How will AI affect the price of real goods and their inputs: lumber, food, electricity, textiles and the like? And will companies pass on the service-based savings to consumers?
The point is that if this article is correct about the assumption that AI is capable enough to reduce the friction of consumers rationally comparing options for a far wider basket of goods, then competition will reduce prices. No company wants to reduce prices if they don't have to -- their hand is forced by declining market share (or, with discounters, price reductions are a deliberate strategy to increase market share and absolute profit).
The bull case for AI and consumer welfare is 1) turning more markets into "perfect competition" like airline tickets, and 2) driving actual prices lower because the marginal cost of production is lower with less labor. Even if real inputs don't change, removing labor will reduce marginal cost (which implies that you'll see the largest price declines in labor-intensive industries).
but doesn't this implicitly assuming the reduction in labour cost and household income is like for like price reduction. When there's probably just as likely (& if not more) that the reduction in household income is lopsided so overall ability to purchase is reduced. You'll then need fewer people to buy more to make up for the loss of some buyers which is unlikely for many goods.
I do think the idea that AI is good for economic growth is for the fairies personally under the current model. I cant square the circle of a consumer based economic model will see higher growth by the apparently significant reduction in said consumers income.
I think the clear point of this piece is that we have the space and opportunity now to ask ourselves as a group: what are we doing? Who actually stands to benefit from the massive devaluation of services in an economy that is buoyed by service-based roles?
There has been so little thought to the multi-order effects of the future we're pushing toward, and even if AI fails to deliver on its lofty promises, it will likely cause an economic crisis in its collapse.
The people saying that AI will rapidly drive costs down are frankly delusional. The things that people actually need to live like food, shelter, and clothes all have inputs that are physical and real. Even if AI somehow can drive the input costs of those things down, it will be delayed, and people will suffer in the interim.
The AI future that I worry about isn't the terminators coming to get us, it is the top 0.1% using this technology to accumulate more wealth. Unlike feudalism, however, the feudal lords will not be dependent on or responsible for the serfs, they can rely on a small minority of humans for production of critical goods for themselves.
These wealthy people don't really hide how they feel either[1], they are clearly stating their contempt for the unwashed masses below them. As Lasch predicted in his "Revolt of the Elites," they are separating themselves entirely from culture in favor of their own insulated fiefdoms. This is already happening: companies more than ever are orienting toward ultra-luxury: from travel, to housing, and everything in-between.
None of the productivity increase of the last 50 years have benefited the workers, there is absolutely no reason for that to change
Is everyone on this website 20 years old? They pulled the same shit with automation, with computers, with internet, with cryptos, and now with ai,... And people keep falling for the same bs over and over again. "the three day workweek", "we'll retire at 45", &c.
You're assuming that the gains from productivity improvements distribute themselves broadly. The last 50 years have clearly shown that this is certainly not the case unless there is political intervention. The elites and the political class attached to them will assign whatever meanger rations they can to avoid revolution but not much else.
Not to mention: the grand majority of the US's GDP is wrapped up into services. If AI can flatten the skill floor so that anyone from anywhere in the world can produce 80% of the output of a US or European skilled worker at a fraction of the cost, what do you think happens? We're doing to US white collar work what offshoring did to manufacturing, but it'll be faster and to the only healthy cohort of economic actors in the US.
AI does not control the inputs of lumber or vegetables.
> We will need good tools to distribute the gains.
There is enormous handwaving happening here. Tools built by whom? The US can hardly pass a budget now, and its dominant political movement is allergic to questions of wealth redistribution. And as I already mentioned, the wealthy class in the US is clearly openly contemptuous of the idea that they owe anything to the broader population.
Right. So then when is the best time for labor to act to ensure those mechanisms are put in place? Before or after AI has eliminated its leverage?
Like I totally realize we're agreed in some sense, but some form of socialism now in the US seems politically untenable, and as soon as AI actually starts making service labor obsolete, we lose our leverage. How do we do something about it?
How do you mean "eliminating the leverage"? White collar jobs go first, robots are right behind. I am relatively certain everyone will be scared shitless of where this is potentially going fairly soon.
The leverage will simply come from existing and being in the same group as roughly ~everyone on the planet.
I mean like: being alive isn't leverage, but controlling the economic output that has made elites rich is.
Mass economic displacement often doesn't produce solidarity, it produces fragmentation, scapegoating, and strongmen. People won't agree on why things are bad.
And the state is way more sophisticated than during the labor movements of the last century or two. If it is aligned with the elites and the masses have no leverage, it has numerous tools for massive surveillance, police and military control, and propaganda that will only get more effective through AI. They might not even need the extreme levers, mass confusion alone might be enough.
There's that old joke:
"I have to admit, I'm always so impressed by Soviet propaganda. You really know how to get people worked up," the CIA agent says.
"Thank you," the KGB says. "We do our best but truly, it's nothing compared to American propaganda. Your people believe everything your state media tells them."
The CIA agent drops his drink in shock and disgust. "Thank you friend, but you must be confused... There's no propaganda in America."
“ We will need good tools to distribute the gains.”
As of now there are basically no such tools. Everything is geared towards letting owners accumulate more and more. We would probably need highly progressive taxes to get some level of distribution
Regardless of the promise of the underlying technology, I do wonder about the long-term viability of companies like OpenAI and Anthropic. Not only are they quite beholden to companies like Nvidia or Google for hardware, but LLM tech as it stands right now will turn into a commodity.
It's why Amodei has spoken in favor of stricter export controls and Altman has pushed for regulation. They have no moat.
I'm thankful for the various open-weighted Chinese models out there. They've kept good pace with flagship models, and they're integral to avoiding a future where 1-2 companies own the future of knowledge labor. America's obsession with the shareholder in lieu of any other social consideration is ugly.
I think google ends up the winner. They can keep chugging along and just wait for everyone else to go bankrupt. I guess apple sees it too since the signed with google and not
OpenAI.
In addition to that, Google and Apple are demonstrated business partners. Google has consistently paid Apple billions to be the default search engine, so they have demonstrated they pay on time and are a known quantity. Imagine if OpenAI evaporated and Siri was left without a backend. It'd be too risky.
