Microsoft and Google adding LLMs to their ecosystem makes sense. But I don't see Apple adding something similar to MS Copilot to Numbers, Pages, etc. In a way, Apple seems to signal lack of hope for getting into the AI game. Soon, Apple will just be a "commodity" producer[0]: the company that makes the machines in the background, while other pioneers actually provide cutting edge technologies for those machines (e.g., local LLMs). Something like the oil companies, chip industry, etc. Apple will be making the iPhones and Macs and that's about it. They're too conservative to get out of their comfort zone now.
In a weird way, it's actually still aligned with their "think different" motto: when everyone is thinking about getting on the AI bandwagon, think different and avoid doing that!
[0]: Which is fine—the world needs boring companies as much as pioneers who break things.
Their track record shows if they can't tightly control a technology, they won't touch it. LLMs can be bent in different ways and Apple wouldn't allow that. I own Apple devices, but I just don't see Apple play in the Google/MS/ClosedAI's field by saying "Our model is 20% larger than the best models out there...".
Also, Apple should first fix Siri, or get rid of it for good. It's a joke at this point.
I would argue that Siri is an example of a lightly controlled technology. My bet is that Apple will appear behind for 2-4 years and out of nowhere will launch a Spike Jonze's Her-style "Samantha" that was in R&D for years.
I don't think this is possible. There aren't that many AI researchers and they talk to each other. If Apple was making a serious AI push it would be obvious.
I feel like any company which has chatbots as a "core" part of their systems (like Siri is often advertised to be) would be looking very heavily into generative AI. It makes way too much sense and would clearly make Siri way more useful. Imagine if Siri was current ChatGPT with plugins for interacting with the OS -- this immediately solves basically all the complaints about Siri. Plus Apple has efficient local AI acceleration already, which could make it even better.
It feels so obvious it'd be weird if they weren't trying to make a better Siri with gen AI! I have a feeling we'll hear about it either at WWDC or the iPhone event in the fall.
The thing about these technologies is that you can't repeat the same recipe for iPhone. Even 6 months is too long a time for a company to stay behind. The rate of progress in this field, combined with the insane VC money driven up by the hype around LLMs, means that no company can afford "surprising" everyone by waiting just a bit longer. Google did that and see where they are now.
You're right, but we're talking about Apple. While they may not be able to introduce a ChatGPT competito, Siri basically has a captive audience. With OS-level integrations, you can easily imagine a much more powerful and interactive Siri which becomes popular just because it's available out of the box and offers OS level integrations which a 3rd party chatbot can't.
It's also not like AI tools are why people buy Apple products. So "waiting" probably doesn't cost them much money -- it doesn't impact their core hardware and service revenue. It's "just" a value add for now.
If MS offers Copilot for office only on Windows, I for one will have an incentive to switch from Mac. And if Bing AI gets integrated into Windows, I’ll most likely switch.
The wild bit about this is this: Imagine reading this comment 6 months ago. A person would probably spit their drink out..but it is true. Apple will spend WWDC talking about XR like it is a 2021 Facebook Earnings Call.
Even wilder bit: imagine reading any of this 6 months ago. ChatGPT just came out. It was a fun toy with some potential, but mostly seen as another example of the copilot-y "might make some jobs somewhat different in the mid-term future" tech demo.
Who would've thought that in just four months, GPT-4 will completely transform people's perception of AI progress timeline and impact.
Apple has always been a consumer product-focused company. Even if computing continues to become more invisible and omnipresent, even if we no longer use laptops and phones, there will still be some kind of physical "product" at the end of the day, and thats where Apple will continue to exist. For LLMs, I don't see Apple opening up a general-purpose LLM or search engine like ChatGPT or Bing, its difficult to differentiate themselves in a meaningful way. I can see them developing a unique "personality" to Siri and integrating its behavior into existing devices and APIs. There's still a lot of room between where we are now and a personal AI-buddy like in "Her", so I imagine Apple is at least interested in moving in that direction.
