I think we will see a lot of small labs with funding. The funding amounts may be small but it is no secret that labs from around the world want to build usable (tiny/small/large) language models. Even if investors are not sure how or when the returns will come, they want to see traction and usage. Enough people have enough money to spend. It may sound odd but I think this is the largest spend of private wealth in research that we have seen globally. I see small investor willingness all over the place and I am far away from tech capitals.
I live far away from all this and reading what has been shared all over social media, I find it hard to believe that there isn't a plot.
Jensen has done horrible things with regards to the affairs. But it is clear that people around him knew this for quite a while. So why this sudden and swift action? Does Jensen support a future that is against people who want to make massive profits from AI?
This is the reason I am hopeful with LLMs. Yes there are massive risks but I do not think the average person has ever been empowered by tech in quite this way. An LLM agent can write automation for much of home/farm/business hardware.
I live in a farm. I am software engineer and gave up city life few years ago. Now, all I dream of is automating little things in my chicken farm and growing the business. Not in the mechanical monstrous way that makes me hate eating meat. But a balance in between with actual daily physical work and care. And tiny little automation ideas. Measure food intake, an automatic weight check, door closer/opener, heating control... so many things. And I do not want to buy the best "Chicken Star 3000" model, I want to make these out of microcontrollers with tiny mechanical parts that I can assemble. AI helping me all the way.
I also have a small sewing machine. I want to get a larger one. I want to make "patch" clothes, jackets. I love them. From scrap clothes, etc. What a time to tinker!
There is a lot of value from LLMs but it is something that I can switch easily. Yes, some frontier models are much better but users can still chose to use cheaper models for most of the work. In fact most applications, through which people will interact with LLMs anyway, will not need frontier models unless they offer "chat with AI". Anything that solves specific problems (even multiple problems) does not need frontier LLMs.
Anthropic and OpenAI will have to become more of the end user application stack. Files, calendars, emails, photos, and much more have to pass through them for them to keep making money. Else they cannot make money like the way they have been in this early era of LLMs.
Switching models is easy as a solo developer. Switching LLM providers as a company is not, once all your employees have it integrated with GitHub, their IDE, etc. I think Anthropic will look a lot more like a cloud provider like AWS or GCP. Same-looking product, very high switching cost that would have to be driven at the CIO-level because of all the integrations that have to be redone, IP concerns etc. And ultimately they are selling compute capacity.
Edit: I should clarify that I don’t think this justifies their valuation. It puts them in competition with much more entrenched players with established customer bases, and who aren’t in massive debt to provide their compute. I’d also like to distinguish between a model provider as someone who provides weights vs a model provider who connects their GPUs and models to your codebase and internal systems - I think the latter is where the sustainable business is.
Incidentally this is why I still feel Google is well-positioned in the long-term AI race. If they can remove their head from their ahh and make a coherent product offering across Workspace, GCP and Gemini, they could become as entrenched as MS.
I would assume AWS is probably the most interested in acquiring Anthropic out of the three major cloud providers.
> Switching LLM providers as a company is not, once all your employees have it integrated with GitHub, their IDE, etc
I don't think this is entirely true. Companies are deploying open-source LLM gateways to be able to do that very easily now. Today, you can switch Claude CLI to any model besides Anthropic via https://github.com/BerriAI/litellm and there are major companies doing that.
This works with your favourite ADEs such as Orca as well so I disagree with this premise. The whole experience of switching is pretty seemless based on personal experience.
We saw this play out in the cloud era, notably with OpenStack. Even through the core abstractions (chat for LLMs, compute and storage for clouds) can be almost perfectly abstracted over, providers will race to add niche, enterprise-y features that are harder and harder to put behind a gateway or proxy. Those features aren't the core value of the platform but they are important enough to create lock in.
When I write a system against a particular cloud storage API, I end up implicitly depending on a bunch of deterministic subtleties of exactly how this storage API works.
AI is already so non-deterministic and fuzzy, I'm not sure the same issues apply.
Agreed, deploying AI gateways has become extremely common in enterprise settings. A lot of admins think they are going to use it for guardrails, but I think the real value is in centralized cost tracking and ability to (relatively) seamlessly aggregate providers.
We have like a dozen different LLM tools available to us and it's as easy as clicking on a dropdown to switch in every single one of them, and we have several thousand employees. Self-hosting with the cheaper OSS models is also becoming more and more attractive for larger orgs.
That's just inference. Claude.ai is a platform that enterprise IT teams connect up to all kinds of other applications, and gets use beyond the claude code CLI. Non-developers get used to this functionality - whether it's their form-building stuff, artifacts, slides/presentations, etc... they're building an offering far beyond just inference. I agree the latter is a commodity now, though companies will compete on privacy options (ZDR, guaranteed confidentiality, etc)
Many large enterprises don't use Claude.ai. All their traffic goes through an internal gateway that routes to Bedrock. In addition, there is a whole ecosystem of LLM related infrastructure/tools deployed by "enterprise IT teams" where Anthropic barely has any presence.
But many do, so price them according to that. This is like saying "my office doesn't use Microsoft". Sure, lots don't. Lots do, no one is questioning that Microsoft has a very viable long-term market there.
For infra sure, but its dominance in other areas is quickly eroding. My org is super heavy Microsoft/Azure and we have all of their tools. Everyone making anything useful is using Claude + Databricks because it is a better experience for both builders and users by every measure and it isn't even a remotely close competition. Power BI and Copilot are a total wasteland. The only Microsoft program with any staying power is Excel.
> Switching LLM providers as a company is not, once all your employees have it integrated with GitHub, their IDE, etc
Once the cost reaches a threshold, everything is doable. We ditched Datadog in less than a month. Someone I know had their company ditch slack in a quarter and google cloud usage of 100+ million in two. Everything is doable.
I already mentioned the timelines. Cost is a beautiful line item in your balance sheet that can get things done pretty quickly, especially when it’s as big as AI. It took less than 2 quarters for us to go from “hey use ai in everything or get fired” to “hey we’re putting in default limits that are intentionally very low, if you need more, you’re going to need approvals”.
The stickiness of AI as a whole, let alone a specific provider is not that high as people are making it out to be.
Working for a 10k+ corporate and we made it pretty easy to switch to any provider, people here can use Anthropic's, OpenAI's, and our own self-hosted open weight models very easily. It's one drop-down or `/` command away to switch between them.
I think other companies are probably doing the same, the moat of a AWS/GCP/Azure is much deeper with how contrived it is to migrate between them even if you abstract as much as possible with something like Kubernetes.
Vendor lock in is quite well understood, most larger companies (the one's I have worked with anyway) are aware of the risk and take steps to actively avoid it. Once bitten, twice shy.
> Switching LLM providers as a company is not, once all your employees have it integrated with GitHub, their IDE, etc.
I'm not so sure about this. I guess it depends on how well people did their job there.
Ideally, whatever people built is loosely coupled to that third-party resource, meaning that switching cost will stay low.
After all, it's probably just prompts + an API layer. Unless one micro-optimizes their prompts for claude (which then breaks once a new checkpoint drops, hence is a bad idea regardless), pointing them towards a different API endpoint should be doable.
