1) A lot of companies simply won't trust Chinese models
2) These companies will still trust a direct frontier provider over a non Chinese company hosting Chinese models.
3) Frontier model providers working with cloud hosting providers for data residency.
4) Hosting your own models is expensive (hardware, people to setup and maintain, risk of little to no return.) And smaller companies likely won't go this route.
5) As mentioned in the article, a good but cheap model can still be expensive if it takes longer to do a task. I imagine I could save days doing work with Fable vs using Chinese models.
ETA: And loads of demand. Not sure the pricing situation is yet settled. As Chinese models get more popular, they may have to raise their prices as well. US models may struggle more with meeting demand rather than getting margins squeezed.
Whenever I see the cost indicator in my harness while building some probably useless thing, I'm reminded that probably everyone else is doing the same thing. Spending loads of imaginary money (according to the harness, I spent $10 worth of tokens for this single issue, but I have a $20 sub) building something of imaginary value. And then I go on social media and see a wall of slop posts, many talking about skills, systems, agents, and "Karpathy Wiki Systems" to build more useless things. Then I hear about all jobs going to AI, and I figure surely someone has to be the sane one to direct people not to build useless things. Surely you can't leave that up to the AI sales guy. Every single idea I have ever passed by GPT is BRILLIANT! I don't know where I'm going with this. At least it wasn't written by AI. ;)
Edit to add: Just use Deepseek Flash 4. You can hit those servers all day for next to nothing and still scratch the itch to build useless things. ;)
I imagine at least part of the reason this doesn't make sense to you is because you're not a decision maker working at that scale. It made sense to someone at a high level. The calculations that person is doing is likely different than yours.
This is also why it's difficult to make money selling to consumers. They run different calculations than enterprise buyers.
I never had a desire to post questions there. Comments or answers only if it were something I felt a personal stake in (community that we were trying to bring up) or the rare case where I would come across something uncommon that was unanswered that I had recently ran into and figured out. It's not the users there who kept me away, I just that I like quiet (high signal to noise) developer spaces. ;)
I imagine a huge number of people were just browsing for quick answers and then bailed as I did.
> I saw a LinkedIn post a week or two ago by a senior engineer with 25 years’ experience in the industry...
And how many years have we had capable AI? Maybe it's going to take a similar timeline for people to figure out how to be good with an AI assisted workflow (if not fully automated.)
Yeah, punk was a bit of a rejection of the polish of the big bands of the time. In a sense, the "horrible" was sort of the point. And for the shock value. But did that really mean they were horrible? Probably everyone kind of sucks at first. But it's hard not to improve your skills once you have got to a point where you have done a certain number of shows because you created a sustainable cash flow to support it.
And you could say similar for the transition between painting and photography.
ETA: It's interesting how the bottleneck may reveal the real skill in the thing. Architecting the code. Having a eye for interestingness in creating an image / painting of something, etc.
Is that what they're wagering on though? Everytime I open X, it's all about the first to some hand-wavy definition of AGI is going to win everything, and it's the only thing that can get us through the painful transition period from massive job loss to abundance.
One of my favorite movies as a kid was Explorers (1985) where kids built a spaceship from a Tilt-A-Whirl and other parts. It was an inspiration. Like you, I enjoy programming, but I haven't built a spaceship yet. Hehe
Didn't read the articles, but at least the planners know and understand a map.
SO... a map is static reference. A calculator is deterministic computation. An LLM is probabilistic generation
In high-stakes environments like military planning, tools that generate new claims rather than reference known data introduce a different class of risk.
Yes, everyone is responsible for their own decisions. But then circle back to risk. How can the planners be sure they aren't dealing with hallucinations, questionable data, differing outputs based on prompts, and a long list of other things...
Okay, I was mistaken. The tooling I was speaking of uses Claude Code rather than the SDK. One uses the Zed ACP protocol. I'm not sure about the other. I should have said Claude Code rather than the SDK. For example, I can run a session through one of the tools, and then access that session directly in Claude Code. It's still Claude though. It seems the important element is that you're not using OAuth tokens from a sub to use in a different tool. If you go through Claude Code, then Claude Code is handling everything and giving your tool the output. Thanks for the correction.
That should be fine, because it's still using their tooling. And this seems like the better way to go. I have a couple of tools that work like this. I think the issue is mostly 3rd party harnesses that seek to do the same as Claude Code. And it seems reasonable that Anthropic decides how you can use the subscription, because it's heavily subsidized. Get a Claude $200 sub and max out the usage limits, then compare that usage to the cost of using their API. The difference is significant, which is why people are getting multiple $200 subs rather than paying for API usage (and I have seen reports where they are cracking down on this as well.)
I would argue the chips don't even matter (important, but not as a reason for defending Taiwan.) It's a strategically important location that is a stone's throw from Japanese islands. If Japan feels the need, then nukes may be on the table. If that were to happen, S. Korea may not be far behind. And the cycle spirals.
Do you though? I guess it depends on how you define family. There's family that you rarely see and you call them family because of the social (even if weak) ties. And then there's family you grew up knowing. The impact of family early in you, never goes away. Your family early in life shapes us in ways we probably can't comprehend. Reading Scott's work was a family ritual at the breakfast table. I'm sure his work had some part in shaping me in a way that I can't delete.
At my age, he was about as close to family as you can get without being physically there. I grew up reading his comics in our newspaper while eating family breakfast. His work was a part of our family morning ritual. His work was part of pre-internet America when our channels were limited. Our thought and worldview were to some degree shaped by these limited channels.
1) A lot of companies simply won't trust Chinese models
2) These companies will still trust a direct frontier provider over a non Chinese company hosting Chinese models.
3) Frontier model providers working with cloud hosting providers for data residency.
4) Hosting your own models is expensive (hardware, people to setup and maintain, risk of little to no return.) And smaller companies likely won't go this route.
5) As mentioned in the article, a good but cheap model can still be expensive if it takes longer to do a task. I imagine I could save days doing work with Fable vs using Chinese models.
ETA: And loads of demand. Not sure the pricing situation is yet settled. As Chinese models get more popular, they may have to raise their prices as well. US models may struggle more with meeting demand rather than getting margins squeezed.