It's strange to me that you think Apple will become a "commodity" producer when it has the largest profit margins in the industry. Which is the exact opposite of how commodities function in economics. Yes Apple's focus is hardware as opposed to software, but there is commodity hardware and non-commodity hardware. Apple is firmly in the non-commodity camp and shows no signs of changing. Things like M2 chips are not "commodity".
I have no idea what Apple will do with LLM's. But Apple is already leveraging ML in things like built-in local dictation and OCR, and we can probably expect that it will leverage things like LLM's at the OS level (iOS/macOS) rather than focusing on their iWork suite. (Honestly at this point I don't understand Apple's strategic purpose with iWork, or even Final Cut Pro.)
They definitely come from the hardware, I'm talking about profits from the sale of hardware alone.
In April 2023, Apple has a 30% market share of smartphones worldwide [1]. And in the first 3 months of 2023, Apple accounted for 50% of global smartphone revenue... and 80 percent of the industry's profits [2]. These is smartphone revenue, not services revenue. (Apple makes money from the App Store too but that's separate.)
80% of industry profit from 30% market share is certainly not a commodity. Cheap Android phones are a commodity. iPhones are the polar opposite.
Huh? The reality is "cutting edge" doesn't necessarily mean it's a great product for consumers.
Also frankly, productionalizing LLMs is very different from making research strides on them. Even inside Google researchers are realizing that a bulk of the improvements are now coming from the OSS community.
Apple will implement these technologies when it makes sense to do so.
> Apple will be making the iPhones and Macs and that's about it. They're too conservative to get out of their comfort zone now.
Why would Apple even pursue LLMs? They are building hardware: computers, phones, screens, accessories. In a sense, that’s what you said. Just phrasing a question to your thought.
Considering the cost of running these LLMs behind the scenes, and the giant customer base of Office, Windows, Bing, Edge, etc. I wonder how Microsoft will keep things profitable. I guess they can just raise subscription price of Microsoft 365 at some point.
Microsoft has tons of cash and on-going profits to be a loss leader for a while, also I imagine they will eventually limit some features to Microsoft 365 users only. They already charge extra for Teams with Copilot, etc; that's not a standard Teams feature. They also announced a "copilot" stack today, which I imagine they will charge for API usage on top of what they are going to be making from ChatGPT.
With the operating costs of AI at this time, this is going to be the big question for everyone. One of these you can't afford not to have it, but your can't afford to have it either. With all this said, Microsoft/Google/Apple have been sucking up stupid profits for decades now and it really is time for their unlimited money machine to have some pushback.
What makes me most excited about this is that it uses LLMs in a way that the result (a computer program) is still understandable and editable (hackable).
Instead of having an AI model output some kind of "black-box conclusion" on a data set / question, it can output code that we can inspect, improve, and learn from. Using LLMs in this way still keeps us as users "in control" - and I think it's a promising path for responsible AI.
I've done some experiments[1] adding ChatGPT-like functionality to my open source Typescript Notebook environment[2] back in december, and hope to revisit this soon :)
I completely agree that the best LLM output is something that is still understandable/hackable -- and maybe even _needs_ further hacking. That's what I tell people when they say "I heard AI is replacing programmers."
Been using copilot with vim for a few weeks and my only issue has been when the file i'm editing is large the node process goes to like 100% and my typing is severely lagged... i quickly do :Copilot disable and move on with my day but really hoping the performance can improve, it's helpful when it works.
You may want to check out Codium[0]. I haven't used Copilot with Vim, but I've never experienced a CPU issue with Codium. Bonus is that it's free! (You would be right to be concerned about the something for nothing here. Personally I was satisfied by their blog post on the subject[1], but of course you need to decide what's right for you.)
For me the most awesome thing here is the 100MB file upload function. May sound like nothing, but this is groundbreaking when we've been so far constrained by token limits. This means we don't need to endure an arduous back-and-forth process of giving it tiny pieces of data, code, and increasingly creative prompts. It can now just injest an entire dataset or software project and make swathing changes of all manner. This is going to disrupt everything (and already is..).