The minute Apple chose Google, OpenAI became a dead duck. It will float for a while but it cannot compete with the likes of Google, their unlimited pockets and better yet their access to data
I think it points to OpenAI trying to pivot to leveraging their brand awareness head start and optimizing for either ads or something like the Jony Ive device- focusing on the consumer side.
For now people identify LLMs and AI with the ChatGPT brand.
This seems like it might be the stickiest thing they can grab ahold of in the long term.
Consumer AI is not going to come close to bailing them out. They need B2B use cases. Anthropic is a little better positioned because they picked the most proven B2B use case — development — and focused hard on it. But they'll have to expand to additional use cases to keep up with their spend and valuation, which is why things like cowork exist.
But I tend to agree that the ultimate winner is going to be Google. Maybe Microsoft too.
Unless you're totally dumb or a super genius, LLMs can easily provide that kind of monthly value to you. This is already true for most SOTA models, and will only become more true as they get smarter and as society reconfigures for smoother AI integration.
Right now we are in the "get them hooked" phase of the business cycle. It's working really damn well, arguably better than any other technology ever. People will pay, they're not worried about that.
It would have to be $60-$80/mo. in value over and above what you could get at the same time with cheap 3rd party inference on open models. That's not impossible depending on what kind of service they provide, but it's really hard.
The value is well worth over $60-$80/mo. But conflating that with the market condition is very different.
In the world where you cheap open weight models and free tier closed sources models are flooding the market, you need very good reason to convince regular people to pay for just certain models en masse in b2c market
After 30 years with a shit operating system known as Windows, Linux still cannot get over 5% adoption. Despite being free and compatible with every computer.
"Regular People" know ChatGPT. They know Gemini (largely because google shoves it in their face). They don't know anything else (maybe Siri, because they don't know the difference, just that siri now sucks). I'm not sure if I would count <0.1% of tokens generated being "flooding the market".
Just like you don't give much thought to the breed of grass growing in your yard, they don't give much thought to the AI provider they are using. They pay, it does what they want, that's the end of it. These are general consumers, not chronically online tech nerds.
> After 30 years with a shit operating system known as Windows, Linux still cannot get over 5% adoption. Despite being free and compatible with every computer.
You need to install linux and actively debugging it. For ai, regular people can just easily switch around by opening an browser. There are many low or 0 barrir choices. Do you know windows 11 is mostly free too for b2c customers now? Nobody is paying for anything
> "Regular People" know ChatGPT. They know Gemini (largely because google shoves it in their face). They don't know anything else (maybe Siri, because they don't know the difference, just that siri now sucks). I'm not sure if I would count <0.1% of tokens generated being "flooding the market".
You just proved my point. Yes they are good, but why would people pay for it? Google earns money through ads mostly.
> Just like you don't give much thought to the breed of grass growing in your yard, they don't give much thought to the AI provider they are using. They pay, it does what they want, that's the end of it. These are general consumers, not chronically online tech nerds.
That's exactly the points, because most of the internet services are free. Nobody is paying for anything because they are ads supported
It's nothing to do with Windows but with the applications (including games) that just run on it and the fact that most companies just run it it by default.
I don't see that. I've used LLMs and I've seen very little direct value. I've seen some value though Photoshop etc. But nothing I'd pay for a direct subsciption for.
It doesn't matter. I firmly believe both OpenAI and Anthropic are toast. And I aay this as someone that uses both Codex and Claude primarily.
I really dislike Google, but it is painfully obvious they won this. Open AI and Anthropic bleed money. Google can bankroll Gemini indefinitely because they have a very lucrative ad business.
We can't even argue that bankrolling Gemini for them is a bad idea. With Gemini they can have yet another source of data to monetize users from. Technically Gemini can "cost" them money forever, and it would still pay for itself because with it they can know even more data about users to feed their ad business with. You tell LLMs things that they would never know otherwise.
Also, they mostly have the infrastructure already. While everyone spends tons of money to build datacenters, they have those already. Hell, they even make money by renting compute to AI competitor.
Barred some serious unprecedented regulatory action against them (very unlikely), I don't see how they would lose here.
Unfortunately, I might add. i consider Google an insidiously evil corporation. The world would be much better without it.
They also have tons of data on the users' habits and desires which they can use to inform the AI with each specific user's preferences without them having to state them. Because so many people use Google maps, Gmail etc. It's not just about training data but also operational context. The others lack this kind of long-term broad user insight.
I'm not using Google services much at all and I don't use Gemini but I'm sure it will serve the users well. I just don't want to be datamined by a company like Google. I don't mind my data improving my services but I don't want it to be used against me for advertising etc.
Yes, I think that’s their plan. Remember when Altman got fired from OpenAI? Msoft was right there with open arms. Msoft is probably letting OpenAI do the dirty work of fleecing investors and then when all their money is gone doing the R/D, MSoft scoops up the IP and continues on.
sort of not really but effectively yes. their deal with OpenAI gave them unrestricted use of all of OpenAI Models and IP (source code, weights, patents probably any other data) except the eventual hypothetical end products of Artificial General Intelligence which would belong to OpenAI alone but Microsoft would still have everything leading to it so could probably make that jump on their own (not a great deal on OpenAI's part as it doesn't give them much of a moat). so when OpenAI runs out of money well Microsoft won't own the IP but will have unrestricted use of it some one else could buy it at bankruptcy but microsoft could still use it. As for the staff well they already showed a willingness to jump ship to Microsoft back when the OpenAI board tried firing Sam without giving a reason, and if OpenAI dies Microsoft would probably hire any of the top talent that applied. So kinda sort of but on paper no but yeah they would have everything of value they would choose have.
I hope I see Anthropic and OpenAI shutter within my lifetime.
Google has been guilty of all of the same crimes, but it bothers me to see new firms pop up with the same rapacious strategies. I hope Anthropic and OpenAI suffer.
You better hope Anthropic and OpenAI thrive, because a world in which Google is the sole winner is a nightmare.
Google's best trick was skirting the antitrust ruling against them by making the judge think they'd "lose" AI. What a joke.
