Really enjoyed this post. I was wondering, as I read it, "Why not both?" AWS's origins were serving some of Amazon's consumer-facing business lines. Amazon was (is?) AWS's first and best customer, which ensured immediate demand for its most advanced features.
What would prevent a similar framing for OpenAI? Ie consumer-facing offerings in the short term push the backend forward as fast as possible, then rolling those features out at the platform level makes them widely available for recombination by developers. (And in the short term, keeping things more closed allows OpenAI to get a better handle on safety.)
Agree. The current product offerings are their marketing. Like Elon's other companies, they don't do marketing. They just build stuff that markets itself. These are that. They're the reason everyone in every role is imagining how to use this in their company all at once.
My guess is that was initially true, and they wanted to build simple apps to show off the tech. But given how much it's taken off, I'm less confident that's still the case. If nothing else, the people running it are human. Attention and hype is addictive, and if it's apps that get them that, there's going to be some pressure to keep building it.
Agreed. And rationally so in some ways: Aggregating consumer demand by essentially becoming the next operating system would be an insanely powerful position. Maybe more powerful than being a shared infra layer. If nothing else it’ll be appealing to pursue both till we see how this all starts to shake out. So many unknowns.
Yeah, and as I said here, it took me all of about an hour to start questioning the premise of this post: "Maybe OpenAI isn't the next AWS, nor is it the next iPhone that hosts all of our apps. It could actually be the next web browser: We stop going to websites, but instead, just chat with a thing that does our bidding. Every web interface is replaced with a conversational one. Physical stores were replaced with web products; web products are replaced with voice commands. It's the Star Trek "computer" prompt: A universal interface into everything, with no need for anything else."
This is a great read, Benn. But the logic breaks down when training and running an LLM stops being as expensive as building data centers :).
I am curious to hear what you think about OSS LLMs like Databrick’s Dolly that can do instruction following, brainstorming, and summarization just like ChatGPT.
BUT….
can be created for $30 using one server for 3 hours on a small dataset using a 2 year old open source base LLM as opposed to training for 100,000 GPU hours at ~$10 million price point.
I saw - this is stuff is moving so fast one that within two hours of publishing this, Databricks went out and proved it all wrong.
To me, there are two possibilities here. The first is the experiential gap between the small models and the big models is too big for the small models to make sense. For example, if we come to rely on an AI bot to do everything for us (eg, the Star Trek "computer" https://www.youtube.com/watch?v=x74qzr0z3B0), a pretty good LLM may not be anywhere close to good enough. Maybe the cheap ones get there eventually, but if it doesn't happen fast enough, the ecosystem moat around OpenAI, Google, and the other big players may be too big to overcome.
The other possibility is these things are good enough, in which case it seems like the entire game is the utilities you build on top. How easy is it to interact with, to develop on, etc etc. That's a more interesting result to me, and much more chaotic, in large part because I don't think we have any idea how this stuff will get used yet. In other words, the killer utilities are almost impossible to predict because we don't know what we're building yet.
Hi Benn, I enjoyed reading about your thoughts in this post, and have two questions I hope you can address.
1. Along the lines of what "smakubi" is saying, what about the argument that LLMs will become commodities? The Dolly example is a good one, but even before that, James Currier of NfX wrote about how LLMs are likely to become commoditized and (at least some) open sourced (link here: https://www.nfx.com/post/generative-ai-tech-5-layers). Do you tend to agree with this LLMs as commodities future?
2. On the application layer, our collective imagination seems to be constrained so far, and while the apps are pretty nifty, we haven't seen the truly breakthrough ones. Some have compared our current apps to calculator apps in the early days of iPhone, and the Ubers and DoorDashes are no where to be seen. My view on this is that we will see a big leap in creativity AFTER advancements in hardware like AR/VR/XR and 5G. If we use the GPT as iPhone analogy, then we can also say that few people if any thought they'd disrupt the taxi industry when the first iPhone came out, and it was only after 4G's wider adoption did that become a killer app idea. What are your thoughts on when the truly killer apps will emerge in this new technological shift?
On 1), I was just talking about this with someone yesterday. I'm not sure how that plays out, but I could see it being kind of analogous to labor markets. You have "unskilled" LLMs that handle rote tasks like support tickets, and then highly skilled (and expensive to create and run) ones that do more advanced stuff. You also could have, as that post says, really specialized ones that do particular creative tasks, and so on. So it starts to become a kind of mirror image of how people work today. So in that way, I suspect some LLMs are commoditized and some aren't.
On 2), I very much agree that all the current apps are kind of just early guesses at what'll work, and the real killer things will come later. I have no idea what those things are though (and if I did, I wouldn't be a blogger yelling into the corner). But it feels pretty wide open right now.
Really enjoyed this post. I was wondering, as I read it, "Why not both?" AWS's origins were serving some of Amazon's consumer-facing business lines. Amazon was (is?) AWS's first and best customer, which ensured immediate demand for its most advanced features.
What would prevent a similar framing for OpenAI? Ie consumer-facing offerings in the short term push the backend forward as fast as possible, then rolling those features out at the platform level makes them widely available for recombination by developers. (And in the short term, keeping things more closed allows OpenAI to get a better handle on safety.)
