_______ __ _______
| | |.---.-..----.| |--..-----..----. | | |.-----..--.--.--..-----.
| || _ || __|| < | -__|| _| | || -__|| | | ||__ --|
|___|___||___._||____||__|__||_____||__| |__|____||_____||________||_____|
on Gopher (inofficial)
URI Visit Hacker News on the Web
COMMENT PAGE FOR:
URI I think Anthropic and OpenAI have found product-market fit
trash_cat wrote 1 hour 58 min ago:
Discovering that Enterprise customers tolerate higher prices compared
to retails consumers is discovering demand elasticities not PMF. Am I
missig something?
j45 wrote 3 hours 25 min ago:
There is always the next, or other product market fit to find.
kamazaki wrote 4 hours 1 min ago:
The product-market fit question feels premature when most usage is
still developers building on top of the APIs, not end users paying
directly.
pandoro wrote 4 hours 18 min ago:
If those companies found PMF we should expect to see the effects of all
those tokens burned on the products we all know and use. In my own
experience: yes new features can be delivered much more quickly (by
orders of magnitude) but knowing what to build and how to evolve a
product with taste and direction does not; so ultimately the
user-facing gains are marginal.
I suspect that once the technology has been tamed and the hardware and
software has been commoditized, the impact will be much less dramatic
than we expect and we will realize the importance of a shared vision,
experience, taste, intuition and discernment in building good products.
motbus3 wrote 4 hours 26 min ago:
this is an opinion and i would love to debate :) dont hate me too hard
i think the article is momentaneously correct but there are some things
that smells to me about the situation overall (not the article!)
(1) i believe gpt 5.5 and opus 4.7 are not as good improvements as
their predecessors and there has been not enough evolution in those
models to justify the price increases. Unless something big changes in
the next few years they will not be able to track these costs (unless
A)
(2) they might not be able to keep up improving the data centers if
their tech keeps demanding more and more hardware capability. even
nvidia does not show signals that they will be able to beat too much
their big GPUs and at some point this increased price will be passed to
them anyway which will need to be put in the token price (unless A)
(3) i've been trying deepseek v4 and other models and honestly, they
are more useful than gpt 5.5 and opus 4.7... i mean, there is still a
difference, but it is so little that it does not make sense the cost of
opus and gpt 5.5 (unless they are going to A too)
(A) I have the impression is that the plan was the whole time to
sideline common folks like us and focus on gov, big techs and military.
they only let us play with the toys to gather data and for people to
not get mad because they stole the internet from us.
m3kw9 wrote 4 hours 54 min ago:
I think the product market fit is still how you interface with
increasingly powerful AI. Right now is Codex/Claude code, and they
just added computer use/web browsing into them. The product is the
speed at which you can interface it and to be able to serve it with
infrastructure.
Traditional product market fit isn't really applicable with tech
evolving so fast.
ludwigvan wrote 5 hours 10 min ago:
Just to be clear, enterprise companies are using the enterprise option
for security/auditing/enforcement reasons, not because OpenAI or
Anthropic forbids them from buying a 200$/month plan for each engineer,
right? Do they have any clause forbidding using those individual plans
in a company?
rhubarbtree wrote 6 hours 6 min ago:
PMF is such a bullshit term. Thereâs no precise definition. Itâs
bs.
The fact is that investment is at a scale so large that current
trajectories are nowhere near going to provide ROI.
The model companies are trying to milk every drop of ARR they can
before IPOing. That is entirely the current narrative. You can hurt ARR
before it churns if your goals are short term.
Anthropic being profitable is laughable. Sure, by some accounting
measure that no one seriously uses. But looking at revenue and what
theyâre raising you can see the true story.
The truth is we need a revolutionary step up in capability to justify
capex spend. Itâs possible that might come - Opus 4.5 brought one -
but failing that weâll see the bubble pop once the IPO pumping is
done.
frevib wrote 6 hours 21 min ago:
âthe reality of AI right now is that it only works for coding.â
Kind ofâ¦
jreynar wrote 6 hours 29 min ago:
I may be biased because I work on an AI powered enterprise productivity
product, but while I agree they have PMF right now, I wonder whether
people's use will evolve in ways that undercut the current PMF.
Chatting with an assistant is great if there's no product with tailored
UI available that also has the AI capabilities. But once there is, I
suspect people may switch, or more importantly enterprises may switch
because they'll get the benefits of AI without the clunkiness of a chat
only interface. We may see another DOS -> GUI-like shift.
More specialized products will consume tokens but their builders will
be incented to optimize token use and switch models as costs and
capabilities change. And if search engines become more AI capable, and
Google is clearly striving for this, then they may have pressure from
two sides that could squeeze the number of use cases for AI chat. AI
coding isn't going anywhere and nor is the need for AI in general but I
wonder if the products will have to evolve significantly to maintain
the current levels of PMF. And then there's the question of
profitability...
oxedom wrote 6 hours 11 min ago:
MCP-Apps could be the solution for tailored UI.
throwatdem12311 wrote 7 hours 25 min ago:
If theyâve spent all this cash, and all we have to show for all this
AI baloney are crappy coding agents that constantly make the same
mistakes no matter how much you try to guide them.
All the slop content, all the bots, all the misinformation and fake AI
images and videos.
All of the social and economic disruption from datacenter buildouts.
The massive nosedive in reliability on the worldâs software
infrastructure.
After all of that and all we get is a code bot so a few incompetent
loser devs can bloviate about not writing their own code and brag about
never reading it.
Burn it all to the gd ground. Destroy this new Tower of Babel.
gcr wrote 7 hours 38 min ago:
For an alternate albeit somewhat contrarian view, also see Ed
Zitronâs piece that add context to Anthropicâs profitability: [1]
TL;DR Ed argues that the deal between Anthropic and xAI could have been
negotiated in such a way as to make Anthropic only appear profitable
during its âramp-upâ period in June, which incidentally is also the
month that Anthropic is making tons of other pricing changes.
URI [1]: https://www.wheresyoured.at/anthropics-profitability-swindle/
dev_l1x_be wrote 8 hours 1 min ago:
I think the third company (likely Google) is going to make LLMs
financially feasible with:
- dedicated hardware ( [1] )
- optimized models ( [2] )
URI [1]: https://cloud.google.com/tpu
URI [2]: https://research.google/blog/turboquant-redefining-ai-efficien...
rldjbpin wrote 9 hours 12 min ago:
money talks, and being a broky from outside the valley/west coast, the
"product-market fit" here are the neighbours of the service providers.
the economics simply don't work unless you make six figures, at least
to just give it a go blindly. the providers are also still figuring out
what they can get away by charging, and they are getting a similar
treatment from those under the stack.
the caps and limits are not very transparent, and it is quite difficult
to know what is "enough". the current rate does not stay the same and
the contract is changed way too often to dedicate for the long term.
regardless, the subsidized rates should not be sustainable forever.
make hay while the sun shines i suppose.
einpoklum wrote 9 hours 53 min ago:
Sounds like mostly-baseless praise for the LLM behemoths / capital
investment black holes.
> I currently subscribe to the $100/month Max plan from Anthropic and
the $100/month Pro plan from OpenAI
... which already indicates a bias.
> If you are a heavy user of coding agents these plans are a fantastic
deal... thatâs $2,180.16 worth of tokens for $200ânot bad at all!
Thank you for the sales pitch. Perhaps go buy a car and tell us how
[insert your manufacturer] has "found product-market fit"?
> Anthropic are strongly rumored to be about to have their first
profitable quarter.
First, the strong rumor is a claim by Anthropic itself. But even
assuming that's true - it's an "operating profit", i.e. disregarding
the massive capital expenses for years, and may also disregards ongoing
capital expenses, if those happen not to be taken this particular
quarter.
> 1 Trillion .. companies spending $200+/month/user will get you there
a whole lot faster
First note the use of the first person plural to talk about Anthropic
and OpenAI.
But that aside - most companies aren't paying $200 USD/user-month. But
even if they were - if we take the 30 Million SW developers mentioned
in trjordan's comment as subscribers, that's 2400/user-year * 30
Million = 72 Billion USD / year. And this is already rather optimistic,
but - want to double that number of subscribers? Fine, make it 150
Billion / year. Still not there with a rosy outlook and assuming the
hype and enthusiasm continue for many years.
