Gemma4 in my view is good enough to do things similar to Gemini 2.5 flash, meaning if I point it code and ask for help and there is a problem with the code it’ll answer correctly in terms of suggestions but it’s not great at using all tools or one shooting things that require a lot of context or “expert knowledge”
If a couple more iterations of this, say gemma6 is as good as current opus and runs completely locally on a Mac, I won’t really bother with the cloud models.
Yep, and to be honest we don't really need local models for intensive tasks. At least yet. You can use openrouter (and others) to consume a wide variety of open models which are capable of using tools in an agentic workflow, close to the SOTA models, which are essentially commodities - many providers, each serving the same model and competing with each-other on uptime, throughput, and price. At some point we will be able to run them on commodity hardware, but for now the fact that we can have competition between providers is enough to ensure that rug pulls aren't possible.
Plus having Gemma on my device for general chat ensures I will always have a privacy respecting offline oracle which fulfils all of the non-programming tasks I could ever want. We are already at the point where the moat for these hyper scalers has basically dissolved for the general public's use case.
If I was OpenAI or Anthropic I would be shitting my pants right now and trying every unethical dark pattern in the book to lock in my customers. And they are trying hard. It won't work. And I won't shed a single tear for them.
I agree. At first I was really turned off by the Gemma 4 line of models because they didn’t function with coding agents as well as the qwen3.5 line of models. However, I found that for other use cases Gemma 4 was very good.
EDIT: I just saw this: “”Ollama 0.20.6 is here with improved Gemma 4 tool calling!”” I will rerun my tests after breakfast.
Even Gemma 4 E2B is more useful than you'd think if you give it the right harness. I've been running it on Android via llama.rn and it handles function calling natively — the model outputs structured tool calls without any prompt engineering. Won't replace Opus for hard reasoning but for a mobile app that needs to pick a tool and run it, the cost math is hard to argue with. $0/query forever.
I think the difference is that with LLMs, in a lot of cases you do see some diminishing returns.
I won't deny that the latest Claude models are fantastic at just one shotting loads of problems. But we have an internal proxy to a load of models running on Vertex AI and I accidentally started using Opus/Sonnet 4 instead of 4.6. I genuinely didn't know until I checked my configuration.
AI models will get to this point where for 99% of problems, something like Gemma is gonna work great for people. Pair it up with an agentic harness on the device that lets it open apps and click buttons and we're done.
I still can't fathom that we're in 2026 in the AI boom and I still can't ask Gemini to turn shuffle mode on in Spotify. I don't think model intelligence is as much of an issue as people think it is.
100% agree here. The actual practical bottleneck is harness and agentic abilities for most tasks.
It's the biggest thing that stuck out to me using local AI with open source projects vs Claude's client. The model itself is good enough I think - Gemma 4 would be fine if it could be used with something as capable as Claude.
And that's gonna stay locked down unfortunately especially on mobile and cars - it needs access to APIs to do that stuff - and not just regular APIs that were built for traditional invoking.
The same way that websites are getting llm.txts I think APIs will also evolve.
GPT 3.5 was intelligent enough to understand that command and turn it into a correct shaped JSON object: the platforms don't have tight enough integration to take advantage of the intelligence
I think security is the issue-ai is good at circumventing this. For example , ai can read paywalled articles you cannot. Do you really want ai to have ‘free range’.?
I mean to me even difference between Opus and Sonnet is as clear as day and night, and even Opus and the best GPT model. Opus 4.6 just seems much more reliable in terms of me asking it to do something, and that to actually happen.
It depends what you're asking it though. Sure, in a software development environment the difference between those two models is noticeable.
But think about the general user. They're using the free Gemini or ChatGPT. They're not using the latest and greatest. And they're happy using it.
And I am willing to bet that a lot of paying users would be served perfectly fine by the free models.
If a capable model is able to live on device and solve 99% of people's problems, then why would the average person ever need to pay for ChatGPT or Gemini?
But even other tasks, like research etc, where dates are important, little details and connections are important, reasoning is important, background research activities or usage of tools outside of software development, and this is where I am finding much of the LLMs most useful for my life.
Even Opus makes mistakes with dates or not understanding news and everything correctly in context with chronological orders etc, and it would be even worse with smaller and less performing models.
My experience is very different than yours. Codex and CC yield very differenty result both because of the harness differencess and the model differences, but niether is noticeably better than the other.
Personally, I like Codex better just because I don't have to mess with any sort of planning mode. If I imply that it shouldn't change code yet, it doesn't. CC is too impatient to get started.
Look at the whole history of computing. How many times has the pendulum swung from thin to fat clients and back?
I don't think it's even mildly controversial to say that there will be an inflection point where local models get Good Enough and this iteration of the pendulum shall swing to fat clients again.
Assuming improvements in LLMs follow a sigmoid curve, even if the cloud models are always slightly ahead in terms of raw performance it won't make much of a difference to most people, most of the time.
