There is a possibility this may not end at simply nerfing the model. The idea of manipulating the behavior of a model depending on the prompt given to it can extend to
1. Detecting if employees from competing companies are using it and sabatoge their work, even not LLM-training related
2. Direct users to outcomes that would justify higher compute spend. Deliberately coding a project to 95% completion but designed to be losing a critical step right before one's weekly rate limit is expended
3. Reduce the quality of writing when a person is writing an essay where the argument is against the interests of the model company, or steering the user using the model for brainstorming in a direction which causes them to waste time or abandon their train of reasoning
etc. etc. The possibilities are enormous. Many people use AI daily for their job, personal advice, companionship. A model company that steers the behavior of the model towards a deliberate outcome could develop a controlling interest in human behavior and productivity at large, even with subtle influence would compound enormously over its millions of users.
The ad-supported alternative suffers from the same principle-agent problem. What's to stop an ad-supported model from declining to refer you to products that would be better for your use case but who's vendors haven't paid the model's provider?
Ultimately if you can't trust the provider it is game over and you don't have an alternative other than to move to self hosted and open source solutions.
> They come from walking around and talking to people. Looking them in the eye, sussing out what their real requirements are, and figuring out how to address their concerns with empathy.
This claim depends on humans remaining the bottleneck to high-leverage information, necessitating a human<->human interfacing role that solicits requirements, ascertains intent, etc.
I don't deny that that is very important and cannot be done by AI now. However, my concern is that AI will be much better at any domain of information processing, and organizations that gate important decisions being made by a network of barriers and information silos dependent on "talking to people" will be outcompeted by largely autonomously run AI agent organizations, which have, by their very design, far higher throughput, auditing, memory, parallelization, etc.
It's kind of like saying that machines could never make fabric because it is impossible for a robot to replicate the complicated motion of a human threading a needle. The industrial revolution was prompted by creating machines which redesigned the entire process to account for machine limitations, and allowed the superior speed and scale of machines to drive higher productivity, delegating humans to a role of maintenance and simply feeding the machines their needed input.
Yes, at some point AI will be fully integrated into society so that there are entire autonomous sections. But society doesn't move at the same pace as technology.
The output of agents has to have economic value, and for the foreseeable future this means someone is going to have to buy something.
Right now, it is humans who ultimately make the economic decision. Even if you have fully autonomous agentic organizations selling to each other, there are one or more humans at the end of the of the chain who agree to exchange money for value.
Science Fiction writers have envisioned futures where some currency other than money is used to track value, but so far as I can tell we are nowhere near moving to anything like that.
It's crazy that so often I see articles, here on HN and elsewhere, where some pundit claims that there is no AI job crisis, AI isn't replacing any jobs, that layoffs are actually due to post-pandemic ZIRP overhiring, etc.
But then people who work in actual tech companies come in and explicitly say they are not hiring any juniors anymore specifically because AI is good enough to do most of what juniors do, and that senior engineers can now write 3x as much code, etc.
There seems to be a desire for a narrative that AI really just can't replace productive work, and that it's all a mirage. However it seems just like common sense that if an AI can do junior-engineer-level coding work, that a company has less reason to hire a junior engineer.
"But then people who work in actual tech companies come in and explicitly say they are not hiring any juniors anymore specifically because AI is good enough to do most of what juniors do, and that senior engineers can now write 3x as much code, etc."
If you want an anecdote: the media company I work for just started hiring interns and juniors in software career tracks again after a lengthy hiatus.
I have a suspicion that Twitter laying off so much of the software staff was very influential for people with hiring abilities. The company didn't crash, and they (relatively slowly) began shipping new features again. I think that coincided with the pandemic-era overhiring, and we've been working against that combo ever since.
Every now and then, I actively try to make an LLM replace my tasks, and fully do greenfield projects I would accept- I don't see it. It's very good, no doubt. But I have or have been given the project parameters, and just like with a junior, failures in communication inevitably lead to breakdowns in execution.
It does your job, but not completely autonomously so you don't see it?
