From what I've seen trying to play around with Claude Tag, it uses shared credentials. You set it up with Github org access, add it to public channel, and all random people in that channel can now ask it to do stuff with code. It's opposite of what Cursor does - each Slack member needing to connect their own Github account. It's a security nightmare basically.
Because I am getting the call to fix it when it breaks. I don't have to fix assembly by hand because compilers are deterministic and I have maybe encountered a single real compiler bug in my whole career. Compilers have earned my trust. LLMs are eroding that trust more and more every day I work with them. I encounter LLM-created problems in basically every single diff they surface for me, just over the months the diffs are getting bigger and harder to review and uncover the problems.
LLMs are not an abstraction(not even a bad one) because by design what they are doing is disambiguation. Compilers are not doing that, what you put IN the compiler has to be unambiguous in the first place.
Disambiguation is not a functionality of an abstraction layer. A good abstraction layer is the one I don't have to understand and can trust, if I have to understand its inner workings to use it it ceases to be an abstraction. Except with LLMs you can't even do that, they are a black box you can have no hope of understanding.
And it is not to say LLMs and agentic coding tools are not useful, they are absolutely very useful. They are just not an abstraction layer.
If someone is trying to bend the rules of my passively managed index fund to their will, are they trying to actively manage my passively managed index ETF ?
The developers are literally on the bleeding edge here, it might be the most developed of the AI use cases right now. The most advanced tooling for LLMs revolves around SWE work, there are multiple prolific benchmarks that the labs are actively targeting in this area, and new ones are being built, whole product categories being spawned, software companies bleeding money for tokens.
It's the other professions that are to follow once the training data is in place to go reach for their livelihoods. SWEs got the early taste of what is coming. And the blender news is exactly that.
Average is only a tombstone of someone having failed to do better. And settling for average means pulling down.
When it comes to bs dashboard where "average is all you need", maybe the "better than average" result would be asking yourself if it's even worth doing in the first place?
Probably the same way other models learned to surpass human ability while being bootstrapped from human-level data - using reinforcement learning.
The question is, do we have good enough feedback loops for that, and if not, are we going to find them? I would bet they will be found for a lot of use cases.
I am vibe-porting an old game, Knights & Merchants(actually its Delphi rewrite - KAM Remake) to WASM. It's going well, I even have multiplayer working, will release it publicly at some point.
Learned more about WASM, OPFS, JSPI and other exotic browser stuff more than ever, also learned more about pascal than I ever wanted to, but it's been immensely fun.
Not having a code review process is archaic engineering practice at this point(at any point in history, really), be it for human written or AI written code.
Perhaps the problem is that you RL on one patch a time, failing to capture the overarching long term theme, an architecture change being introduced gradually over many months, that exists in the maintainer’s mental model but not really explicitly in diffs.
It doesn't matter, neither of those scenarios makes the effort impressive in this case. The vibe coded thing might even be useful - that does not make it impressive though. Effort does.
You know what is an insult? Supermarket on my street putting on display sloppy ads with ramen bowl that has 3 different thickness chopsticks and cartoon characters with scrambled faces. Now that is an insult, because there was a human being doing that job, and I am sure there was a great "productivity boost" related to that change.
I am a heavy AI user myself, and sure as hell I am not putting my foot in that place again.
Is that even possible? Last time I checked it wasn't, while it was possible with OpenAI. Since that moment(early this year) - OpenAI has removed that option and their "Project budget" feature turned from being a hard limit into an email notification.