Currently we are building agentic systems for various business functions like Legal and Finance. How these models perform in those tasks is paramount, and things like cost per run is only a part of the equation. It can become very subjective, and things like brand recognition comes into play. The cost of a counsel or a VP finance overshadows the savings you get by micro-optimizing for a cheaper model. We tend to use the best-in-breed LLM for the various tasks - mainly Claude and Gemini, and sometimes GPT.
> How these models perform in those tasks is paramount,
Then you have no choice but to use your own hosted model because that's the only way you can be sure the model isn't going to be removed or quantized/nerfed behind your back.
If your system can widstand the fluctuations of closed models over time, then it can most likely support switching between multiple providers in the first place.
I should be more clear. The "performance" I'm talking about is not entirely objective in these use-cases, and the bigger part is something you can't reliably measure through tests or evals. The value proposition of the frontier labs is their post-training taste, how reliably the model handles ambiguity and infers intent, while at the same time progressing at the leading edge. Self-hosting trades that for output stability, and absorbing minor drift within one provider is still far easier than switching labs when your workflow and UX are tuned around a specific model's quirks.
> The value proposition of the frontier labs is their post-training taste, how reliably the model handles ambiguity and infers intent,
That's also literally what frontier labs break all the time, be it by removing the model outright (remember the outrage over the forced replacement of GPT4o with GPT5) or by tampering with some configuration (aggressive quantization or whatever bullshit occurs right before new models are released).
If you ever need to scale to millions of customers, then cost per task is paramount. If it’s a one-off vibe coded custom project for one client, then best-in-breed is mandatory without it falling apart.
Ur case actually strength the argument for open weight self hosted model. You would have precision control over not only the technical issues, but also the legal issues.
Once the open weight models are good and cheap enough, there is no reason not to use them
Switching LLM providers as a company is not, once all your employees have it integrated with GitHub, their IDE, etc.
Speaking as someone who has been at professionally this since the 1980s, I suggest that while sometimes switching costs and network effects happen "organically," in this century businesses devote enormous amounts of attention to artificially building switching costs.
It's such a dirty practice that we have a four-letter word for it, a "moat." But of course, the moat isn't really to keep competitors out of a market, it's actually to keep customers trapped within it.
I fully expect every one of these AI providers to put as much work into trying to trap their customers within their ecosystem that they do to improve their capabilities.
Since siege warfare is about bringing down the fortifications and a moat is part of those fortifications, cyber siege may be something akin to patent trolling or IP copyrights? Fire off dozens or hundreds of claims against your target and drown them in legal costs.
Then you have the experience to know that every land grab in technology so far has followed the same enshittification path: At first it's all about competing with other entrants to provide value and win customers. Then, inexorably, it shifts from winning customers to defending their turf. Then, inexorably, it shifts from protecting turf to maximizing extraction from their turf.
If you believe that this will be different, why not? But for now, I'm expecting this to go the same way. Yes, LLMs are something completely new. But also, the people behind them are exactly the same as the people behind every previous land grab.
I can't wait for anthropics next big drop to be that it can order you a pizza, steering towards benign, token burning features that you didn't ask for and eye real-estate abuse for ads.
It's the end of an era. Lets pray open weight models give us a real choice soon.
In the meantime, Stonks up Anthropic. I'll always remember the good old days when you championed developers and changed what was possible.
Switching LLM providers as a company is even easier than a solo developer, most large companies I know are using AWS Bedrock and selecting another model is basically just a checkbox for them
As far as I know, our corporate has deals with all of them: Anthropic, OpenAI and GitHub Copilot.
They stroke a great discount deal with all, so different teams/departments have different provider, and those using Copilot are even free to switch between models as they please (although with lower usage limits).
You could make the argument that it's not worth switching SAAS providers for enterprises when the costs were like 10 dollars a month per developer and switchign would require 1 team to work for half a year... but with AI providers it's more like paying 1000 dollars a month per developer vs paying 200 dollars if you strike a good deal with different provider and swtiching costs are more like 4 to 8 hours of work for each developer.
Indeed, and I think AWS is in a perfect position to consolidate the entreprise market rather than Anthropic here, they already have the rest of the server stack so this is just another addition.
My company switched from direct access with providers to AWS Bedrock. Now we just pick which model we want from the list exposed by Bedrock. We can also leverage tools like T3 which allow hand off from one model to another.
Easy to say, not so easy to actually get your mechanics, your project managers, and your business operations staff to get over that hurdle unless you’re willing to handhold them through the process of changing directories or explaining what an environment variable is to them. (This isn’t just boomers btw. Lots of young people have had the command line abstracted away from them their entire working lives)
There are many perfectly successful companies out there with limited computer literacy among much of their employees, let alone AI literacy. Those companies are great targets for Anthropic’s enterprise sales team, if they can provide an intuitive user interface that allows said employees to make use of AI tools, allowing them to leapfrog the need for computer skills.
We’re not talking about developers, we’re talking about people who just want the technology to let them do their job. What even is AI literacy? 2 years ago it was prompt engineering
Weird take. Tech boomers started with tape or commandline. The ide only became the development environment of choice in the late 90s, firmly with genx and millennials.
Claude code is a terminal interface not a command line tool. More like vim than sed.
Visual BASIC was 1991. QuickBASIC, Borland Turbo Pascal, SmallTalk-80, NeXT InterfaceBuilder, all pre-1990, several well-pre-1990. Boomers absolutely had, and indeed devised, the IDE.
My F500 company switched ~10k employees of Claude to copilot in 30 days. Google did similar between Jetstream and Claude. The moat just isn’t there right now
Model switching behavior is going to be far more complicated.
First, everyone and their mom is selling AI now. It is being bundled in nearly every product. Buy CRM to run your company - they will upsell you their bespoke AI - mostly RAG/chatbot using the data inside the platform. This is an important area because today lot of companies are trying to build their own AI bespoke workflows. But companies like SAP, Oracle will spend money on improving this aspect. And people who buy/use AI for this purpose don't care if the underlying AI is Claude or not.
Then there are cloud providers. You can get Claude, OpenAI from AWS Bedrock etc. They will take the heavy lifting of changing integrations, IP concerns etc to keep people on their platform. With time expect them to make it even easier, not harder to switch models. Also, they will push for open source models because it will be more profitable.
Lastly buying from direct providers. This is here the lock in comes in. But I believe fewer and fewer enterprise will buy directly - many preferring to buy against their cloud commitments.
I really think you overestimate the level of lock in here. If a companies tells the devs to use X because it cost 1/10 of the price they will just have to suck it up and switch.
I didn't particular want to work on a "all in agentic project" but that the nature of being employed.
The same LLM we use day to day will make it trivial to switch out providers in our infrastructure.
One thing that makes easier to switch with LLMs as opposed to other tech is the text-in text-out nature of the LLMs. All specs and data are in text files mostly in natural language which is patent free. This makes it harder to lock in content.
