bensyverson

8 ms·

61.
by bensyverson 2mo ago
It's more complicated than that. Lots of providers use an "OpenAI-ish" API, but many of them have subtle differences in things like tool calling or thinking blocks. OpenRouter normalizes the wire format. OpenRouter does more
62.
by bensyverson 2mo ago
OpenRouter is a proxy, not an automatic router. Rather than building API clients for five different AI providers, you build one client to OpenRouter, and switching models become extremely easy. This matters when new models are coming out vi
63.
by bensyverson 2mo ago
Sure, in the same way that nano is better than Microsoft Word. Unfortunately, sometimes nano is not enough. In the case of this client project, we are only now getting to enough functionality to be generally useful. The complexity that exis
64.
by bensyverson 2mo ago
It’s not about speed-ups, it’s the fact that I’ll attempt a project far more ambitious than I would have in the past. Over the past two months I’ve built out a very complex web app for a client. Without the aid of AI it would be far smaller
65.
by bensyverson 2mo ago
Can OpenCode dispatch background subagents yet? I tried it a week ago and saw nothing. This is 99% of my workflow at this point.
66.
by bensyverson 2mo ago
Unless you have `cargo fix` hooked up to automatically run on any compiler error, every error is still is a round trip to the LLM. It may be trivial for the model to fix, but it's added latency and cost.
67.
by bensyverson 2mo ago
Fair—though SwiftUI has influenced/required new Swift features like Result Builders. I've found generating Swift to be a mixed bag. The LSP consistently reports stale errors which the model has to ignore, handling strict concurren
68.
by bensyverson 2mo ago
Yeah, C# is probably the closest direct comparison. I like that Go emits a simple binary, whereas it seems like that needs to be configured with a .NET project. Probably just a matter of taste/preference!
69.
by bensyverson 2mo ago
I do not think "good error messages" is an even trade for "fast compile times." What happens is the LLM catches the error, then may hit another error, and try again. This leads to more tokens and more latency, and then a
70.
by bensyverson 2mo ago
Probably both, but I have to add: I agree with the post. I've done a ton of agentic development using Go over the past 6 months, and it hasn't let me down. You may ask "why not Rust, or Zig, or ____?" The reasons boil do
71.
by bensyverson 2mo ago
Yes but how often in practice does the same information stay accessible via 30x redirects after one reorganization, let alone two or three?
72.
by bensyverson 2mo ago
I would assume that training pales in comparison to inference when it comes to energy use, taking into account the number of inference users
73.
by bensyverson 2mo ago
That’s because no one wants to admit that their product is vibe coded… But there are plenty of vibe coded products out there making money from actual users.
74.
by bensyverson 2mo ago
This is super cool! I like the idea of expressing the game declaratively, with a language that an LLM can easily understand. Do you have any other demos of what the language is capable of? I'm wondering if this could do a platformer or
75.
by bensyverson 2mo ago
What about alphabetical order would lead "Section 9" to sort above "Section 8"?
76.
by bensyverson 2mo ago
I'm not sure what's going on with the ordering of the sections, but it's definitely creative. - 01: Section 1 - 02: Section 11 - 03: Section 10 ... - 10: Section 3 - 11: Section 2 - 12: Section 12
77.
by bensyverson 2mo ago
one week later: "It says 'You're absolutely right! Let me look at the seams so I'm checking, not guessing—' and I guess that's when my SSD melted."
78.
by bensyverson 2mo ago
And is this using conventional model loading (all in VRAM), or are you streaming it in some way?
79.
by bensyverson 2mo ago
Are people getting decent tokens/second throughput? Some of these demos crawl at 1 tok/s or worse, which limits their utility.
80.
by bensyverson 2mo ago
I do think these “run a bigger model than will fit in VRAM” projects are necessary steps, but are they functionally useful or helpful to anyone currently? For example, is anyone out there running a big Qwen for coding on a 16-32GB machine w
81.
by bensyverson 2mo ago
Yes, the writing just has a disproportionate impact on this outcome.
82.
by bensyverson 3mo ago
How does one even purchase an AMD MI355X?
83.
by bensyverson 3mo ago
People said this kind of thing about MiniDV cameras, and then iPhones. The reality is that technology hasn't been a barrier for filmmakers for a very long time. Writing is still the hard part.
84.
by bensyverson 3mo ago
I think it just gets really into the weeds and assumes you have full context. I do find myself saying "You have more context than me—can you restate the question in simple terms?" It does well then. To be fair, I don't find t
85.
by bensyverson 3mo ago
The machines really are learning
86.
by bensyverson 3mo ago
New music genre just dropped
87.
by bensyverson 3mo ago
On the one hand, they say "for experimental purposes only," but on the other hand, hey, no cancer warning!
88.
by bensyverson 3mo ago
Yes, I do this all the time, and also check in the co-authored project plans which drove the commits. For a project that is transparently only possible due to agentic coding, I don't see any reason to conceal the methods.
89.
by bensyverson 3mo ago
Needs more "invariant," its latest catchphrase
90.
by bensyverson 3mo ago
> models don't have a good sense of their own capabilities. It actually had some memorized knowledge about LLMs being bad at confidence intervals, then it did a cursory web search. But it didn't need to do any of that; I alread

More ›