Meanwhile they're camping everyone's trademarks, turning them into lucrative bidding wars because they own 92% of the browser URL bars.
Try googling for Claude or ChatGPT. Those companies are shelling out hundreds of millions to their biggest competitor to defend their trademarks. If they stop, suddenly they lose 60% of their traffic. Seems unfair, right?
I understand that Google is an extraordinarily bloated monopoly.
What I mean is that I am so bitter about OpenAI and Anthropic's social media manipulation and the effects of AI psychosis on the people around me that I would gladly accept a worse future and a less free society just to watch them suffer.
Also, Sam Altman (at least) gives the impression of being a bit of a manipulative psychopath. Even if there are others out there like him, who are just more competent at hiding their tendencies, I really don't want him to win the "world's richest man" jackpot; it'd be a bad lesson to others. Steve Jobs hero-worship is bad enough.
I'm waiting to see a more egregious company than openai and a bigger scammer ceo like altman. no, thank you. i hope openai goes bankrupt. especially since the ousting of ilya.
Honestly at this point, I don't care which company lives or dies.
Because recent open source models have reached my idea of "enough". I just want the bubble to burst, but I think the point of the bubble burst is that Anthropic and OpenAI couldn't survive whereas Google has chances of survival but even then we have open source models and the bubble has chances of reducing hardware costs.
OpenAI and Anthropic walked so that Google or Open source models could run but I wish competition and hope that maybe all these companies can survive but the token cost is gonna cost more, maybe that will tilt things more towards hardware.
I just want the bubble to burst because the chances of it prolonging would have a much severe impact than what improvements we might see in Open source models. And to be quite frank, we might be living an over-stimulus of "Intelligence", and has the world improved?
Everything I imagined in AI sort of reached and beyond and I am not satisfied with the result. Are you guys?
I mean, now I can make scripts to automate some things and some other things but I feel like we lost something so much more valuable in the process. I have made almost all of my projects with LLM's and yet they are still empty. Hollow.
So to me, the idea of bursting the bubble is of the utmost importance now because as long as the bubble continues, we are subsiziding the bubble itself and we are gonna be the one who are gonna face the most impact, and well already are facing it.
in hindsight, I think evolution has a part in this. We humans are so hard coded to not get outside of the tribe/the-newest-thing so maybe collectively us as a civiliazation can get dis-enchanted first via crypto now AI but we also can think for ourselves and the civilization is built from us in my naive view.
So the only thing we can do is think for ourselves and try to learn but it seems as if that's the very thing AI wants to offload.
Don't want to sound rude, but anytime anyone says this I assume they haven't tried using agentic coding tools and are still copy pasting coding questions into a web input box
I would be really curious to know what tools you've tried and are using where gemini feels better to use
It's good enough if you don't go wild and allow LLMs to produce 5k+ lines in one session.
In a lot of industries, you can't afford this anyway, since all code has to be carefully reviewed. A lot of models are great when you do isolated changes with 100-1000 lines.
Sometimes it's okay to ship a lot of code from LLMs, especially for the frontend. But, there are a lot of companies and tasks where backend bugs cost a lot, either in big customers or direct money. No model will allow you to go wild in this case.
My experience is that on large codebases that get tricky problems, you eventually get an answer quicker if you can send _all_ the context to a relevant large model to crunch on it for a long period of time.
Last night I was happily coding away with Codex after writing off Gemini CLI yet again due to weirdness in the CLI tooling.
I ran into a very tedious problem that all of the agents failed to diagnose and were confidently patching random things as solutions back and forth (Claude Code - Opus 4.6, GPT-5.3 Codex, Gemini 3 Pro CLI).
I took a step back, used python script to extract all of the relevant codebase, and popped open the browser and had Gemini-3-Pro set to Pro (highest) reasoning, and GPT-5.2 Pro crunch on it.
They took a good while thinking.
But, they narrowed the problem down to a complex interaction between texture origins, polygon rotations, and a mirroring implementation that was causing issues for one single "player model" running through a scene and not every other model in the scene. You'd think the "spot the difference" would make the problem easier. It did not.
I then took Gemini's proposal and passed it to GPT-5.3-Codex to implement. It actually pushed back and said "I want to do some research because I think there's a better code solution to this". Wait a bit. It solved the problem in the most elegant and compatible way possible.
So, that's a long winded way to say that there _is_ a use for a very smart model that only works in the browser or via API tooling, so long as it has a large context and can think for ages.
You need to stick Gemini in a straightjacket; I've been using https://github.com/ClavixDev/Clavix. When using something like that, even something like Gemini 3 Flash becomes usable. If not, it more often than not just loses the plot.
Every time I've tried to use agentic coding tools it's failed so hard I'm convinced the entire concept is a bamboozle to get customers to spend more tokens.
My guess is that Google has teams working on catching up with Claude Code, and I wouldn't be surprised if they manage to close the gap significantly or even surpass it.
Google has the datasets, the expertise, and the motivation.
I've had the same experience with editing shaders. ChatGPT has absolutely no clue what's going on and it seems like it randomly edits shader code. It's never given me anything remotely usable. Gemini has been able to edit shaders and get me a result that's not perfect, but fairly close to what I want.
have you compared it with Claude Code at all? Is there a similar subscription model for Gemini as Claude? Does it have an agent like Claude Code or ChatGPT Codex? what are you using it for? How does it do with large contexts? (Claude AI Code has a 1 million token context).
I tried Claude Opus but at least for my tasks, Gemini provided better results. Both were way better than ChatGPT. Haven't done any agents yet, waiting on that until they mature a bit more.
Gemini 3.1 (and Gemini 3) are a lot smarter than Claude Opus 4.6
But...
Gemini 3 series are both mediocre at best in agentic coding.
Single shot question(s) about a code problem vs "build this feature autonomously".
Gemini's CLI harness is just not very good and Gemini's approach to agentic coding leaves a lot to be desired. It doesn't perform the double-checking that Codex does, it's slower than Claude, it runs off and does things without asking and not clearly explaining why.
(Claude Code now runs claude opus, so they're not so different.)
>it's [Gemini] nowhere near claude opus
Could you be a bit more specific, because your sibling reply says "pretty close to opus performance" so it would help if you gave additional information about how you use it and how you feel the two compare. Thanks.