It requires a lot of guidance, luckily, I mean thank god otherwise we would be goners. The job itself hasn't become easier, but it did change. If you come across failures you update the spec, the guard rails, whatever you use to guide it. It's not a "it produced bad code and now it's forever useless" type of situation.
I hate to be "that guy" but you're crazy if you don't see AI being able to do greenfield projects you'd accept.
I mean, in your defense, last year I think you would have been right, but right now? Codex rocks, as does Claude. I am literally making money shipping a greenfield project to a customer right now. I'm basically a cheap consultant that is incrementally adding features to make something exactly what they want for way cheaper than it would be to do it the old way.
There are hiccups, outages, things to fix etc. but the customer is happy with the output, and the reduced price means they get bespoke solutions rather than some BS one size fits all SaaS app. Then my job is maintenance and effectively "ITSM." Which kind of sucks in some ways, I miss writing real code for real projects, but this is the future going forward. If you want something for your business, you'll generate it rather than pay for it and for now at least, getting beyond localhost requires someone who knows a bit about computers or is willing to learn. Most small businesses aren't willing to learn.
Now, to your point. Is the code all that clean? Nope (and in your defense sometimes I read through the codebase and shudder)... but who cares? Like, for awhile I would go through and frantically edit it, but why? It worked. Not only that, but there's going to be a new model in 3 months or whatever that can clean it up and make it less shitty. I've literally done that a couple times since I started doing this in January.
The customer ain't reading the code. They don't care as long as the the functionality works - that's what counts. The gazillion tests I have keep it stable as I push code, and the CI/CD pipeline removes a ton of the ass pain I'd have without it.
The biggest thing I'm worried about when it comes to clean code and good design is trying to make sure I keep the token count down on these projects so I can actually do meaningful work without burning through a week's worth of tokens in a single day. That, and I like to try to keep a sort of architectural bird's eye view on what's happening...
Like, I'm not sure what niche of the industry you're in, but for the stuff I'm using it for, stuff is working really well with LLMs.
So to try and understand your position - you are hired by a small? Company in which sector? And you built an app that does what? And I think most importantly - how did you find the gig? Were they explicitly hiring a AI capable person to do X?
I run my own thing since the start of the year. I’m building little tools for an industry I’m highly familiar that needs very specific scheduling software, data tools, tracking tools etc. My old job was a boring as a government bureaucrat.
I started this by doing some work for an old employer that asked me to start by modernizing an excel spreadsheet into an app I made for them like 10 years ago? They kept asking for more though since, and they’re my biggest customer right now. Which is good because I only have the bandwidth for like one more place right now.
I’ve had a few sort of one off things with other places? But I’m working on getting another company in the same industry right now and I’ll be able to adapt most of the code I’ve built here for another company if they end up deciding to use me.
But my biggest value to companies is “I already know this industry extremely well.”
I assume they are working on low stakes software. Does it really matter if you use an LLM to code a scheduling app for a hair salon or veterinary clinic?
Think about all the geneticists that complain about excel re-writing DNA sequences.
It doesn't matter if its high stakes or low stakes. People use the software that generates the results they like. Not the software that is "correct to use".
There's no rule that a phenomenon needs 1 cause - the decline could've been junior cuts to start after '22 since they have less seniority and impact, followed by a substitution effect after broad LLM adoption. I don't doubt that remote had an additive effect on productivity via mentoring etc as well.
> There seems to be a desire for a narrative that AI really just can't replace productive work, and that it's all a mirage.
Yes, and juniors aren't known for their productive work in the beginning. That's not their purpose. Their purpose is to do the mundane work, because it is important for them to become less junior and more senior.
That is robbed of them.
Which in 5-10 years means the need for senior developers is gonna shoot through the roof.
> All the software engineers we need for the next 20-30 years are already in the current workforce.
There's no way of knowing what the industry will look like decades from now. Even assuming the prediction that seniors become 10x more productive, that would mean software becomes much cheaper to produce. Does that induce enough demand for additional software that keeps employment levels high? Could be, who knows.
Alternatively, maybe software hits a saturation point where there just isn't as much new ground to cover and employment levels crater. That could happen too.