The local models have their own advantages (privacy, no -as-a-service model) that, for many people and orgs, will offset a small performance advantage. And, of course, you can always fall back on the cloud models should you hit something particularly chewy.
(All IMO - we're all just guessing. For example, good marketing or an as-yet-undiscovered network effect of cloud LLMs might distort this landscape).
My thinkpad is nearly 10 years old, I upgraded it to 32GB of ram and have replaced the battery a couple of times, but it's absolutely fine apart from that.
If AI which was leading edge in 2023 can run on a 2026 laptop, then presumably AI which is leading edge in 2026 will run on a 2029 laptop. Given that 2023 was world changing then that capacity is now on today's laptop
Either AI grows exponentially in which case it doesn't matter as all work will be done by AI by 2035, or it plateaus in say 2032 in which case by 2035 those models will run on a typical laptop.
Local models seem somewhere between 9 and 24 months behind. I'm not saying I won't be impressed with what online models will be able to do in two years, but I'm pretty satisfied with the prediction that I won't really need them in a couple of years.
> If a couple more iterations of this, say gemma6 is as good as current opus and runs completely locally on a Mac, I won’t really bother with the cloud models.
> That’s a problem.
While improvements should continue rolling in, and might even match current SOTA in benchmarks down the line, is it "good enough"?
Hard to believe we have reached that stage with current models, which would continue to stretch beyond what we can economically run. Call it skill issue, or try to fix it with a revolutionary harness, it seemingly takes a village to get it all working. Maybe by then we will have good enough ecosystem in these layers too, but if current capabilities is the benchmark, it might need more time in the oven.
> if I point it code and ask for help and there is a problem with the code it’ll answer correctly in terms of suggestions
could I ask how you do that? I installed openclaw and set it to use Gemma 4 but it didn't act in an agent mode at all, it only responded in the chat window while doing nothing, and didn't read any files or do anything that you wrote (though I see you do mention that it's not great at using all tools). What are you using exactly?
But that difference atm is the difference between it being OK on its own with a team of subagents given good enough feedback / review mechanisms or having to babysit it prompt by prompt.
By the time gemma6 allows you to do the above the proprietary models supposedly will already be on the next step change. It just depends if you need to ride the bleeding edge but specially because it's "intelligence", there's an obvious advantage in using the best version and it's easy to hype it up and generate fomo.
When that happens, you'll have fomo from not using opus 5.x. The numbers that they showed for Mythos show that the frontier is still steadily moving (and maybe even at a faster pace than before)
There is a cognitive ceiling for what you can do with smaller models. Animals with simpler neural pathways often outperform whatever think they are capable of but there's no substitute for scale. I don't think you'll ever get a 4B or 8B model equivalent to Opus 4.6. Maybe just for coding tasks but certainly not Opus' breadth.
This is the classic apple approach - wait to understand what the thing is capable of doing (aka let others make sunk investments), envision a solution that is way better than the competition and then architect a path to building a leapfrog product that builds a large lead.
People can correct me if I'm wrong, but I think the core logic behind OpenAI's valuation was essentially that AI would work like search. Google had the best search engine, it became a centre of gravity that sucked everything in and suddenly network effects meant it was the centre of the universe. There seem to be 2 big problems with that though. The first is that for search, queries are both demand for the product and a way of making the product better. The second, is that Google was genuinely the best product for a very long time.
Maybe point (1) was unclear at some point, but I think it's mostly clear today that's not happening. Training the model is modestly distinct from inference.
Point (2) is really funny - because sure, at some point OpenAI was the best, and then Sam Altman blew the place up and spawned a whole host of competitors who could replicate and eventually surpass OpenAI's state of the art.
It now looks like AI is a death march. You must spend billions of dollars to have the best model or you won't be able to sell inference. But even if you do, a whole host of better funded competitors are going to beat you within months so your inference charges better pay off extremely quickly. When the gap between models starts to drop, distribution becomes king and OpenAI can't compete in that field either.
Google can do that. Meta can do that. MSFT probably can do that. Amazon can do that. OpenAI cannot. They do not have the cash to do it.
Apple aren’t in the business of building chatbots to impress investors (other than some WWDC2024 vaporware they’d rather not talk about any more). They’re in the business of consumer hardware.
Consumers want iPhones and (if Apple are right) some form of AR glasses in the next decade. That’s their focus. There’s a huge amount of machine learning and inference that’s required to get those to work. But it’s under the hood and computed locally. Hence their chips. I don’t see what Apple have to gain by building a competitor to what OpenAI has to offer.
What I don't get about Apple is when everyone else was giving up on yet another VR attempt, moving into AI, they decide AI isn't worth it, and it was the right time for a me too VR headset.
So no VR, given the price and lack of developer support, and late arrival into AI.
Nvidia restricts gamer cards in data centers through licensing, eventually they will probably release a cheaper consumer AI card to corner the local AI market that can't be used in data centers if they feel too much of a threat from Apple.