On top of every version of Gemini, you also get both Claude models and GPT-OSS 120B. If you're doing webdev, it'll even launch a (self-contained) Chrome to "see" the result of its changes.
I haven't played around Codex, but it blows Claude Code's finicky terminal interface out of the water in my experience.
It is a rather attractive view, and I used to hold it too. However, seeing as Alphabet recently issued 100-year bonds to finance the AI CapEx bloat, means they are not that far off from the rest of the AI "YOLO"s currently jumping off the cliff ...
They have over $100B in cash on hand. I can't pretend to understand their financial dealings, but they have a lot more runway before that cliff than most of the other companies.
This is the conclusion I came to as well. Either make your own hardware, or drown paying premiums until you run out of money. For a while I was hopeful for some competition from AMD but that never panned out.
Google has proven themselves to be incapable of monetizing anything besides ads. One should be deeply skeptical of their ability to bring consumer software to market, and keep it there.
They don't have the know how (except by proxy via OpenAI) nor custom hardware and somehow they are even worse at integrating AI into their products than Google.
They don’t need to. Just like Amazon they are seeing record revenues from Azure because of their third party LLM hosting platforms only being gated because no one can get enough chips right now
I was thinking about that (I definitely agree with you on the software and data angle).
But when you think about it it's actually a bit more complex. Right now (eg) OpenAI buys GPUs from (eg) NVidia, who buys HBM from Samsung and fabs the card on TSMC.
Google instead designs the chip, with I assume a significant amount of assistance of Broadcom - at least in terms of manufacturing, who then buys the HBM from the same supplier(s) and fabs the card with TSMC.
So I'm not entirely sure if the margin savings are that huge. I assume Broadcom charges a fair bit to manage the manufacturing process on behalf of Google. Almost certainly a lot less than NVidia would charge in terms of gross profit margins, but Google also has to pay for a lot of engineers to do the work that would be done in NVidia.
No doubt it is a saving overall - otherwise they wouldn't do it. But I wonder how dramatic it is.
Obviously Google has significant upside in the ability to customise their chips exactly how they want them, but NVidia (and to a lesser extent) AMD probably can source more customer workflows/issues from their broader set of clients.
I think "Google makes its own TPUs" makes a lot of people think that the entire operation in house, but in reality they're just doing more design work than the other players. There's still a lot of margin "leaking" through Broadcom, memory suppliers and TSMC so I wonder how dramatic it is really is
My take is it's the inference efficiency. It's one thing to have a huge GPU cluster for training, but come inference time you don't need nearly so much. Having the TPU (and models purpose built for TPU) allows for best cost in serving at hyperscale.
Yes potentially - but the OG TPUs were actually very poorly suited for LLM usage - designed for far smaller models with more parallelism in execution.
They've obviously adapted the design but it's a risk optimising in hardware like that - if there is another model architecture jump the risk of having a narrow specialised set of hardware means you can't generalise enough.
Prefill has a lot of parallelism, and so does decode with a larger context (very common with agentic tasks). People like to say "old inference chips are no good for LLM use" but that's not really true.
NVidia is operating with what, 70% gross margin? That’s what Google saves. Plus, Broadcom may be in for the design but I’m not sure they’re involved in the manufacturing of TPUs.
Just by removing nvidia's profits Google gets their TPUs for like 25-30% of what they cost to get them from nvidia, assuming similar cost structures. Google's cost structures are probably higher than nvidia's so realistically theyre probably paying around 50% of nvidia charges, but thats still billions of dollars a year and allows them to create a product that is tailor made for their needs.
Yeah this is a bummer. If it goes south everyone in power will also have perfect hindsight and say they saw it coming because obviously you shouldn't have this much built on such a small footprint. And yet...
> Yeah this is a bummer. If it goes south everyone in power will also have perfect hindsight and say they saw it coming because obviously you shouldn't have this much built on such a small footprint. And yet...
It'll be true, everyone does see it coming (just like with rare earth minerals). But the market-infected Western society doesn't have the maturity to do anything about it. Businesses won't because they're expected to optimize for short-term financial returns, government won't because it's hobbled because biases against it (e.g. any failure becomes a political embarrassment, and there's a lot of pressure to stay out of areas where businesses operate and not interfere with businesses).
America needs a lot more strategic government control of the economy, to kick businesses out of their short-term shareholder-focused thinking. If it can't manage that, it will decline into irrelevance.
When USSR fell there was a lot of talk about how it was meant to be since the US system is the best of all the terrible systems. It deserved to win and the USSR system deserved to die.
>America needs a lot more strategic government control of the economy, to kick businesses out of their short-term shareholder-focused thinking. If it can't manage that, it will decline into irrelevance.
If it is meant to be then its meant to be. If the US decides to cling to its old system and it fails well then we would know that it wasn't the best system after all. Humankind will keep moving forward even if it means that another continent controls the show.
While I think Gemini is the worst of the three big competitors, Waymo is an superb example of this talent. Kudos to Google engineers for producing so many diamonds despite producing many terrible flops over the years. We might find out their system of organization was the best one after all.
downvote all you want. google has all the money to keep up and just wait for the others to die. apple is a different story, btw, can probably buy openai or anthropic, but for now they're just waiting like google, and since they need to provide users AI after the failure with Apple Intelligence, they prefer to pay for Google and wait for the others to fight against each other.
openai and anthropic know already what will happen if they go public :)
That’s not a well informed argument. Even if Apple could finance the $1T+ it would cost to buy Anthropic - they’re not making that money back by making the iPhone a little better. The only way to monetize is by selling, as Anthropic does, enterprise services to businesses. And that’s not Apple’s “DNA,” to use their language.
Google is vulnerable in search and that already shows as we see a decline as many parallel paths emerge. At the beginning it was a simple lookup for valid information and it became dominant - then pages of pay ranked preference spots filled pages that obscured what you wanted = it became evil.
We see no such thing. Google just announces review revenue and profit and Apple hinted at it not seeing any decline in revenue from their search deal with Google which is performance based.