And those rosy estimates are more likely than not unrealistic. I am
reminded of this review of some empirical research regarding the
benefit of LLM/AI use:
URI [1]: https://cmr.berkeley.edu/2025/10/seven-myths-about-ai-and-prod...
emsign wrote 10 hours 6 min ago:
They've finally forced the system to fake profits. Attacking personal
computing, sucking the rest of the industry dry, yeah eventually you
can make people pay for your shit. I still pray these vultures will
implode without too-big-to-fail safety nets. And big daddy Trump can't
help them anymore by fixing the system for them, no matter how much
billions they are paying into his personal pockets.
The future are small models, nobody really needs big compute in the
long run, that's why big tech is going for our personal hardware. So we
won't become their competitor in their rent only economy. True
competition is eliminated, natural evolution is being fixated by the
government. This is not going to end well for the USA.
dep_b wrote 11 hours 26 min ago:
Itâs funny how a US$30 ReSharper license that would give me a 20%
performance boost in programming was out of the question, while me
burning US$3000 per month on tokens for a 40% boost in performance
never was questioned
mrpopo wrote 9 hours 32 min ago:
I feel this so much. What was all that fuss about needing 3 levels of
approval for any software license? Or having to provide a compare
chart against open-source alternatives?
The money would be so much better spent that way as well, supporting
individual programmers.
mchusma wrote 11 hours 48 min ago:
If we define product market fit as profitable with a trillion dollar
valuation, I think the term has lost its helpfulness.
I do agree with the author that these companies seem much stronger
financially recently though.
tornikeo wrote 12 hours 9 min ago:
And I think that's amazing. I'd like to keep using the subsidized
coding tools, especially Codex, since I've given up on Claude.
Hopefully the PMF allows the subsidy to continue. Would hate to have to
move to the next coding harness again.
crakhamster01 wrote 12 hours 11 min ago:
> As further evidence that enterprise agents represent product-market
fit for these companies, consider their open job listings.
PMF is one interpretation, but it could also be read as desperation.
In my opinion, we've been at PMF for quite a while now. The November
inflection point that's often referenced definitely changed how we
interface with models, but as far as coding goes, I feel like Cursor
had proven itself useful for at least a year prior to that.
The demand has always been there, the outstanding question is still -
how do you build a business on top of these products? None of the
frontier models have emerged as uniquely capable, but open weight
models are now catching up in capability as well. The explosion in
go-to-market roles feels more like an attempt to lock customers into
contracts so that they don't consider alternatives.
I assume the hope is that during this 12-month contract they will
develop real integrations, something deeper than just a CLI harness. If
you've ever worked in procurement or dev tooling at a reasonably sized
company, you'll know that this is exactly what teams try to avoid.
It's anyone's guess what will happen this time, but I'm excited to see
how the IPOs go.
harrouet wrote 12 hours 12 min ago:
Product-market fit, but what about customer retention?
It is quite trivial to switch from using one model or another.
Likewise, in a few years we'll have affordable laptops to run today's
frontier models.
What's their plan to let us keep subscribing?
simonw wrote 9 hours 10 min ago:
Right now the main plan for that appears to be having those
enterprise accounts commit for a year at a time.
attentive wrote 12 hours 38 min ago:
> and both companies now have measures to lock their enterprise
customers (who tend to sign year-long deals) at those API prices
why would enterprises do that if they can just use bedrock or vertex?
CAX wrote 13 hours 3 min ago:
$50/month tops is like 60B
vfalbor2 wrote 13 hours 17 min ago:
He is talking about tokenization of different llm and comparing, and
this is inacurate because you cannot compare if you dont how it
tokenice the words, some tokenicer could be cheap with more tokens than
other because of their features. Not agree with Simón speech.
jimnotgym wrote 13 hours 40 min ago:
They have the whole software industry addicted to AI. They will be
pushing the price continually up, and like addicts they will find a
way to pay
kkarpkkarp wrote 13 hours 43 min ago:
> I currently subscribe to the $100/month Max plan from Anthropic and
the $100/month Pro plan from OpenAI. If you are a heavy user of coding
agents these plans are a fantastic deal.
Guys, what - in your opinion - does "heavy user" mean? I thought I am
heavy user (I am using AI to code every day 8hr a day + side projects)
but 20 USD/month Cursor plan is always enough. What should I be doing
to extend my license to higher level?
whazor wrote 13 hours 23 min ago:
My agents are easily busy 30 min to a hour independently.
Implementing their plan, building and running tests, verifying
deployment, linting. So I switch between multiple agents. Each agent
has its own branch and worktree.
mock-possum wrote 13 hours 27 min ago:
How many sessions are you running simultaneously, how careful are you
being about token consumption and context window management, and what
level of reasoning are you using (with which models?)
8hrs a day doesnât really mean anything without a lot additional
qualifiers.
fwiw lately Iâve been straddling 2 or 3 claude codes and one Claude
cowork, primarily on 4.7 with high effort - the companyâs paying
for it, so Iâm doing my best to burn as many tokens as I have the
mental capacity to manage. At that rate, the 100 account is
completely necessary, I was blowing through my 4-hour limits
consistently before requisitioning an upgrade.
kkarpkkarp wrote 1 hour 9 min ago:
at most I run 2-3, but usually one.
About token consumption and models: that's the thing. I hear this
question often, but my answer is always the same. As I am Cursor
user, I run it always in Auto mode (so Cursor decides which model
to use, I don't even know which is in use).
Sometimes I switch from Auto to defined model but I found it
quickly triggers "you are out of tokens" notices so nope: I stick
to Auto :) 20 usd / month and that's it
jeswin wrote 13 hours 33 min ago:
I'm out of credits codex pro credits (200 plan) in 4-5 days, and then
I switch to claude (another large plan) mostly or just go out more
and get Vitamin D. :)
I spend most of my time designing and tweaking tests suites, and
improving test performance. These commits are almost entirely Codex:
[1] - but it's possible only because there's a very large test suite
attached to it.
All of that is very token intensive. If OpenAI gave me 3x my limits,
I'd find ways to eat it up in a week.
What do these tokens give me? Well, I think in a week or two, I hope
to port the TypeScript-Go compiler back into TypeScript, but compiled
to native code. It's probably not particularly useful for most ppl,
but it's a hobby project that I've spent the last 7 months on.
URI [1]: https://github.com/tsoniclang/tsonic/commits/main/
iancmceachern wrote 13 hours 50 min ago:
It reminds me of this Steve Jobs Clip: [1] He said, so many years ago,
that there will become a time where computing power is so prevalent
that we will stop using the person to make the computers job easier and
start using the computer to make the humans job of interfacing with it
easier.
But in this context, it would mean the other side of increase
productivity is decreased time to do the same work. These are the same
thing.
URI [1]: https://youtu.be/0lvMgMrNDlg?si=QkkOnngYTjaSPlIy
Maojer wrote 13 hours 58 min ago:
Man what a disaster article
DeathArrow wrote 13 hours 59 min ago:
I think we are overly fixating on Anthropic and OpenAI.
Since there are lots of models that are competitive and have a much
better pricing, both OpenAI and Anthropic seem inefficient. I don't get
why someone would want to buy shares after IPO apart from fomo and
artificially built enthusiasm.
Anthropic and OpenAI may well be the Altavista and the Yahoo of the AI
age.
jsemrau wrote 14 hours 3 min ago:
Anthropic for sure. It's a useful professional product that I find many
use-cases for in my professional and private life. OpenAI not so much.
willsmith72 wrote 14 hours 25 min ago:
can we stop throwing around the term "profitable"? didn't they say
"operating profit"? so so different to actually making a profit
RevEng wrote 15 hours 49 min ago:
I usually agree with Simon but I think he is overlooking an important
factor.
There is a lot of AI usage happening not because it shows benefits, but
because the business has mandated its ubiquitous use. Companies having
dashboards for token usage and rewarding people for using more tokens
is a real thing. I just spoke with someone today who works at Microsoft
and they are required to use AI for all of their work - they have to
make a special request with justification if they decide not to use AI
for even a single PR. This kind of demand isn't driven by value from
either the company itself or from its workers; it is the kind of
artificial demand you get from make-work projects to keep people
employed during hard times.
