think one thing that could explain it is that some of the big tech companies might have been big customers of theirs, and now they have alternatives that, according to recent news, their employees are being pushed to use instead. Some of those reports mentioned internal usage getting cut in half. If this is true, it doesn't necessarily mean their growth has stalled.
He says a lot of things, but one specific thing I remember from weeks ago is that he said Anthropic growth slowed on Q3, and that's why they decided to push for the IPO (and use only Q2 numbers).
It reminds me of the Apple TV rumors. For ~4 years the new AppleTV was just around about to be announced for <reasons> and you should wait to buy one. Then it's finally announced and people are all "yeah, but a few weeks before the AppleTV was launched they said it was going to have more RAM!".
Sure, and they said similar kinds of things for 4 years. Why act like the broken clock is suddenly worth noting just because the time was finally right this moment?
Zitron's literally just some normie that got pushed by the algo a bit harder than other shocked faces with flames and red arrows and no industry insight.
How so? Or is this just vibes based analysis? I dont really know what you mean by classical engineering, but sure if something can be done on a computer it can infinite monkey theorem its way to a solution. This isnt exactly useful its just an enablement of laziness.
Idk how many people need to hear this, but not everything 'is computa',no matter what Trump says.
Much of the worlds problems are not verifiable with a compiler that can be trained or brute forced against.
Llms are not as useful as ya'll think they are, not everything exists in a web browser, which is why Anthropic et al have saturated their TAM and are now frantically hiring sales people to convince managers in other verticles.
I use classic engineering to refer to the (engineering) disciplines where mother nature messes your day up rather than obscene levels of man made (or machine made) pure complexity ruining your day.
It's easy to understand how LLMs got so good at coding, they can rapidly iterate, self test, and it all happens in their digital universe.
These models now starting to spit grey beard knowledge of precision analog electronics, lightweight fastener design, and chemical surfactant systems is something very very recent. A few years ago I think we were pretty confident that the tech would hit a wall at the digital divide. Well, I can tell you now that it did not. It seems it kinda blasted right past it.
> The real question is how many people haven't figured out how to use AI to make their job easier/more productive.
I think the answer, to a reasonable approximation, is “everyone”. Outside software engineering most people aren’t using it for anything useful, and I keep encountering technical people including programmers who don’t use it either, which never fails to surprise me.
One of my projects is a transportation management system (TMS) and of course I have added a lot of features to it including a very capable chat-based UI that can automate a lot of tedious data entry and other tasks for dispatchers and accounting and so on. It gets used but really not very much. I’m starting to realize that I’ve probably taken updates as far as they can go in terms of the ability of people to absorb change. Most people just carry on doing things the way they always have.
We had one person on the admin team who started using Cowork to automate all kinds of stuff. She loved it. Then she left. It hasn’t caught on with anyone else.
This is not to say that things won’t change but I think it will take a while.
my brother uses it to do ocr on his books but, lately, claude has been refusing because of copyright or whatever so i've had to do it for him (using ai, yes, but not anthropic's)
I think this dynamic is underappreciated. Another thing I see in working with different companies (consulting) and discussing with colleagues at different jobs is that we as humans tend to do a pretty bad job of modeling business outcomes and the dynamics that govern them. There are lots of outcomes where you can identify one contributing role, increase the productivity for that role by a significant factor, and still ultimately not move the needle much, because the constraints span multiple functions.
As an example outside of software engineering, a lot of people in marketing roles have found agents to be incredibly useful for particular tasks. I've talked to multiple people who wouldn't standardly be classified as "technical hires" who are using Claude to manage things like digital advertising. It's valuable here because managing campaigns on a platform like Google or LinkedIn involves a ton of granular tweaking and analysis. Every performance marketer makes tradeoffs around what they're realistically able to spend time doing, and what makes the biggest impact. So if you have 250 campaigns that you manage solo, you likely don't have the time to really manually optimize the bidding and budget for each campaign every couple of days based on new data. You can only try so many new angles, because creating new media is time consuming. You don't have time to construct custom experiences on-page for every particular campaign. There's a whole industry of tools that turn a tidy profit automating parts of this for this reason. And agents are really good at it, because a lot of it is just applying the same analysis over and over again.
But for a lot of startups, increasing traffic from ads 20% while decreasing spend is a good outcome, but it doesn't actually turn into revenue. The product itself might just be limited, and already being discovered by the narrow segment of potential users who find it valuable enough to pay. Or the sales process might be kind of busted, meaning the team will really only close a certain volume of new deals in a given month, no matter how much inbound they have. Or some combination of all of the above.