>But then people who work in actual tech companies come in and explicitly say they are not hiring any juniors anymore specifically because AI is good enough to do most of what juniors do, and that senior engineers can now write 3x as much code, etc.
Yup, that's reflected in the data as well, no need to invoke "vibes" or whatever.
In my experience, AI raises the ceiling but also lowers the floor. It can make experienced people more productive: they can vet the output. The juniors? Not so much. I've worked with guys who write prompts that are literally "fix it." The result is about as good as you expect.
"Productive junior" is vastly different than a productive software developer with a couple decades of experience. What they produce will differ in quality significantly, AI or no AI.
And now it becomes so good, that you want to have it. Which means companies have to deal with the budget increase which they will recupe.
How? By making a 5 person team a 4 person team + AI.
If i think about my co-workers (not excluding myself) from last 15 years, there was always someone you would accept just because it was better than not having that one person. If i can now replace them with more tokens/better models, man i wouldn't hesitate (of course i know what this means on a person to person level :/)
I think that's coming at it backward. The article lacks data. Nobody should be expected to run around writing the article that's not there. As it stands, it's unsubstantiated junk.
> However it seems just like common sense that if an AI can do junior-engineer-level coding work, that a company has less reason to hire a junior engineer.
I mean, if we want to not talk about economics that’s fine, but can the AI actually do junior work at the same price? What if we don’t look only at quarterly reports, and instead include the value of having people knowing about the business having to explain it to others, who then learn it and can improve it over time?
I think it’s clear the AI is strong, there’s no doubt about that, but that’s not the whole picture.
This is a collective action problem. Its too easy to claim the answer is "we'll hire our competitors juniors when they become seniors" and then every company wants to do that and not train their own juniors. Soon no one is willing to train juniors because theyll just leave for the companies that dont train juniors.
The point is not whether it’s a mistake to not hire juniors, but whether it’s actually factoring into the hiring/layoff decisions at tech companies. Many claims are being made that no companies are actually changing hiring/layoff decisions on account of AI, and are using it as a distraction to avoid admitting their mistakes over hiring. That may be true, but many managers and execs actually do seem to consider AI a sufficient replacement for much of their engineering team, and are stalling hiring/prompting layoffs because of it.
It's not so much a question of whether AI is strong or not. It's a question of whether the tradeoffs (theft of intellectual property, coal burning, lack of transparency, stealing water, rising energy prices, global surveillance, etc.) are worth the outcomes. It's not even a serious question.
If AI was truly strong, we would be seeing signs in the job market. And we would certainly be seeing a lot more subscriptions and demand for these services. For most people, AI does not improve their lives. For a lot of them, especially younger people, it makes their lives much harder and sadder.
> include the value of having people knowing about the business having to explain it to others, who then learn it and can improve it over time?
It's always been difficult to put a number on that value, which is the problem for the MBAs running the show. There's no number on the P&L assigned to tribal knowledge, and improvements that can be made by those with that knowledge and experience within the business.
It's a mistake that businesses keep repeating, over and over again, yet never actually learning the lesson. And now the industry is going to repeat it again, until there's enough pain that they realize that lost value and start rehiring again.
well said. my wife's company has been having layoffs particularly from the overhiring during COVID. They went from 10 billion down to 1 billion thus far.
> At this point Anthropic is a pure marketing and PR company. Super catchy names like Opus, Mythos and Fable trying to get you to think that these software products are actually super-human
Lol anti-AI bias on HN is crazy. Simply giving your product a quirky name is now being considered manipulative advertising. Is just doing normal PR and marketing something AI companies aren't allowed to do?
when they keep saying “oooh this new model is too big and crazy and totally can’t be released” or “this new model is a 10x game changer totally unlike our previous iterations” it feels sort like boy crying wolf. yes they’re still pretty clearly improving models, but when you’ve hit diminishing returns / more incremental gains and you’re still saying this is sounds like pure PR hype from a company that previously been the “honest good guys” in the room
Their model did find thousands of security vulnerabilities across the companies they previewed Mythos with via project Glasswing. Is it not sensible that, given that emergent level of capability, that they do this gated release structure, as all those vulnerabilities would be exploitable by anyone using a Mythos-level model?