We have to wait for the hype to settle down and people start making
business decisions based on results before we can really value these AI
products.
no-name-here wrote 15 hours 36 min ago:
I think token leaderboards are idiotic, however⦠many companies are
requiring AI use, with by far the most likely reason being that they
truly believe AI-assisted development gets (or will after the time
needed to get employees acclimated) better returns.
andrewflnr wrote 2 hours 4 min ago:
That mere belief by companies is not enough on its own to
substantiate durable market fit. Companies believe all kinds of
silly things. We're looking for long-term trends, and the returns
have to be real for AI development to be more than a dream
companies will eventually wake up from.
ggttk wrote 15 hours 38 min ago:
Correct the question is one of real vs fake demand curves.
mark_l_watson wrote 16 hours 1 min ago:
Well, good for them that they are charging enterprises API rates. Why
in the world not do something similar for consumers? Use for free a few
times a day, have a $5 dollar plan for light use, and perhaps $10 to
$15 for heavier use. If 90% of consumers pay nothing then âdrop
themâ except for letting them have an account and a few queries a
day.
It is easy for me to change providers. Right now I use the open source
Claud Code harness with two paid API venders for DeepSeek v4 (flash and
Pro). I like seeing how much each session costs.
jwpapi wrote 16 hours 19 min ago:
Unfortunately Simon drank from the AI Cool Aid.
I know everything youâve done for the tech community, but I please
you to take some time off and reflect on this article. Itâs not on
par with ur usual level, but the tendency has been visible from the
last couple of articles.
simonw wrote 16 hours 8 min ago:
Genuine question: what's wrong with it?
I thought this was one of my best pieces of writing this year.
(In case you missed it, the title was meant as a subtle burn on those
two companies - it's pretty absurd for them to only just be finding
product-market fit when they're already valued at over a trillion
dollars.)
jwpapi wrote 9 hours 44 min ago:
I see you trying really hard to make things right and are
everywhere in the comment. I feel a bit bad in formulating my
comments a bit polemical.
Also my highest respect for responding so calmly without lowering
your debating level to mine. I try to best to explain what I think
is wrong.
To start I didnât interpret the article as a burn.
I think itâs interesting to explain whatâs wrong with it,
because it seems similar to whatâs wrong with AI. The issues are
subtle.
- Anthropic doesnât have a profitable quarter, itâs financial
engineering ( [1] )
- The first argument about your subscription price, doesnât has
anything to do with the overall claim of the article. It would if
at all be a weak argument to support the opposite. Subsidizing
prices signals a lack of PMF.
- That you hire sales people after you had a billions of funding is
nothing surprising and doesnât indicate PMF or not.
- AI Implementations are fresh and of course AI Failures are thin,
but so are AI successes. I havenât seen any companies creating
billions of shareholder value because theyâve massively invested
in AI and their competitiors didnât You really can look at these
things in 5 to 10 years and it is multi-faceted including cultural
acceptance.
- That they need to buy more compute to satisfy the requests is
probably the strongest argument in the artical, but donât
conclusive. The product is been sold heavily subsidized and in hype
cycle. And again both OPENAI and Anthropic have to show growth in
order to justify the IPO.
- Regarding the part about revenue I refer to the linkedm article
above, as it does explain it very well.
The conclusion is reasonable given the arguments, but not the
title.
However it is missing all the real discussion points that are
actually in observation at the moment.
Local models as alternatives, IPO finanical engineering, how AI
implementation actually will perform over years... Letâs all not
forget crypto. Itâs been full of "use cases" just a bunch of
years ago. I like the idea of crypto(btc,eth) and Iâm still
invested, but 99.99% of coins have died on promies.
So this is not a piece of critical thinking, but this reads like a
twitter thread to sell me a course :/
URI [1]: https://www.wheresyoured.at/anthropics-profitability-swind...
showurwerk wrote 16 hours 22 min ago:
Not sure I can agree with this. Sure, we have enterprise paying, but
are we actually getting anything new and interesting made from LLMs
compared to if we just didn't have them at all?
Maybe acceleration in smaller teams. We still seem in the era of the
early internet where what questions LLMs change hasn't exactly emerged.
signalbright wrote 16 hours 33 min ago:
That might be an understatement
jvaqueiro wrote 16 hours 38 min ago:
I'm surprised by how good coding agents are. I think PMF was reached
last year or the year before (at least for coding).
With current limits my 100/mo codex subscription is more than enough
for the work I do.
However, I do worry about when does current subsidies are going to end?
I can see myself paying up to 300/mo, but more than that will be
prohibitive.
What's the long term plan here? Are OpenAI's and Anthropic's costs
expected to increase/decrease?
chipsrafferty wrote 16 hours 42 min ago:
I think my company is going to cancel our subscription once it realizes
we're spending more money in a week that a developer gets paid in a
month.
wg0 wrote 16 hours 56 min ago:
This requires hard evidence which isn't available. Circumstantial
evidence is all otherwise.
Bloggers are having AI psychosis too.
simonw wrote 16 hours 54 min ago:
I thought the links I provided were pretty solid. A lot more solid
then you'll see in most commentary about this stuff!
I agree with this person, let's use AI psychosis for when using an
LLM gives someone psychosis, not for when we think, what, that a
blogger made some poor assumptions?
URI [1]: https://news.ycombinator.com/item?id=48296794#48303200
cedws wrote 17 hours 21 min ago:
My employer went from the Max plans to Enterprise this month which was
utterly stupid. We went from paying $200 per head to like $500 for some
people, even more for others. For the same product. Oh well, guess
weâre doing our part to prop up this bubble.
Zafira wrote 17 hours 29 min ago:
The amount of caveats and vibes that this article is powered by could
run New York City for a decade.
claysmithr wrote 17 hours 33 min ago:
Iâm happy with spending zero dollars on ai lol..
raincole wrote 17 hours 38 min ago:
The article is at least one year too late. Claude Code has been the
product-market fit. It's so obvious retrospectively that questioning it
at this point would be very silly.
The problem is not whether they have PMF (they do) but how they're
going to compete against on-prem and Chinese providers.
Having PMF != printing money forever.
The author claim:
> Thatâs $2,180.16 worth of tokens for $200
No matter what it means, rebuild the same thing you built with these
$2,000 tokens with DeepSeek Pro V4 and let's see if Claude has a chance
to survive.
MaxPock wrote 17 hours 55 min ago:
The writer is on to something. Musk saw the writing on the wall and
bought cursor.
AI money is in coding plans
try-working wrote 18 hours 44 min ago:
Product market fit for the models, sure. But their cost structure is
going to kill them as token costs drop this year, following DeepSeek
and Xiaomi and soon more providers.
jwpapi wrote 18 hours 59 min ago:
I have to give them kudos. This whole thing is the greatest swindle of
all time.
AI has some use cases, but not at the price itâs currently priced at.
Iâve been on AI since GPT-2 with a lot of heavy users. Every user has
the same story, curiosity, surprise, hype, hate, realization.
Enterprise is usually a bit behind and are right now at hype cycle,
thatâs where they sold all the deals and do the IPO.
Itâs really a VC masterclass.
Donât get me wrong there is are useful cases of AI, but not the way
the want it to be. Quite similar to Blockchain. The idea of
decentralized money has right to exist. 99% of other coins not.
AI is a faster, but still less accurate search engine. AI is great in
finding bugs, itâs great at ruber duck debugging.
The reason I call it a swindle is, because along with the marketing it
gives tons of people in the world the impression, they can now build
their own startup, game, infra etc without the need to learn it
themselves. This leads to millions of abandoned and low qualiy projects
and products, because the vast majority has never built the mental
modal necessary to solve the problem thoroughly. In the end theyâve
wasted months and money (but burnt tokens). This is what I call a
swindle.
All early adaptors I know have not drastically winded down their usage,
not because of money, but because there is no new case. If you want to
explore a new project you can get onboarded quickly learn a lot and
then switch to documentation and live testing. For me usage is the
lowest it has been the last 2 years.
I would not let AI touch my code. I have anxiety around it, because it
will gripple back up. I will let it read my code and let me know what I
did wrong so Iâm sharpening myself.
100s of companies including open source solution can offer that for me.
All my non-tech friends are now in hype cycle and share their hype and
fore forseeable frustration with me.
I have to say Iâm in a way impressed in how AI has been rigorously
vc-utilized (conciously or not-conciously) to generate these vast
companies with the whole world watching.
aspenmartin wrote 18 hours 51 min ago:
Trying to parse the raw claims here maybe you can help me out
- itâs a swindle because ROI of tokens for coding models is not
positive? As in it doesnât bring enough value to charge like the
$100/mo?
