I read the article and while it doesn’t say this nor imply it, this is my takeaway, though correct me if I’m wrong:
Model innovation is effectively converging and slowing down considerably. The big companies in this space doing the research are not making leap over leap with each release, and the downstream open source projects are coming closer to the same quality or in fact can produce the same quality (e.g DeepSeek or LLAMA) hence why it’s becoming a commodity.
Around the edges model innovation - particularly speed ups in returning accurate results - will help companies differentiate but fundamentally, all this tech is shovels in search of miners, IE you aren’t really going to make money hand over fist by simply being an LLM model provider.
In another words, this latest innovation has hit commodity level within a few short years of going mainstream and the winners are going to be the companies that make products on top of this tech, and as the tech continues to become a commodity, the value proposition for pure research companies drops considerably relative to application builders.
To me this leaves a central question: when does it hit a relative equilibrium where the technology and the applications on top of it have largely hit their maximal ability to add utility to applicable situations? That’s the next question, and I think the far more important one
One other thing, at the end of the article they wrote:
>Ultimately, businesses won’t rearrange themselves around AI — the AI systems will have to meet businesses where they are.
This is demonstrably untrue. CEOs are chomping at the bit to reorganize their business around AI, as in, AI doing things humans used to do and getting the same effective results or better, thereby they can reduce staff across the board while supposedly maintaining the same output or better.
Look at the leaked Shopify memo for an example or the trend of “I can vibe code with an LLM making software engineers obsolete” that has taken off as of late, if LinkedIn is to be believed
I agree with this but I think it's still an open question if anyone can build a successful product on top of the tech. There will likely be some but it feels eerily similar to the dot com boom (and then bust) when the vast majority of new products built on top of this (internet) technology didn't produce and didn't survive. Most AI products so far are fun toys or interesting proofs, and mediocre when evaluated against other options. They'll need to be applied to a much smaller set of problems (that doesn't support the current level of investment) or find some new miracle set of problems where they change the rules.
Businesses are definitely rearranging themselves structurally around AI - at least to try and get the AI valuation multiplier and Executives have levels of FOMO I've never seen before. I report to a CTO and the combination of 100,000 foot hype combined with down in the weeds focus on the "protocol de jour" (with nothing in between that looks like a strategy) is astounding. I just find it exhausting.
The dot com boom is an apt analogy: the internet took off, we understood it had potential, but the innovation didn't all come in the first wave. It took time for the internet to bake, and then we saw another boom with the advent of mobile phones, higher bandwidth, and more compute per user.
It is still simply too early to tell exactly what the new steady state is, but I can tell you that where we're at _today_ is already a massive paradigm shift from what my day-to-day looked like 3 years ago, at least as a SWE.
There will be lots of things thrown at the wall and the things that stick will have a big impact.
other than constantly feeling gaslit about the quality of these tools, I can tell you where we are _today_ is basically the same in my day to day as it was three years ago.
oh except, sometimes someone tells me I could use the bot to generate a thing, and it doesn't work, and I waste some time, and then do it manually.
I would agree with this and also say that it's been clear this is true for at least a year. Innovations like Deepseek may not have been around a year ago, but it was very clear that "AI" is actually information retrieval and transformation, that the chat UI had limited applicability (nobody wants to "chat with their documents"), and that those who could shape the tech to match use cases would be the ones capturing the value. Just as SaaS uses databases, but creates and captures value by shaping the database to the particular use case.
> This is demonstrably untrue. CEOs are chomping at the bit to reorganize their business around AI, as in, AI doing things humans used to do and getting the same effective results or better, thereby they can reduce staff across the board while supposedly maintaining the same output or better.
Nah. Maybe tech CEOs. Companies are blocking AI carte blanche at the direction of their security teams and/or only allowing an instanced version of MS Copilot, if anything. Other than write emails, it doesn't do much for the average office worker and we all know it.
The value is going to be the apps that build on AI, as you said.
It certainly isn't maybe, look at the recent Shopify memo leak, and the way that lots of companies are talking about AI.
Any company with any sort of large customer service presence are looking at AI to start replacing alot of customer service roles, for example. There is huge demand for this across many industries, not only tech. Whether it actually delivers is the question, but the demand is there.
Claiming these AIs "don't do much" overlooks the very real productivity gains already happening – automating tedious tasks and accelerating content creation. This isn't trivial and will lead to the deeper integrations and streamlined (read: downsized) workforces. The reorganization isn't a distant fantasy; it's already here.
I don't disagree, but your average excel jockey isn't going to build out these automation workflows and likely neither will IT. I'm not saying AI isn't useful. I'm saying the average person doesn't know what to do with it.