If it's just one marketer using an agent to optimize their advertising, there's not a lot they can do about the other constraints. And so what you end up with is one employee who feels really productive and sings the praises of Claude, but not a meaningful lift in the higher level outcomes you care about, and the analysis gets messy. In an ideal world, everyone realizes where the other bottlenecks are, but that's a genuinely hard thing to do and people have complicating incentives.
I don't think this is a particularly new problem. It feels identical to any other situation where a department asks for a tool, loves it, and it doesn't really make a difference to the bottom line or make anyone else's life easier in noticeable ways.
> very capable chat-based UI that can automate a lot of tedious data entry
Because rewieving chat bots data entry is even more tedious then doing it yourself. And as long as they have to review it, it does not even matter that it makes only few mistakes.
I see it with JIRA and Rovo. It just sux using it, despite it being reasonably competent at creating tables. If I could do markdown as used to be normal, it would be appreciated a lot more.
> The real question is how many people haven't figured out how to use AI to make their job easier/more productive.
There are two people, of course. Developers who can use AI to make their jobs more productive, more comfortable, less stressful. Then there are developers who can use AI to make themselves fired. (Not FIRE)
That there is a glut of high-paid, high-performing programmers who can kick back and, huh let me tighten the belt for two months and buy my own AI rig, isn’t surprising to anyone. Even those of us in the rural outbacks of Europe see them here every day. For them it’s just a consequence-free sabbatical.
It's a combination of natural language being an awful programming language and frontier ai companies optimizing for longest time horizon at 50% success. I mean what did they expect? Length of time to not check in and still bumble around to turn a single prompt into a literal half-assed solution.
> I don't think the "tokens consumed" metric was a very good one.
When mid or high level tech leadership pushes from top-down new things, that look dumb to the engineers, the leadership is either stupid or there are hidden incentives that regular people don't see. In my experience mid level usually matches with the former, and top level with the latter.
Completely anecdotal but I did find the better engineers tended to take to the AI tools quickest. At the beginning, tokens was probably a good way to pick them out. “Good” being entirely subjective around a product focus/mindset. If you ever find a good metric like that, it is critically important to never tell anyone, else it will quickly become not good.
From what I could tell, it probably even distracted from the metrics that traditionally matter. I seem to remember tokenmaxxing dashboards being built, and other "shovels."
It may have something to do with the summer intern season ending. I know one of the interns hired by my employer used up the allotted AI spend by doing “fun” things, like making Claude write an entire SQL-to-Pandas transpiler in 10,000 lines of Python.
Your machinery was not written elegantly. It was sloppily assembled via evolutionary pressures.
A gene gets duplicated, mutates, and suddenly you have new functionality and new fitness criteria.
Try something bad and you die without reproduction.
Our code isn't clean. There's redundancy and extra packing and slop all over the human body.
Embrace it. Stochastically hill climb.
Code is DNA. Evolve it, don't intelligently design it. Fit the curves of your environment quickly and use tests to cull bad behavior and the slopes of your domain.
I don't have a few billion years to deal with stochastic evolution of my software, so I think I'll stick to intelligent design, thanks. Humans would be a lot better in a lot less time if they had been intelligently designed, too.
Interesting angle to blame the slopmaxers when the AI companies have been pushing maximum usage for a while now.
I assume they’d hoped that the value of the spend would manifest itself and they could raise prices. The opposite happened and companies are instead cutting back.
It sounds to me like they chose a poor marketing and pricing strategy for a product that is very difficult to make cost effective.
When was the last time anyone used the term "tokenmaxxing" right now?
Exactly, No-one and as predicted it was a fad. [0]
Vibe-coders don't want to risk getting their accounts suspended if you use their models in a certain way.
We are seeing the result of those subsidizations and discounts that Anthropic made this year and how much of this "tokenmaxxing" hype was used to make their revenue numbers look "good".
With the arrival of more cheaper models which are good enough, it doesn't make much sense to use Anthropic's models.
So don't be easily fooled by their revenue projections. Rather question if it is sustainable for them. (Spoiler, it isn't)
Everyone trying to vibecode their next million dollar idea only discover the web is dead. Nobody is looking for anything because that would require leaving the safety of the same handful of sites everyone else is on.
Because not only is there no guarantee that the peddled slop will be online or maintained or supported for any amount of time after its debut, but it's safer and more effective to just reap the slop as context into something slightly less sloppy.
Lmao a colleague who works elsewhere said some goon at work created an AI leaderboard. He spun up research agents. Top of board achieved. Granted the guys a very solid developer, but he did it to prove how dumb this is.