- enterprise customers are too dumb to see this
- IPO to max out the CEO profits for what is ultimately blockchain
vaporware
Am I getting that right? Or am I putting words in your mouth?
> it gives tons of people in the world the impression, they can now
build their own startup, game, infra etc without the need to learn it
themselves.
I canât speak for peoples beliefs and motivations, but this seems
to be strawmanning, no? AI is a powerful tool to force multiply
people. You canât just prompt âbuild me an enterprise SaaS app
worth $1Bâ or âbuild me GTA6 and donât hallucinateâ but is
that your impression of whatâs happening? Dario and Sam are saying
âif you buy our coding agent subscription you can build a game with
zero skill and one shot and then be richâ?
If you donât find value in AI agents I can see reasons why that
could easily be true today. Also if it just gives you the
heebie-jeebies. But to say itâs a swindle on par with the
blockchain I think that contradicts an enormous amount of signals and
also the actual dialogue (not just headline sound bytes) around what
these systems are capable of today and what we expect them to do say
at the end of the year.
jwpapi wrote 17 hours 1 min ago:
You kind of got it right, but the biggest loser of them all are the
investors especially the index investors. They donât even decide
what they invest in but the savings that goes in funds need to
invest in these stocks.
Itâs quite an elaborate swindle obviously. But you generate hype
with underselling your core product, you claim way more usability
then there is. Users will experience usability initially.
Everything multiplies with each other and then you put it on the
market. Everybody involved makes money and youâve succesfully
extracted money from everyone whoâs invested in NASDAQ index
funds at the very least.
> Dario and Sam are saying âif you buy our coding agent
subscription you can build a game with zero skill and one shot and
then be richâ?
Thatâs Anthropics marketing, yes.
Also their offering is not uniqe that justifies a 1 trillion
valuation. The first companies are already rowing back. Itâs a
really certain time window that they are about to hit now with
their IPOs
The companies that have signed these enterprise deals havenât
done a ROI analysis. They had Fomo.
aspenmartin wrote 3 hours 52 min ago:
> you generate hype with underselling your core product, you
claim thereâs way more usability than there is
Isnât this a contradiction?
> Everything multiplies with each other and then you put it on
the market. Everybody involved makes money and youâve
succesfully extracted money from everyone whoâs invested in
NASDAQ index funds at the very least.
Sorry I may have totally missed what youâre saying here. Anyone
in S&P has already made a lot of money thanks to effects from
these companies. No one has to invest in an index fund. Markets
have riskâ¦
> Thatâs Anthropics marketing, yes.
Show me.
> Also their offering is not uniqe that justifies a 1 trillion
valuation. The first companies are already rowing back.
That there is competition doesnât imply they arenât worth 1
trillion.
> The companies that have signed these enterprise deals havenât
done a ROI analysis. They had Fomo.
Alsoâ¦wrong. I have seen the data at my company, everyone at
scale tracks this.
brazukadev wrote 18 hours 22 min ago:
> Dario and Sam are saying âif you buy our coding agent
subscription you can build a game with zero skill and one shot and
then be richâ?
They are saying profitable companies should replace the engineers
that built their systems with a subscription (while they are
hiring).
aspenmartin wrote 3 hours 50 min ago:
Can you point to any quotes or claims? I donât think this is
accurate
sothatsit wrote 13 hours 20 min ago:
I donât remember ever hearing Dario or Sam recommend replacing
people. Rather they say that smaller groups of people can do more
work, so hiring will slow because small teams can do more.
The only times when people talk about actual full replacement of
people is always when they are talking about some âfuture
AGIâ that is far more capable than the tools we have today.
protocolture wrote 19 hours 0 min ago:
I feel like we should just ignore all LLM news until after they IPO,
lots of positive sentiment astroturfing bots.
jcmfernandes wrote 20 hours 19 min ago:
These folks have no lasting moat and they know it. We are still close
to the November 2025 inflection point, so they had a clear advantage
during these past few months. That will soon fade as open-weights
models become really good, which is arguably already the case with
DeepSeek V4.
aenis wrote 20 hours 46 min ago:
The end game here is going back from a model where a bunch of product
and tech management people sit in the U.S. or Europe, and try to manage
thousands of mediocre talent sitting somewhere far away. The new model
is you give those coding tools to good engineers colocated with your
product people, and you ship good stuff much faster. If you can achieve
such a setup, the token costs can be $50k per seat per month and you
still run circles around the legacy IT models in terms of efficiency.
Giving everyone the API keys and not changing the way products are
managed is not going to work.
overgard wrote 19 hours 42 min ago:
Good lord, what company would want to spend 600k per employee just to
go maybe %20 faster (what the studies seem to show is a realistic
estimate for productivity gains),
I'm building a product right now with some AI coding (despite my
negative sentiment about AI in general they are useful). I am both
the product person and the engineer, and I'm pretty decent at using
it, so according to the hype I should be seeing like a 10x speedup. I
am not seeing that. It's definitely faster, but there are also days
where I'm stuck cleaning up things after going too fast for too long,
or periods where I need to put the software in front of people to get
real feedback, or even periods where I just need to use it
extensively myself to find the pain points and bugs. I just don't see
this "running circles" once you get past an MVP and you actually need
to build something secure and not embarassingly broken.
grttq wrote 19 hours 26 min ago:
To me the question is, can the frontier labs make the variance of
output lower + make the output of higher quality to justify their
prices?
If not lower priced chinese offerings will be better as its cheaper
per token - giving you more attempts to offset the variance.
My feeling on the former is no... I believe they tried really hard
but they've settled on pure marketing now to attempt to fight off
the chinese with perceived superiority in quality.
noddingham wrote 21 hours 4 min ago:
I feel like there's a bit of AI psychosis in this particular post.
>"These are tools which burn vastly more tokens, but are also quickly
becoming daily drivers for the work carried out by extremely
well-compensated professionals."
>"Somehow this fragment turned into headlines like Uberâs COO says
itâs getting harder to justify the money spent on AI tokenmaxxing,
because the market for stories about AI failures remains enormous."
Yes, it's just the yearning for AI failures. It couldn't possibly be
runaway costs, record revenues, and massive layoffs. It couldn't
possibly be that these tools are lighting dollars on fire by people
already paid significantly well and not producing any increase in
"value" for it (I recognize that output is 100x but outcomes are flat
by all measures). [1]
URI [1]: https://cmr.berkeley.edu/2025/10/seven-myths-about-ai-and-prod...
URI [2]: https://futuretech.mit.edu/publication/crashing-waves-vs-risin...
BloondAndDoom wrote 5 hours 7 min ago:
I just cannot imagine zero net productivity gain from AI usage by a
somehow experienced developer. Even if itâs happening right now
itâs matter of time for people to figure out how to utilize it
right. Only exception would be if people are getting the same output
but working less, which is still a net positive (just not in terms of
output).
spacechild1 wrote 8 hours 46 min ago:
Please let's not dilute the meaning of 'AI psychosis'. It is a real
phenomenon that involves actual psychosis.
mock-possum wrote 13 hours 19 min ago:
You canât just redefine âAI Psychosisâ to mean âsomebody
whose opinion about ai I disagree withâ
MichaelDickens wrote 17 hours 27 min ago:
Being pedantic, but I don't want to lose the meaning of the term: "AI
psychosis" doesn't refer to someone who thinks AI is really good. It
refers to someone who develops symptoms of psychosis from talking to
an LLM, e.g. believing they have developed a new Grand Unified Theory
of physics.
noddingham wrote 4 hours 33 min ago:
Fair and I would edit if I still had time. How about "AI brain
fry"?
pydry wrote 3 hours 4 min ago:
It's definitely a form of psychosis - contact with reality has
been lost in both cases.
mold_aid wrote 7 hours 16 min ago:
I don't know, "workaday professionals will find $200/month a
particularly good deal, such that there will be widespread
adoption" sounds either credulous enough to support the diagnosis
or dishonest enough to dismiss. I am a "knowledge worker" who is
doin' fine, has a lot of templated written work/report writing, and
there is no way in hell I am justifying that kind of spending to my
boss or my family.
infecto wrote 4 hours 30 min ago:
I think is the classic dilemma where people donât know how to
value their time.
