Comment by api | Hacker News Reader

Comment by api | Hacker News Reader

Parent Full thread

I've tried a bunch of LLMs, and I agree that Claude is currently the best at coding, but the gap isn't huge. Gemini 3.8 Flash is surprisingly good and I'm looking forward to trying Gemini 4. The open weights models like Kimi, GLM, etc. are a little behind in quality I think but not massively.

With harnesses I see little difference between what Zed or Pi gives me and what Claude Code gives me. Enterprise features? The AI can help code those in a week.

AI itself is a great breaker of moats.

The question isn't whether they lead today. It's whether there's any kind of moat. Without a moat, frontier model development is an industry like chip fabrication or large volume manufacturing. Those industries are a red queen's race where if you stop innovating for even a short while you sink. Ask Intel what happened when they decided to coast for a few years.

Chip fabs and bulk manufacturers make a lot of money, sure, but they don't command such a multiple. Anthropic is priced like they're going to be to AI what Google was to web search or Apple is to (US at least) mobile and higher-end PCs.

I don't see it. I think they'll be a big player but that big?

Of course... maybe if they suddenly have that much cash they'll be able to carpet bomb the entire world with AI doom Rationalist ideology and buy their regulatory capture and outlaw their competition.

They made it convenient to get value for the specialist masses and I don't think anyone else managed to achieve it yet. That's their moat.

I've seen firsthand attempts at integrating open source in place of it and they all failed. Time is money and it costs premium. Touching the ground running is a huge advantage. That's why cloud services stuck.

time is money so when the subsidies stop or they magically make inference price an afterthought like cloud did to traffic then it's going to be much more economically better to spend a one-time price setting up a local or a self hosted open model than paying $2000 for a subscription

The question isn’t whether Anthropic makes a good product. The question is whether it can rationally sustain multiples like this.

Yes, they have the same problem WeWork had on the multiple front. Coworking was a thing long before WeWork was a thing. It’s an important but boring business that trades at low multiples. WeWork tried to IPO on the basis that somehow it was worth vastly more, but the market woke up and realized it’s just subleasing office space.

The same is unfolding with Anthropic. It’s doing important work but everyone is realizing it’s not got a serious edge over anyone else. It’s important but ultimately a boring business. It’s still likely worth a few hundred billion, which is massively impressive for any startup, but trillions is pipe dream.

Great comparison.

The thing that makes the frontier labs frontier is mostly access to huge amounts of training data and compute. In the earlier days of AI access to top AI talent was a moat, but that moat is rapidly eroding as more and more people familiarize themselves with the "black magic" of LLMs and training. The success of the Chinese models is evidence of this diffusion of knowledge.

Give me a team of, say, 20 good engineers with a deep knowledge of AI/LLMs and training and give me a billion dollars worth of GPU compute and a huge amount of training data and I'll give you a frontier model.

No moat.

Reply on news.ycombinator.com