vessenes

18 ms·

361.
by vessenes 6mo ago
I do not think this is true. You will continue to get smaller, cheaper-to-host models in the free tier that are distilled from current and former frontier models. They will continue to improve, but I’d be very surprised if, e.g., 5.4-mini (
362.
by vessenes 6mo ago
Hi Mike! Long time - super nice to see your name in my HN feed. I’ll fight you on profit. The major labs are super profitable. If you replace “profitable today” with “cashflow positive today” then I think you’re correct. They are clearly no
363.
by vessenes 6mo ago
I agree that a full linux distro compile as a matter of practice is a waste of time. But, doing it a few times is good if you want to understand your tools. I don’t believe that top tier engineers just skip learning things because they migh
364.
by vessenes 6mo ago
Do not use the free plan. It is not good.
365.
by vessenes 6mo ago
Interesting little bit of history; this pre-Chinchilla paper proposed MoE training longer would improve performance. Good idea. They also proposed using a hash function to choose experts rather than training a routing layer and showed it ma
366.
by vessenes 6mo ago
‘Have to’ and ‘every other’ are both doing so much work here that I think your worldview on this is likely just incorrect. The decisions to mobilize a large rural base toward manufacturing and the central bank goals to keep the yuan cheap a
367.
by vessenes 6mo ago
This is only good advice if you don’t have the need to understand what’s happening on the edge of the frontier. If you do, then you’ll lose on compounding the knowledge from staying engaged with the major developments.
368.
by vessenes 6mo ago
I’ll note that it’s common and dangerous , in that there’s a generation of engineers who are at risk of leading each-other astray as to the economics and therefore probability distribution of outcomes for some firms that will massively imp
369.
by vessenes 6mo ago
That story changed my mind on uploading a connectome. Super dark, super brilliant.
370.
by vessenes 6mo ago
Yay. 5.4 was a frustrating model - moments of extreme intelligence (I liked it very much for code review) - but also a sort of idiocy/literalism that made it very unsuited for prompting in a vague sense. I also found its openclaw engag
371.
by vessenes 6mo ago
Buried in here is the very nice idea of an AI dividend for CLI users - as we keep designing interfaces for LLMs directly, those interfaces are text-oriented, and afford us a look at a future with the command line being central. I like that.
372.
by vessenes 6mo ago
Not only is it almost certainly compute (“services”) it’s likely priced at Anthropic rack-rate, or at least what Cursor’s been paying Anthropic. The cluster’s already paid for, so likely in the $2B range for operating cash needs. Not more t
373.
by vessenes 6mo ago
This just feels like unbridled, blind distaste for Anthropic. Everyone's entitle to their opinion, but the blog post doesn't add much to the internet's general knowledge in my opinion. The smoking gun he wants to dig in on is
374.
by vessenes 6mo ago
why is it that speculative decoding lowers quality? My understanding of it is that you use a small/distilled fast model to predict next token - when it doesn't match, you generate more. Checking against the large model is quick.
375.
by vessenes 6mo ago
Any materially false public statement by one of the foundation lab CEOs is a huge foot fault. I'm not saying they would never lie, but it would be a very, very dumb thing to do. That public information can be relied on by their private
376.
by vessenes 6mo ago
4.7 is super variable in my one day experience - it occasionally just nails a task. Then I'm back to arguing with it like it's 2023.
377.
by vessenes 6mo ago
I like the final idea, but I think I'd like it better on a log scale -- so many of the colors get basically one or two pixels of height.
378.
by vessenes 6mo ago
I mean, my openclaw instance was billing $200 a day for Opus after they banned using the max subscription. I think a fair amount of that was not useful use of Opus; so routing is the bigger problem. but, that sort of adds up, you know! At
379.
by vessenes 6mo ago
Google china solar deployments to read about the logistics end of it
380.
by vessenes 6mo ago
With Anthropic buying compute in dark alleys I’d assume that day is coming..
381.
by vessenes 6mo ago
I don't think so. GB200 prices are GOING UP. A100s are still expensive. This implies massive utilization and demand, no? These machines are not sitting idle, or prices would drop in the very competitive hyperscaler environment.
382.
by vessenes 6mo ago
It's ... really long, according to Dylan Patel on the Dwarkesh Podcast. The supply chain is extremely deep and complex.
383.
by vessenes 6mo ago
You misunderstand. "I built a ship to go to the Indies and bring back tea." "Bro, the ship cost 100,000 pounds sterling and only brought back 50,000 pounds of tea. I don't care if you paid 12,500 pounds for the tea itsel
384.
by vessenes 6mo ago
ASML only makes a certain number of machines a year that can do extreme ultra-violet lithography. Also - turbine blades limit power, according to Elon. Between them - we cannot chip fabs past a certain rate, and we cannot stand up the datac
385.
by vessenes 6mo ago
It seems very possible that we have at least five years of real limitations on compute coming up. Maybe ten, depending on ASML. I wonder what an overshoot looks like. I also wonder if there might be room for new entrants in a compute-scarce
386.
by vessenes 6mo ago
Mm, don't you just need to retrain the embedding layer for the new tokenizer? I agree it seems likely this is like a stopgap new model release or a distillation of mythos or something while they get a better mythos release in place. Bu
387.
by vessenes 6mo ago
there's a line inside claude code mentioning to care about this. combined with new stronger instruction following behavior, you're going to be seeing it a lot unless you patch it out. or wait for a fix.
388.
by vessenes 6mo ago
Oh don't worry. They have Mythos and the extremely dystopian-named "helpful only" series which is internal only and can do all the things.
389.
by vessenes 6mo ago
Interesting. The model card mentions 4.7 is much more attentive to these instructions and suggests you will need to review and soften or remove or focus them at times.
390.
by vessenes 6mo ago
I use and pay for both. Currently I use 4.6 (well as of yesterday) to do broad strokes creation. I use codex for audit. Generally first two or three audit cycles claude completes. There is often a subtlety that only codex can fix, but I usu

More ›