Imagine a future where Nvidia sells the exact same product at completely different prices, cheap for those using local models, and expensive for those deploying proprietary models in data centers.
The best part is that it’ll all run on your device, instead of siphoning off your data to the provider. Local first AI.
I think the creatives will also turn around their seething hatred of AI for Apple AI because they use more ethical training data and it feels more like they own their AI, no one’s charging them a subscription fee to use it and then using their private data for training.
Thing is, Apple never considered racing against LLM runners. Apple's success comes from human-centered design, it is not trying to launch a me-too product just because it increases their stock price.
iPod was not the first MP3 player.
iPhone was not even 3G at launch -- in the middle of 3G marketing craze.
They sure got lucky that unified memory is well-suited for running AI, but they just focused on having cost- and energy-efficient computing power. They've been having glasses in sight for the last 10 years (when was Magic Leap's first product?) and these chips have been developed with that in mind. But not only the chips: nothing was forcing Apple to spend the extra money for blazing fast SSD -- but they did.
So yes, Apple is a hardware company. All the services it sells run on their hardware. They've just designed their hardware to support their users' workflows, ignoring distractions.
With that said, LLM makes the GPU + memory bandwidth fun again. NVidia can't do it alone, Intel can't do it alone, but Apple positioned itself for it. It reminds me how everyone was surprised when then introduced 64-bit ARM for everyone: very few people understood what they were doing.
Tbh there are NVidia GPUs that beat Apple perf 2x or 3x, but these are desktop or server chips consuming 10x the power. Now all Apple needs to do is keep delivering performance out of Apple Silicon at good prices and best energy efficiency. Local LLM make sense when you need it immediately, anywhere, privately -- hence you need energy efficiency.
I think introducing the MacBook Neo now, at that price point, was a genius move. While they're playing the waiting game on AI, they're cementing the next generation into the Apple ecosystem and getting them to not sway towards whatever device(s) OpenAI is cooking up.
The MacBook Neo feels like the iPod of this generation.
> Pure strategy, luck, or a bit of both? I keep going back and forth on this, honestly, and I still don’t know if this was Apple’s strategy all along, or they didn’t feel in the position to make a bet and are just flowing as the events unfold maximising their optionality.
Maximizing the available options is in fact a "strategy", and often a winning one when it comes to technology. I would love to be reminded of a list of tech innovators who were first and still the best.
I wouldn't characterize Apple's AI strategy as "smart". "Accidental" is a perfect descriptor here. "Apple Intelligence" and "Liquid Glass" show they are asleep at the wheel. I wrote an email to Tim last year imploring him to leverage Apple Silicon and its unified memory for private AI. I didn't tell him that I had dumped 95% of my Apple shares.
Apple is almost 2 years out from their announcement of Apple Intelligence. It has barely delivered on any of the hype. New Siri was delayed and barely mentioned in the last WWDC; none of the features are released in China.
In other news, people keep buying iPhones, and Apple just had its best quarter ever in China. AAPL is up 24% from last year.
When using Siri recently it really struck me how much worse it feels after using ChatGPT. It struggles to understand what I say correctly and you have to give commands in more of a 'computer-friendly' form.
I hope they can at least fix this, as I really only use it as a hands-free system while driving.
Honestly, I think part of the reason Apple hasn't jumped deep into AI is due to two big reasons:
1) Apple is not a data company.
2) Apple hasn't found a compelling, intuitive, and most of all, consistent, user experience for AI yet.
Regarding point 2: I haven't seen anyone share a hands down improved UX for a user driven product outside of something that is a variation of a chat bot. Even the main AI players can't advertise anything more than, "have AI plan your vacation".
I’m confused why he keeps calling out “the Mac Mini craze after claw went viral”. I thought the various versions of claw used remote models, not local models, and I thought the point of using a Mac mini was that it can send and receive iMessages, not anything about the hardware.
This seems mistaken to me. The core idea is that LLMs are commoditizing and that the UI (Siri in this case) is what users will stick with.
But... what's the argument that the bulk of "AI value" in the coming decade is going to be... Siri Queries?! That seems ridiculous on its face.
You don't code with Siri, you don't coordinate automated workforces with Siri, you don't use Siri to replace your customer service department, you don't use Siri to build your documentation collation system. You don't implement your auto-kill weaponry system in Siri. And Siri isn't going to be the face of SkyNet and the death of human society.
Siri is what you use to get your iPhone to do random stuff. And it's great. But ... the world is a whole lot bigger than that.
Any field with abstraction becomes susceptible to ai disruption. In fact, ai susceptibility is proportional to the amount of abstraction. In this sense, the more abstraction then the more ai will displace people (my observation). This turns the millenia old model upside down. Traditionally more abstraction required more schooling and experience and was rewarded with more financial rewards. Until robots and world models become safe, affordable and ubiquitous, the financial apex of careers will be those that are abstraction resistant (technicians, emts, trades, etc) and those protected by requlation and the requlators(politicians, ceos)
Why is Nvidia so central to LLMs? Because they embraced ML a decade ago. Apple did as well, machine learning is central to so many things in the iPhone. Its not so surprising then, that a strong showing in ML sets you up good for LLMs..