I know many IP-heavy and health-centric companies are blocking AI use severely. For example, pharma depends on huge amounts of secrecy and does not want any data leaked to OpenAI, and often has barely-competent IT and security staff that don't know what "threat model" means. Those who deal with controlled health data also block with a heavy hand.
I imagine it'll take time for any of this tech to permeate and the lower barrier of entry will see adoption faster - as is usually the case with new tech - but it'll make its way eventually. On premise AI will be a thing
>Model innovation is effectively converging and slowing down considerably. The big companies in this space doing the research are not making leap over leap with each release, and the downstream open source projects are coming closer to the same quality or in fact can produce the same quality (e.g DeepSeek or LLAMA) hence why it’s becoming a commodity.
You're just showing how disconnected from the progress of the field you are. o3/o4 aren't even in the same universe as anything from open source. Deepseek R1, LLama 4? Are you joking?
Depends on how they're applied. I've had success using LLama and while we check to see if OpenAI or Google's Gemini would give us any noticeable improvement, it really doesn't for our use case.
While certainly newer models are more capable on the whole, it doesn't mean I need all that capability to accomplish the business goal.
This is kind of a useless statement. If your use case is so easy that "old" models work for it then obviously you won't care about or be following the latest developments but its just not accurate to say that Deepseek R1 is equivalent to o3 or Gemini 2.5.
Producing quality results is not the same thing as saying Deepseek R1 is the equivalent of o3 or Gemini 2.5
Again, its not about capabilities alone, (on this, many models lag behind, I already said as much). I follow these developments quite closely, and I purposely said results as to not say they're equivalent in capability. They aren't.
However, if a business is getting acceptable results from older models or cheaper models than capability doesn't matter, the results do. Gemini 2.5 can be best of breed but why switch if it shows no meaningful improvement in results for the business?
If I need more capability or results are substandard, I can always upgrade, but its like saying there's no room for cheaper processors and you'd be out of your mind not to be using only the latest at all times no matter the results.
That's not what GP was saying though. To stay with that analogy, the assertion was that "all processors are kinda the same, there's no real qualitative difference", which sounds pretty strange. It's somewhat accurate if your use-case is covered by the average processor and the faster one doesn't benefit you. They're not equal, but all of them surpass your needs.
> If I need more capability or results are substandard, I can always upgrade
You wouldn't be able to upgrade (and see improved results) if the model you use today was close to equal to the top of the line.
That wasn’t the assertion. The results - not the models themselves, not strictly speaking their over all capabilities - if they have no meaningful improvement by moving to a newer model, why then would I want to switch if I’m not getting any tangible improvement in results?
> The big companies in this space doing the research are not making leap over leap with each release, and the downstream open source projects are coming closer to the same quality or in fact can produce the same quality (e.g DeepSeek or LLAMA) hence why it’s becoming a commodity.
This was the assertion. "Open source is close/equal in quality", not "open source is enough for plenty of use-cases, not everyone needs the top of the line".
This take doesn't really highlight the fact that the most competitive foundational model companies are innovative application builders. Anthropic and OpenAI are vying for consumers to use their models by building these sort of super applications (ChatGPT, Claude) that can run code, plot graphs, spin up text editors, create geographic maps, etc. These are well staffed and strategically important areas of their businesses. There's competition to attract consumers to these apps and they will grow more capable and commoditize more compliments along the way. Who needs Jasper when you can edit copy in ChatGPT, or an AI python notebook app, or, now, Cursor?
One thing this article gets wrong is how OpenAI isn’t an application layer company, they built the original ChatGPT “app” with model innovation to power it. They’re good at UX and actually have the strongest shot at owning the most common apps (like codegen).
I don't disagree. But that's a pretty good reason to make sure you're making something other than the obvious common apps if you want a big chunk of acquisition money.
I think the UX of chatgpt works because it's familiar, not because it's good. Lowers friction for new users but doesn't scale well for more complex workflows. if you're building anything beyond Q&A or simple tasks, you run into limitations fast. There's still plenty of space for apps that treat the model as a backend and build real interaction layers on top — especially for use cases that aren’t served by a chat metaphor
I wouldn't call it familiar, it's a weird quasi-chat. They didn't even do the chat metaphor right, you can't type more as the AI is thinking. Nor can you really interrupt it when it's off over explaining something for the 20th time without just stopping it.
It's missing obvious settings, has a weird UX where every now and mysterious popups will appear like 'memory updated', or now it spews random text while it's "thinking", it'll every now and then ask you to choose between two answers but I'm working so no thanks, I'm just going to pick one at random so I can continue working.