As you have written, the agentic harness is a slightly different matter with plugins and features. But the core tech makes replicating software a lot easier so you can in principle use an agent to make it easier to switch from its own harness. There are caveats to this since in practice the quality and feature sets of harnesses are not the same. But in my opinion, it is far easier to switch agents than switching other tech that came before them that used proprietary formats and apis to lock you in.
> Switching LLM providers as a company is not, once all your employees have it integrated with GitHub, their IDE, etc. I think Anthropic will look a lot more like a cloud provider like AWS or GCP.
This might be true today because Anthropic/OpenAI/Google have good sales AEs, but as soon as company's start doing FinOps on their LLM spend, the game is over for them. The model will become fungible and hot-swappable (as it already is) and it'll be a race to the bottom.
I think the future will be a lot different they we expect, AI is going to be the infrastructure layer while there will be toolings for each processes, so switching between model is a built-in feature, i.e. the tools will inject specific prompt or context based on the models, and all these are designed by people selling the tools, hence, a moat?
Let's say that the face for largest foundation model slows down enough to make the companies cloud-like, instead of whatever they are mow. Imagine the free credits that OpenAI would consider giving you as long as they are spent on migration out of Anthropic. They could even classify conversations by themselves.
Depending on how the company leverages them, it's super easy.
For us it's as easy as clicking a dropdown. The copilot vs code extension allows for this ez pz, the same is true for companies that implement an AI gateway for other tooling.
This is the logic behind B2B SaaS's revenue multiples, and yet AI itself is causing those multiples to rerate based on a change in expectations about the durability of an entrenched SaaS provider.
Totally agree with your Google thesis. I would never count them out. I say that as someone that uses Claude (and a little chatgpt), but not Google's AI at the moment.
With ad money Google never had to learn how to ship a product that makes money. An obvious place they fail is that when something breaks there's no one to talk to to fix it.
I've observed that there are some companies and some folks that are very fanatical about a specific provider. "Only Claude" is very common. "Only GPT" less common, but still see it.
A lot of the non-technical and semi-technical folks are only comfortable using a specific client like the Claude desktop app. A lot of teams wrote skills and eval'd/benchmarked their skills specific to a model family and harness.
The "I can use any model, any harness" or "I only use a meta harness" (Orca, Paseo, Herdr, etc.) crew is relatively small IRL. Most people are still single-session, single platform user mode.
Most orgs have an enterprise contract with a single provider so all of their internal assets are built for the provider. It goes beyond skills, but also the telemetry, enterprise policies, and managed configuration.
For a company though, the idea that a few disgruntled employees will shift massive purchasing decisions is a bit of a stretch.
We have a developer on our team that keeps asking for a claude code enterprise license, we keep telling them "no, it's way too expensive, get used to using the copilot extension/cli".
For consumers, sure, but the bulk of the revenue opportunity is enterprise. Anthropic/OpenAI have a hard row to hoe IMO.
As an only-clauder, there's another elephant in the room. Anthropic is by no means perfect, but:
* using OpenAI supports the Altman/Thiel/Musk/Ellison/Karp/Trump billionaire Orwellian worldview and also ads
* using Grok supports a ketamine-addled open fascist sympathizer jizz-spewing 12-year-old-edginess memelord and his similarly broken worldview
* Google's models are pretty good but every single product they have ever had is ultimately either suborned or killed by the advertising division
* local models pretty much only half-exist because Jensen Huang wanted to cheat on his wife; future is unclear
One can argue for harnesses and so forth being lock-in, but the ultimate lock-in is morality and wanting to have a civilization at the end of all of this; I can't recommend either enough
China is literally banning Nvidia chips, over 600K Nvidia chips sitting collecting dust in Hong Kong warehouses due to China blocking their import. The notion that open weights half exist due to smuggling or Huang is a sick joke.
It’s weird that people don’t seem to get this. LLMs are the most easily switched-out technology I’ve ever used as an engineer. The input is literally just free-form text.
> It’s weird that people don’t seem to get this. LLMs are the most easily switched-out technology I’ve ever used as an engineer. The input is literally just free-form text.
It depends, if you're automating business processes (honestly probably where the money is), then you'll need a really comprehensive set of evals and moniotoring to ensure stuff doesn't break if/when you change providers.
If you have none of that, then switching will be very expensive (but it was ever thus).
> you'll need a really comprehensive set of evals and moniotoring to ensure stuff doesn't break if/when you change providers
Unless your self-hosting open weight models, you need this even if you don't change providers, since you have no control of what's on the other side of the API.
That just means the business automation is not deterministic, and highlights the issue with LLMs. The solution is not to switch providers but instead make the LLMs generate something that is deterministic given the same input.
> That just means the business automation is not deterministic, and highlights the issue with LLMs. The solution is not to switch providers but instead make the LLMs generate something that is deterministic given the same input.
I'm thinking of stuff that would typically be outsourced (like compliance stuff for identity verification et al), where it's really really hard to have deterministic rules.
LLMs are definitely a step up in dealing with fuzzy inputs, and enables you to mostly automate stuff that really wasn't feasible before.
But you do need really good evals and monitoring to make this work effectively.
Ah that would make sense then. I was thinking strictly in terms of business logic for handling input of data from an external source; sort of glorified data validation and then a decision based on the input data.
I suspect that's what some people are using it for, because they've been sold on this panacea of AI, when in reality it is far less capable and needs extremely strict boundaries to be reliable.
Yeah, it's the absolutely weirdest thing about all of this. People keep telling me I need to up my AI game yet no matter how advanced the tools get I'm still typing stuff into a text box. My skill is typing English. I was already pretty good at that.
People seem very unimaginative about the ways in which somebody equipped with a lot of intelligence and compute, and control over both, might make money.
The recent Anthropic push into biomedical research gives a clue.
I am not unimaginative about Dario's megalomania. I just don't believe he has the ability to execute. He has no prior experience to Anthropic and nothing he has done as CEO would create confidence that he is actually capable of operating at this scale.
Additionally, I think people are being unimaginative about what future generations of frontier models might be able to do, and the business value of having a 6 month head start on competitors.
For example, people are now content with small models for agentic coding, but even frontier models were mediocre at agentic coding just 12 months ago! Frontier adopters have had more time to transition their systems and processes to what we're already taking for granted.
4-6 months is nanoseconds in field of AI, just for reference, Opus 4.8 was in June. If I had a free Opus 4.8 for rest of my life I would be more than content for coding.
This is equally true for how risky it is to build such a business on top of a 3rd party black box. Model nerfing is widely speculated and the only true way to guarantee that the models you're using are stable and consistent is to run your own hardware. Not to mention the recent Navier-Stokes controversy.
The moment you find a money-making formula on top of someone else's black box model, there's nothing stopping them from taking it for themselves.
For govt or science research, reproducibility and transparency is paramount, more than results because if it’s opaque box and can’t be reproduced, its not not reliable.
Model nerfing is rampant from people using subsidized plans.
Stories are much fewer and further between for people using the API. Which makes sense, the subsidized plans are "You get what we give" whereas the API is "You get what you pay for".