And Gemini is already integrated into the results page and gives useful answers instantly, alongside advertising... What problem for google are you seeing?
Google is the new Open AI.
Open AI is the new Google. Guess who wants to shove advertisements into paying customers' face and take a % of their revenues for using their models to build products? Not Google.
Google's main revenue source (~ 75%) is advertising. They will absolutely try to shove in ads into their AI offerings. They simply don't have to do it this quickly.
> Guess who wants to shove advertisements into paying customers' face and take a % of their revenues for using their models to build products? Not Google.
A search engine which you leech for free, not pay and get ads shoved onto your face. The core argument here is getting ads into your face despite having paid for the product. And also, no one is forcing you to use Google when you have so many alternatives.
OpenAI is not viable. OpenAI is spending like Google without a warchest and they have essentially nothing to offer outside of brand recognition. Nvidia propping them up to force AI training to be done on their chips vs. google in-house cores is their only viable path forward. Even if they develop a strong model the commitments they've made are astronomically out of reach of all but the largest companies and AI has proven to be a very low moat market. They can't demand a markup sufficient to justify that spend - it's too trivial to undercut them.
Google/Apple/Nvidia - those with warchests that can treat this expenditure as R&D, write it off, and not be up to their eyeballs in debt - those are the most likely to win. It may still be a dark-horse previously unknown company but if it is that company will need to be a lot more disciplined about expenditures.
OpenAI and Anthropic don't have a moat. We will have actual open models like DeepSeek and Kimi with the same functionality as Opus 4.6 in Claude Code <6mo IMO. Competition is a good thing for the end-user.
The open-weight models are great but they're roughly a full year behind frontier models. That's a lot. There's also a whole lot of uses where running a generic Chinese-made model may be less than advisable, and OpenAI/Anthropic have know-how for creating custom models where appropriate. That can be quite valuable.
I would not say a full year... not even close to a year: GLM-5 is very close to the frontier: https://artificialanalysis.ai/
Artificial Analysis isn't perfect, but it is an independent third party that actually runs the benchmarks themselves, and they use a wide range of benchmarks. It is a better automated litmus test than any other that I've been able to find in years of watching the development of LLMs.
Benchmarks are always fishy, you need to look at things that you'd use the model for in the real world. From that point of view, the SOTA for open models is quite behind.
If benchmarks are fishy, it seems their bias would be to produce better scores than expected for proprietary models, since they have more incentives to game the benchmarks.
No... benchmarks are not always "fishy." That is just a defense people use when they have nothing else to point to. I already said the benchmarks aren't perfect, but they are much better than claiming vibes are a more objective way to look at things. Yes, you should test for your individual use case, which is a benchmark.
As I said, I have been following this stuff closely for many years now. My opinion is not informed just by looking at a single chart, but by a lot of experience. The chart is less fishy than blanket statements about the closed models somehow being way better than the benchmarks show.
That's a lot now, in the same way that a PC in 1999 vs a PC in 2000 was a fairly sizeable discrepancy. At some point, probably soon, progress will slow, and it won't be much.
I just did a test project using K2.5 on opencode and, for me, it doesn’t even come close to Claude Code. I was constantly having to wrangle the model to prevent it from spewing out 1000 lines at once and it couldn’t hold the architecture in its head so it would start doing things in inconsistent ways in different parts of the project. What it created would be a real maintenance nightmare.
It’s much better than the previous open models but it’s not yet close.
>various open-weighted Chinese models out there. They've kept good pace with flagship models,
I don't think this is accurate. Maybe it will change in the future but it seems like the Chinese models aren't keeping up with actually training techniques, they're largely using distillation techniques. Which means they'll always be catching up and never at the cutting edge. https://x.com/Altimor/status/2024166557107311057
> they're largely using distillation techniques. Which means they'll always be catching up and never at the cutting edge.
You link to an assumption, and one that's seemingly highly motivated.
Have you used the Chinese models? IMO Kimi K2.5 beats everything but Opus 4.6 and Gemini 3.1... and it's not exactly inferior to the latter, it's just different. It's much better at most writing tasks, and its "Deep Research" mode is by a wide margin the best in the business. (OpenAI's has really gone downhill for some reason.)
I have been using a quorum composed of step-3.5-flash, Kimi k2.5 and glm-5 and I have found it outperforms opus-4.5 at a fraction of the cost
That's pretty cutting edge to me.
EDIT: It's not a swarm — it's closer to a voting system. All three models get the same prompt simultaneously via parallel API calls (OpenAI-compatible endpoints), and the system uses weighted consensus to pick a winner. Each model has a weight (e.g. step-3.5-flash=4, kimi-k2.5=3, glm-5=2) based on empirically observed reliability.
The flow looks like:
1. User query comes in
2. All 3 models (+ optionally a local model like qwen3-abliterated:8b) get called in parallel
3. Responses come back in ~2-5s typically
4. The system filters out refusals and empty responses
5. Weighted voting picks the winner — if models agree on tool use (e.g. "fetch this URL"), that action executes
6. For text responses, it can also synthesize across multiple candidates
The key insight is that cheap models in consensus are more reliable than a single expensive model. Any one of these models alone hallucinates or refuses more than the quorum does collectively. The refusal filtering is especially useful — if one model over-refuses, the others compensate.
Tooling: it's a single Python agent (~5200 lines) with protocol-based tool dispatch — 110+ operations covering filesystem, git, web fetching, code analysis, media processing, a RAG knowledge base, etc. The quorum sits in front of the LLM decision layer, so the agent autonomously picks tools and chains actions. Purpose is general — coding, research, data analysis, whatever. I won't include it for length but I just kicked off a prompt to get some info on the recent Trump tariff Supreme Court decision: it fetched stock data from
Benzinga/Google Finance, then researched the SCOTUS tariff ruling across AP, CNN, Politico, The Hill, and CNBC, all orchestrated by the quorum picking which URLs to fetch and synthesizing the results, continuing until something like 45 URLs were fully processed. Output was longer than a typical single chatbot response, because you get all the non-determinism from what the models actually ended up doing in the long-running execution, and then it needs to get consensus, which means all of the responses get at least one or N additional passes across the other models to get to that consensus.