Typical tech worker costs a company around $100/hour minimum.
That $200 subscription cost can look mighty attractive if it
saves some time or mental load.
I donât think there is anything about addiction or spooky with
that math. I suspect a lot of this is coming from tokenmaxxing
firms but on the flip side on our small team, we end up spending
about $200 per person per month for tokens using tools like
Cursor. We feel the spend is justified with measurable value.
arcticfox wrote 5 hours 10 min ago:
You can't justify $200/month in spending to your boss? Many
people charge more than that per single billable hour. I would
put your salary side by side with that number, which is your
boss's perspective, and reconsider.
mold_aid wrote 5 hours 2 min ago:
>You can't justify $200/month in spending to your boss?
No. What? Of course not.
>Many people charge more than that per single billable hour.
hrmmmm not so sure about the work that "many" is doing there
rogerrogerr wrote 4 hours 51 min ago:
Fully loaded costs for an average employee at a bigco are
scary. Not $200 but significantly higher than the number on
your W2 by the time the company pays vacation, benefits,
unemployment insurance, etc.
mold_aid wrote 3 hours 57 min ago:
I've seen people try to argue for resources using
reasonable but abstract arguments, and it just never works.
The fact of the matter is that I want $200/month, on top of
my other asks, and that comes out of somebody's budget. I
just don't see folks snapping their fingers and $200 a
month (discounted for now!) appears. Good for the other guy
I was replying to if that's the case! I just don't see it
though.
kazga wrote 7 hours 0 min ago:
"I firmly believe this technology will create business value" is
so obviously and categorically different from "Humanity has
birthed a silicon god that I have also developed romantic
feelings for" that I'm not sure if your comment is even trying to
be in good faith
mold_aid wrote 5 hours 50 min ago:
What sort of bad faith would even apply here? idgaf if poster x
or y has psychosis or not. "$200 a month is classic addict
behavior" seems pretty spot-on to me though, I just don't want
to have to pay it, too.
simonw wrote 21 hours 2 min ago:
What's the psychosis?
Npovview wrote 2 hours 54 min ago:
Hi simonw, can you write your thoughts (hivemind is yet to catch up
with this idea), the idea of distilling a large Opus 4.7 model into
a purely reasoning core with plugin like architecture (a
programming sub-model, a literature submodel, a history submodel, a
geography submodel). Why is Russian and Chinese data part of my
model training process, its costs more to train and do inference. I
want a core model and specialized models to which Core Reasoning
model can talk to. This kind of innovation is what Mistral team
should be doing. Is it fundamentally impossible to do?
lelanthran wrote 10 hours 40 min ago:
Hi Simon; while it's true that the token providers have found that
their product is 90% useful to devs and 10% useful to everyone
else, this is something they found out in the first quarter of 2025
anyway.
It's not exactly news, is what I'm saying. And even with the PMF
they found, the product is still only a commodity i.e `tokens`,
which is what every other provider on the planet is also providing.
All their other products boil down to "harnesses", which does not
look viable as a product in the sense of PMF - you cannot sell it,
you cannot lock it to your own subscription, API, etc. so you can't
use it to generate revenue any more than the free harnesses do.
PMF has a specific meaning, and "code harness" or "coding model"
does not satisfy the commonly accepted meaning. Maybe Mythos (or
similar) will.
brazukadev wrote 19 hours 35 min ago:
> because the market for stories about AI failures remains enormous
How enormous? 1 trillion dollars, 2, 10 trillion enormous?
shimman wrote 19 hours 58 min ago:
That leaders are completely fine with impoverishing vast swaths of
American workers because of "progress?"
aspenmartin wrote 18 hours 47 min ago:
Iâm so curious: letâs say youâre the president or the CEO
of a major tech corporation building frontier AI systems.
What are your directives?
shimman wrote 6 hours 27 min ago:
I would never be in this position because I'm not an anti-human
extremist that wants to hoard wealth and misallocate capital.
aspenmartin wrote 4 hours 52 min ago:
Whoâs an anti-human extremest? Doesnât that sort of
warrant some sort of facts to justify such an extreme
position? What makes you feel this is a misallocation of
capital OR somethingâ¦anti human?
simonw wrote 19 hours 54 min ago:
If they are then yes, that's psychotic. Not sure how it's
relevant to my article about Anthropic and OpenAI's enterprise
pricing though.
yokoprime wrote 20 hours 49 min ago:
Sometimes it feels like theres this opposite AI psychosis, where
anything AI is bad and boils the ocean, takes our jobs and makes
RAM expensive. Its a component in the current economy, but things
like tariffs, closing the strait of hormuz etc is equally bad for
the economy. Anyway, just find it strange to be so militantly anti
a certain tech.
3form wrote 6 hours 25 min ago:
>Anyway, just find it strange to be so militantly anti a certain
tech.
You know, that's fair. I'm much more against super-rich investing
hundreds of billions in the things they don't understand,
creating massive disruptions in their wake.
AI didn't create stupidity and greed. In the end, it's just
another tech. I'm just tired, time and again, of people who I
would hope to know better (and repeatedly they prove me wrong).
girvo wrote 7 hours 50 min ago:
> takes our jobs and makes RAM expensive
I mean it is doing both of those, so thats fair to be honest.
dmix wrote 20 hours 9 min ago:
Thatâs the modern internet. What sells is the most overdramatic
doom and gloom take possible.
Gigachad wrote 19 hours 36 min ago:
It's more the tech leaders than the internet. Sam Altman, Jeff
Bezos, Eric Schmidt and such get up on stage or interview
regularly with a shit eating grin telling us all about how they
are coming for our jobs, will make us obsolete, and there is
nothing you can do about it.
It's a natural response for society to despise these people who
have such contempt for us. It's almost embarrassing these days
being at a social function and telling people I work in
software, it's got a negative stigma almost like working in
gambling or the military.
throwthrowuknow wrote 8 hours 29 min ago:
That sounds more like a you problem.
biztos wrote 13 hours 24 min ago:
I donât know about gambling, but if âworking in the
militaryâ has a stigma, I humbly suggest seeking out
different social functions.
Gigachad wrote 11 hours 50 min ago:
I don't have to because I don't write software for child
seeking missiles for Palantir
dgellow wrote 12 hours 43 min ago:
I would say the exact opposite
keeda wrote 15 hours 46 min ago:
I think it's a matter of perception because I didn't
interpret any of them as being gleeful about it. If you think
about it, "AI will take your jobs and maybe destroy the
world" is horrible, horrible marketing -- like, your comment
is a perfect illustration of how it is received everywhere --
and yet these CEBros can't stop saying it, which indicates to
me that they actually believe it.
Oh, now that their IPOs are nigh they're changing their tunes
( [1] ) but to me that looks more like they've decided to let
$$$ prospects override what they really think.
URI [1]: https://archive.md/s9EO3
Gigachad wrote 15 hours 12 min ago:
The general public is not their customer base, they don't
have nearly enough money to spend on AI. Going to the media
and saying "this product can automate so many jobs" is
marketing to other businesses who want to use it to cut
their workers out.
There was crazy clip of Eric from Google telling a crowd of
university students that in the future AI will do
everything, and after the whole audience boos him he keeps
pushing the point that they better accept it and get on
board. The mentality these guys have is sickening. They
have no humility and no humanity.
aagha wrote 21 hours 18 min ago:
Ed Zitron begs to differ
overgard wrote 21 hours 49 min ago:
Anthropic isn't actually profitable from what I'm reading, a discount
briefly pushed them into the black. This guy makes the case well: [1]
I'm skeptical that their current price raise is sufficient, and I'm
also skeptical that most users/businesses will accept more significant
price raises that will be needed. Especially for individual users, $200
a month is already incredibly expensive, I really don't think most
people are going to be willing to pay more like $1000 a month.
URI [1]: https://www.wheresyoured.at/anthropics-profitability-swindle/
empath75 wrote 3 hours 19 min ago:
If you're an investor in Anthropic you probably _don't want them_ to
be profitable right now. They should be spending literally 100% of
their money and then some on training and compute.
bob1029 wrote 4 hours 23 min ago:
I think we could eventually get to a point where $1000/m+ consumption
would be considered a bargain.
I currently pay ~$150/m in tokens to cover my 1099 work. If I had to
pay $1000 for those same tokens each month, I would still be
massively positive on my margins. I bill customers based upon # of
completed features & bugs, not total time spent at the computer.