This is the same argument we have heard many times with SaaS that would presumably get replaced by open source equivalents. In practice it does not happen because there is a lot of value in having someone else deal with a problem.
Nobody trusts Anthropic's ZDR, in the limited cases where they even offer it. The key thing is that weight activation sequences from a query are not "customer data" they are internal Anthropic data. So in Anthropic's legal theory they are training on their own internal data, not "customer data," and that is perfectly legal, even though it is the exact same thing in practice as training directly on customer data.
obviously that is not the bar. but microsoft and amazon are much more trusted than anthropic. with anthropic it is just assumed that they will steal your data.
It’s literally a dropdown button to switch LLM models in Copilot, Azure foundry, or even in OpenRouter.
The closest equivalent is changing your credit card on checkout. It’s different from SaaS which holds your data hostage, transferring out has high cost and labor, these models are storage and data agnostic, it works on top of the data. It’s completely different
The problem here is having a stable access to LLMs with strong trust the provider will respect ZDR, won't poison the models and will provide a stable platform you can depend on.
a big move into end-user software would mean that they have to compete heavily with their own largest customers. The economics of offering SaaS are also under meaningful pressure and don't warrant anywhere near the valuation that is being suggested
The biggest moat with applications like MS Office is compatibility. Sharing a slide deck or spreadsheet with others is a simple decision, you know they have it or have software that can read/write it.
I don't get the need to have Anthropic build their custom versions of MS Office apps, I can tell Claude to generate a word doc from a wiki page and it does a pretty damn good job. Anthropic's statistical median level of software that they had Claude create isn't going to replace Office apps to the point where companies cancel/stop renewing their MS Office subscriptions/licenses.
The other element is trust. Do end users trust their data in Anthropic's systems? A lot of people get creeped out when they search for content from one device in their household and then nearly instantaneously they start seeing ads from other devices on the same topic. How creepy will AI get if it starts consuming all your docs, spreadsheets, personal finance info, emails, etc.?
I beg to differ I already am stopping using MS office recently but that's mainly because I really dislike how they lock up the apps if your late on your annual payment. It's a grating value capture model. I'm sure I will have to get back to it at some point but in the meantime google is providing great optionality and I don't have a concern that the software stack has glaring security concerns.
On the trust thing - I don't trust anthropic and I don't think MS has been a great steward over the years either.
The moat of backwards compatibility and plugins is why MS Office has survived decades of competitors without any appreciable loss of market share.
It's hard to see Anthropic replacing it, and even if so - why bother, the revenue is miniscule. They need 100 MS Offices for the compute contracts to pencil out.
End-user software is going to be under a lot of pressure. If you're Microsoft, Apple, or Google: you can easily copy products or integrate ideas into their platform. I see companies like Adobe, Salesforce, and Oracle having to downsize significantly to stay profitable. Having said that, my in-laws still pay for an AOL email so I don't think their cashflow is going to zero soon.
Adobe seems to be in a tough spot, because so much of the lock-in is client side, I think Salesforce is fairly well protected, it's just so impossibly difficult to remove once a company has fully adopted it. Similar to the staying power of the ERPs of Oracle and SAP.
It's the same really. There have always been alternatives that can work with Adobe's various file formats. Even more aren't going to change much, at least for their enterprise key customers. This is not only the case for Adobe, but many others as well, e.g. MS office is still a thing these days, and so are things like 3Ds Max or Maya, albeit having many compatible cheaper or even free alternatives. It's not so much about features, pricing or even stability. It's about customer support. Do I want to ask an LLM and test my luck or write a GitHub issue and cross my fingers that someone will read it within the next weeks? No, I will pick up my phone or write a mail and get a response within hours if not minutes.
I ask my llm to write custom software and love the freedom to customize it to my needs. Today I do a presentation on Ubuntu with a new "Presenter" app that works offline, runs 100% in CUDA on GPU and streams videos, starts in 2 seconds, lets me demo live robotics policies in 3D simulations and Opus 5.5 wrote it yesterday. This may sound very unusual, but I think it will be common in the future: moats disappear one by one.
The big problem is Photoshop. I have no love for Adobe, I proudly moved my department off of all Adobe products a few years ago. I am a big fan of open source solutions and I am always impressed at what the GIMP team (and some others) have accomplished.
But…
Photoshop is king. Everything else is behind by a solid decade. If you’re a professional the $40 a month or whatever it is now means nothing to you as well. It has the professionals are all on it, and because the price isn’t astronomical, that’s what everyone else uses who needs to use that kind of program more than once in a blue moon.
Premier? Replaceable. Lightroom? Pretty replaceable though still somewhat ahead. Photoshop? There’s only one Photoshop. And as much as I’m loathe to give Adobe any credit, it’s the best by a mile. There are too many essential features in QoL things that other programs are missing.
Luckily my department doesn’t need it, but if we did, we would simply have to accept Photoshop.
> If you’re a professional the $40 a month or whatever it is now means nothing to you as well.
Exactly this, but also for an employer, paying $40/month/user is good value in order to ensure you have a room full of productive designers or image editors.
No employer is going to dump Photoshop and force the designers to re-learn everything they've learnt over years or decades. There would be such an enormous loss of productivity.
If you've ever watched a designer at work you would see that. They just breeze through Photoshop, hundreds of things in a few minutes. Super productive. All thanks to muscle memory, knowing all those shortcuts, knowing blindfolded where all the toolbar buttons are, knowing how the filters behave, knowing where everything is in the dialog boxes.
A similar comparison is all those "just write it in Rust" people. Well, sure. But if I've spent decades coding C or whatever, I'm not exactly going to be able to switch to Rust overnight am I .... realistically its going to take a timeframe measured in years to reach the same level of coding proficiency. And no, vibing with an LLM is not the same thing.
Also Photoshop has native integration with Adobe Indesign and Illustrator, so there's also the cross-tool aspect too.
I'm not sufficiently skilled with Photoshop to gauge how well this replaces it, but you may be interested in this supposed clean-room reimplementation of Photoshop: https://github.com/storytold/photocraft
Heard about this! Haven’t had a chance to check it out yet. I’m sure once I have an actual project that needs me to roll up my sleeves with Photoshop I’ll give it a shot. I probably should have mentioned it in my original comment as a possible option
Not a graphic designer, so, truly asking out of curiosity: what are the QoL things and essential features in Photoshop? I've been slumming it with GIMP for years now and not felt anything very obviously missing, but again I'm not a pro.
Compatibility issues, which can be a huge deal partially due to near-vendor lock in. If you historically been on Photoshop, bringing your working/former projects into another software is not easy, which matters because revisiting work is common.
Not all tools are non-destructive in GIMP. Things like raw support, smart objects, retouching, etc. all either are super thin compared to Photoshop or require workarounds/a lot of manual effort. There are just lots of tools you expect in Photoshop that either don’t exist or exist in an inferior form. I can’t elaborate too much because I don’t really touch Photoshop much these days, but I work with people who do and worked in production long enough to know the expectations. GIMP is great but Adobe is a massive company with deep pockets, it’s hard to keep up. Their user experience is miserable but the tool itself is just far more robust, and a client simply isn’t going to accept that you won’t do something that Photoshop could do quickly because your alternate program can’t do it or does it with a lot more effort. That’s the reason I ultimately dropped Final Cut!
