girvo

Applying Heretic to your model weights is a very straightforward thing, though.

Its possible (though very hard to do well, and needs a fair bit of compute) to improve the local models we have, too. Continued training is possible as time goes on.

I still remember when TravisCI giving free CI to FOSS projects was revolutionary

I said last year to people that I swear BDD is going to make a comeback. Hasn’t yet, but I still reckon it might!

And reads like AI generated text, to boot. But then that might just be because it’s marketing, and LLMs write in non stop marketing speak

I use tabs as ephemeral bookmarks, yeah

Maybe there’s a better way

This is how I’ve got very very good results from the latest models. If it can follow what I want and how I build, it does very well. If it can’t, it’s more likely to go off the rails and do something brittle or nonsense.

It’s getting a lot easier these days. The microcontrollers these are built on are not and have never been actually secure. But I know that because I’ve built firmware for them in production, I guess

They’re very much trained to do this, in fact. Even Step 3.5 Flash, a medium sized open model from way earlier this year is trained to use python to calculate answers on the fly :)

As someone who was a die hard Xbox fan since the original all the way up to the One X… Microsoft are just terrible. Getting rid of exclusives, constantly focusing on the wrong things, it’s just a mess. Terrible leadership.

Scientific calculators are surprisingly hard to get right. SpeedCrunch itself just fixed some calculation errors (usually its order of operations based, but not always). I’d be interested to see how accurate yours is, though I also wonder if there are lower level libraries to leverage to make it a moot point

That is sort of the point I thought? There’s no low skill exploit for the script kiddies to use even with AI, otherwise they’d happily use it

The pros have RCEs, but won’t burn it for something this small.

It’s a real vulnerability, but a complex one

> or have people really just started putting out slop everywhere due too much usage of AI?

Yes, they really have been. It’s exhausting at work.

Yes, but in practice it’s more interesting than that

The massively lopsided compute makes speculative execution techniques (MTP/DFlash etc) way more effective, which makes modern MoE models quite usable

I’m running Qwen 3.8 Flash Next with Nota AI’s conservative NVFP4 quant at 30tk/s lowest decode, usually more around 45-55 for coding, and with prompt processing at a solid 2000tk/s

And that’s without very optimised kernels or anything, mostly standard vLLM

No, the AMD ones he’s talking about allow the GPU to address that memory as well: they’re “APU’s” that share system RAM as GPU memory

That said the compute power of the GB10 is awesome, my DGX Spark-like has been better than I expected.