We can finally dispel the myth that management is competent when they, almost as one across industries, tied continued employment to setting fire to company money
AI put a massive spotlight on all the required but unspoken decision making and coordination that must occur with stakeholders outside their own teams during the development process.
Dysfunctional middle management has always traditionally neglected it or actively undermined it, blaming others when things go wrong and taking credit when things go right.
Naive developers are always more than happy to do what they think is "picking up the slack" not realizing it's a trap.
Perhaps most organizations are just now more wise to the fact that using AI to "accelerate the process" is like arming the monkeys with machine guns?
If you ever wonder why Anthropic is going all out to try and fearmonger their way to an open weights ban, this graph will provide all the answer you need.
156 comments
[ 1.4 ms ] story [ 37.3 ms ] threadThat debt will transfer to public shareholders at the IPO but also will double or triple in the process, as the investors expect to be paid very, very well. So the current shareholders are looking to extract something over $400 billion from the IPO.
With the current revenue they might pull off that P/E. But they’ll still be extracting some of that value from the customer base, because shareholders will demand it.
Subscriptions are a growing minority of revenue.
Then why all the rush to IPO then?
The strongest chinese models from DeepSeek, Kimi and z.ai are already good enough to run for hours on most coding tasks cheaply.
I think the first implying the second might be a bit of a stretch. At some point, companies are going to start looking at their AI spending and start negotiating deals. OpenAI is right behind them and Anthropic is around 8x more expensive (Sol 6.1 vs Opus 5.5 based on Artificial Intelligence's cost per task; cost per token is a somewhat useless metric when different models use vastly different numbers of tokens for the same thing). Chinese models like Deepseek 4.1 Flash are cheap, can be self-hosted, and while certainly behind Claude Opus 5.5, are still good.
Lots of companies buy "inferior" products like Microsoft Teams. Over the next 5 years, it seems likely that many models will get really good to the point that people often won't be able to tell any difference. Even if something isn't "definitively" better, companies might say "our engineers can't tell the difference and we're cutting our costs by 85%."
It's not absurd to think that a company would use its own harness or an open source harness and/or put a proxy in to route requests variably. For two months, they have engineers using different models unbeknownst to them and they measure things like how often the engineers have to correct or redirect the model. At the end of two months, they might find little difference in performance and switch to a much lower cost model.
It's not like it'd be crazy hard to analyze workers frustration as they did their jobs and figure out if one model was actually better. They could log all the conversations and it's not like AI can't easily determine things like "no, that's completely wrong" as a phrase showing that the agent was really off base.
I think a big part of the next 5-10 years is going to be about cost/efficiency.
1. Frontier models are so good that the token cost of an average task becomes miniscule, so users are incentivized to use frontier models because their sheer intelligence makes them cheaper to use.
2. There is a pareto optimal point where even if you spend more on tokens overall profits decrease because token costs rise faster than revenue does.
The "no lab has a clear lead" world leads to 2.
That's a great product idea!
There seems to be another magic ingredient for improving agents beyond model size & efficiency, and so far only OpenAI, Anthropic, and maybe Google know what it is.
"I have a girlfriend for real she just goes to a different school"
I ping-pong between different agents all day long and don't necessarily agree that Anthropic is clearly in the lead on coding at this point.
Has anyone else seen this?
I wonder if more people using Haiku coming from more expensive models may have an impact in those metrics.
Accidentally leave one on medium when it should have been high for a specific task has an outsized impact compared to the variation in the models.
For me, that's amazing, but maybe it explains these numbers and maybe it'll finally change when Fable 5.5 forces me back into high usage territory. If I even really need it's abilities as more than an advisor (words I didn't think I'd imagine myself saying back when Fable 5 was dropping).
The total market for code generation might just not be enough. The investments levels really demand getting rid of whole classes of professions. Not just 80% of programmers.
AI is just a compute commodity with a value add.
[0] https://x.com/jumperz/status/2112226308868436154
Even in next 3-4 years, the company that has the best ecosystem of researchers, culture, shared knowledge and processes will be able to churn out models in a way that can't be replicated.
the second part is not sexy - but we see all a.i labs having fde arms e.g ode, microsoft frontier company etc but then again that's the part you can't hype.
You can go very far in life with a thing that isn't futzy
How much easier my life would be if Chinese open models were even just 1 year behind outside of X posts and one-shots.
My easy litmus test is would the year old model (Opus 4.5) being open sourced today would eat the traffic of every single Chinese model... and the answer is pretty much yes.
