411–420 of 736 posts
Earlier quoted context omitted.
> The damage is done. You cannot build a business critical function on top of American SOTA frontier model. Especially not with the current crew in charge. The switching costs of changing LLM providers is as low as it gets. All the individuals and startups I know try different models all of the time, even down to the level of choosing which provider to use based on the task. Bigger companies move slower but only beca…
For real production I find the switching cost is not as trivial as you portray. Even going to a new model version in the same model family, say GPT-4o to GPT-5.2, a transition I just finished on a not too complicated application, requires extensive retesting and tweaking of prompts, guardrails and parameters.
Earlier quoted context omitted.
I measured an ~8x increase in my project's commit count after AI, and I'm painstakingly reading, reviewing, understanding and editing everything the models write. It's gotten to the point I'm trying to slow down in order to let the new knowledge crystallize. I'm manually writing articles about what I'm doing as I go. I can only imagine what people are doing at their jobs with unlimited token budgets.
> I measured an ~8x increase in my project's commit count after AI, That's irrelevant. What's the increase in revenue?
Earlier quoted context omitted.
> Mistral is just not competitive enough Does anyone know why? I was really excited when they emerged, but their models and targets don't seem to be quite in the same market.
Most probably lack of capital and talent. At the end of the day they have to compete with other giants for the chips to train the models.
Earlier quoted context omitted.
Maybe we will in a year, but then we'll just complain about China copying protectionism and censorship from the USA. If that's a comfortable position for you, all good. We held the US in higher regards, that's all.
[flagged]
I don't think anyone is making the same mistake with China, as open weight models can't be Thanos'd away.
The damage is done. You cannot build a business critical function on top of American SOTA frontier model. Especially not with the current crew in charge. Now whether AI tech is in the same league as say Nuclear tech and therefore by any reasonable standard should be regulated is a different question. We hit the slippery slope on a random day in June 2026 and there is no putting the genie back in the bottle. Any exec…
I mean, this was already pretty clear before. But it surely didn't help!
Earlier quoted context omitted.
I'm a small software business owner in Europe. I have to assume my competition is willing to pay for any business advantage they can get. And so I also have to pay for the SOTA model, whatever it is.
The good news (for you and most everyone other than the current leading AI companies), the gap between the SOTA and the near-frontiers is getting smaller every week or two. The leading Chinese models are only a few months behind now (GLM 5.2 tickles the tail of GPT 5.3 or 5.4 and Opus 4.6, according to benchmarks and the vibes among heavy users who've spent some time with it), where they were a couple of years behind…
Earlier quoted context omitted.
I use whatever model is SOTA. I switch between them in order to avoid lock in.
>I use whatever model is SOTA. I switch between them in order to avoid lock in. What's your competitive edge here? Shaving off an hour of a feature delivery? Not having to see the code that is produced?
For a change, I let DeepSeek V4 Pro implement it on Max thinking level. Nothing too out there - some DB migrations, some Django back end changes and Vue SPA front end changes.
Implementation time in total including tests was a few hours, so nothing too egregious. However, one of the migrations would break with pre-existing data, one of the column references in the entity was wrong, the API endpoint wasn't made consistently with the others in adjacent code (e.g. permission checks) and the front end had a Pinia state related issue and submitting one of the forms didn't work.
Tooling was run: ruff, ty, Oxfmt, Oxlint, also Docker build was green across the board, but the overall feature just didn't work. In both cases, sub-agents with clear context would review the code for serious/critical issues, at least three in parallel and do review loops until they spot nothing. The harnesses both has LSP integration.
Opus spent another hour fixing it, needed a few iterations, because I couldn't be bothered there.
> What's your competitive edge here? Shaving off an hour of a feature delivery? Not having to see the code that is produced?
The difference largely was not needing to waste time in fixing all sorts of subtle bugs that sub-optimal models will produce, worse yet if it was some sort of a serious project and those wouldn't have been spotted but instead that slop would have gotten shipped.
That said, Opus isn't ideal either and messed up a whole bunch when I was training some neural nets and try to process a bunch of satellite data and configure Garage to store them so that tiles can be served from a slow HDD and stuff like that. Obviously, it also needs a lot of babysitting in regards to UI looks, but it's better at the rest of development.
I think that DeepSeek V4 Pro and GLM 5.2 are cool though, it's just that you want as many checks and tests as you can throw at any given problem, or use languages that make shipping completely broken code increasingly likely.
Earlier quoted context omitted.
The good news (for you and most everyone other than the current leading AI companies), the gap between the SOTA and the near-frontiers is getting smaller every week or two. The leading Chinese models are only a few months behind now (GLM 5.2 tickles the tail of GPT 5.3 or 5.4 and Opus 4.6, according to benchmarks and the vibes among heavy users who've spent some time with it), where they were a couple of years behind…
This is nonsense. The gap between Chinese models and American frontier models is estimated at 10 months by Anthropic themselves, and it's growing. China has no flywheel for long-form agentic traces like Claude Code and its telemetry over its userbase (no one uses the Chinese harnesses yet). Most Chinese models are forced to price themselves significantly below cost to compete with the huge demand for bootleg claude t…
#1 I've had use cases where it was clearly obvious the Chinese models were behind.
#2 I've also had use cases where I couldn't tell a difference at 1/20th of the price.
The problem is - the #1 is the use case where American frontier is gated behind saboteur classifiers and is tiny minority anyway. Vast majority of work is #2.
The gap doesn't matter anymore.