“The early bird might get the worm, but it’s the second mouse that gets the cheese.”
Apple is hanging back, waiting for mousetraps to be triggered as AI companies make mistakes that could not have been reliably foreseen. Then it’ll swoop in, adopt the best bits, and put out a product that is immensely polished and easy to use.
That has been, after all, one of its most important strategies over the years. They realized that early adopters only became industry leaders if their error rate remained low enough to keep ahead of those who let others make mistakes for them.
the core thesis of the article and comparison with non-"losers" seem vague at best. comparing model providers with hardware vendor with software ecosystem is like differentiating between apple to oranges.
if hardware moat was to be discussed, then compare with nvidia, amd and google's tpu division perhaps. in-house intelligence is best left alone for apple. they are relying on the "peers" for underlying capabilities as is. [1] [2]
outside of inference and (pro/con)sumer space, there is little to offer for the enterprise or the people developing the lowest end of the stack. even the recent tinygrad egpu is shockingly slow [3]. which might made gb10 look much more capable for in-house training.
regardless, most of the industry "moat" does not appear sustainable at best. only time will tell how it will turn out for everyone but on a positive note, apple does not put all its eggs in this basket, which is probably wiser.
there are always three elements in the equations of business model:
1. marginal cost
2. marginal revenue
3. value created
for llm providers, i always believe the key is to focus on high value problems such as coding or knowledge work, becaues of the high marginal cost of having new customers - the token burnt. and low marginal revenue if the problem is not valuable enough. in this sense no llm providers can scale like previous social media platforms without taking huge losses. and no meaning user stickiness can be built unless you have users' data. and there is no meaningful business model unless people are willing to pay a high price for the problem you solve, in the same way as paying for a saas.
i am really not optimistic about the llm providers other than anthropic. it seems that the rest are just burning money, and for what? there is no clear path for monetization.
and when the local llm is powerful enough, they will soon be obsolete for the cost, and the unsustainable business model. in the end of the day, i do agree that it is the consumer hardware provider that can win this game.
> Think about the App Store. Apple didn’t build the apps, they built the platform where apps ran best, and the ecosystem followed.
As far as I remember Apple basically got forced into opening the platform to 3rd party developers. Not by regulation but by public pressure. It wasn't their initial intention to allow it.
Looks like Apple fell into a winning/winnable position in the AI wars. Their privacy/safety first culture is the cause of them not embracing AI as effectively as other more maverick styles. Their AI was always hindered by privacy, and local first AI is their savour.
Apple is a marketing company now, not an innovative tech company. They will wait until model progress is not moving in such quick progressive strides then create a new flagship product by rebadging [top model] to match their aesthetic and increase the price 3x.
It's the same everywhere: great fundamentals pay off. It's true of martial arts, dance, and absolutely about software platforms. You just have to trust that process and invest in it, which Apple does (although frustratingly not enough!).
But why do I feel like the quality of the software from Apple declined sharply in recent years? The liquid glass design feels very unpolished and not well thought out throughout almost everywhere… seems like even Apple can’t resist falling victim to AI slop
The whole premise is that if you don't get to AGI first then you loose.
The idea is that Anthropic with AGI could build a better version of Apple, or whatever it wants.
This was the conversation like 1 year ago. What has changed?
> Then Stargate Texas was cancelled, OpenAI and Oracle couldn’t agree terms, and the demand that had justified Micron’s entire strategic pivot simply vanished. Micron’s stock crashed.
Well.. no. The Stargate expansion was cancelled the orginally planned 1.2MW (!) datacenter is going ahead:
> The main site is located in Abilene, Texas, where an initial expansion phase with a capacity of 1.2 GW is being built on a campus spanning over 1,000 acres (approximately 400 hectares). Construction costs for this phase amount to around $15 billion. While two buildings have already been completed and put into operation, work is underway on further construction phases, the so-called Longhorn and Hamby sections. Satellite data confirms active construction activity, and completion of the last planned building is projected to take until 2029.
> The Stargate story, however, is also a story of fading ambitions. In March 2026, Bloomberg reported that Oracle and OpenAI had abandoned their original expansion plans for the Abilene campus. Instead of expanding to 2 GW, they would stick with the planned 1.2 GW for this location. OpenAI stated that it preferred to build the additional capacity at other locations. Microsoft then took over the planning of two additional AI factory buildings in the immediate vicinity of the OpenAI campus, which the data center provider Crusoe will build for Microsoft. This effectively creates two adjacent AI megacampus locations in Abilene, sharing an industrial infrastructure. The original partnership dynamics between OpenAI and SoftBank proved problematic: media reports described disagreements over site selection and energy sources as points of contention.