I'm left wondering what the role of a software eng is and all I can come up with is a glorified quality analyst. Nothing wrong with that, but it's just such a transformation, and so sudden. I'm excited, scared, and exhausted all at the same time!
I could imagine that we will probably be checking generated code for correcteness while integrating larger parts of generated code. And due to the sheer complexity will need an en engineering understanding to make it all work. I’m excited for that, because it means I can do more with my time. But I also understand how that is not everyone’s cup of tea.
> Unless, of course, you want to be a software engineer.
> I could imagine that we will probably be checking generated code for correcteness while integrating larger parts of generated code.
So we're all to become systems integrators? I'm genuinely happy that excites you. Personally, I find the prospect depressing.
> It’s time to adapt to a changing world.
How do you adapt to no longer being able to do what you love to do? Saying "it's time to adapt" doesn't address the fundamental issue here at all. Obviously, we have to adapt. The question is how to do that and, if happiness is no longer on the table, at least be able to maintain some amount of career satisfaction.
If I can't write software, then my career as a dev is over, no?
When I made this comparison I was trashed in the comments. Not at the hyperbole of it (I don't think it is), that they are in poverty now while we are not. That we should have saved money. Which, shocker, not all of us work in California with those salaries... At this point as a SWE my only goal is to try and find IT employment to support loved ones and keep my visa till I get permanent residence. Fuck everything, just survive at this point. I loved programming, if that is ripped from me, then survive is all I want to do. That or jump in front of a train.
I don't mean to be so dark... it is just I get so worried about the future. It might be age, it might be that i clawed my way to a tiny bit of stability. Things could be great, I actually have seen you in a few threads and you always say the same thing that you love programming because you can focus on global ideas now. That to me sounds fun too. I like both. I am just worried you need less of those people which means more unemployment. For me when I say programming is ripped from me I mean it in all sense of the word.
Hey, I'm really sorry this is happening to you. I hope you are able to find that same meaning in your life, whether it's a continuation of your coding career, or another problem-solving field.
Change is hard, but it is possible, and one thing humans are good at that machines still can't do is to make our own meaning.
I've been in this industry for a long time, and have been through a few fundamental shifts, and of course, I have adapted to and valued change throughout. But I have never felt nearly as pessimistic/doubtful about the future of software development as I do now.
Nothing will make me happier than to be wrong about what I see coming.
Feel free to substitute it with print journalists, who over the last few decades have watched their careers vanish and replaced with something quantifiably less valuable to society.
> How do you adapt to no longer being able to do what you love to do? Saying "it's time to adapt" doesn't address the fundamental issue here at all. Obviously, we have to adapt. The question is how to do that and, if happiness is no longer on the table, at least be able to maintain some amount of career satisfaction.
Forget “love” how are we supposed to adapt when jobs seem to be increasingly specialized requiring ridiculous education costs and companies don’t want to train people anymore.
I do not know the answer to your question, since for me programming has become a lot more enjoyable being able to throw outputs of LLM's together and then being the glue in between with my expertise. Coming from my perspective, going back would throw me into a career with prospects which I started to lose passion for. Doing the same thing over and over again, spending weeks on stuff I have already solved many times over in different flavors, frameworks and languages.
I understand that some engineers/developers struggle with this change, but it also seems to be a personal take on ground breaking changes which are shaking up this career path; which are positive for some, and negative for others. Maybe we are on two different part of the spectrum, because I can't wait to have my job reduced to prompting and building even bigger things. IF it was not that way, I would have been close to a career change. No doubt, these changes will introduce new engineering challenges, which will be equally as exciting and in need of an engineering mindset as the previous paradigms have been.
> I understand that some engineers/developers struggle with this change
This isn't an issue of struggling with adapting to change. If it were just a matter of doing my job in a different way, that would be totally fine. This is an issue of seeing the sort of work that I enjoy -- the entire thing that drew me to this career -- evaporate completely.
But this all makes me sound more upset than I actually am. I'm very troubled, but I also only have a decade or two left in my career, so I think it's likely that I'll be able to find niche positions where I can keep doing what I love to do.
I do worry about younger devs who are in the business for similar reasons as me, though.