People had copy pasta templates they dropped into every chat with no way of savings Ng thatz they they added a sort of ability to save that but it worked in a inscrutable and confusing manner, but then they released new models that didn't support that and so you're back to copy pasta, and blurgh.
It's a success despite the UI because they had a model streets ahead of everyone else.
There's a lot of opportunity to apply leading edge AI models to specific business applications, but success here is determined more by experience with those business domains than with AI generally.
An AI startup could still be a useful "resume" to get acqui/hired by one of the big players.
The LLM space was never going to be kind to those without deep pockets. And right now there's no point getting in it because it's hit a wall. So yeah, startups should steer clear of trying to make frontier LLM models.
On the other hand, there's a ton of hype and money looking for the next AI related thing. If someone creates the next transformer, or a different AI paradigm that pushes things forward, they'll get billions.
Every startup should generate a shitty Ai wrapper product, write one or two lines of code, generate hype and have 2025's version of Softbank give you a billion $'s.
Frankly it's bordering on irresponsible to not be targeting acquisition in this climate.
This focuses on case where the acquirer seeks to capture the value of the startup's business. But this is not always the case, sometimes the startup is dubious, but a cash-rich enterprise can purchase startups simply to eliminate potential avenues of competition. They may not be interested in adding a better product to their portfolio, only in quashing any nascent attempts at building the better product so they can keep selling their own mediocre one.
Also, "model innovation" strikes me as missing the point these days. The models are really good already. The majority of applications is capturing only a tiny bit of their value. Improving the models is not that important because model capability is not the bottleneck anymore, what matters is how the model is used. We just don't have enough tools to use them fully, and what we have is not even close to penetrating the market, while all the dominant tools are garbage. Of course application innovation is the place to be!
I read the article and while it doesn’t say this nor imply it, this is my takeaway, though correct me if I’m wrong:
Model innovation is effectively converging and slowing down considerably. The big companies in this space doing the research are not making leap over leap with each release, and the downstream open source projects are coming closer to the same quality or in fact can produce the same quality (e.g DeepSeek or LLAMA) hence why it’s becoming a commodity.
Around the edges model innovation - particularly speed ups in returning accurate results - will help companies differentiate but fundamentally, all this tech is shovels in search of miners, IE you aren’t really going to make money hand over fist by simply being an LLM model provider.
In another words, this latest innovation has hit commodity level within a few short years of going mainstream and the winners are going to be the companies that make products on top of this tech, and as the tech continues to become a commodity, the value proposition for pure research companies drops considerably relative to application builders.
To me this leaves a central question: when does it hit a relative equilibrium where the technology and the applications on top of it have largely hit their maximal ability to add utility to applicable situations? That’s the next question, and I think the far more important one
One other thing, at the end of the article they wrote:
>Ultimately, businesses won’t rearrange themselves around AI — the AI systems will have to meet businesses where they are.
This is demonstrably untrue. CEOs are chomping at the bit to reorganize their business around AI, as in, AI doing things humans used to do and getting the same effective results or better, thereby they can reduce staff across the board while supposedly maintaining the same output or better.
Look at the leaked Shopify memo for an example or the trend of “I can vibe code with an LLM making software engineers obsolete” that has taken off as of late, if LinkedIn is to be believed
I agree with this but I think it's still an open question if anyone can build a successful product on top of the tech. There will likely be some but it feels eerily similar to the dot com boom (and then bust) when the vast majority of new products built on top of this (internet) technology didn't produce and didn't survive. Most AI products so far are fun toys or interesting proofs, and mediocre when evaluated against other options. They'll need to be applied to a much smaller set of problems (that doesn't support the current level of investment) or find some new miracle set of problems where they change the rules.
Businesses are definitely rearranging themselves structurally around AI - at least to try and get the AI valuation multiplier and Executives have levels of FOMO I've never seen before. I report to a CTO and the combination of 100,000 foot hype combined with down in the weeds focus on the "protocol de jour" (with nothing in between that looks like a strategy) is astounding. I just find it exhausting.
The dot com boom is an apt analogy: the internet took off, we understood it had potential, but the innovation didn't all come in the first wave. It took time for the internet to bake, and then we saw another boom with the advent of mobile phones, higher bandwidth, and more compute per user.
It is still simply too early to tell exactly what the new steady state is, but I can tell you that where we're at _today_ is already a massive paradigm shift from what my day-to-day looked like 3 years ago, at least as a SWE.
There will be lots of things thrown at the wall and the things that stick will have a big impact.
other than constantly feeling gaslit about the quality of these tools, I can tell you where we are _today_ is basically the same in my day to day as it was three years ago.
oh except, sometimes someone tells me I could use the bot to generate a thing, and it doesn't work, and I waste some time, and then do it manually.