Tokens would have to more than 10x in cost before I would start to
have a problem on my end.
QGQBGdeZREunxLe wrote 3 hours 51 min ago:
How are you billing clients when using AI? Do you pass on costs or
do you multiply time saved by your hourly rate and bill that?
ralphael wrote 6 hours 59 min ago:
this - I'm wondering why this isn't getting picked up more in the
analysis? Its an accounting trick, which is not sustainable. Time
will tell.
vonneumannstan wrote 21 hours 58 min ago:
HN is the least agi-pilled place on the internet
reducesuffering wrote 21 hours 0 min ago:
HN has been on every wrong side of AGI predictions since the founding
article on OpenAI...
marcosdumay wrote 19 hours 13 min ago:
You mean that HN has been predicting AGI will happen, or you mean
the Singularity happened earlier today and I missed it that
humanity is already all killed?
sandeepkd wrote 22 hours 4 min ago:
If we take out the circular interests and investments here then there
is no way that this is a feasible business in current state.
Zizdefense wrote 22 hours 13 min ago:
My Costs without CC Max+Caching over past 2 months: $112K.
Ran `ccusage` on my Claude Code logs.
- Total tokens: 22.2B
Without current Claude deals, my personal cost would have been
*~$112,000*.
cmiles8 wrote 22 hours 26 min ago:
Article doesnât mention on-prem and on device models. Almost
guaranteed that there are a range of killer enhancements on these
fronts waiting to drop until IPOs get closer to inflict maximum chaos
on the valuation games.
While the big guys will argue theyâre worth trillions expect others
to drop chaos booms showing their NPV may be effectively zero.
Gravityloss wrote 22 hours 35 min ago:
I see two really good ideas for monetizing the free tier for consumers.
Firstly, if the user is asking for things where AI can link to products
or services to buy, there's a very good relevancy, much higher than in
other types of ads.
Secondly, since the AI often takes time to compute answers to user's
questions, they could be shown ads while waiting. People could perhaps
be less annoyed by this than some other commercials since they know the
break has to be there anyway.
(First idea is something I came up when asking Claude to compare some
products, or ask for help in lawn care. Second idea was by a
colleague.)
vishalrad wrote 22 hours 37 min ago:
This is great analysis but my first reaction was - is this trolling?
The fact that we have to think about whether a $1TN company has
achieved product-market fit both articulates 1/just how high the
valuations are 2/How hard it is to pin down EXACTLY what PMF is.
As a pre-revenue startup, I am laser focused on PMF and frankly if this
is the bar, no one will achieve it. But OTOH its heartwarming that
people are willing to value you at $1TN before you reach that. Guess
everything is in the eye of the beholder?
simonw wrote 22 hours 32 min ago:
The title was indeed meant to throw subtle shade at the idea that
supposedly $1tb companies were only just now starting to figure out
product-market fit.
epolanski wrote 22 hours 51 min ago:
Wait till people figure out they can swap Claude code for DS4 Pro and
spend a fraction in API billing (actually, significantly less than
$100) while barely noticing a difference.
bambax wrote 23 hours 13 min ago:
> Thatâs $2,180.16 worth of tokens for $200
So the author claims he's getting $2000 per month worth of frontier AI
free of charge. Ok. If he's been doing that for 6 months that's $12k.
What has this produced concretely? For $12k you can find a used car in
decent condition. Heck for $1200 (his actual out-of-pocket spend) you
get a brand new ebike! (on which you could put a pelican and make a
photo of both if that's your fancy). But here it's unclear what has
come of it.
aspenmartin wrote 23 hours 9 min ago:
I would be very curious what kind of answer would satisfy you here.
Simon isn't building a product, where $200 is a line item on a
balance sheet. If he tells you what sort of analyses or time savings
$200/mo on coding agents have enabled him, do you honestly think that
would satisfy you?
simonw wrote 23 hours 11 min ago:
I've written a great deal of code - code that would have taken me
years of work to produce without LLMs.
(It's mostly open source, you're welcome to dig around in [1] and [2]
if you like.)
My time as an experienced software engineer is worth a lot of money -
a whole lot more than $12,000 for the past six months.
URI [1]: https://github.com/simonw
URI [2]: https://github.com/datasette
nevertoolate wrote 22 hours 40 min ago:
> My time as an experienced software engineer is worth a lot of
money - a whole lot more than $12,000 for the past six months
From this I assume you think that what the llm has generated is as
valuable as your own work generally is. How do you even calculate
this?
bambax wrote 22 hours 44 min ago:
> code that would have taken me years of work to produce without
LLMs
As you might suspect, this is what I have an issue with. Without
LLMs, isn't it possible or even likely that that code wouldn't have
been written at all, and wouldn't have been missed? If LLMs are
mostly used to produce throwaway prototypes then it's a stretch to
say that's money well spent.
If indeed it let you advance your main product much faster then
sure it's a different story. You're the judge of that. It's hard to
see the impact from the consumer side; everything is still broken
and no extraordinary app seems to be emerging. Maybe it's just a
question of time. We'll see.
odyssey7 wrote 20 hours 49 min ago:
Iâm watching to see what happens to big enterprise software
contracts. Why pay some vendor $800k annually for something a
couple mid-level devs can replaceâ-and tailor closely to your
needsââby leveraging AI.
Open source software changed the world. AI that will cheaply
write whatever you want in a few days will also change the world.
simonw wrote 22 hours 35 min ago:
I've thought about this a lot. I am very confident that the way I
use LLMs is both accelerating progress on my core projects
(here's a substantial, reviewed PR I landed just yesterday [1] )
and helping me create plenty of projects that otherwise would not
have existed.
URI [1]: https://github.com/simonw/datasette/pull/2741
tredre3 wrote 19 hours 57 min ago:
The point being made by GP was that your projects have no value
and their non-existence wouldn't be a negative to this world.
And that is likely a fair assessment, though I understand
perfectly the feeling that you have that you are accomplishing
great(er) things thanks to AI.
bambax wrote 12 hours 17 min ago:
I was certainly not saying that all the author's projects, in
general, have no value! That would be rude, mean and most of
all, incorrect.
But yes, it's likely that the ease of which code can now be
outout lets us produce lots of unnecessary code just because
we can, and the author says as much in a below comment
URI [1]: https://news.ycombinator.com/item?id=48303890
jonluca wrote 18 hours 17 min ago:
This is the economic theory of value creation though -
arguably the world is better off because new projects can be
created, and they are marginally cheaper than they would have
been previously
simonw wrote 19 hours 52 min ago:
I certainly hope that's not true, given that I've dedicated
7+ years to my main open source projects at this point.
I take some reassurance from knowing that they are indeed
used by real people to solve real problems though.
jwpapi wrote 16 hours 4 min ago:
Now what percentage of the 200$ have been used on the
useful stuff and how much on exploration or other stuff.
How long would it taken you to do it yourself? How much
longer will the next task take you, compared to when you
wouldâve written the code yourself. How is the mental
model compared to when you wouldâve written it yourself?.
Iâm not saying youâre wrong, again there are use cases.
But the calculation is not plain and simple it goes deep
into our perception, perceived productivity versus actual
productivity.
Iâve 2 months maxxed out all 6k of Claude Code and bought
Antigravity on top. My codebase became 140k lines. I
introduced tons of bugs and spent another 2 months,
deleting 80k of code. I wish I wouldâve just chatted with
AI and not let agents touch my codebase. I wouldâve saved
approx 300$ subscription prices a month and 2 months of my
life.
simonw wrote 16 hours 0 min ago:
I had that experience back in January! I used it for a
joke in a talk recently... did anyone really need a
half-finished buggy slow implementation of a JavaScript
interpreter in Python? They did not. [1] It's taken me
the best part of this year to readjust and find a pace
and level of ambition that fits.
URI [1]: https://github.com/simonw/micro-javascript
jwpapi wrote 9 hours 39 min ago:
It was challenging for me as well, but my pace and
level is now that I use it for these cases at the
moment:
New language, infrastructure, general level of
understanding of something I barely have an idea of.
Rubber duck debugging, if i dont know the correct
solution
Checking my code for issues and bugs.