Every Chinese model would throw away their work just to post-train or distill Opus 4.5 and get a more reliable model.
As it stands, even the people closest to this technology have barely figured out how to use it effectively.
I don’t think this era is going to end with us all unemployed, but I do plan to be retired before the next cycle.
At some point someone moved the goalposts for this to include LLMs that emphatically are not that. I think it was Altman when they came out with Astra?
they aren't as intelligent as a human, but they can throw shit at the wall faster, and get something that sticks faster. they are superhuman in areas that require trial and error, like hacking.
everyone stopped talking about the Turing test once ChatGPT came out and broke right through it while still being very very dumb. AGI is just too blurry to define.
I have to ask, what humans do you know that have more general intelligence? This isn't a dream team where you can say "Magnus Carlson can beat it at chess, and Linus can write better C code", this is a comparison across all domains and levels of intelligence. Even its IQ is estimated at 141, which is seriously higher than the average human. The average human's reading level is atrocious. The average human cannot write a standard 5 paragraph essay about a random topic.
https://www.aiiq.org/labs/anthropic/
My understanding is that models will often still forget games rules etc unless you design a harness specifically to assist them and prevent them going off the rails. I was going to use chess as an example but it sounds like they are good enough to do that now, albeit I'm not sure if they have a special harness for it.
It's not about how "smart" they are, or how good they are at recall. It's about whether I can teach a model to understand and execute an arbitrary, general task regardless of whether they have been trained on that task before.
Disregarding the above, the fact models still require a human prompter/decision-maker is defacto proof they aren't AGI in my view. If they were humans would be obsolete in decision-making.
I don't get what you're saying here- AGI doesn't need intelligence? You just want autonomy?
If these were AGI we wouldn't be talking about them helping engineers, engineer jobs would just not exist. The AI would be superior in every way that mattered. That's not the case yet, so IMO we don't have AGI.
If you left an AI alone for a long time and when you came back they had built up a working society and learned new useful things, then that's the proof it's AGI. From this definition you only know it's AGI after you let it loose.
Nvidia's Jensen Huang says "We've achieved AGI."
Keep in mind this growth has mostly been sustained by low-level business automations, coding, and chatbots. With the adoption of products like personal agents, e.g., Meta's Muse, I am pretty optimistic there will be demand and growth around this level will be sustained for a few quarters. Doubtful we will ever see another ~150% growth quarter.
So that’s ignoring labor and perhaps facilities? That’s like saying we’ve hit net positive fusion by only considering the power produced by the laser and ignoring the 80% power losses in the feed-in.
moreover, Anthropic's GAAP operating margin drop from Q2 was largely driven by buying ~$5B in short-term surge compute capacity, not because of fundamental issues with the business.
Most likely the combined revenue growth continued just as fast, but OpenAI took a lot of market share with a model that's ~as good.
If true, then the current market sentiment is too pessimistic because demand for AI (and chips) has continued growing.
- hmmm, why 30 days? Make it 60 days and I'll take the penny.
The effort needed to double everyday, or month is realistic at first. But then you can't just keep doubling without swallowing the universe along the way.
~5 quadrillion after 60
Discussed here a few weeks ago: https://news.ycombinator.com/item?id=49997161
Sure, and they said similar kinds of things for 4 years. Why act like the broken clock is suddenly worth noting just because the time was finally right this moment?
Especially with this latest crop of models, they seem to be pretty damn competent in classic engineering rather than just software engineering.
Idk how many people need to hear this, but not everything 'is computa',no matter what Trump says.
Much of the worlds problems are not verifiable with a compiler that can be trained or brute forced against.
Llms are not as useful as ya'll think they are, not everything exists in a web browser, which is why Anthropic et al have saturated their TAM and are now frantically hiring sales people to convince managers in other verticles.
Turns out most people dont need or want AI.
It's easy to understand how LLMs got so good at coding, they can rapidly iterate, self test, and it all happens in their digital universe.
These models now starting to spit grey beard knowledge of precision analog electronics, lightweight fastener design, and chemical surfactant systems is something very very recent. A few years ago I think we were pretty confident that the tech would hit a wall at the digital divide. Well, I can tell you now that it did not. It seems it kinda blasted right past it.
I think the answer, to a reasonable approximation, is “everyone”. Outside software engineering most people aren’t using it for anything useful, and I keep encountering technical people including programmers who don’t use it either, which never fails to surprise me.