I would agree with this and also say that it's been clear this is true for at least a year. Innovations like Deepseek may not have been around a year ago, but it was very clear that "AI" is actually information retrieval and transformation, that the chat UI had limited applicability (nobody wants to "chat with their documents"), and that those who could shape the tech to match use cases would be the ones capturing the value. Just as SaaS uses databases, but creates and captures value by shaping the database to the particular use case.
so when do we get to the point where AI apps are just CRUD apps essentially? RAG kinda feels like a better version of those to me
now!
> This is demonstrably untrue. CEOs are chomping at the bit to reorganize their business around AI, as in, AI doing things humans used to do and getting the same effective results or better, thereby they can reduce staff across the board while supposedly maintaining the same output or better.
Nah. Maybe tech CEOs. Companies are blocking AI carte blanche at the direction of their security teams and/or only allowing an instanced version of MS Copilot, if anything. Other than write emails, it doesn't do much for the average office worker and we all know it.
The value is going to be the apps that build on AI, as you said.
It certainly isn't maybe, look at the recent Shopify memo leak, and the way that lots of companies are talking about AI.
Any company with any sort of large customer service presence are looking at AI to start replacing alot of customer service roles, for example. There is huge demand for this across many industries, not only tech. Whether it actually delivers is the question, but the demand is there.
Claiming these AIs "don't do much" overlooks the very real productivity gains already happening – automating tedious tasks and accelerating content creation. This isn't trivial and will lead to the deeper integrations and streamlined (read: downsized) workforces. The reorganization isn't a distant fantasy; it's already here.
I don't disagree, but your average excel jockey isn't going to build out these automation workflows and likely neither will IT. I'm not saying AI isn't useful. I'm saying the average person doesn't know what to do with it.
> Companies are blocking AI carte blanche at the direction of their security teams
What companies?
I know many IP-heavy and health-centric companies are blocking AI use severely. For example, pharma depends on huge amounts of secrecy and does not want any data leaked to OpenAI, and often has barely-competent IT and security staff that don't know what "threat model" means. Those who deal with controlled health data also block with a heavy hand.
I imagine it'll take time for any of this tech to permeate and the lower barrier of entry will see adoption faster - as is usually the case with new tech - but it'll make its way eventually. On premise AI will be a thing
Once upon a time, they blocked docker too. Things change.
One of the possible alternative routes is this:
Model providers and model labs stop opensourcing/listing their innovations/papers and start patenting instead.
>Model innovation is effectively converging and slowing down considerably. The big companies in this space doing the research are not making leap over leap with each release, and the downstream open source projects are coming closer to the same quality or in fact can produce the same quality (e.g DeepSeek or LLAMA) hence why it’s becoming a commodity.
You're just showing how disconnected from the progress of the field you are. o3/o4 aren't even in the same universe as anything from open source. Deepseek R1, LLama 4? Are you joking?
Depends on how they're applied. I've had success using LLama and while we check to see if OpenAI or Google's Gemini would give us any noticeable improvement, it really doesn't for our use case.
While certainly newer models are more capable on the whole, it doesn't mean I need all that capability to accomplish the business goal.
This is kind of a useless statement. If your use case is so easy that "old" models work for it then obviously you won't care about or be following the latest developments but its just not accurate to say that Deepseek R1 is equivalent to o3 or Gemini 2.5.
Producing quality results is not the same thing as saying Deepseek R1 is the equivalent of o3 or Gemini 2.5
Again, its not about capabilities alone, (on this, many models lag behind, I already said as much). I follow these developments quite closely, and I purposely said results as to not say they're equivalent in capability. They aren't.
However, if a business is getting acceptable results from older models or cheaper models than capability doesn't matter, the results do. Gemini 2.5 can be best of breed but why switch if it shows no meaningful improvement in results for the business?
If I need more capability or results are substandard, I can always upgrade, but its like saying there's no room for cheaper processors and you'd be out of your mind not to be using only the latest at all times no matter the results.
That's not what GP was saying though. To stay with that analogy, the assertion was that "all processors are kinda the same, there's no real qualitative difference", which sounds pretty strange. It's somewhat accurate if your use-case is covered by the average processor and the faster one doesn't benefit you. They're not equal, but all of them surpass your needs.
> If I need more capability or results are substandard, I can always upgrade
You wouldn't be able to upgrade (and see improved results) if the model you use today was close to equal to the top of the line.
That wasn’t the assertion. The results - not the models themselves, not strictly speaking their over all capabilities - if they have no meaningful improvement by moving to a newer model, why then would I want to switch if I’m not getting any tangible improvement in results?