But not for:
- writing my code
- agentic coding (help me)
The inference has reduced drastically. Itâs basically
just chatting. I donât let it write anything, but
sometimes I purposely use the browser window instead of
them sitting in my codebase, because I know it gets
things subtly wrong and migth focus on the wrong
things.
The same way people used to say donât copypaste code
at least write it out I think itâs still true. It
helps to buidl the mental model and to find the right
abstractions.
ex-aws-dude wrote 23 hours 7 min ago:
And what was your return on investment?
simonw wrote 23 hours 5 min ago:
As I commented elsewhere, I'm still bad at making money from my
open source work: [1] (I have a feeling if I could say "and I
closed $2m in sales with the software I wrote!" people would find
a way to say that didn't mean anything anyway, because how can I
prove I wouldn't have made those sales writing it by hand?)
URI [1]: https://news.ycombinator.com/item?id=48296794#48298909
brazukadev wrote 18 hours 11 min ago:
> "and I closed $2m in sales with the software I wrote!"
Given the audience you are reaching, that is actually the
expectation. Github stars is not a great metric.
atleastoptimal wrote 23 hours 23 min ago:
I think this was obvious since the birth of ChatGPT
Intelligence is a universal good, it can apply to anything, and no,
"human intelligence" is not the only form that is useful nor special.
There are limitations to AI but also huge advantages, and its obvious
that the advantages are worth paying for, given their revenue.
Szpadel wrote 23 hours 35 min ago:
> but as far as I can tell those credit costs are an exact match for
the API token costs listed for those models.
it is only true for USD.
for example if you pay in euro, this is actually more expensive. kind
of makes no sense, because it translates to $1 = â¬1
wewewedxfgdf wrote 23 hours 38 min ago:
Simon Willison just hit the "Publish to top of HN" button.
simonw wrote 23 hours 31 min ago:
Wish I'd hit that one the other day on this one, which I cared a lot
more about:
URI [1]: https://news.ycombinator.com/item?id=48228321
vb-8448 wrote 23 hours 44 min ago:
> Thatâs $2,180.16 worth of tokens for $200ânot bad at all!
Just imagine how funny it will be if it comes out that big labs were
doing some fancy maths to count the 2k$/month in their forecasts ...
vb-8448 wrote 23 hours 48 min ago:
I'm a huge fan of agent coding but kinda dislike this "llm evangelism".
There are still several open points (eg.: code churn, maintainability,
subtle bugs human will never do) that everyone with a minimal
programming knowledge that seriously used a LLM agent knows about but
somehow none of these "big influencers" never mention (or just saying
"it's your fault").
dnnddidiej wrote 23 hours 53 min ago:
Is PMF enough. It is such a dynamic self-disrupting wave that it is
like predicting physical chaos. These aren't early Googles in a blue
ocean. Maybe a blue ocean full of pirates and dragons!
This isn't me being a doomer I just don't know. Can we look at Q2
profits and draw hockey sticks yet?
Remember people are boasting how much their expenses are. That is where
we are in the bubble/new paradigm.
hintymad wrote 1 day ago:
The real timing is that we don't have strong enough new business needs
for now and we have accumulated enough tech assets, so our work has
been increasingly incremental. That means we can build reliable
features on top of vast amount of past work - where AI really shines.
So, with or without AI, companied would hire fewer software engineers
if majority of our work is incremental: add a feature here, fix a bug
there, tweak a configuration and etc, then we wouldn't need as many
software engineers anyway. AI just accelerated such squeeze.
In contrast, imagine if we had the same AI 20 years or so ago. Could AI
really write Jersey? I guess not as people were still trying to
understand JAX-RS. Could AI really answer all the questions about
React? I guess not as React was just invented. Would we use 10x fewer
people to build out infra on the public cloud or the entire so-called
Big Data platforms? I guess not, as they were still rapidly evolving
and we'd need so many engineers to explore so many different
possibilities? Could we use AI to build our ML ecosystem with 10X fewer
people? I highly doubt so. Heck, 20 years ago R was all the rage and
Python's ecosystem was not mature at all. Oh, and mobile computing,
could AI lead to 10X fewer people to build all the mobile apps and the
underlying infra?
aniceperson wrote 22 hours 14 min ago:
> Could AI really answer all the questions about React?
yes, due to ICL
> Would we use 10x fewer people to build out infra on the public
cloud or the entire so-called Big Data platforms?
No, cannot solve core problems, makes a mess at scale
You are right about the incremental work. But most of the work is
historically incremental imo, only few positions are R&D.
no-name-here wrote 15 hours 16 min ago:
ICL = In-Context Learning where LLMs learn to perform new tasks by
reading examples directly within the prompt, rather than requiring
permanent model retraining.
stego-tech wrote 1 day ago:
The big assumption with all of these sorts of analyses is that things
will continue as they are for the foreseeable future.
In hype-driven markets, you cannot be certain of that.
Let's take a view that the author is right: coding agents and their
associated harnesses were the inflection point for some degree of
profitability and widespread consumption, and that these tools are now
yet another SaaS subscription or API bucket expense to bake into every
single developer (or developer-adjacent) in the organization alongside
your collab suite, HR seat, CRM seat, design seat, etc. To be fair I
honestly think that's a safe assumption to make for highly technical
firms whose image is derived from remaining on the cutting edge of
things.
That begs the following questions, which we won't know until IPOs start
happening:
* Are subscriptions profitable, or just API consumption?
* What's the run rate when we just consider subscription-based usage
like Claude Code and Codex? What about API calls?
* Is there any profitable pathway forward at which enterprises can get
unlimited usage but at fixed rates via subscription?
* What does customer churn look like for subscription users versus API
users?
We also have a number of questions for customers that I suspect we'll
start seeing receipts for in the coming months, at least from the early
adopters:
* What was the net gain (loss) from leveraging coding agents?
* What's the cost of a developer with or without access to a coding
agent + harness? Is it cheaper to hire an outsourced worker with a
coding agent subscription, or a domestic worker without one?
* At what point does further AI spend result in diminishing returns,
i.e. where's the 'sweet spot' for spend?
* Did AI boost actual revenue and outcomes, or did it just gamify KPIs?
* What roles or work did AI actually replace, versus merely displace
during the hype cycle?
Not to mention the questions regarding the technology itself:
* Will we develop the means to run foundational/frontier models at edge
using less resources through some existing (e.g. distillation) or new
technology, thus cutting off the profit centers of these firms?
* When the market mismatch between supply and demand is resolved, won't
it be more affordable for consumers and companies to operate their own
AI infrastructure rather than support further centralized buildouts?
* Will coding agents improve to the point of being able to bootstrap
and self-orchestrate on edge/consumer hardware without substantial
technical expertise, or at least improve to the point that traditional
IT teams can securely operate them internally without an expensive
subscription or API token bucket?
All of these will influence the long tail of this bubble, because it is
a bubble at this point. Even if these companies are indeed profitable
thanks to the coding agent inflection point, there's still so many
unanswered questions about utility beyond coding that it's impossible
to extrapolate a future. If coding agents are indeed the extent of
utility for profitability, then there's no possible way these entities
will recoup the investment already sunk into their infrastructure
buildouts. Even if more profitable uses are discovered, does this
offset or replace the firms disappearing due to AI speculation and
their associated contributions to the economy as a whole (RE: the
consumer compute industry at present, higher energy costs due to
datacenter builds, opportunity cost from harms to local infrastructure
from haphazard builds, etc)? Should these firms indeed be runaway
successes and immensely profitable to the point of paying off their
investors and growing the larger economy, does this end up stifling
innovation in a world where most new ideas are fed into LLMs for R&D
that are then controlled by only a handful of companies and immensely
wealthy people, via systems that are easily surveilled and stolen from
without recourse?
So many, many questions yet to be answered. Betting the farm because of
coding agents is one hell of a gamble.
cj wrote 1 day ago:
> Coding agents really did change everything. These are tools which
burn vastly more tokens
The assumption here is that this is a positive thing.
But this very well could end up being a major negative long term by
increasing the cost per user, reducing margins.
More usage = more cost = less profit.
It's not obvious that more usage is good. It's only good if revenue per
user increases more than cost does. I'm skeptical about that.
simonw wrote 23 hours 54 min ago:
> It's only good if revenue per user increases more than cost does.
That's why it's so important for these labs that they're selling API
tokens for more than the compute+energy costs needed to generate
them.