I would say that Apple's Create Studio can move toward feature parity faster with Adobe. My point if you're not a platform provider it's going to be harder to make a profit.
The risk for Adobe isn’t that we replace their software with some AI generated alternative. It’s that we bypass that class of software altogether and have an AI do the work.
Surely not just prompting like this and it will one shot it. The usefulness of the flash authoring tool, was it was battle tested and polished.
But I did gave it a shot to took up developement again of abandoned www.wickeditor.com (a web reimplementation of flash). Good results, but limited time and energy for that right now.
> If you're Microsoft, Apple, or Google: you can easily copy products or integrate ideas into their platform
This argument doesn't hold water. Google tried like no one else to have a chat app and it failed miserably more than a handful of times. Same for social networks.
I'm not sure corporate procurement has ever selected the cheapest and best option, so I think they may be fine. Plenty of reasons to avoid completely locking into your cloud provider as well.
Adobe creative suite is $70/m. From most businesses perspective it's not something worth looking for cheaper alternatives and the alternatives has existed for a long time now.
now imagine that said customer/competitors have been uploading/generating their source code, problem reports, marketing collateral, and other IP _to the LLM company_ for the last few years.
Why haven't people switched then? Same could be said about Google search, Facebook, X etc. Claude is the de-facto SWE model now, yes you could use an open model but you would always be wondering if Claude would do better, like using Duck Duck or some other Search engine
> Claude is the de-facto SWE model now, yes you could use an open model but you would always be wondering if Claude would do better,
In my testing (incredibly similar workflows, moving around between claude code and cursor), I haven't found that much difference between Sonnet and GLM-5.3/flash.
Sonnet seems lazier, but has better prompt adherence, and is constantly telling me it can't do things, while GLM will just Find A Way to do whatever I've asked for.
I really appreciate being able to see GLM's thinking tokens, as I can head it off quicker. Claude's a little harder to steer, but the integration in Claude code makes it much easier to leave it running on things overnight, which is hard to get working properly in Cursor.
Overall, I'm not sure Anthropic have that much of a moat, in coding particularly. This seems weird to me, as they have so much more usage data that's hard to replicate.
If Google and Microsoft have gone dormant in this space, I have to think it’s because they haven’t been seeing any ROI, either. Possibly because customers have been so underwhelmed.
I’ve been using Office Copilot more or less since it was released, and the general opinion around the office is that it’s a huge waste of time. The auto-transcripts produce 3 pages of fluff when previously someone would send 3 sentences’ worth of what actually mattered. The auto-drafted documents rarely add anything meaningful beyond what was already in the prompt. The question answering and research tools are only as good as Sharepoint/Confluence/Jira/etc’s pre-existing search feature, and therefore untrustworthy, etc.
What made Anthropic exciting was they had locked onto a use case that reliably generates positive value: coding. Not much else has proven, from a business perspective, to be anything more than an exceptionally buzzworthy money pit.
I have been using Claude privately for coding a few weeks now, and my company encouraged us to use company paid Copilot recently. I tried coding with it and the result made me irate i had to close it it was so infuriating. The idea that i need to point out it’s bugs and have Copilot tell me “you are absolutely right!” Haha never again
> Meanwhile, Google and Microsoft, which already dominate that area, seem to be dormant.
Google owns 15% of Anthropic I think.
And Google offers Argon 4 to paying Workspace and Google One customers now. How's what was the number 1 model on nearly every single bench a few weeks ago being "dormant"?
I know people all in in the Google ecosystem (getting credits from Android phones they buy etc.) who only swear by the Google models and who are now using Argon 4: I asked them and a few days ago they had access (but they were already using other Google models before that).
Doesn't GMail have billions account by now?
I'd tend to say with billions of users on GMail and a model like Argon and 15% of the Anthropic shares, Google is not where you think it is.
It's easier to use claude to do stuff insider workspace than it is to use the integrated gemini. Claude has more power to do stuff than gemini, in it's own house. That is such an egregious fumble that it's honestly hilarious.
It's almost like the workspace devs are anthropic plants that are purposely underminig google.
Integrated gemini stuff sucks but it's nothing that special about Claude, general agentic tools are just better than app-specific ones. Agy is also much better at doing things in workspace than the bespoke in-app gemini stuff
The sad thing is Google could totally own this space.
They had a huge head start, more raw data (from search) than anyone, a huge installed base on Workspace and consumer Gmail, tons of compute resources...
Right, because Google actually has assets and a responsibility to keep their customers secure, whereas Claude can do things that are completely insane from a security perspective because they have no liability.
Yes, that's the direction I've been witnessing for Claude too. But it's a short to middle-term strategy IMO, by the time companies are ready to drop all these apps they will need something else entirely, not Claude's slides or spreadsheets or Salesforce.
The fact that Frontier AI can become Google Workspace at a flip of a switch is what is crazy valuable. Anthropic and OpenAI can enter and disrupt any market, especially software, faster than any company ever could, because users want the flow to start and end in their chat interfaces.
I think eventually almost all non-creative software will, essentially, be a wrapper around a session.
This pattern has already happened before, in the translation industry. Someone figured out how to learn from and recall context from ordinary translation work (the "translation memory"). It started as an add-on to Microsoft Word, and matured into a standalone software, that was ultimately a wrapper around a document editor. Then language management systems plugged into that wrapper, hey presto, every translator at your company does the same work they were always doing, and you've produced an entire corpus of quality translations to reuse everywhere.
Now _translate_ that idea over. Every time any human works with an agent on anything, simple heuristics find everything true and reusable, save it to the everything-memory as skills, refences, glossaries.
Anyone does any work, all agents can now that work at that organisation. Self-fine-tuning models everywhere, and it fits into corporate intellectual property expectations.
I have done something like this at a medium-sized listed business.
Helping my wife this past week try and setup an automation in her workspace account revealed how dormant google is here.
Gemini is unable to take an attachment from gmail, read it, then place it in the correct drive folder and update a spreadsheet with what it found in the attachment.
For whatever reason, Google decided to firewall "Gmail Gemini" from "Drive Gemini". Gmail Gemini can send files to drive, but not edit anything, and Drive Gmail can edit files, but cannot be initiated from gmail.
Meanwhile claude code with Google extensions can do this no problem.
One thing I have noticed; as a project matures frontier AI is less useful essential.
Qwen3.8 and Meta Muse Glimmer are great at "fill in the middle" of frontier architectures and plans. Frontier generated plenty of great context already to keep them on rails.
Delegating like a CEO down a pipeline of expensive experts to cheaper local small models could become a common trope, especially if Apple M7 and beyond live up to rumors as LLM tuned, and chip competition exists to keep forward progress, SaaS models are cooked.
The goal for chip corps too is the right mix of generic model enabling logic; not necessarily a full concrete model but enough of the abstract to efficiently boot strap and run a model; move common patterns to silicon not data.