I really want to try the code interpreter. I got access to the plugins early but there was no browsing and interpreter. Then I got access to browsing when they released it widely last week. Does anybody know how to get access to the code interpreter? I have lost track of all the ${LLM_Service} waitlists I have signed up for.
In shameless attempt to raise awareness about my open source gpt-4 coding tool "aider" [1], I replicated the author's US Census data analysis example from the command line. This doesn't require the beta ChatGPT plugins, just a gpt-4 api key.
You can see in the chat transcript [2] that I had GPT-4 tell me how to download census data, suggest some hypotheses to test, test one, summarize the results and plot a graph. As we went along, I had it write the code for all that as a python script in my local git repo.
Getting an API key has been just a waiting game here... Not been approved through work or personally yet after almost 3 months. I already have plugin access though, it's...alright. The web browsing leaves a ton to be desired.
Ya, I'm not sure why I got lucky to receive one. For coding tasks, gpt-4 really is significantly better than 3.5. I've been building tools for coding with GPT for a few months now, and they got way more effective and reliable with gpt-4.
The thing I've used browsing for the most is linking to lengthy markdown files for large initial prompts with heavy detail. It works alright. Not perfect but decent.
I wonder about the copyright status of code put out by ChatGPT.
I presume anything ChatGPT generates belongs to the user who asked for the code to be generated. Is there any clause in the EULA that says this is the case? Or is code generated by ChatGPT automatically BSD-licensed? Or can ChatGPT claim ownership or joint-ownership of code generated using their tools?
Another question is the code it was trained on. Presumably all of it was open source. BSD-licensed code typically requires attribution. GPL code requires anything copied to remain GPL. If the GPT is reproducing small stanzas of code from memory it probably doesn't matter. If GPT reproduces larger blocks of code from stuff that it was trained on, there is a possibility of failing to comply with copyright.
> I presume anything ChatGPT generates belongs to the user who asked for the code to be generated.
Machine generated works cannot be copyrighted, only human-created works. The part here that has yet to be determined is whether or not AI output is considered "machine generated". An argument could be made that writing a prompt is sufficient control to consider the output human-made for the purposes of copyright. If that's the case, then copyright will go to the human who wrote the prompt (or their employer).
This will, at some point, have to be decided in court. Right now, the only answer is "nobody knows".
Right, which is why I think that courts will end up deciding that writing a prompt is sufficient control. But as of this moment, none of this has been decided -- so nobody actually knows.
I think in this case the product is the scene and not the picture though. If one could copyright the result of a prompt then I assume no one else could use the same (or very similar) prompt for profit as it is likely to produce similar code whereas a scene can be photographed for profit by anyone regardless if a similar picture has already been taken, or so I assume. If Andy Warhol just photographed Prince in a same/similar pose and clothes would there be the lawsuit case ? I don't think so.
So a binary created by a compiler isn’t copyrighted ?
Last I checked , that’s not the case. Otherwise the entire proprietary software universe would collapse (not that having that outcome isn’t desirable per se )
A binary created by a compiler is exactly as much a deterministic result of the input and choice of compiler and settings without additional human creativity as, say, an image created by a prompt and a diffusion model, and the logic of the Copyright Office determination that prompt (if sufficiently a work of authorship) but not the output in the latter case is subject to copyright would, if applied consistently, seem to apply to the former case equally.
I'm a little confused by the responses I'm getting to my comment. People seem to be thinking that I'm saying that copyright cannot apply to AI produced output. I said no such thing.
What I am saying is that this is not a legally settled matter. It will certainly end up in a courtroom, and will then become a settled matter. That day isn't here yet.
If the level of determinism is truly the same then I think that might play into the argument that AI output is indeed a statistical collage. And yes, human output might be as well but the burden of proof would be on the other side of the court. I'd really be careful with the parallels because I think we can agree that compilers don't work like LLMs. The output of a compiler isn't related to the previous programs it compiled in form of training data - so there is no fair use or not fair use question here. Dealing with something without precedence means we can't point to something else and expect an exact match.