> The big companies in this space doing the research are not making leap over leap with each release, and the downstream open source projects are coming closer to the same quality or in fact can produce the same quality (e.g DeepSeek or LLAMA) hence why it’s becoming a commodity.
This was the assertion. "Open source is close/equal in quality", not "open source is enough for plenty of use-cases, not everyone needs the top of the line".
[flagged]
This take doesn't really highlight the fact that the most competitive foundational model companies are innovative application builders. Anthropic and OpenAI are vying for consumers to use their models by building these sort of super applications (ChatGPT, Claude) that can run code, plot graphs, spin up text editors, create geographic maps, etc. These are well staffed and strategically important areas of their businesses. There's competition to attract consumers to these apps and they will grow more capable and commoditize more compliments along the way. Who needs Jasper when you can edit copy in ChatGPT, or an AI python notebook app, or, now, Cursor?
One thing this article gets wrong is how OpenAI isn’t an application layer company, they built the original ChatGPT “app” with model innovation to power it. They’re good at UX and actually have the strongest shot at owning the most common apps (like codegen).
I don't disagree. But that's a pretty good reason to make sure you're making something other than the obvious common apps if you want a big chunk of acquisition money.
I personally find their UX frustrating, basically a junior developer's attempt at doing a front end. What do you think is so good about it?
It's also janky as hell and crashes regularly.
I think the UX of chatgpt works because it's familiar, not because it's good. Lowers friction for new users but doesn't scale well for more complex workflows. if you're building anything beyond Q&A or simple tasks, you run into limitations fast. There's still plenty of space for apps that treat the model as a backend and build real interaction layers on top — especially for use cases that aren’t served by a chat metaphor
I wouldn't call it familiar, it's a weird quasi-chat. They didn't even do the chat metaphor right, you can't type more as the AI is thinking. Nor can you really interrupt it when it's off over explaining something for the 20th time without just stopping it.
It's missing obvious settings, has a weird UX where every now and mysterious popups will appear like 'memory updated', or now it spews random text while it's "thinking", it'll every now and then ask you to choose between two answers but I'm working so no thanks, I'm just going to pick one at random so I can continue working.
People had copy pasta templates they dropped into every chat with no way of savings Ng thatz they they added a sort of ability to save that but it worked in a inscrutable and confusing manner, but then they released new models that didn't support that and so you're back to copy pasta, and blurgh.
It's a success despite the UI because they had a model streets ahead of everyone else.
The fact that a VSCode-based GPT-wrapper is being offered $3B tells you how desperate the LLM companies are.
Anthropic and xAI will also make similar acquisitions to increase their token usage.
There's a lot of opportunity to apply leading edge AI models to specific business applications, but success here is determined more by experience with those business domains than with AI generally.
An AI startup could still be a useful "resume" to get acqui/hired by one of the big players.
I think too many people are focused on the idea of AGI instead of doing what you're suggesting, which is where the real value-add is for customers.
I don't need God in a datacenter. I need help diagnosing an Elastic Search problem.
Its just it is not easy to come into specific pre-occupied space for outsiders.
The LLM space was never going to be kind to those without deep pockets. And right now there's no point getting in it because it's hit a wall. So yeah, startups should steer clear of trying to make frontier LLM models.
On the other hand, there's a ton of hype and money looking for the next AI related thing. If someone creates the next transformer, or a different AI paradigm that pushes things forward, they'll get billions.
Every startup should generate a shitty Ai wrapper product, write one or two lines of code, generate hype and have 2025's version of Softbank give you a billion $'s.
Frankly it's bordering on irresponsible to not be targeting acquisition in this climate.
This focuses on case where the acquirer seeks to capture the value of the startup's business. But this is not always the case, sometimes the startup is dubious, but a cash-rich enterprise can purchase startups simply to eliminate potential avenues of competition. They may not be interested in adding a better product to their portfolio, only in quashing any nascent attempts at building the better product so they can keep selling their own mediocre one.
Also, "model innovation" strikes me as missing the point these days. The models are really good already. The majority of applications is capturing only a tiny bit of their value. Improving the models is not that important because model capability is not the bottleneck anymore, what matters is how the model is used. We just don't have enough tools to use them fully, and what we have is not even close to penetrating the market, while all the dominant tools are garbage. Of course application innovation is the place to be!
1) Collect Underpants
2) ?
3) Profit!
Work just to be a part of it. This is the most consequential time in history.
It's the best time ever to build. Don't work on anything that could have been done two years ago.
Learn the current tools - so that you can adapt to the new tools that much faster as they come out.
[dead]
[flagged]