I don't think we're far enough into this game to know what winning even looks like. If the US frontier labs are losing, it may have been their own missteps on handling such massive change of being the fastest growing apps ever. Until China can ever show me something new rather than just cheaper, then I just see efforts from that region as a force of commodification. OpenAI will go down in history as the company who brought what many understand to be AI to the first billion people.
It's hard to explain, but AI mostly doesn't get the nuance of my thinking right. I'll read a response and right away know it's a non-answer for my purposes. If I get a "hit" then it's often by luck. And so, I use it mostly for anything other than direct answers, such as research, etc. Beyond that, it's not much different than old school Google result. So, no, I don't offload my thinking.
The article is describing how every harness functions. If you're already using Codex, Claude Code, or Antigravity then you're already doing this. And you get a lot of usage. I generally don't hit my limits with GPT if I stick to reviews of code that cheaper models create. Google gives you a low number of free requests. OpenCode has a generous free tier for open source models (and a plan that is $5 / month for the first month.)
I think the more interesting thing is to actually get Claude to stop saying it. Or to get an AI to do things you want, but more precisely. Ran into an issue the other day where I couldn't get a model to "just copy this and change the details!" without doing way too much thinking and then bloating the thing 2x.
My solution to this is to dump it into an LLM and a prompt that roughly does something like...
Hand wavy list just to get a general idea...
1) Give me a condensed summary
2) Is this adding anything to what we already have? (I save good articles along with annotations and whatever notes I may write to go along with it.)
3) Locate any upstream ideas on this (often AI articles are rehashing much better written ideas.)
...
Something like that. Not that I have some great system for it. I find these articles are so full of fluff that I have lost patience to attempt to get through them. So, I pull out the AI to parse the AI. I know that the AI may miss some hidden gems, but I'm okay with that.
And speaking of technology, it would have been great to see the Phase Plasma Rifle at a 40 watt range make an appearance, if only by mention of name again.
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...