Cost-wise, these three models are all either free-tier or pennies per million tokens. The entire session above (dozens of quorum rounds, multiple web fetches) cost less than a single Opus prompt.
When you say quorum what do you mean? Is it like an agent swarm or using all of them in your workflow and independently they perform better than opus? Curious how you use (tooling and purpose - coding?)
I have not heard of step-3.5-flash before. But as the other commenter asked, I would love to hear about your quorum technique. What type of projects are you building with the quorum?
For certain use-cases, sure it doesn't matter. but that doesn't make those models cutting edge. Some use-cases are adversarial, and 1% lower efficacy matters a lot.
I can't shake the feeling that the RAM shortage was intentionally created to serve as a sort of artificial moat by slowing or outright preventing the adoption of open weight models. Altman is playing with hundreds of billions of other people's dollars, trying to protect (in his mind) a multi-trillion dollar company. If he could spend a few billion to shut down access to the hardware people need to run competitor's products, why wouldn't he?
From what I understand the RAM producers see the writing on the wall. They’re not going to invest in massively more capacity only to have it sit completely idle in 10 years.
RAM shortage is probably a bubble indicator itself. That industry doesn’t believe enough in the long term demand to build out more capacity.
It's very difficult to "intentionally create" a real shortage. You can hoard as much as you want, but people will expect you to dump it all right back onto the market unless you really have a higher-value use for the stuff you hoarded (And then you didn't intentionally create anything, you just bought something you needed!).
Plus producers will now feel free to expand production and dump even more onto the market. This is great if you needed that amount of supply, but it's terrible if you were just trying to deprive others.
Anthropic, at least, has gone to lengths to avoid hardware lock-in or being open to extortion of the nvidia variety. Anthropic is running their models on nvidia GPUs, but also Amazon Trainium and Google's TPUs. Massive scale-outs on all three, so clearly they've abstracted their operations enough that they aren't wed to CUDA or anything nvidia-specific.
Similarly, OpenAI has made some massive investments in AMD hardware, and have also ensured that they aren't tied to nvidia.
I think it's nvidia that has less of a moat than many imagine they do, given that they're a $4.5T company. While small software shops might define their entire solution via CUDA, to the large firms this is just one possible abstraction engine. So if an upstart just copy pastes a massive number of relatively simple tensor cores and earns their business, they can embrace it.
Openai is just playing catchup at this point, they completely lost thier way in my view.
Anthropic on the other hand is very capable and given the success of claude code and cowork, I think they will maintain their lead across knowledge work for a long time just by having the best data to keep improving their models and everything around. It's also the hottest tech conpany rn, like Google were back in the day.
If I need to bet on two companies that will win the AI race in the west, it's Anthropic and Google. Google on the consumer side mostly and Anthropic in enterprise. OpenAI will probably IPO soon to shift the risk to the public.
If anthropic continues getting their foot in the enterprise door then maybe they can tap into enterprise cloud spending. If Athropic can come up with services and things (db, dns, networking, webservers, etc) that claudecode will then prefer then maybe they become a cloud provider. To me, and I am no business expert btw, that could be a path to sustainable financials.
Edit: one thing I didn’t think about is Anthropic more or less runs at the pleasure of AWS. Of Amazon sees Anthropic as a threat to AWS then it could be lights out.
Yes, they depend on AWS for compute and Amazon also owns a big chunk of Anthropic (it used to be close to 30%, probably less now with the recent raises). I think it's a good partnership since for the most part they focus on different things and I don't see Anthropic going after AWS - they are an AI company first and foremost. Amazon has their own AI stuff for enterprise but no one uses it so I don't think they take it seriously. They know they cannot compete here.
I think that OpenAI and Microsoft is a more challenging partnership with much more overlap.
Anthropic at least seems to be doing well with enterprises. OpenAI doesnt have that level of trust with enterprise use cases, and commodization is a bigger issue with consumers, when they can just switch to another tool easily
Yeah, Anthropic is inarguably in a better position, but I don’t see how they justify their fundraising unless they find some entrenched position that is difficult for competitors to replicate.
Enterprise switching costs aren’t 0, but they’re much less than most other categories, especially as models mature and become more fungible.
The best moat I can think of is a patentable technique that facilitates a huge leap that Anthropic can defend, but even then, Chinese companies could easily ignore those patents. And I don’t even know if AI companies could stick to those guns as their training is essentially theft of huge portions of copyrighted material.
> Yeah, Anthropic is inarguably in a better position
OpenAI's biggest problem as well as their biggest advantage is that they have way, way more users than anyone else. Unfortunately for them the users they have dont pay for shit and they dont serve ads so more users = more money wasted right now. But its unlikely that will always be the case. If OpenAI turns on ads, most users will not leave because retail users hate change, and suddenly their massive user base is a boon instead of a problem.
To take the other side of this, as computers got commodified there still was a massive benefit to using cloud computing. Could it be possible that that happens with LLMs as well as hardware becomes more and more specialized? I personally have no idea but love that there’s a bunch of competition and totally agree with your point regulation and export controls are just ways to make it harder for new orgs to compete.
I do think the models themselves will get commoditized, but I've come around to the opinion that there's still plenty of moat to be had.
On the user side, memory and context, especially as continual learning is developed, is pretty valuable. I use Claude Code to help run a lot of parts of my business, and it has so much context about what I do and the different products I sell that it would be annoying to switch at this point. I just used it to help me close my books for the year, and the fact that it was looking at my QuickBooks transactions with an understanding of my business definitely saved me a lot of time explaining.
On the enterprise side, I think businesses are going to be hesitant to swap models in and out, especially when they're used for core product functionality. It's annoying to change deterministic software, and switching probabilistic models seems much more fraught.
Think LLM by itself is basically a commodity at this point. Not quite interchangeable but it’s more of artistic differences rather than technological. I used to think it was data and that would give companies like Google a leg up.
I've tried deepseek a few months ago and asket about the Tiananmen square protests and massacre.
At first the answer was "I can't say anything that might hurt people" but with a little persuasion it went further.
The answer wasn't the current official answer but way more nuanced that Wikipedia's article.