Agree. The current product offerings are their marketing. Like Elon's other companies, they don't do marketing. They just build stuff that markets itself. These are that. They're the reason everyone in every role is imagining how to use this in their company all at once.
My guess is that was initially true, and they wanted to build simple apps to show off the tech. But given how much it's taken off, I'm less confident that's still the case. If nothing else, the people running it are human. Attention and hype is addictive, and if it's apps that get them that, there's going to be some pressure to keep building it.
Agreed. And rationally so in some ways: Aggregating consumer demand by essentially becoming the next operating system would be an insanely powerful position. Maybe more powerful than being a shared infra layer. If nothing else it’ll be appealing to pursue both till we see how this all starts to shake out. So many unknowns.
Yeah, and as I said here, it took me all of about an hour to start questioning the premise of this post: "Maybe OpenAI isn't the next AWS, nor is it the next iPhone that hosts all of our apps. It could actually be the next web browser: We stop going to websites, but instead, just chat with a thing that does our bidding. Every web interface is replaced with a conversational one. Physical stores were replaced with web products; web products are replaced with voice commands. It's the Star Trek "computer" prompt: A universal interface into everything, with no need for anything else."
https://www.linkedin.com/posts/benn-stancil_some-personal-news-i-just-wrote-this-blog-activity-7045092475899629569-XFng
This is a great read, Benn. But the logic breaks down when training and running an LLM stops being as expensive as building data centers :).
I am curious to hear what you think about OSS LLMs like Databrick’s Dolly that can do instruction following, brainstorming, and summarization just like ChatGPT.
BUT….
can be created for $30 using one server for 3 hours on a small dataset using a 2 year old open source base LLM as opposed to training for 100,000 GPU hours at ~$10 million price point.
I saw - this is stuff is moving so fast one that within two hours of publishing this, Databricks went out and proved it all wrong.
To me, there are two possibilities here. The first is the experiential gap between the small models and the big models is too big for the small models to make sense. For example, if we come to rely on an AI bot to do everything for us (eg, the Star Trek "computer" https://www.youtube.com/watch?v=x74qzr0z3B0), a pretty good LLM may not be anywhere close to good enough. Maybe the cheap ones get there eventually, but if it doesn't happen fast enough, the ecosystem moat around OpenAI, Google, and the other big players may be too big to overcome.
The other possibility is these things are good enough, in which case it seems like the entire game is the utilities you build on top. How easy is it to interact with, to develop on, etc etc. That's a more interesting result to me, and much more chaotic, in large part because I don't think we have any idea how this stuff will get used yet. In other words, the killer utilities are almost impossible to predict because we don't know what we're building yet.
Hi Benn, I enjoyed reading about your thoughts in this post, and have two questions I hope you can address.
1. Along the lines of what "smakubi" is saying, what about the argument that LLMs will become commodities? The Dolly example is a good one, but even before that, James Currier of NfX wrote about how LLMs are likely to become commoditized and (at least some) open sourced (link here: https://www.nfx.com/post/generative-ai-tech-5-layers). Do you tend to agree with this LLMs as commodities future?
2. On the application layer, our collective imagination seems to be constrained so far, and while the apps are pretty nifty, we haven't seen the truly breakthrough ones. Some have compared our current apps to calculator apps in the early days of iPhone, and the Ubers and DoorDashes are no where to be seen. My view on this is that we will see a big leap in creativity AFTER advancements in hardware like AR/VR/XR and 5G. If we use the GPT as iPhone analogy, then we can also say that few people if any thought they'd disrupt the taxi industry when the first iPhone came out, and it was only after 4G's wider adoption did that become a killer app idea. What are your thoughts on when the truly killer apps will emerge in this new technological shift?
Thank you
Hey Kevin, thanks!
On 1), I was just talking about this with someone yesterday. I'm not sure how that plays out, but I could see it being kind of analogous to labor markets. You have "unskilled" LLMs that handle rote tasks like support tickets, and then highly skilled (and expensive to create and run) ones that do more advanced stuff. You also could have, as that post says, really specialized ones that do particular creative tasks, and so on. So it starts to become a kind of mirror image of how people work today. So in that way, I suspect some LLMs are commoditized and some aren't.
On 2), I very much agree that all the current apps are kind of just early guesses at what'll work, and the real killer things will come later. I have no idea what those things are though (and if I did, I wouldn't be a blogger yelling into the corner). But it feels pretty wide open right now.
I learned so much in this one today - thanks Benn! Also, the "agents of chaos" being 3 different YouTube video links really got me.
There were so many other options too:
- https://www.youtube.com/watch?v=qhxlZC8BZJ4
- https://www.youtube.com/watch?v=ONFj7AYgbko
- https://www.youtube.com/watch?v=j92SjjcUe9o
I wonder if OpenAI will switch from non-profit to for-profit
Yeah, it's kind of already a weird hybrid (https://www.axios.com/2023/01/10/how-a-silicon-valley-nonprofit-became-worth-billions). I'd bet that they try to split it in two at some point, where the models are supposedly "owned" by some non-profit, but there's a corporate entity that makes money from it.