I think the article is missing a whole aspect on how Apple is ensuring to not face actual competition while they're "playing it safe":
Even if the investment is overblown, there is market-demand for the services offered in the AI-industry. In a competitive playing field with equal opportunities, Apple would be affected by not participating. But they are establishing again their digital market concept, where they hinder a level playing field for Apple users.
Like they did with the Appstore (where Apple is owning the marketplace but also competes in it) they are setting themselves up as the "the bakn always wins" gatekeeper in the Apple ecosystem for AI services, by making "Apple Intelligence" an ecosystem orchestration layer (and thus themselves the gatekeeper).
1. They made a deal with OpenAI to close Apple's competitive gap on consumer AI, allowing users to upgrade to paid ChatGPT subscriptions from within the iOS menu. OpenAI has to pay at least (!) the usual revenue share for this, but considering that Apple integrated them directly into iOS I'm sure OpenAI has to pay MORE than that. (also supported by the fact that OpenAI doesn't allow users to upgrade to the 200USD PRO tier using this path, but only the 20USD Plus tier) [1]
2. Apple's integration is set up to collect data from this AI digital market they created: Their legal text for the initial release with OpenAI already states that all requests sent to ChatGPT are first evaluated by "Apple Intelligence & Siri" and "your request is analyzed to determine whether ChatGPT might have useful results" [2]. This architecture requires(!) them to not only collect and analyze data about the type of requests, but also gives them first-right-to-refuse for all tasks.
3. Developers are "encouraged" to integrate Apple Intelligence right into their apps [3]. This will have AI-tasks first evaluated by Apple
4. Apple has confirmed that they are interested to enable other AI-providers using the same path [4]
--> Apple will be the gatekeeper to decide whether they can fulfill a task by themselves or offer the user to hand it off to a 3rd party service provider.
--> Apple will be in control of the "Neural Engine" on the device, and I expect them to use it to run inference models they created based on statistics of step#2 above
--> I expect that AI orchestration, including training those models and distributing/maintaining them on the devices will be a significant part of Apple's AI strategy. This could cover alot of text and image processing and already significantly reduce their datacenter cost for cloud-based AI-services. For the remaining, more compute-intensive AI-services they will be able to closely monitor (via above step#2) when it will be most economic to in-source a service instead of "just" getting revenue-share for it (via above step#1).
So the juggernaut Apple is making sure to get the reward from those taking the risk. I don't see the US doing much about this anti-competitive practice so far, but at least in the EU this strategy has been identified and is being scrutinized.
For the love of all that's holy - folks please stop using AI to publish smart sounding texts. While you may think you are "polishing" your text, you are just disrespecting your readers. Write in your own words.
384 comments
If a couple more iterations of this, say gemma6 is as good as current opus and runs completely locally on a Mac, I won’t really bother with the cloud models.
That’s a problem.
For the others anyway.
Plus having Gemma on my device for general chat ensures I will always have a privacy respecting offline oracle which fulfils all of the non-programming tasks I could ever want. We are already at the point where the moat for these hyper scalers has basically dissolved for the general public's use case.
If I was OpenAI or Anthropic I would be shitting my pants right now and trying every unethical dark pattern in the book to lock in my customers. And they are trying hard. It won't work. And I won't shed a single tear for them.
EDIT: I just saw this: “”Ollama 0.20.6 is here with improved Gemma 4 tool calling!”” I will rerun my tests after breakfast.
I won't deny that the latest Claude models are fantastic at just one shotting loads of problems. But we have an internal proxy to a load of models running on Vertex AI and I accidentally started using Opus/Sonnet 4 instead of 4.6. I genuinely didn't know until I checked my configuration.
AI models will get to this point where for 99% of problems, something like Gemma is gonna work great for people. Pair it up with an agentic harness on the device that lets it open apps and click buttons and we're done.
I still can't fathom that we're in 2026 in the AI boom and I still can't ask Gemini to turn shuffle mode on in Spotify. I don't think model intelligence is as much of an issue as people think it is.
It's the biggest thing that stuck out to me using local AI with open source projects vs Claude's client. The model itself is good enough I think - Gemma 4 would be fine if it could be used with something as capable as Claude.
And that's gonna stay locked down unfortunately especially on mobile and cars - it needs access to APIs to do that stuff - and not just regular APIs that were built for traditional invoking.
The same way that websites are getting llm.txts I think APIs will also evolve.
But think about the general user. They're using the free Gemini or ChatGPT. They're not using the latest and greatest. And they're happy using it.
And I am willing to bet that a lot of paying users would be served perfectly fine by the free models.
If a capable model is able to live on device and solve 99% of people's problems, then why would the average person ever need to pay for ChatGPT or Gemini?
Even Opus makes mistakes with dates or not understanding news and everything correctly in context with chronological orders etc, and it would be even worse with smaller and less performing models.