Token vending is little more than behaving as a "wrapper app" over electricity. There is zero technical moat to their business.
I have tried so many times to get off Chat GPT. But it just... knows me? Switching to another LLM seems like a huge loss. For coding projects especially... I've set MD files to the max, explaining how i like code to be written, etc. etc. i've tried to tell it not to create a bunch of files and such, but i can never use an IDE. I feel like i'm the only caveman coding in a chat window... but it still seems the best to me? The only time it's been worth it for me to use an IDE LLM like codex or cursour is when i'm prototyping a website, but then i just go and pay someone to implement it in wordpress because it's so much easier to maintain that way.
I believe the models have been effectively commoditized. It's believed by many that the lead of the frontier models over open weight models is only 6-9 months. That was one of the pillars holding up the SpaceX/Anthropic/OpenAI valuations.
The other pillar is hardware but I think that's only going to last a few years at most.
These trillion dollar valuations were based on the idea of "winning" and "owning" AI (IMHO) and the open weight models have killed that outcome. Now it's a rush to cash out before it all comes tumbling back to Earth.
They are fungible in the sense that they can accomplish the same task up to a certain acceptable quality.
If I needed to implement a simple function that is well described, all LLMs (for the most part) are fungible. They will have almost identical implementations (sans variable names) and all be fit for purpose.
Why? Given that there is a limit of things that they can do, why would inventing a new "AI-native OS" be an interesting target? They could instead focus on solving actual existing problems in science and engineering, finding solutions to problems that they know people already care about and would pay money for, instead of putting a lot of effort into maybe inventing a middle layer, that might be used to solve problems or maybe not.
There will also be a lot of competition in that space, because everyone can take a stab at building an "AI OS" (whatever that means) using AI for basically free.
One more piece - say Google Gemini gets way better than OpenAI or Anthropics offering, or Microsoft Copilot etc gets way better...
If Microsoft or Google's AI is killing it, and it's built into the OS, how will Anthropic or OpenAI compete unless we run into the same old antitrust/competition issues with OS?
At that point, they might just say - "fine, it's easier to just have our big AI brains and money make a new OS we control and we can compete too."
Again, why bother why IT? Just do the things that IT is built to help build. Solve medicine, energy and materials directly, instead of competing at AI integration.
I see your point. You're proposing they can just do an end run around the "business and world problems" directly.
I guess that could be a thing, but, that puts them in competition with the people who they might want to be their customers, whereas I don't necessarily know that they would see other tech companies as customers unless they were super interested in partnerships.
According to "the normal progression of companies" though, you might be right because after a company has becoming big enough and exhausts consuming its own market, it often expands into verticals - we've seen that with traditional big 5 consulting companies or IBM, where they expanded into "working embedded with their customers on their problems directly" many years ago, and now every company in the industry approaches it that way.
That is corporate partnership and field work, not so much consuming the other person's work/business... whereas I could see them just consuming the OS because "it's no big deal when you can do everything you need with AI".
- at some point, look at what they have they can do to grow to satisfy shareholders - AI companies have been acquiring companies that make developer tools - that is the build side - but once they've exhausted that, if they want to keep growing, what do they go after? Bigger and better models at ad infinitum? I don't think they can go on like that forever, and, if optimization of AI and competition keeps cutting into their main business line of LLMs
etc, they have to diversify.
- If everyone has the power to make apps, there is only so far you can go there - my guess is that the pressures to reinvent regular things like office suites, email clients, calendars to be integrated into their desktop tooling is part of this
- who is in the best position to be the platform/front door for everything, if everything depends on AI?
- when have these kinds of companies ever said "this boundary is enough"? It never happens, they keep going, keep trying to find ways to expand, capture, build moats, etc.
- look at what their desktop apps are turning into - they are already an application layer and becoming moreso every day, at some point they will ask - why are we sitting on these other OSes... we have enough money and we can get into more "hard" engineering things, like building hardware using their own tech (which they are already experimenting with), and if you're building the hardware, why not build the OS?
- and, what is an OS in the future? Have you considered what a post-AI OS even looks like? Does it need a start menu? Does it need any of the legacy stuff? Is it a local software factory?
There is someone already experimenting in this space - the guy who founded Basecamp and is building Omarchy [1] Linux... he sees it.. how long until these companies do?
If they don't build this, what will Microsoft do? Both Microsoft and Apple are well down their way, Apple is probably further ahead, but neither of them are the frontier model companies. Don't forget Google has a stake in OS too.
"Everybody will try to build an AI OS", but which ones are the most likely to survive in our current economic system?
If our economic system was based on partnership, I might agree there isn't much point in these companies building OSes, but ours is based on competition and domination generally... so unless the AI bubble truly does pop and the funding stops (which it may of course), how else do they keep fueling their stock prices to dominate the market?
I think something is going to have to give somewhere in the chain... otherwise the ease with which any software can be replaced will be too high and I suspect they will be forced to compete.
One last thought - between Microsoft/Apple or Anthropic/OpenAI - long term - who is in a better position? The debate arises from that question, in my opinion.
I hate to even put this out there, but "OS-as-a-subscription" (OAAS) could be a way. Your Claude payment covers all the software running 'your' computer, plus the ability to develop new 'apps' for specific needs on the fly telling the OAAS computer in human language what you need. Throw in services like email and cloud storage and some customers might find such a bundle attractive for the right price. It would suck revenue away from MS, Apple, Google. Enough to justify this valuation? No idea.
Based on how shoddy their customer facing applications are, I wouldn't trust any of these AI companies for an OS. It's like Google II - can't trust them to support a product longer than it took to develop it.
I agree with you on quality... but if AI/LLM/etc continues down the path of continual improvement, it's only a matter of time until the products become better or different enough that it isn't an issue.
And making products fast and first over quality has never stopped these companies from trying, when they have so much funding....
I do not share your same hope of improvement given time. The commitment should be on quality first, not speed of development. This is why we are all drowning in a proliferation of poopy modern software.
The fact that their desktop offerings are so pap is an indictment of their engineering prowess and a demonstration of their apparent inability to use their own (actually very good!) LLMs to generate decent software.
It baffles me how a technology so good at spewing out software is apparently unused by the people who are developing the technology.
Fair enough. It will be interesting to see if the world returns to sanity and focuses on proper engineering discipline, planning, quality etc...
I've lived through enough hype cycles in tech myself that I do know that eventually it tones back down and people go back to work, perhaps altered from what they did before with new approaches, but back to work nonetheless.
I just felt myself that this particular technology, and this particular cycle, might be a little different - let alone in how it caused changed in our society whether it worked or not, but that also, there is still room to go.
I'd be fascinated to know what they say on the inside, if there is still room to go and to improve, or if we're nearer some sort of plateau in progress.
> Yes, some frontier models are much better but users can still chose to use cheaper models for most of the work. In fact most applications, through which people will interact with LLMs anyway, will not need frontier models unless they offer "chat with AI"
I am finding some really routine tasks to be borderline with frontier models. Performing E2E testing of enterprise web apps would clearly disintegrate with local models. You need that reasoning black box on top. I don't care what your favorite benchmark says. I care about what actually fucking works.