I think many of the jobs which aren't completely automated, but could be automated based on a explicit reading of their job requirements, are due to many implicit requirements being part of the job.
For cashiers, beyond simply ringing up customers, they serve the function of:
1. Validating IDs
2. Preventing theft
3. Creating a positive atmosphere
4. Helping customers bag groceries
5. Resolving issues/questions about products/the store
For waiters, likewise they have the job of
1. Creating a positive atmosphere
2. Physically bringing food to the table and setting the meal
3. Upselling items, providing recommendations, catering to specific unusual guest needs
etc. Basically all these jobs have a huge soft-skills dependent interface which no technology currently can replicate what humans can do.
I don't think that every job can be trivially automated by a large language model, but any job where the inputs/outputs are entirely via a computer, LLM's are approaching the point where they are equivalently enabled to a human, and there is no "real human body in-person" soft-skills interface.
When I worked at Disney, there were some jobs where people's entire jobs were compiling reports and following up with various departments. Like taking lists of security vulnerabilities from scans and getting commitment dates to fix them. They would take the data out of one system and put it in a spreadsheet. Then they would reach out and create Jira tickets for the teams responsible and then schedule meetings if necessary to discuss. These roles are definitely at risk.
Is the implication that currently it is rare to get a well-explained, fully detailed account of what someone wants, necessitating you as a "translator" of poorly specified requirements to features that actually solve the problems people are having?
My question is, is that thing which you are doing, ascertaining the subtle concerns, soliciting requirements, etc. truly out of the range of what an LLM or LLM-guided system could do?
What if the best "person" to operate an LLM is an LLM itself, or more precisely an agentic loop driven by an LLM? DO you believe this won't be the case?
> The CEO of the company is not going to be directing the LLM
Why not? Many CEOs are prompting LLMs and coding agents directly. What happens when the CEO -> LLM interaction is more efficient than CEO -> product person -> LLM?
> If you want to argue for fully autonomous companies then thats drifting into sci-fi
Why couldn't a software company be completely automated with sufficiently powerful LLM-run agents? What fundamentally is the barrier, if the models are intelligent enough?
The music would have a risk of "stopping" if these deals were backed by a speculative entity. However AI actually has real value/revenue, and is not a speculative product (i.e. people aren't buying tokens to resell them, a token is "consumed" at moment of inference)
Enron collapsed due to legitimate fraud. To imply Enron is an apt comparison requires assertion that AI companies are actually cooking the books. Is that what you are saying?
The ARR were fine but showing skewed quarterly profitability numbers by slowing down research due to hitting compute capacity suggests otherwise.
I am certain Anthropic spent less on building the next model this quarter if they make it to profitability due to the shear fact that they don't have enough compute.
Which solves the profitability problem with relative ease momentarily.
Also just to confirm, AI subscriptions are definitely being sold at a loss how big I don't know but these models are much harder to run.
API is definitely being sold at a decent profit.
So if you rate limit users and do usage billing + lower research costs which is a money pit temporarily.
(Proof is the fact that we don't have a new pre training run since 4.5 yet, they used to do one every 2 releases)
4.9 will probably be the same.
Next model Mythos doesn't seem to have a successor yet and was trained previous quarter most likely, they don't seem to have pre trained another one just improved Mythos if at all.
As much as I am into AI these attempts to show that there can be a profitable quarter seem like cooking the books, even if we assume no shady dealings otherwise.
Unless one of the Labs can say for certain training is going to stop they can't be profitable and I don't think training can stop because marginal gains is all they have.
8-12 months behind narrative for Chinese labs literally is going to kill the company that stops training first.
If we assume only a 3-6 month gap once China has more compute, then well then even if they keep training the lack of ability to arbitarily scale data centers in US, will kill them first.
DeepSeek V5 might actually just end the AI race for good.
Also given Mythos is atleast a 10x model compared to Opus, then it's pricing is likely going to be 10x as well so well token prices are likely never coming down, especially if these companies want to IPO.
Why would V5 kill the AI race? Do you believe that there are diminishing returns on model intelligence when applied to real-world tasks?