One of my projects is a transportation management system (TMS) and of course I have added a lot of features to it including a very capable chat-based UI that can automate a lot of tedious data entry and other tasks for dispatchers and accounting and so on. It gets used but really not very much. I’m starting to realize that I’ve probably taken updates as far as they can go in terms of the ability of people to absorb change. Most people just carry on doing things the way they always have.
We had one person on the admin team who started using Cowork to automate all kinds of stuff. She loved it. Then she left. It hasn’t caught on with anyone else.
This is not to say that things won’t change but I think it will take a while.
As an example outside of software engineering, a lot of people in marketing roles have found agents to be incredibly useful for particular tasks. I've talked to multiple people who wouldn't standardly be classified as "technical hires" who are using Claude to manage things like digital advertising. It's valuable here because managing campaigns on a platform like Google or LinkedIn involves a ton of granular tweaking and analysis. Every performance marketer makes tradeoffs around what they're realistically able to spend time doing, and what makes the biggest impact. So if you have 250 campaigns that you manage solo, you likely don't have the time to really manually optimize the bidding and budget for each campaign every couple of days based on new data. You can only try so many new angles, because creating new media is time consuming. You don't have time to construct custom experiences on-page for every particular campaign. There's a whole industry of tools that turn a tidy profit automating parts of this for this reason. And agents are really good at it, because a lot of it is just applying the same analysis over and over again.
But for a lot of startups, increasing traffic from ads 20% while decreasing spend is a good outcome, but it doesn't actually turn into revenue. The product itself might just be limited, and already being discovered by the narrow segment of potential users who find it valuable enough to pay. Or the sales process might be kind of busted, meaning the team will really only close a certain volume of new deals in a given month, no matter how much inbound they have. Or some combination of all of the above.
If it's just one marketer using an agent to optimize their advertising, there's not a lot they can do about the other constraints. And so what you end up with is one employee who feels really productive and sings the praises of Claude, but not a meaningful lift in the higher level outcomes you care about, and the analysis gets messy. In an ideal world, everyone realizes where the other bottlenecks are, but that's a genuinely hard thing to do and people have complicating incentives.
I don't think this is a particularly new problem. It feels identical to any other situation where a department asks for a tool, loves it, and it doesn't really make a difference to the bottom line or make anyone else's life easier in noticeable ways.
Because rewieving chat bots data entry is even more tedious then doing it yourself. And as long as they have to review it, it does not even matter that it makes only few mistakes.
I see it with JIRA and Rovo. It just sux using it, despite it being reasonably competent at creating tables. If I could do markdown as used to be normal, it would be appreciated a lot more.
There are two people, of course. Developers who can use AI to make their jobs more productive, more comfortable, less stressful. Then there are developers who can use AI to make themselves fired. (Not FIRE)
That there is a glut of high-paid, high-performing programmers who can kick back and, huh let me tighten the belt for two months and buy my own AI rig, isn’t surprising to anyone. Even those of us in the rural outbacks of Europe see them here every day. For them it’s just a consequence-free sabbatical.
When mid or high level tech leadership pushes from top-down new things, that look dumb to the engineers, the leadership is either stupid or there are hidden incentives that regular people don't see. In my experience mid level usually matches with the former, and top level with the latter.
Your machinery was not written elegantly. It was sloppily assembled via evolutionary pressures.
A gene gets duplicated, mutates, and suddenly you have new functionality and new fitness criteria.
Try something bad and you die without reproduction.
Our code isn't clean. There's redundancy and extra packing and slop all over the human body.
Embrace it. Stochastically hill climb.
Code is DNA. Evolve it, don't intelligently design it. Fit the curves of your environment quickly and use tests to cull bad behavior and the slopes of your domain.
I assume they’d hoped that the value of the spend would manifest itself and they could raise prices. The opposite happened and companies are instead cutting back.
It sounds to me like they chose a poor marketing and pricing strategy for a product that is very difficult to make cost effective.
We are seeing the result of those subsidizations and discounts that Anthropic made this year and how much of this "tokenmaxxing" hype was used to make their revenue numbers look "good".
With the arrival of more cheaper models which are good enough, it doesn't make much sense to use Anthropic's models.
So don't be easily fooled by their revenue projections. Rather question if it is sustainable for them. (Spoiler, it isn't)
[0] https://news.ycombinator.com/item?id=49330214
Behold, the beauty of the modern internet.
Dysfunctional middle management has always traditionally neglected it or actively undermined it, blaming others when things go wrong and taking credit when things go right.
Naive developers are always more than happy to do what they think is "picking up the slack" not realizing it's a trap.
Perhaps most organizations are just now more wise to the fact that using AI to "accelerate the process" is like arming the monkeys with machine guns?