The people who don't understand the economic situation are those who are in the pure play space with AI and multi agent circuses. Why bother spending any money when the occasional failure has literally no consequences? Who in their right mind would pre-pay for tokens?! I wish I could see the statistics on what % of HN uses AI for fun vs for work. It is clearly a kind of Factorio entertainment experience for a lot of people. If I was just using it for entertainment this whole IPO business would look really silly to me too.
We all love moats, but I mean, it is still possible to make money in an industry simply by being the best at what you do and charging for it.
Anthropic has shown a consistent ability to deliver the best general purpose models. Can that continue indefinitely? Will models become "good enough"? Will open weight models catch up to SOTA models? Will some genius start a new company with better models?
And, on Anthropic's side, will they be able to use their market power that comes from being the best to push through regulations that will create a moat; instituting some sort of mandatory safety checks that coincidentally are exactly what Claude is trained to pass, regulating the amount of compute or trying to claw back data centers to prevent new training, etc.
Or maybe they just have the magic sauce and they're going to stay on top of the game and people (like me) will be happy to continue paying money for their product.
What work do you find frontier models significantly better for?
I have been playing about with different models having started a new job recently, the medium priced ones don't seem any worse so far. Admittedly my tasks have been fairly straightforwards so far.
Not only can you switch easily, they all seem to do the same thing, fail in the same ways, make the same bad linguistic choices, write the same code... I think at the end of the day, LLMs are a commodity. They're like a grease, or oil. They're trying all the tricks with branding cute names, version numbers, benchmarks, thinking levels, but it all just seems the same stuff to me: I put in tokens, I get more tokens back, and if I didn't know ahead I wouldn't be able to distinguish if they'd come from Anthropic or OpenAI or wherever. Maybe some people can but I can't.
I'm fascinated by exactly how an investment in the company that first reaches AGI (or whatever we decide to call it) might pay off... It feels like the stocks would absolutely be worth owning in the immediate term because they'd be hot property at the point where people realise it's happening.
The thing that I'm less clear on is, if and when they achieve some kind of recursive improvement exponential growth thingy, isn't that going to heavily disrupt the entire economy / stock market? In that environment, would Anthropic's stocks really be a good bet, given they have to operate within the economy?
>> that first reaches AGI (or whatever we decide to call it) might pay off
It is a myth. They are chasing science fiction. They either dream of an omniscient overseer who will take control of society and nominate its shareholders-parents into the governing elite, or they envision a chatbot that will solve every problem, again turning shareholders into the power brokers. They forget that nobody will ever be allowed to "win capitalism" and take over the planet. Companies have are torn down for far far less. Every AI model is data, ones and zeros. It therefore can be copied/shared/leaked. Intellectual Property law is not a ladder for assent to lordship. It can and will be rewritten or outright ignored rather than allow capitalism to collapse.
> The thing that I'm less clear on is, if and when they achieve some kind of recursive improvement exponential growth thingy, isn't that going to heavily disrupt the entire economy / stock market? In that environment, would Anthropic's stocks really be a good bet, given they have to operate within the economy?
That's a long-term problem to be dealt with later, greed dictates to pursue the former as long as the music is playing and everyone is dancing. Similar to 2007-2008 when it was clear that things were breaking but the incentives were to play the game for as long as possible before it imploded.
It's clearer to me why this works for generic investors than for the AI companies themselves - unless their plan is just for key employees to cash out and live on a private moon just before Consequences arrive.
> It feels like people are still trying to bet on the horse that will achieve gAI.
I think that at this point gAI is just a marketing blurb, and the actual race is to dominate current practical applications involving basic RAG and ReAct loops. This can go from anything between software development, business intelligence, and market research. A company whose top brass can use a team of agents to do what nowadays they assign to mid- and top-level management is an executive with tighter control over the whole business process. For this you don't need gAI.
I've noticed they've actually stopped harping on about this AGI nonsense, I suppose because no one was buying it to begin with and they themselves didn't really have any idea wtf they were talking about when talking about "AGI". Now it's all about a bajillion agents running autonomously or whatever
No no, Anthropic just published an article claiming "jagged AGI," another marketing term. In it the writer spells out why they haven't achieved AGI though. But the marketing is still going strong, now with even more vague terms!
Easily in the context of the corporate world is to pay for something, that something gets installed and then just works, and if it doesn’t work there’s always someone you can call 24/7 to come and fix it. I don’t see how this is possible with open models.
What do you mean? We use Bedrock at our company and I can switch to literally any model, including the open ones, by just typing /model and picking it, even including inside Claude Code. It's ridiculously easy to switch, and imho GLM 5.3 is good enough for 99% of my daily tasks already.
the answer is to use a gateway that all the prompts go through. There are many proprietary and open source solutions for this. You get the benefit of making it trivial to switch between proprietary providers in case you are unhappy with the performance or squeezed on price. The AI Gateways put cap on LLM providers' margin upside.
Additionally, the gateway allows for caching, additional security checks, centralized monitoring. I think most corporations have a multi-provider setup anyway.
You pay a service provider, who manages the open models… there is absolutely nothing special about it. The hardware can be something you own. Or you rent. Or you just pay the service provider to access the open model apis. Lots of options here, it’s not different from pretty much any other software service in the corporate world
- work: I need frontier, cost of mistake is huge
- home: i do entire renovation with it, cost of mistake is big. monitor baby health, development, etc: cost of mistake is big , taxes: cost of mistake is big
casual researches about certain facts, shopping, booking: cost of mistake is medium
non of this fits : i will run an expensive laptop hidden from the baby, since i'm on the move, there is a small percent of casual queries that normally will go into finish products: google ai, muse, etc.
But you’re wrong ser about small or cheaper models. Choosing to work or not having a choice to work with cheaper models means you work slower and more methodically because you have to plan out the work and size it to fit. This has huge implications on the value of the work you can produce as an employee or solo developer, and it’s a skill working with these SOTA models - there is specific crafting that must be done to be effective with each and their capabilities.
You will see the old eyeball real estate game, forcing more of your experience into their apps and platform so you can see ads or “new slop features”. You will see the SOTA models more closely guarded in an effort to bait you into that platform, and yes, you will see them doing things they never should be doing, moving towards an open AI like model where it can order a pizza, where monetization it was always important, but is now the golden rule, before all else.
I hope open models can at least give me opus 4.8 experience soon, I can work with that. But nothing out there today can do it and that terrifies me.
I hope open models can at least give me opus 4.8 experience soon, I can work with that. But nothing out there today can do it and that terrifies me.
You wouldn't be able to tell you weren't running Opus 4.8 if you sat down at my keyboard with DS4.1F 1017 running on the other end. Unless you asked it to do something that would get you yelled at by Claude, of course. Then you might notice the difference.
Anyone who can afford a decent car can afford their own personal Opus. Fable/Carbon/Astra, no, not yet. Opus, yes.
Enshittification has been already running full speed even before the IPO via guardrails and attempts at vendor lock-in. Now they will just push it to 110%. It's good that we have other options.
