girvo

11 ms·

271.
by girvo 6mo ago
I don’t have the mental capacity to do that at that scale. It produces more than I can review.
272.
by girvo 6mo ago
I do the same, but 3-4 is simply too many. I would guess you’re smarter than me, because more than 2 and I’m already at a fixed bottleneck of reviewing it (and reviewing my teams output on top)
273.
by girvo 6mo ago
> the bag holders are going to be utterly doomed Good news, the plan is for us to be the bag holders as they rush to IPO.
274.
by girvo 6mo ago
Two decades professionally here too (and nearly three in terms of programming as a whole), and I still use ‘em. Reviewing and adjusting, they make for quite a good experience even in agent-first development with the various nice extensions.
275.
by girvo 6mo ago
Correct, signed another vegan :) Unscientifically, it feels like your gut microbiome adjusts to it after a while!
276.
by girvo 6mo ago
Just like NFTs, just like the blockchain before that, in some ways kind of like the Web 2.0 craze (though we at least built some things then and the tight financing at the time kept a lid on it). This LLM stuff at least has some real possib
277.
by girvo 6mo ago
For me/my team, I use it to fix DevProd pain points that I would otherwise never get the investment to go solve. Just removed Webpack for Rspack, for example. Could easily do it myself, which is why I can prompt it correctly and review
278.
by girvo 6mo ago
Ah but it looks cool and I can put it on my stack ranking perf eval
279.
by girvo 6mo ago
Output is what the compute is used for above all else; costs more hardware time basically than prompt processing (input) which is a lot faster
280.
by girvo 6mo ago
Everyone’s ideas of what batteries should be included differ
281.
by girvo 6mo ago
That's easily explained by those being two different people with two different opinions?
282.
by girvo 6mo ago
Budgets are already happening
283.
by girvo 6mo ago
Tbf the Sparks usefulness isn’t for inference IMO. Its memory bandwidth is too low for that. But on the other hand, running Qwen 3.5 122B A10B locally on it using ~110GB of memory and getting 50tk/s generation and quite excellent prefi
284.
by girvo 6mo ago
Not that you asked but I’m getting ~20 tokens/s on my DGX Spark (Asus actually) using an Int4 AutoRound quant, MTP 1 and some other tricks
285.
by girvo 6mo ago
One of the projects I'm most proud of is still running ten years later, and has processed over a billion AUD through it in that time, with very minimal maintenance. I recently consulted on it, and sure enough it's still ticking al
286.
by girvo 6mo ago
Apparently if you believe them it was a “test”
287.
by girvo 6mo ago
According to https://claude.com/pricing it no longer is. They’ve explicitly updated that page to exclude it (it was on there before)
288.
by girvo 6mo ago
Go to the pricing page: https://claude.com/pricing
289.
by girvo 6mo ago
Claude Pro no longer includes Claude Code
290.
by girvo 6mo ago
Anthropic literally just removed Claude Code from their Pro plan today, so you're even more right than you know.
291.
by girvo 6mo ago
I assume someone knows someone, backroom deal perhaps? I'm not sure either, when Cursor has a lot of risk and not that much moat.
292.
by girvo 6mo ago
Exactly the same thing as coding, where “we don’t want to lower the quality bar” platitudes are repeated, while in actual fact that’s exactly what they want from us with AI output, consequences be damned. The stock market will reward us for
293.
by girvo 6mo ago
Indeed. And that’s with Anthropic hiding reading traces unlike these other comparisons.
294.
by girvo 6mo ago
You can use GLM’s coding plan in Pi, just use the anthropic API instead of the OpenAI compatible one they give.
295.
by girvo 6mo ago
Claude Design into PenPot via its MCP was a really neat flow, for something generic looking anyway. With the correct prompts and it even built out reusable PenPot components and design system tokens etc
296.
by girvo 6mo ago
> And, if you are using AI, the work is still treated as less valuable, even if it brings value. This is true, some people do react this way. But I've noticed its far more pronounced if people try to hide the AI usage for code. AI
297.
by girvo 6mo ago
GLM 5.1 is absolutely on par with Sonnet 4.5, sometimes better in practice (it holds abstractions over longer context windows better) It’s about the only one that is at that level though to be fair. They’re all still useful, still!
298.
by girvo 6mo ago
I didn’t see the old one, but the current page made it clear straight away!
299.
by girvo 6mo ago
FedRAMP will be the death of me.
300.
by girvo 6mo ago
On the other hand, teaching taste is quite hard, and is what people respond to and what designers learn. Teaching programming is a bit of mostly solved problem, today anyway.

More ›