Scheduling, planning, researching products, shopping, trip plans, etc...
My experience is very different than yours. Codex and CC yield very differenty result both because of the harness differencess and the model differences, but niether is noticeably better than the other.
Personally, I like Codex better just because I don't have to mess with any sort of planning mode. If I imply that it shouldn't change code yet, it doesn't. CC is too impatient to get started.
The world has moved on, that code-golf time is now spent on ad algorithms or whatever.
Escaping the constraint delivered a different future than anticipated.
I don't think it's even mildly controversial to say that there will be an inflection point where local models get Good Enough and this iteration of the pendulum shall swing to fat clients again.
The local models have their own advantages (privacy, no -as-a-service model) that, for many people and orgs, will offset a small performance advantage. And, of course, you can always fall back on the cloud models should you hit something particularly chewy.
(All IMO - we're all just guessing. For example, good marketing or an as-yet-undiscovered network effect of cloud LLMs might distort this landscape).
My thinkpad is nearly 10 years old, I upgraded it to 32GB of ram and have replaced the battery a couple of times, but it's absolutely fine apart from that.
If AI which was leading edge in 2023 can run on a 2026 laptop, then presumably AI which is leading edge in 2026 will run on a 2029 laptop. Given that 2023 was world changing then that capacity is now on today's laptop
Either AI grows exponentially in which case it doesn't matter as all work will be done by AI by 2035, or it plateaus in say 2032 in which case by 2035 those models will run on a typical laptop.
> If a couple more iterations of this, say gemma6 is as good as current opus and runs completely locally on a Mac, I won’t really bother with the cloud models.
> That’s a problem.
While improvements should continue rolling in, and might even match current SOTA in benchmarks down the line, is it "good enough"?
Hard to believe we have reached that stage with current models, which would continue to stretch beyond what we can economically run. Call it skill issue, or try to fix it with a revolutionary harness, it seemingly takes a village to get it all working. Maybe by then we will have good enough ecosystem in these layers too, but if current capabilities is the benchmark, it might need more time in the oven.
If someone gets a really great axe and are happy with it, that’s great for them.
But then, other people will be on bulldozers.
They can say they are happy with the axe, but then they are not in the competition at that point.
> it’s not great at using all tools
Glad it wasnt just me - i was impressed with the quality of Gemma4 - it just couldnt write the changes to file 9/10 times when using it with opencode
> if I point it code and ask for help and there is a problem with the code it’ll answer correctly in terms of suggestions
could I ask how you do that? I installed openclaw and set it to use Gemma 4 but it didn't act in an agent mode at all, it only responded in the chat window while doing nothing, and didn't read any files or do anything that you wrote (though I see you do mention that it's not great at using all tools). What are you using exactly?
Looking at current advancements - this is the horse I would bet my money on.
By the time gemma6 allows you to do the above the proprietary models supposedly will already be on the next step change. It just depends if you need to ride the bleeding edge but specially because it's "intelligence", there's an obvious advantage in using the best version and it's easy to hype it up and generate fomo.
Maybe point (1) was unclear at some point, but I think it's mostly clear today that's not happening. Training the model is modestly distinct from inference.
Point (2) is really funny - because sure, at some point OpenAI was the best, and then Sam Altman blew the place up and spawned a whole host of competitors who could replicate and eventually surpass OpenAI's state of the art.
It now looks like AI is a death march. You must spend billions of dollars to have the best model or you won't be able to sell inference. But even if you do, a whole host of better funded competitors are going to beat you within months so your inference charges better pay off extremely quickly. When the gap between models starts to drop, distribution becomes king and OpenAI can't compete in that field either.
Google can do that. Meta can do that. MSFT probably can do that. Amazon can do that. OpenAI cannot. They do not have the cash to do it.
Consumers want iPhones and (if Apple are right) some form of AR glasses in the next decade. That’s their focus. There’s a huge amount of machine learning and inference that’s required to get those to work. But it’s under the hood and computed locally. Hence their chips. I don’t see what Apple have to gain by building a competitor to what OpenAI has to offer.
So no VR, given the price and lack of developer support, and late arrival into AI.
When I open up JIRA or Slack I am always greeted with multiple new dialogues pointing at some new AI bullshit, in comparison. We hates it precious
Imagine a future where Nvidia sells the exact same product at completely different prices, cheap for those using local models, and expensive for those deploying proprietary models in data centers.
I think the creatives will also turn around their seething hatred of AI for Apple AI because they use more ethical training data and it feels more like they own their AI, no one’s charging them a subscription fee to use it and then using their private data for training.
They sure got lucky that unified memory is well-suited for running AI, but they just focused on having cost- and energy-efficient computing power. They've been having glasses in sight for the last 10 years (when was Magic Leap's first product?) and these chips have been developed with that in mind. But not only the chips: nothing was forcing Apple to spend the extra money for blazing fast SSD -- but they did.