> Anthropic and OpenAI will have to become more of the end user application stack. Files, calendars, emails, photos, and much more have to pass through them for them to keep making money. Else they cannot make money like the way they have been in this early era of LLMs.
This is above my pay grade but since LLMs have become the 24/7 code writers, I started to wonder if our processing architecture could see much more innovation. I see light-enabled processing, analog processing and other headlines on posts/YouTube. Most are focused on inference though. But there is also the need to run generated code on cheaper, efficient chips.
I guess these are all really expensive and need VC funding. But perhaps there are existing architectures that allow baking code into silicon, or new architectures for NNs and so on. If costs are low, there is a massive demand coming for code that we want to write and forget - keeps on running on low powered last-mile devices. From factories, farms to home equipment.
I really do not understand the value of running large models locally. I am absolutely fond of local models but for specific tasks. The agents need to do a lot of work in this case: orchestration, big or small model split of planning and work, split into small tasks, merging, etc.
LLMs with <30B parameter models seem like the sweet spot. I would rather focus on multiple tiny LLMs that are tuned to specific tasks (SQL queries, finance/accounting, backend/frontend dev, etc.).
For larger models, a $10-40 ish Neuralwatt/Opencode Go or any similar subscription goes so far for the average user that it is ridiculous. We are still in the "youth" era of agents. There are many patterns that work well with smaller LLMs. Or with a big-small LLM combination.
My context:
I own the tiny 16GB M4 Mac Mini and I use LLMs (coding agents) to build software and a few of my products use LLMs directly. Full-time user of LLMs, do not read or write code manually. I run the tiny Qwen series (<10B params) sometimes but a couple $20 subscriptions (Claude, Neuralwatt) take care of my full-time agent usage.
I can imagine a future where we look back at these times of relying on big opaque hosted services to run inference and find it quaint. Perhaps the same way nowadays we find it quaint that you couldn't always just compile software on a fanless portable computer.
How far away that is, or whether it will ever happen, I don't know. I agree though that the tooling is still so young it's hard to say what shape this type of computing will look like and therefore what kinds of resources will be in-demand.
Same context as yourself. I run 8 - 12Bish models locally on a 24GB M4 Macbook Air and also have a couple subscriptions (OpenCode Go, Mistral Pro) for access to DeepSeek etc. to power through yucky big bloated codebases at work.
True that. But ZDR is also not that expensive. Although for many businesses, buying a Mac with tons of memory is not expensive. Running a dedicated agent on top of private data is certainly valuable.
ZDR is just a "bug" or "rogue AI agent" away from being all data retention. ZDR is not for privacy but for corporations that need to check a compliance box.
Yeah. I for example use local inference for tagging up my documents that I self-host in a local Paperless instance, and I don't really feel like sending my medical invoices to clammy Sammy.
Model/harness that runs locally, which you can connect from your phone to do non-dev queries (looking stuff up, recommendations, planning and etc.) can save ~$300/year on various subscriptions.
Agreed but I feel we will see even cheaper annual or prepaid options. Say, a $20 upfront and get the cheapest inference, comparable to Sonnet with subscription. For a personal assistant that can run on laptop and phone, these will be mainstream. $20 may last a couple months for vast majority of users.
I have the same Mac M4 16gb. Do you have the Qwen model assist the larger models for menial tasks/tool calls during coding it can handle? As in have them work in tandem to save on tokens/calls on the large models or are your local agents just meant for their own specific side tasks?
A mix of both. I have been experimenting building agents that focus on small/tiny LLMs. For example, I built a web-app prototyping harness which used Qwen 4/9B models but separate agents (tool calls, prompts and actual code to parse) for CSS management, TSX management, Router, state, etc. It used SolidJS only.
There was a task manager which was given context of the tech stack so it could split the incoming requests. Then branch out the tasks to appropriate agents. Combine the generated code - this was a file/module/import manager.
The other way is let tiny models (again I would use agents for different use cases) do the busy work - tool calling, finding evidence, etc. I usually setup agents to re-write tool call history and create fresh API requests without using multi-turn chat with tool call history.
I really appreciate the response, that gives me good ideas on where to start going about all this. It seems so silly to have API models doing so much file exploration and tool calling.
I saw this a while ago which shows small models and focuses on the ones that make the least (or no) wrong tool calls and it got me intrigued on the possibilities in this realm, especially for lower end devices. https://github.com/MikeVeerman/tool-calling-benchmark
The addressable market is enthusiasts. The cheapest and least loyal consumer that exists. But I guess it is just a small side project for apple which is fine.
Idk. Once they say our AI app is facilitating AI needs of big number of DAUs, the stock price will likely go up even if there’s no revenue path. It’s a pretty tulip.
When I hear Apple and AI, I think of Siri. And then of my attempts to use it in my smart home where it would turn off all smart sockets instead of turning on the lights in one room. Or its response latency. Or how I was excited about Apple Intelligence and then removed it completely and am hesitant to upgrade to OS 27.
I understand that we’re talking about local inference with an Apple inference engine for Apple Silicon uniform memory. And I know it’s a stretch of an analogy. But somehow I’ve grown skeptical of Apple’s AI capabilities. Not to say Microsoft’s or Google’s weren’t disappointing.
For me Siri hasn’t improved a bit since release and it’s just the fact that she can’t understand my voice/words. Constantly mishears or asks me to repeat things. I have a pretty “regular” or “generic” US English accent and other voice agents have no problem so it’s definitely Siri. So I have never really been able to ascertain if it’s any good without an acceptable interface. I bounce immediately when I encounter the friction of having to yell, try to enunciate differently, think how to word my statements, etc. It’s like ordering food at a drive thru that has a broken mic and speaker and the employee is distracted not even trying to pay attention to my order. All to say, the spirit of your analogy seems appropriate to me.
subagents use mostly the same allocation of memory, so you can do orchestration and planning easily with many subagents just like you do with Claude Code with anthropic's cloud models
Pi has a subagents package to enable in its harness
The idea that agent can built and see it is very important and how I have been building my own GUI framework https://github.com/brainless/akar
I have been using an internal screenshot feature along with a dump of the component tree. Multi-modal LLMs can be used to read the screenshot and fix issues. The dump can be read by any text based LLM. I made my life much easier.
I saw the Blender UI based framework shared on HN the other day and this one today. I am not sure if I want to continue building my own (although all code is written by agents) or if I should switch. I can take inspiration from the two other frameworks and continue with my own.
LLMs will eat software. Much, not all, of software is taking ideas into a strict language, compile to a stricter language and check if the processing unit will run it. There are a ton of algorithms and many can be combined with many others.
A token prediction system that has built a multi-dimensional model of all tokens on the Internet should be good at generating a bunch of algorithm combinations. Then it is about a loop to test, find out other combinations, find a winner.
I don't think they are 'just a tool', just like drugs aren't just a tool but a method to subjugate. Sometimes, with tools of sufficient complexity, the intention of the tool cannot be separated from the tool itself so much so that the tool and the intention and human values behind the tool cannot be separated.
reply