Every indicator I've seen is that they do have a positive margin on
that. If they don't, they're screwed.
grttq wrote 20 hours 2 min ago:
No this is not what matters.
The customers of these tokens need to see returns on their projects
that exceed the cost of financing.
Laying people off only goes so far.
If enough said firms donât see enough value given the price of
frontiers they will cancel and consume open source. This is the
risk the frontier labs are exposed to.
mattas wrote 23 hours 36 min ago:
What's an example of an indicator? Genuinely curious!
simonw wrote 23 hours 33 min ago:
Insider tips from Google and AWS telling me that they run
inference at a profit (though that was over a year ago now).
Dario telling Dwarkesh three months ago that they have a margin
on inference:
URI [1]: https://www.dwarkesh.com/p/dario-amodei-2?timestamp=3528...
Denzel wrote 17 hours 1 min ago:
Unless your insider is the CFO, I wouldnât trust these
sources to have the access, knowledge, or insight to determine
whether theyâre running inference at a profit.
Simple test: can they get their hands on a data center
contracts and financials?
Search isnât anywhere near as high profile as the
profitability of AI inference, and yet, even aspects of the
Search org were walled off from the rest of the company such
that other employees couldnât see what we see.
simonw wrote 16 hours 56 min ago:
I don't think the fact that inference runs at a profit is a
particularly deeply kept secret inside these companies.
It might be worth keeping it secret from the staff if they
were running at a loss.
infinitezest wrote 22 hours 46 min ago:
Are these sources not incentivized to say exactly this,
regardless of whether it's true?
simonw wrote 22 hours 39 min ago:
The insiders tips? I don't think so, they were people I
trust.
They had all the incentive in the world to say "I'm not going
to talk about that."
_verandaguy wrote 1 day ago:
With respect to Simon, whose writing I've usually agreed with in the
past and whose insights I've liked: this is a bad take that overlooks
the extent to which corporations are imposing the use of AI on
employees, and in particular ICs, who make up a majority of the
AI-using workforce by headcount.
Many of us are either openly having our performance reviews tied to AI
use, especially at larger enterprises. Whether that's measured by sheer
token count or just "how many of your tasks are you using AI for these
days" (combined with the implication that question carries at many orgs
which are heavily invested in AI).
simonw wrote 23 hours 59 min ago:
Are you saying that Anthropic's huge leaps in revenue are caused by
stupid company policies and token leaderboards, and the moment
companies stop imposing AI on their employees revenue will drop to a
point where Anthropic are unlikely to be profitable?
I don't think that's the case. I think the token leaderboard thing
(which is clearly ridiculous) affects a tiny portion of companies and
is already going out of fashion.
_verandaguy wrote 23 hours 53 min ago:
I'm saying that the truth lies somewhere in between, and that
Anthropic's current revenue is being, in part, propped up
artificially.
We're also in a place where a lot of the usage guidance around
these tools is still nascent. People are cowboying a lot of stuff,
even as larger companies start to organize AI
policy/safety/responsible use working groups to try and policy
around the shortfalls of the technology.
IMO: if this technology persists, and if we figure out a way to use
it in a broadly safe way, the value proposition will probably trend
down rather than up, at least on the code generation front.
As a research tool, it shows some promise, though I still find the
ethics of the technology disgusting.
mtrifonov wrote 1 day ago:
They certainly have, but it relies entirely on the assistant frame,
which is a problem in and of itself for the trillion-dollar economics.
Anthropic and OpenAI have shown people want a tool for task offloading,
driving predictable token consumption and justifying the math, so long
as users stay in that dynamic.
However, knowledge workers using these tools daily are getting
exhausted with them. Outputs come out polished but hollow. Talking to a
frictionless, frame-completing model all day drains you.
If user behavior drifts away from assistant usage because of that,
per-token math implodes. The valuations we're hearing about all the
time rely on usage compounding daily. The fatigue is a timer running
against that compound.
Anthropic's Constitution is the closest hedge out there, I think.
Installing an identity structure into the model through training. But
it's still assistant-first, so the fix there is only partial.
I've spent the last year running a product that flips the architecture
so identity is primary and the assistant role is secondary. Same
frontier models, completely different conversational quality. The
fatigue property doesn't really show up.
Whichever labs figure out how to install real identity natively in the
weights are going to be the ones with PMF in the next phase.
mbesto wrote 1 day ago:
> but I suspect thereâs a more important factor here: I think
theyâve finally found product-market fit
Ahhh the classic startup term that's definition is nebulous. But also,
since when does any definition of product/market fit mean a product is
profitable? And profitable in what sense? Unit economics? Overall
company?
simonw wrote 1 day ago:
Oh I'm absolutely taking advantage of the fact that "product-market
fit" has a bit of a nebulous meaning here.
It's a great hook to build an article around. My core point is more
that April 2026 was the point when Anthropic and OpenAI finally
appeared to have figured out a credible business model.
mbesto wrote 21 hours 31 min ago:
> My core point is more that April 2026 was the point when
Anthropic and OpenAI finally appeared to have figured out a
credible business model.
How so? What's specifically changed? We still don't know what their
unit economics are and everything you've documented is basically
speculation at this point.
simonw wrote 21 hours 22 min ago:
> What's specifically changed?
1. Both Anthropic and OpenAI significantly increased the prices
of their latest models. They're clearly not trying to offer the
lowest-price-possible to drum up demand any more.
2. Both Anthropic and OpenAI no longer let enterprise companies
buy discounted almost-all-you-can-eat subscriptions. Those big
enterprises are now paying full API prices.
3. According to reasonably well-sourced leaks, Anthropic may be
about to have their first profitable quarter.
And I didn't even say "profitable", I said "credible business
model". I think getting companies to spend hundreds of dollars
per month per seat, WITHOUT crazy subscription discounts, is a
credible business model.
mbesto wrote 7 min ago:
> And I didn't even say "profitable", I said "credible business
model".
What's the difference? I would assume a business model that
rewards profit (and therefore can sustain) would be considered
credible...and not much else.
tlongo wrote 5 hours 45 min ago:
> 3. According to reasonably well-sourced leaks, Anthropic may
be about to have their first profitable quarter.
might as well be because of this:
> [...] with capacity ramping in May and June 2026 at a reduced
fee.
And then, we have
> The agreements may be terminated by either party upon 90
daysâ notice.
The timing of the "leak" of profitability is just as superb for
Anthropic as the timing of the $45b deal is for SpaceX.
I wonder how this partnership looks like on August 1st...
lbreakjai wrote 18 hours 34 min ago:
Let's see how it goes in three months. I'm not convinced it's
anything but an attempt to make the numbers look better just
long enough to IPO.
Enterprises will take a while to adjust. We're going along with
Copilot price changes in the short term "just to see what it'd
look like". They will get customers to continue with the all
you can eat mindset at API pricing for a while, but the finance
departments will start asking questions very soon.
overgard wrote 19 hours 37 min ago:
I suspect it might end up being a credible business model for
whoever buys the IP eventually (Microsoft, Google?), but I
think OpenAI and Anthropic are so far in the hole debt wise
that simply having a credible business model won't be enough to
save them - it needs to be wildly successful.
pzo wrote 1 day ago:
> If you are a heavy user of coding agents these plans are a fantastic
deal. I just ran the ccusage tool on my laptop to get an estimate of
how much I would have spent if I were to pay for API tokens in the past
30 days and got
You think this is fantastic deal only because they use similar like
tricks where they inflate the price and tell you something supposed to
cost $1000 but they have this today promo for $100.
I was there too and paying for a while. Few weeks ago I tried DeepSeek
V4 Pro - expected its gonna be shit but its actually pretty good.
The deal is I pay daily ~$1 for DSV4-pro for ~100M API token usage. And
they probably not getting broke because >90% of those token in practice
is cache read and they very well optimized for that.
conradkay wrote 22 hours 23 min ago:
GPT 5.5 is maybe 4x the size of v4 pro, hard to compare price since
cache hit is basically free with Deepseek, but 40x cheaper (with
their 75% off) seems about right.
So ballpark same price per parameter as Simon.
sourcecodeplz wrote 23 hours 10 min ago:
Yep, exactly this. And I have so much less anxiety that I have to use
my 5-hour/weekly usage or I lose it... with deepseek api the credits
never expire, I can use them when I want, how much I want and the
prices are ridiculously low for the quality/intelligence/performance.
DIR <- back to front page