An argument could also be made that If you pulled all the copyright infringements out of the training set for ChatGPT it would be not be half as intelligent.
Well, from a legal point of view, it's not yet clear if training using copyrighted material constitutes copyright infringement. My guess is that courts will rule that it does not.
(I'm not saying that's the right call from an ethical point of view, just that I think that's how courts will interpret the law.)
83 comments
[ 4.2 ms ] story [ 153 ms ] threadIn a weird way, it's actually still aligned with their "think different" motto: when everyone is thinking about getting on the AI bandwagon, think different and avoid doing that!
[0]: Which is fine—the world needs boring companies as much as pioneers who break things.
Also, Apple should first fix Siri, or get rid of it for good. It's a joke at this point.
I feel like any company which has chatbots as a "core" part of their systems (like Siri is often advertised to be) would be looking very heavily into generative AI. It makes way too much sense and would clearly make Siri way more useful. Imagine if Siri was current ChatGPT with plugins for interacting with the OS -- this immediately solves basically all the complaints about Siri. Plus Apple has efficient local AI acceleration already, which could make it even better.
It feels so obvious it'd be weird if they weren't trying to make a better Siri with gen AI! I have a feeling we'll hear about it either at WWDC or the iPhone event in the fall.
It's also not like AI tools are why people buy Apple products. So "waiting" probably doesn't cost them much money -- it doesn't impact their core hardware and service revenue. It's "just" a value add for now.
For acquisition yeah I was thinking along the lines of Siri.
Who would've thought that in just four months, GPT-4 will completely transform people's perception of AI progress timeline and impact.
I have no idea what Apple will do with LLM's. But Apple is already leveraging ML in things like built-in local dictation and OCR, and we can probably expect that it will leverage things like LLM's at the OS level (iOS/macOS) rather than focusing on their iWork suite. (Honestly at this point I don't understand Apple's strategic purpose with iWork, or even Final Cut Pro.)
In April 2023, Apple has a 30% market share of smartphones worldwide [1]. And in the first 3 months of 2023, Apple accounted for 50% of global smartphone revenue... and 80 percent of the industry's profits [2]. These is smartphone revenue, not services revenue. (Apple makes money from the App Store too but that's separate.)
80% of industry profit from 30% market share is certainly not a commodity. Cheap Android phones are a commodity. iPhones are the polar opposite.
[1] https://www.oberlo.com/statistics/smartphone-market-share#:~....
[2] https://www.statista.com/chart/29925/apples-share-of-the-glo....
Also frankly, productionalizing LLMs is very different from making research strides on them. Even inside Google researchers are realizing that a bulk of the improvements are now coming from the OSS community.
Apple will implement these technologies when it makes sense to do so.
Why would Apple even pursue LLMs? They are building hardware: computers, phones, screens, accessories. In a sense, that’s what you said. Just phrasing a question to your thought.
Instead of having an AI model output some kind of "black-box conclusion" on a data set / question, it can output code that we can inspect, improve, and learn from. Using LLMs in this way still keeps us as users "in control" - and I think it's a promising path for responsible AI.
I've done some experiments[1] adding ChatGPT-like functionality to my open source Typescript Notebook environment[2] back in december, and hope to revisit this soon :)
[1] https://twitter.com/YousefED/status/1599805936280907776 [2] https://www.typecell.org/
[0] https://www.codium.com/
[1] https://codeium.com/blog/how-is-codeium-free
I'm left wondering what the role of a software eng is and all I can come up with is a glorified quality analyst. Nothing wrong with that, but it's just such a transformation, and so sudden. I'm excited, scared, and exhausted all at the same time!
Unless, of course, you want to be a software engineer.
If all there is to my career is checking generated code for correctness, then there is nothing of value in my career anymore.
That's not how it will be, though. At worst, there will still be niches I can fit in.
> Unless, of course, you want to be a software engineer.
It’s time to adapt to a changing world.
So we're all to become systems integrators? I'm genuinely happy that excites you. Personally, I find the prospect depressing.
> It’s time to adapt to a changing world.
