Comment by seviu | Hacker News Reader

Comment by seviu | Hacker News Reader

Parent Full thread

This. I have been using anthropic and codex subs, on max. All this changed in June. We are clearly entering an era where we cannot rely on American models. As a solo developer I value reliability over performance. I cannot pay hundred of $, plus a lot of my private time figuring out how to properly use this technology, for it to be taken away within hours.

On top of that, the intelligence is being dialed down. Sonet 5 is a living proof of this. Fable has strong guardrails, but new Sonet is a dumbed down expensive model, which already falls behind GLM 5.2 and Kimi 2.7. I might go back to Claude since I know Fable is just a limited offer, and I am not going to pay for API usage. But what they are signaling with Sonet will also come to Opus. A lobotomized more expensive model.

I am honestly baffled how the current administration is giving the whole world, on a golden plate, to China. And they don't seem too bothered about it. They are living in their own bubble and reality distortion field I guess.

I could go on endless rant about Dario, but I feel I am so strongly biased now that my judgement might be clouded.

Time to move on

I feel like I see this comment every few months and yet in between people keep talking about all of the functionality they’re getting out of anthropic’s offerings. It doesn’t seem to me that people are willing to give up the “shackles” as it were and we’re just going to wind up with what we’re fearing here. On top of that, local models are just not turnkey enough for the average person yet (go ahead and drop somebody into LM studio and tell them to get to work, it won’t go well).

The best models are clearly all from US companies.

But I think what you’ll see is people making sure the model they use can just be plugged into their workflow.

I used to use Gemini-cli until they did a Google and cancelled it in favour of anti-gravity.

That was my fault. Fool me the 10th time shame on me.

So I picked more open source offerings that I can use with any model. Once the other models are good enough, I just need to jump ship.

The reason Anthropic gets away with all of that is that Claude revenues are increasing with no end in sight. People write these entitled rants and silently go back to Anthropic and OpenAI like obedient puppies.

It’s in increasing but they still haven’t gotten into the black. I’m pretty sure nobody has yet, right?

Probably not, but I think the investors care more about users and subscriptions than profits at this point.

That’s the old model for sure but we’ve never seen this much investment this rapidly. They have to be getting cold feet by now. Hence anthropic going public. If they had a strategy for getting in the black AND longterm viability I feel like we’d see them roll private longer. But instead we’re seeing all the signs of “grow, exit, let the next dude figure it out.”

I completely agree. I was just pointing out the current expectations and behavior.

AI tech is clearly a Red Queen's Race[0]. In the long-term, whoever can run the fastest the longest will win. Adverse action that doesn't materially impact the rate of execution has little effect on this outcome except to the extent it reduces the rate of execution of competitors. Historically, American business is exceptional at executing this type of game and patient when it comes to making it pay for itself.

The AI model people choose today has no bearing on the ultimate trajectory of the competition. Both the US and China understand this. EU simply can't move quick enough to be competitive in this type of game, which I think they also recognize.

Everyone is betting that the model you use will be a Hobson's Choice[1] over a long enough time horizon. They are likely correct.

[0] https://en.wikipedia.org/wiki/Red_Queen%27s_race

[1] https://en.wikipedia.org/wiki/Hobson%27s_choice

the intelligence is being dialed down. Sonet 5 is a living proof of this.

Huh? Sonnet 5 is a strict improvement over Sonnet 4.6 at the same price.

The tokenizer is 30% more expensive (it generates 30% more tokens) just they are giving an offer now which will end by end of August. On max effort its way more expensive than Opus and slightly less than fable.

It has also strong guardrails and deliberate reduced training data around cybersecurity. On purpose. If you develop a system and then try to harden it, it will miserably fail.

It’s a neutered model. Pretty sure not what they had in mind, but they have to, since they invested so much VC money in developing it.

Reply on news.ycombinator.com