girvo

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.

Nvidia’s QAD NVFP4 quant is close to this on Blackwell gear.

I’ve been doing this job for two decades, you don’t have to tell me that. But programming isn’t the tedious part of the job to me.

They are pretty up front about that when they hire from what I was told by someone who went through it: that you have to “be okay” with your job not exisiting in the next couple of years

And here I was convinced that they’re burritos!

I will say that explaining functors to my coworkers was always much simpler, and then naturally led to monads

I wish it was…

https://www.reuters.com/world/asia-pacific/china-proposes-ad...

They’ve technically got expedited approval for open weight models I think, but it still needs approval. And frontier level stuff gets extra scrutiny

That is genuinely shocking to me… I guess NYC is at “capacity” and has been for that long?

And yet amusingly it’s the ICs that are being replaced, not management or executives…

> they don't seem to be so sure of golang anymore. I suspect it's going to get rewritten in rust.

With how “cheap” it is to achieve now, I’m pretty sure they will too. Though things like that still worry me, core tools I rely on becoming LLM-ported blobs of code that no human understands is a bit scary. I guess I’m old fashioned like that

China just announced tightened export controls on AI models, for what it's worth.

We already have controls for that: we control access to the reagents and equipment needed. Yknow, like we always have.

Information is not dangerous. This is a ridiculous argument to me, but maybe my chemistry degree has me biased.

It’s absolutely addictive, anyone who says otherwise is suspect

What I take issue with is the original sentence saying it’s “bad for you”, I was expecting new science showing that it’s, well, bad for you rather than “just” addictive

They are? In all my evals at work, Jev seems to win, despite me wanting the open weight versions to.

I imagine if I fine tuned one of the open weight models for my use case it would be possible to beat it, though, but as a general drop in decision model, I’ve not really seen much that beats it especially for the cost

But the first version was “mostly handwritten”?

That’s what I’m seeing too. It’s depressing. I hope I can pay off my house in time to go do something else.

And “bad for you” is left completely unexamined. It feels very dishonest to me, and it’s a shame because I was hoping to learn something new.

It is? Am I getting a different link to you, or is my reading comprehension just terrible at 9.30pm or something.

https://psychotechnology.substack.com/p/a-love-song-to-nicot...

It says the exact opposite to me?

> It turns out, actually, nicotine is addictive, and is bad for you even if you don't smoke it, as you probably already obviously knew.

This is not the point, but I’m very confused by the link that is linked as evidence on that line? It doesn’t match the sentence it’s attached to

Link in question: https://psychotechnology.substack.com/p/a-love-song-to-nicot...

> which Apple finally put in their new device

Not to take away from what you're saying, but I have a 2022 Apple TV 4K with a gigabit ethernet port.

You just had to buy the higher storage model, weirdly.

Model MN893X/A

No, and anyone who says otherwise is being silly.

It’s still damned capable though, I’ve done a crazy amount of good work with 5.3 Flash.

My 2007 Suzuki Swift felt horrible at 130km/h or above!

For now. It is definitely true right now, there's a notable difference between my coworkers depending on their skill driving these models: it mostly seems to be correlated with those who are willing to build their own tools, now that the cost to do so is basically zero.

But the models are getting so good and so fast, that I'm starting to wonder if that will stay true forever... probably not, if I had to guess. I'm just trying to pay off my house as quickly as I can...

Would they not have RL'd Claude such that we picked the one most likely to not complain and emit those stop tokens? Otherwise they wouldn't be useful. Seems flawed...

No, that’s just standard AI generated nonsense marketing prose

Sure, if you think so, but discussing or arguing for/against those points here is far more interesting than the done-to-death stochastic parrot stuff.

I really wish the article in question hadn’t used that phrase, because you’re right. But the rest of the article is also correct in the dangers facing us from this technology as an accelerant regardless, and I think it’s much more interesting to talk about that: but this entire comment thread will dunk on the stochastic parrot paragraphs instead.

The funny thing is though, I think it’s an in group signalling mechanism. You need the anti-AI crowd to know that you’re anti-AI, and this is a rapid way of doing so, to get it upvoted here.

> they also commit to keeping private

Reading between the lines:

"The avid reader of my blog might have noticed that I had previously written two posts that have since been removed. Everyone else might now be wondering which game I am talking about. To both of you I can only say that corporate America was here to ruin our fun."

Keeping it private likely wasn't the original plan...

That’s fascinating, but not that surprising to me. We act like quantisation is free “Q8 is basically lossless” is often said in the local LLM community, but it really isn’t. The trade offs are worth it, personally, and the damage to coding ability seems low: decision model approaches are stricter though

Super cool finding!

Counterpoint: my work has already allowed us to call and test Jev. Those others? Who knows when, if ever.