How do you adapt to no longer being able to do what you love to do? Saying "it's time to adapt" doesn't address the fundamental issue here at all. Obviously, we have to adapt. The question is how to do that and, if happiness is no longer on the table, at least be able to maintain some amount of career satisfaction.
If I can't write software, then my career as a dev is over, no?
Humans gotta realize that we are all replaceable. What happens to the least of us, happens to us all.
Whatever we do about this AI stuff? We should do it for everything else, too.
https://www.nbcnews.com/science/environment/west-virginia-co...
Change is hard, but it is possible, and one thing humans are good at that machines still can't do is to make our own meaning.
Wishing you the very best.
Nothing will make me happier than to be wrong about what I see coming.
I doubt there were that many who actually enjoyed digging coal as opposed to just having a relatively well paying job
Forget “love” how are we supposed to adapt when jobs seem to be increasingly specialized requiring ridiculous education costs and companies don’t want to train people anymore.
I understand that some engineers/developers struggle with this change, but it also seems to be a personal take on ground breaking changes which are shaking up this career path; which are positive for some, and negative for others. Maybe we are on two different part of the spectrum, because I can't wait to have my job reduced to prompting and building even bigger things. IF it was not that way, I would have been close to a career change. No doubt, these changes will introduce new engineering challenges, which will be equally as exciting and in need of an engineering mindset as the previous paradigms have been.
This isn't an issue of struggling with adapting to change. If it were just a matter of doing my job in a different way, that would be totally fine. This is an issue of seeing the sort of work that I enjoy -- the entire thing that drew me to this career -- evaporate completely.
But this all makes me sound more upset than I actually am. I'm very troubled, but I also only have a decade or two left in my career, so I think it's likely that I'll be able to find niche positions where I can keep doing what I love to do.
I do worry about younger devs who are in the business for similar reasons as me, though.
You can see in the chat transcript [2] that I had GPT-4 tell me how to download census data, suggest some hypotheses to test, test one, summarize the results and plot a graph. As we went along, I had it write the code for all that as a python script in my local git repo.
[1] https://github.com/paul-gauthier/aider
[2] https://aider.chat/examples/census.html
I presume anything ChatGPT generates belongs to the user who asked for the code to be generated. Is there any clause in the EULA that says this is the case? Or is code generated by ChatGPT automatically BSD-licensed? Or can ChatGPT claim ownership or joint-ownership of code generated using their tools?
Another question is the code it was trained on. Presumably all of it was open source. BSD-licensed code typically requires attribution. GPL code requires anything copied to remain GPL. If the GPT is reproducing small stanzas of code from memory it probably doesn't matter. If GPT reproduces larger blocks of code from stuff that it was trained on, there is a possibility of failing to comply with copyright.
Machine generated works cannot be copyrighted, only human-created works. The part here that has yet to be determined is whether or not AI output is considered "machine generated". An argument could be made that writing a prompt is sufficient control to consider the output human-made for the purposes of copyright. If that's the case, then copyright will go to the human who wrote the prompt (or their employer).
This will, at some point, have to be decided in court. Right now, the only answer is "nobody knows".
Pushing a button is sufficient to own copyright, in theory.
Last I checked , that’s not the case. Otherwise the entire proprietary software universe would collapse (not that having that outcome isn’t desirable per se )
I'm a little confused by the responses I'm getting to my comment. People seem to be thinking that I'm saying that copyright cannot apply to AI produced output. I said no such thing.
What I am saying is that this is not a legally settled matter. It will certainly end up in a courtroom, and will then become a settled matter. That day isn't here yet.
(I'm not saying that's the right call from an ethical point of view, just that I think that's how courts will interpret the law.)
In addition to data analysis stuff, I've also used it for running comparative benchmarks: https://simonwillison.net/2023/Apr/12/code-interpreter/
You can upload all sorts of weird things to it. I've uploaded .whl wheel files from PyPI to give it extra Python dependencies.
I also uploaded a Deno (and a Lua) interpreter and that gave it the ability to write and execute code in JavaScript and Lua!
https://til.simonwillison.net/llms/code-interpreter-expansio...