I think there are accelerating returns: i.e. a models are still not good enough to be “drop in” remote workers, but once that threshold is passed, the value of each token of inference has a far higher multiplier.
This justifies the buildup. However not everyone agrees that model intelligence will continue scaling thus they assert that eventually the economics will hit a wall.
>Also given Mythos is atleast a 10x model compared to Opus, then it's pricing is likely going to be 10x as well so well token prices are likely never coming down, especially if these companies want to IPO.
I don't know why people say this when cost per unit of intelligence has been going down continuously over the past few years. When Opus 3 was first released, its API cost was $15.00 per million input tokens and $75.00 per million output tokens. Opus 4.8. which is significantly better, is $5.00 per 1 million input tokens and $25.00 per 1 million output tokens
Assuming 2-3 years from now when V5 is out China would have mostly caught up in compute, and honestly that's it China can scale up compute a lot faster than US maybe a few countries can match it, or help match it but won't happen while US Iran thing is going on.
Further the human costs in the loop for AI training are insanely low or atleast substantially lower outside of US, so sure without the Nvidia upcharge I think everyone else who can use Compute from China is at an advantage.
If the assumption is AI is scaling issue then China will win because they can do infrastructure. Maybe if US wasn't in a trade war with rest of the planet there was some hope but I don't think so.
Once Deepseek figures out the new compute and can get it on par with Nvidia's clusters even if by using 4x the energy(cause they can). I don't think OpenAI or Anthropic can maintain a lead, if they don't have a lead the pricing difference will kill the AI race.
The best case scenario is OpenAI and Anthropic are dead in 2-5 years once China is caught up.
The worst case scenario where AI is not a productive boost is that well the thing pops.
Either way I don't see how this works out. Sure US govt could bomb China that's always an option.
>The ARR were fine but showing skewed quarterly profitability numbers by slowing down research due to hitting compute capacity suggests otherwise.
I have to say, I find this really puzzling. We know for a fact that Anthropic are making bank on metered inference. That's their biggest source of profitability, we are seeing software companies start to majorly adopt coding agents over just the last few months.
Right as the biggest driver of enterprise adoption is accelerating, and it's tied to their biggest profit vector, you find it suspect that their profits are increasing significantly?
Also, can you clarify what you mean by "slowing down research" exactly? Do you mean they're not doing big pretraining runs? Less compute available for researchers? Scaled back RL?
>Also just to confirm, AI subscriptions are definitely being sold at a loss how big I don't know but these models are much harder to run.
Maximum usage of AI subscriptions is a loss, but do we actually know how that nets out? Has anyone done any research to try to figure that out?
> can you clarify what you mean by "slowing down research"
He is claiming that they have been investing less in R&D and that this is juicing their numbers in an unsustainable way given how close the competition is to catching up. His evidence is the content and cadence of model releases recently. (I'm not taking a position one way or the other, just clarifying for you.)
> Maximum usage of AI subscriptions is a loss, but do we actually know how that nets out?
They almost certainly don't have to care. All the enterprise accounts use the API pricing AFAIK and that appears to be profitable and is expected to be the vast majority of the usage in the medium to long term (if it isn't already).
> API is definitely being sold at a decent profit.
Where do you get this from?
Enterprise plans are being cancelled or limited all over the place (Uber, Microsoft). I doubt Anthropic would be leveraging a loss leader with their consumer plans, while catastrophically hemorrhaging customers on the enterprise.
They are either operating at a loss (possibly a minor one), or a minor profit (which is chasing customers away).
If they were comfortably profitable they wouldn't need to participate in the circular deal circus.
It would be insane, if they can't serve the models at a profit sure at current GPU prices the profit might be 10% or lower.
But at realistic gpu prices it would have been close to 30-60% based on how big the models actually are and how much they have optimized the stack to serve them.
1T parameter models like Kimi K2.6 can be served for 1/10 to 1/5 of the price of opus 4.8 for perspective.
Sure opus is 2x the size and hosting might be non linearly scaling so still it should be around 50% margin at regular gpu prices.
If it isn't I would be very surprised.