More in the vein of "we don't know for sure", "different versions", "external propaganda", "some officials have lied and been arrested since"
In the end, when I asked whether I should trust the government or ask for multiple source, it strongly suggested to use multiple sources to form an opinion.
Anthropic I feel will be alright. They have their niche, it's good and people actually do pay for their services. Why do people still use salesforce when there's other free CRM's. They also haven't from what I can tell scaled for some imaginary future growth.
OpenAI I'm sorry to say are all over the place. They're good at what they do, but they try to do too much and need near ponzi style growth to sustain their business model.
They'll ban Chinese models, or do something like calling them security risks without proof.
Enterprise customers will gladly pay 10x to 20x for American models. Of course this means American tech companies will start to fall behind, combined with our recent Xenophobia.
Almost all the top AI researchers are either Chinese nationals or recent immigrants. With the way we've been treating immigrants lately ( plenty of people with status have been detained, often for weeks), I can't imagine the world's best talent continuing to come here.
I consider almost all news to be entertainment unless I need its perspective to make a decision (which is almost never). It is a lot safer to remain uninformed on a subject as it settles than to constantly attempt to be informed.
Information bias is unfortunately one of the sicknesses of our age, and it is one of the cultural ills that flows from tech outward. Information is only pertinent in its capacity to inform action, otherwise it is noise. To adapt a Beck-ism: You aren't gonna need it.
To me, tech entrepreneurship looks more like some form of "lemon socialism." It feels more centrally planned than ever, and a company's success has much more to do with your relationships with capital than anything else. It's why we're seeing so much money invested into a bunch of similar takes on AI. Founders with a real vision of the future aren't really accepted into VC that has almost wholly accepted the FOMO strategy of investment.
I used to hold a lot of respect for Paul Graham and his essays, but I've realized his stances on things are pretty elementary, and largely come back to his ego or wealth management. People like Graham and Tan don't seem to really care about human flourishing, and they certainly don't seem to have any coherent vision of the future. Graham, like Andreessen, was technically good enough during a veritable tech gold rush, and Graham's lieutenants like Tan and Altman were lucky more than anything--just in the right place at the right time versus having started anything of value.
I am *absolutely* cynical and jaded when it comes to tech nowadays, so no need to call me out there. These people remind me of the high modernists, that tech will solve all problems, and we don't have to care too much as to how we solve those problems. Just handwave, and AI will solve all problems. But I think how we solve problems matters, and the entrepreneurship meritocracy that Tan and Graham allude to does not exist, and it never did.
I just find it abhorrent that while 15% of American households are food insecure, a company like Anthropic spent millions on a superbowl ad just lamenting OpenAI's ad strategy. Or that the Trump administration dropped a FTC case against Pepsi and Walmart for colluding to price out grocery competition. Or that Facebook and Google have been shown to have pushed for apps to addict people to their slop content. Or that tech capex this year alone rivals the Louisiana Purchase or the amount America spent on building out the railroads[1].
We're not solving the right problems because capital is entirely disconnected from the every day reality of Americans in this country. But by all means, let's aim to replace 50% of white collar workers with AI and handwave that prices will come down.
It's pretty simple: you don't get to that kind of wealth without having a few screws loose in the ethics department. There are some exceptions but they are just there to confirm the rule.
I miss the era of Internet forums. They didn’t need to be federated, just simple deployments of MyBB, vBulletin, PHP, Xenforo and so on.
I made a lot of friends on those communities growing up, and it inspired me to go into software because I saw how it brought people together.
And I still sorely miss the WhatCD forums. While I didn’t make any friends there, it shaped my early experiences with music which still reverberates through me today.
Even with the reinvigoration of new ideas from LLMs, tech feels like it has been languishing for well over a decade at this point. The playbook is to disrupt traditional industry at a loss, then enshittify when competitors are gone. A lot of tech plays really feel like some form of: bring the yellow pages into the digital realm and overcharge for facilitating that access. Finding a firm that even uses AI outside of a chatbot UX is rare.
>And I still sorely miss the WhatCD forums. While I didn’t make any friends there, it shaped my early experiences with music which still reverberates through me today.
Could not relate to this more. Spent my formative years in those forums and they genuinely helped mold many of the tastes and interests that have stuck with me into adulthood. Not to over-romanticize, at the end of the day it was just a forum on a music tracker - but the sense of community and sheer diversity of thread topics made it such an interesting place to peruse.
Discord certainly has its applications. But since it became the defacato community tool, I find it essentially useless. Discussions are ephemeral (from a UX standpoint at least), and much more constrained. Its difficult to lurk and only chime in now and then unless you're regularly online.
A lot of tech's current trends have a lot to do with its inability to see beyond the first order, like:
- How layoff culture backfires: companies that lean into this culture tend to underperform compared to those that do not.
- The deleterious effects of overwork on employees: work carries diminishing returns after a certain number of hours per week, and eventually the mistake rate from exhaustion outweighs the productivity from more hours. Not to mention, this causes burnout which leads to valuable people leaving.
- How AI removes flow: this is something I've seen in myself, but using agents means I do not achieve the cognitive engagement necessary for flow, which is one of the most pleasant states I can get into while working (and it often makes work feel worthwhile).
I'd also note: if you get hired at Rilla for their senior engineer position, and you're able to command the top of their stated band (300K), that is defacto ~165K for 40 hours worked / week.
Many people fought very hard for a long time to secure a 40 hour work week, and it's pretty silly how easily a lot of tech people will throw it away. Time is your most important asset, don't waste your life behind a screen not seeing your family or friends.
Layoffs should happen exactly once for a very long time.
Round-after-round of layoffs craters morale because all workers will think about is how miserable, uncertain, chaotic, and stressful is the business and how clueless and incompetent is their management. It will completely hollow-out a business and most all of the really good talent will leave. The dinosaurs that engage in it are because their leadership is insulated from reality and don't know what they're doing. Zuck is a poster-child for clueless amateurs lacking understanding of business, reality, or real empathy; I'm surprised he's not in the Epstein files, oh, wait.