So yes, Apple is a hardware company. All the services it sells run on their hardware. They've just designed their hardware to support their users' workflows, ignoring distractions.
With that said, LLM makes the GPU + memory bandwidth fun again. NVidia can't do it alone, Intel can't do it alone, but Apple positioned itself for it. It reminds me how everyone was surprised when then introduced 64-bit ARM for everyone: very few people understood what they were doing.
Tbh there are NVidia GPUs that beat Apple perf 2x or 3x, but these are desktop or server chips consuming 10x the power. Now all Apple needs to do is keep delivering performance out of Apple Silicon at good prices and best energy efficiency. Local LLM make sense when you need it immediately, anywhere, privately -- hence you need energy efficiency.
The MacBook Neo feels like the iPod of this generation.
Here's to another 10 years of scuffed Metal Compute Shaders, I guess.
Unlike Apple, they have even more devices in the field PLUS they have strong models PLUS Apple uses Google models.
> Pure strategy, luck, or a bit of both? I keep going back and forth on this, honestly, and I still don’t know if this was Apple’s strategy all along, or they didn’t feel in the position to make a bet and are just flowing as the events unfold maximising their optionality.
Maximizing the available options is in fact a "strategy", and often a winning one when it comes to technology. I would love to be reminded of a list of tech innovators who were first and still the best.
Anyway, hasn't this always been Apple's strategy?
In other news, people keep buying iPhones, and Apple just had its best quarter ever in China. AAPL is up 24% from last year.
I hope they can at least fix this, as I really only use it as a hands-free system while driving.
1) Apple is not a data company.
2) Apple hasn't found a compelling, intuitive, and most of all, consistent, user experience for AI yet.
Regarding point 2: I haven't seen anyone share a hands down improved UX for a user driven product outside of something that is a variation of a chat bot. Even the main AI players can't advertise anything more than, "have AI plan your vacation".
But... what's the argument that the bulk of "AI value" in the coming decade is going to be... Siri Queries?! That seems ridiculous on its face.
You don't code with Siri, you don't coordinate automated workforces with Siri, you don't use Siri to replace your customer service department, you don't use Siri to build your documentation collation system. You don't implement your auto-kill weaponry system in Siri. And Siri isn't going to be the face of SkyNet and the death of human society.
Siri is what you use to get your iPhone to do random stuff. And it's great. But ... the world is a whole lot bigger than that.
“The early bird might get the worm, but it’s the second mouse that gets the cheese.”
Apple is hanging back, waiting for mousetraps to be triggered as AI companies make mistakes that could not have been reliably foreseen. Then it’ll swoop in, adopt the best bits, and put out a product that is immensely polished and easy to use.
That has been, after all, one of its most important strategies over the years. They realized that early adopters only became industry leaders if their error rate remained low enough to keep ahead of those who let others make mistakes for them.
if hardware moat was to be discussed, then compare with nvidia, amd and google's tpu division perhaps. in-house intelligence is best left alone for apple. they are relying on the "peers" for underlying capabilities as is. [1] [2]
outside of inference and (pro/con)sumer space, there is little to offer for the enterprise or the people developing the lowest end of the stack. even the recent tinygrad egpu is shockingly slow [3]. which might made gb10 look much more capable for in-house training.
regardless, most of the industry "moat" does not appear sustainable at best. only time will tell how it will turn out for everyone but on a positive note, apple does not put all its eggs in this basket, which is probably wiser.
[1] https://news.ycombinator.com/item?id=40636980
[2] https://news.ycombinator.com/item?id=46589675
[3] https://www.youtube.com/watch?v=C4KWsmezXm4
> I am actually of the opinion that without some kind of bailout, OpenAI could be bankrupt in the next 18-24 months, but I am horrible at predictions
I find this intriguing.. Does anyone here have enough insight to speculate more?
for llm providers, i always believe the key is to focus on high value problems such as coding or knowledge work, becaues of the high marginal cost of having new customers - the token burnt. and low marginal revenue if the problem is not valuable enough. in this sense no llm providers can scale like previous social media platforms without taking huge losses. and no meaning user stickiness can be built unless you have users' data. and there is no meaningful business model unless people are willing to pay a high price for the problem you solve, in the same way as paying for a saas.
i am really not optimistic about the llm providers other than anthropic. it seems that the rest are just burning money, and for what? there is no clear path for monetization.
and when the local llm is powerful enough, they will soon be obsolete for the cost, and the unsustainable business model. in the end of the day, i do agree that it is the consumer hardware provider that can win this game.
> Think about the App Store. Apple didn’t build the apps, they built the platform where apps ran best, and the ecosystem followed.
As far as I remember Apple basically got forced into opening the platform to 3rd party developers. Not by regulation but by public pressure. It wasn't their initial intention to allow it.
This was really unsurprising [0].
[0] https://news.ycombinator.com/item?id=40278371
This was the conversation like 1 year ago. What has changed?
they wait until the dust settles before making their well-thought-out moves.