Also for enterprises we joke but Google is not paying same rates as us there are big massive enterprise discounts. I have heard upto 20-30%...
OpenAI is supposedly even more generous.
I don't think API is being sold at a loss at the end of the day even if the API profits are marginal 10-20% because of insane GPU prices now.
Please address the primary point first: Selling some product does not disprove speculation.
In the case of Enron, people were obviously speculating in its stock, and that remains true regardless of why it collapsed later, or even whether it collapsed at all.
I say "first" because if you still can't agree that speculation in AI stocks even exists, then it's pointless to discuss what people might be doing to exploit or encourage it.
Speculation exists for every security. However wrt revenue numbers, Anthropic/OpenAI’s revenues are largely made of companies/individuals purchasing tokens. Enron’s was accounting which stated future potential revenue as current earnings. They are not the same. Enron pulled off a lot of shady schemes to hide their accounting practices. All of the “circular deals” AI labs are doing are publicly known and clear to see, so its not like anyone who knows what a circular deal simply knows something everyone doesn’t.
Also to be more specific about our point of disagreement, I think we are referring to speculation in different domains. When I brought it up, I am referring to the fact that any companies whose revenue is driven by a speculative bubble (like what precipitated the 2008 crisis) would be at risk of massive losses "if the music stops". Anthropic/OpenAI aren't flipping assets. It is true that VC funding is based on speculation, but their core business model is producing massive revenue growth on selling tokens.
It's an interesting point that the token revenue will presumably survive a crash in stock prices. But (IIUC) much of the new infrastructure is funded using stock is it not? So it seems like token revenue theoretically surviving doesn't address the risk to the rest of the economy here. And if the economy takes a large enough hit then presumably so will token spend because someone has to pay for that after all.
Sure their actual immediate revenue is driven by concrete numbers but when the rest of the economy is reorganizing itself based on their projected future revenue is the former observation still relevant?
That is true, if all the new data centers don’t produce revenue then there will be a crash. However you’d have to bet that the models won’t stop getting better, or if they still keep getting better, that somehow better models does not translate to increased productivity. Would it be wise to look at how AI has progressed over the last 5 years and make that bet?
It's possible to question the accuracy of the projections without disputing that the numbers are expected to go up. It's not that the new data centers wouldn't produce any revenue but rather that those numbers are where unfounded speculation could be happening. If and when those numbers fail to materialize (or when investors revise their projections) would presumably be the point at which the music stops.
Recall that the exchange earlier called into question the similarity or difference to enron. Sure, the current revenue numbers don't appear to be cooked but if the future revenue numbers are unrealistic and everyone is using those future numbers to make their decisions then isn't the end result roughly analogous? Blatant fraud not withstanding of course.
Note that I'm not claiming the above to be the case. Merely illustrating the commonality and acknowledging the possibility.
Remember when nvidia asked us to stop calling them enron because unlike enron they actually admit to doing all the things enron did so it's not illegal?
Circular dealing or round tripping is a form of cooking books and sometimes results in accounting fraud. Especially when circular revenue is booked without cash flow growth. Do you see cash flow growth on any side of these transactions.
The value of AI companies is speculative just like the railroads were. Railroads also have real value. But you have to have everything ready to use those railroads to make money, or they're just steel bars in the dirt and a big loud heavy thing that moves along the horizon. Too much speculative investment in the railroads (in part) led to the panic of 1873, because just having a promise of a return isn't the same thing as having the return.
1. Detecting if employees from competing companies are using it and sabatoge their work, even not LLM-training related
2. Direct users to outcomes that would justify higher compute spend. Deliberately coding a project to 95% completion but designed to be losing a critical step right before one's weekly rate limit is expended
3. Reduce the quality of writing when a person is writing an essay where the argument is against the interests of the model company, or steering the user using the model for brainstorming in a direction which causes them to waste time or abandon their train of reasoning
etc. etc. The possibilities are enormous. Many people use AI daily for their job, personal advice, companionship. A model company that steers the behavior of the model towards a deliberate outcome could develop a controlling interest in human behavior and productivity at large, even with subtle influence would compound enormously over its millions of users.
reply