I started exploring Christianity from an archetypal or psychological lens last year, and have found it really rewarding. I've put in thousands of hours of westernized Buddhist oriented meditation (I think "Pragmatic Dharma" is the term), and ultimately found it and the communities attached to it cultures of avoidance that loses something in its detachment of meditation technology from its larger context. I also grew up vaguely Presbyterian and hated it, so this was a great moment for me to reclaim my heritage on my own terms.
I started with various books of the Nag Hammadi collection, reading the excellent Meyer translations, and started noticing some metaphors that felt like "hidden signposts" in the text (and had some relevance to some ideas in Buddhism). Gospel of Thomas and especially Gospel of Philip felt like they map quite well to non-dual ideas in Buddhism.
I decided after some explorations of gnostic text to jump back into the gospels, wondering if I noticed the same kinds of hidden signposts there. I started this exploration during a trip to London with my wife, where I went and hunted down a copy of Bruce Rogers's amazing Oxford Lectern Bible at the Church of England reading room. What a beautiful bible -- it's so forward thinking that it feels like it was typeset last year, but while it is a beautiful piece, the King James translation of the bible is pretty incomprehensible. This little journey led me to the Sarah Ruden translations of the gospels, and as soon as I read them I felt the same kind of resonance.
This all eventually led me to Cynthia Bourgeault's amazing "The Heart of Centering Prayer," which explores the non-dual kind of ideas in esoteric Christianity and lays out the practice of centering prayer as a basis of Christian spirituality. And I would remiss if I didn't mention Jacob Needleman: Esoteric Christianity was good, but his "Money and the Meaning of Life," really helped me put my own relationship with money in perspective.
This is all a long winded way of saying: Christianity has a rich set of amazing spiritual resources, but they need to be consumed in a sort of non-literal way, where you're meeting the authors in the same mind as they were when they wrote the text. I'd also note that this kind of reading is not scholarly, the point isn't to find the right answer but to impute a larger meeting by meeting the author with your own struggles.
We live in a time that is committed to a materialist reductionist mindset, but I believe that humans are naturally mystical beings, and that we leave a lot of real meaning on the table when we reduce the world down into solely material order.
Rob Burbea explored these ideas (largely inspired by James Hillman's concept of "soulmaking") in his soulmaking dharma (https://hermesamara.org/), the idea being an extension of emptiness: if all is fabrication, why wouldn't we make meaning that is beautiful?
I'm sure I'm coming off quite a bit rambly, but it's very exciting to see such a resource on the HN front page. If you read my comment and feel any similar excitement, please check my profile and feel free to email me!
> It struck me recently how few of the most successful people I know are mean. There are exceptions, but remarkably few.
*mean to Paul Graham. I’ve worked with a lot of mean people in important positions in my career, and they all have a kind, charismatic side when they need to. Those same people are awful to subordinates or people that can’t do something for them. Paul is high value to many people, so they treat him well.
People like Graham who aren’t often in positions where they’re taken advantage of or humbled like to pretend they and their peers are magnanimous and kind but often enough they’re just not exposed to the forces that make people ugly. All other things being equal: it’s often lack of agency over work and over their own lives—this shows up in work where people are given lots of responsibility but without the freedom to fulfill it.
I often find it concerning how elementary a lot of well off tech peoples’ theory of mind is. People are not acausal personalities, they are functions of their internals and their environments. A person mean at a stressful job might be delightful at a party after.
> I often find it concerning how elementary a lot of well off tech peoples’ theory of mind is.
This is a great way of putting it. I always get surprised when I discover that others aren't constantly modeling a theory-of-mind of other people as a part of interacting with people. That's leaving aside whether those models are accurate, people have varying degrees of skill at it, but it shocks me when I discover that some people don't do it at all, badly or otherwise.
Isn't that why a lot of us went into tech in the first place, because other people's minds are weird and confusing and they keep doing inexplicable stuff like saying things that actually mean something totally different, or not thinking about how something works when trying to use it?
Yeah, and I think that avoidance is the cardinal sin of the modern world. You can’t avoid the dissonance between how a person acts and what they say or want to be. This is especially true within yourself: you can’t ignore what James Hillman might call the “multitudes of the soul” because the pot eventually boils over, and you wonder why you cannot measure up to this ideal version of yourself.
When I’ve been mean at work, and I’ve had many moments, I’m often unsure where it comes from: why did I react that way? It’s not what I wanted. And I’ve been fortunate to be humbled so much in my life and career, now perhaps more than ever, that I’m forced to juggle with the parts of myself that I’ve left in the shadows to my own detriment. I think this is the central paradox of a lot of folks like Graham: when you have capital, even your losses can be manufactured to be wins, and so you become estranged from dissolution and decay and humbling necessary to be born again into someone with fresh perspective.
I think Jung said something like: it is better to be yourself and accept the consequences of being you versus forcing yourself into a mold of what you or others think you ought to be. And often this means engaging deeply with what you’re avoiding and coming to terms with why.
Read more, like biographies and good fiction. Meditate and see into yourself. Understanding minds is a skill like any other, and gets better with practice.
I think I'm pretty decent at understanding people these days, as in predicting how they'll react to things, figuring out what they want, that sort of thing.
But I don't understand it on a fundamental level, the way I understand something like math. I have a decent grasp of the game, but I don't understand why the rules of the game are what they are. I get the impression this is how a lot of people feel about math, or computers. They know you can compute a 20% tip by sliding the decimal point over and doubling, but it's just a procedure they follow, they don't understand it.
I've read and thought and interacted plenty. This has built up the skill just fine, but the fundamental understanding is something I don't think will ever come.
And I’m telling you that’s not a rare sentiment. I do a lot of zen, and one famous teacher talks about feeling disconnected from people for a long time. I’m not saying you have to do anything, it’s your life, but just don’t assume it’s something unique or permanent.
I don't think it's unique, but it doesn't seem to be the norm. Most people seem to just accept how people behave without thinking about it too much. When they respond to "how are you?" used as a greeting, they just do it, they're not calculating the right answer.
As far as permanence goes, I'm probably more than halfway through my lifespan at this point and there's no sign of any improvement there. Like I said, as a skill I do just fine. But that deeper understanding isn't there and I doubt it will be.
reply