Every time they’ve jumped the hype train too quickly it hasn’t worked out, like Siri for example.
> Then Stargate Texas was cancelled, OpenAI and Oracle couldn’t agree terms, and the demand that had justified Micron’s entire strategic pivot simply vanished. Micron’s stock crashed.
Well.. no. The Stargate expansion was cancelled the orginally planned 1.2MW (!) datacenter is going ahead:
> The main site is located in Abilene, Texas, where an initial expansion phase with a capacity of 1.2 GW is being built on a campus spanning over 1,000 acres (approximately 400 hectares). Construction costs for this phase amount to around $15 billion. While two buildings have already been completed and put into operation, work is underway on further construction phases, the so-called Longhorn and Hamby sections. Satellite data confirms active construction activity, and completion of the last planned building is projected to take until 2029.
> The Stargate story, however, is also a story of fading ambitions. In March 2026, Bloomberg reported that Oracle and OpenAI had abandoned their original expansion plans for the Abilene campus. Instead of expanding to 2 GW, they would stick with the planned 1.2 GW for this location. OpenAI stated that it preferred to build the additional capacity at other locations. Microsoft then took over the planning of two additional AI factory buildings in the immediate vicinity of the OpenAI campus, which the data center provider Crusoe will build for Microsoft. This effectively creates two adjacent AI megacampus locations in Abilene, sharing an industrial infrastructure. The original partnership dynamics between OpenAI and SoftBank proved problematic: media reports described disagreements over site selection and energy sources as points of contention.
https://xpert.digital/en/digitale-ruestungsspirale/
> Micron’s stock crashed. [the link included an image of dropping to $320]
Micron’s stock is back to $420 today
> One analysis found a max-plan subscriber consuming $27,000 worth of compute with their 200$ Max subscription.
Actually, no. They'd miscalculated and consumed $2700 worth of tokens.
The same place that checked that claim also points out:
> In fact, Anthropic’s own data suggests the average Claude Code developer uses about $6 per day in API-equivalent compute.
https://www.financialexpress.com/life/technology-why-is-clau...
I like Apple's chips, but why do we put up with crappy analysis like this?
Even if the investment is overblown, there is market-demand for the services offered in the AI-industry. In a competitive playing field with equal opportunities, Apple would be affected by not participating. But they are establishing again their digital market concept, where they hinder a level playing field for Apple users.
Like they did with the Appstore (where Apple is owning the marketplace but also competes in it) they are setting themselves up as the "the bakn always wins" gatekeeper in the Apple ecosystem for AI services, by making "Apple Intelligence" an ecosystem orchestration layer (and thus themselves the gatekeeper).
1. They made a deal with OpenAI to close Apple's competitive gap on consumer AI, allowing users to upgrade to paid ChatGPT subscriptions from within the iOS menu. OpenAI has to pay at least (!) the usual revenue share for this, but considering that Apple integrated them directly into iOS I'm sure OpenAI has to pay MORE than that. (also supported by the fact that OpenAI doesn't allow users to upgrade to the 200USD PRO tier using this path, but only the 20USD Plus tier) [1]
2. Apple's integration is set up to collect data from this AI digital market they created: Their legal text for the initial release with OpenAI already states that all requests sent to ChatGPT are first evaluated by "Apple Intelligence & Siri" and "your request is analyzed to determine whether ChatGPT might have useful results" [2]. This architecture requires(!) them to not only collect and analyze data about the type of requests, but also gives them first-right-to-refuse for all tasks.
3. Developers are "encouraged" to integrate Apple Intelligence right into their apps [3]. This will have AI-tasks first evaluated by Apple
4. Apple has confirmed that they are interested to enable other AI-providers using the same path [4]
--> Apple will be the gatekeeper to decide whether they can fulfill a task by themselves or offer the user to hand it off to a 3rd party service provider.
--> Apple will be in control of the "Neural Engine" on the device, and I expect them to use it to run inference models they created based on statistics of step#2 above
--> I expect that AI orchestration, including training those models and distributing/maintaining them on the devices will be a significant part of Apple's AI strategy. This could cover alot of text and image processing and already significantly reduce their datacenter cost for cloud-based AI-services. For the remaining, more compute-intensive AI-services they will be able to closely monitor (via above step#2) when it will be most economic to in-source a service instead of "just" getting revenue-share for it (via above step#1).
So the juggernaut Apple is making sure to get the reward from those taking the risk. I don't see the US doing much about this anti-competitive practice so far, but at least in the EU this strategy has been identified and is being scrutinized.
[1] https://help.openai.com/en/articles/7905739-chatgpt-ios-app-...
[2] https://www.apple.com/legal/privacy/data/en/chatgpt-extensio...
[3] https://developer.apple.com/apple-intelligence/
[4] https://9to5mac.com/2024/06/10/craig-federighi-says-apple-ho...