nl

12 ms·

91.
by nl 1mo ago
I agree the API prices are profitable (or at least break even) on an operational basis. I don't think that means the subscriptions aren't subsidized though! I expect they are, at a carefully calibrated rate. They want to maximize
92.
by nl 1mo ago
Why this judgemental tone? One person's "asking a question" is another's "demanding an answer" Clearly both search engines and chat interfaces are merging, and clearly they should be because they fulfill the sa
93.
by nl 1mo ago
Yes absolutely. I should have pointed out that I always prefer Sol for OpenSCAD for example.
94.
by nl 1mo ago
AFAIK there is nothing in the ToS that forbids an employee paying for the 20x, $200/month plan. I could be wrong about this in which case it'd be useful to have a link the clause. I think multiple plans are against the ToS, but
95.
by nl 1mo ago
Fair. I was chatting to the maintainers on 2 of the 3 repos this affected and we estimated 4-6 person months work of we were hand coding. The PRs on those 2 repos are 48,000 lines of code (which is a problem in itself!)
96.
by nl 1mo ago
> The US government basically told them they can't sell "Fable" and so they aren't. > You're agreeing with me. No I'm not. They are selling it. I used Fable very extensively before the ban (100% usage i
97.
by nl 1mo ago
> Codex is more token efficient This is very true, especially vs Opus 5. > tends to get better results than Fable with less need for extreme token burning shenanigans like 18 hours of subagents. The strength here was less the code qua
98.
by nl 1mo ago
Yes, and 4 months ago I'd have done it this way. I'd estimate that generating the specs would have been maybe 2 weeks work? It's across 3 repos, and I'm only really familiar with one. > done in 1/5th of the time
99.
by nl 1mo ago
It's a differential privacy framework. It's a fairly large code base split across 3 repos. The good thing was that it is fairly easy to verify: we have a working (but slow) version that uses Spark, with lots of existing unit tests
100.
by nl 1mo ago
I have no idea what you think violates the TOS here? I was doing work at work on a work task using a plan paid for by work.
101.
by nl 1mo ago
I think it's actually Opus 5 that is comically slow! And yes probably that was a factor. It was a lot of code too though.
102.
by nl 1mo ago
I've retweeted and posted this here before too. I think there is probably some truth to it. Worth noting that Fable (ie, Mythos) is actually nice to interact with.
103.
by nl 1mo ago
> The US government basically told them they can't sell "Fable" and so they aren't.... They either had to lock it up (Mythos), neuter it (Fable), But this isn't what happened? It was only a couple of months ago!
104.
by nl 1mo ago
Wow the sentiment here is so negative. I'm on the $200 plan (work pays) and I also have the $20 OpenAI plan (I pay) and keep a balance on OpenRouter. There is nothing as good as Fable, not even close. I recently had it run a 18 hour
105.
by nl 2mo ago
How good is GPT-5.6 Luna! Such a cheap model, and Sonnet levels of performance.
106.
by nl 2mo ago
This would be great to be able to use via Web Midi. That size LLM will run well in a browser.
107.
by nl 2mo ago
Stacktraces are almost the ideal context for a LLM
108.
by nl 2mo ago
There's a theory going around on Twitter which goes something like this: Internal Anthropic employees have been using Mythos since February to orchestrate their (Opus) sub-agents. This works well, and subsequent RL runs have used inter
109.
by nl 2mo ago
In real world comparisons K3 is somewhere between Sonnet and Opus. Fable is just a completely different (higher) level. The long tail of tasks and queries is where you see the difference.
110.
by nl 2mo ago
We had a VMS system at Uni in the 90s and we could tell what lab someone was in via the Finger command. Apparently it caused significant load on the system.
111.
by nl 2mo ago
The open models don't really match Opus. For example I regularly do Fable+Opus agentic coding runs over 24 hours without intervention. I think I've had GLM do a run that was a few hours. That's the closest I've had an op
112.
by nl 2mo ago
It's not an Opus match. The difference is very visible in long tail applications. Exactly where you'd expect parameter count to matter.
113.
by nl 2mo ago
Yes. One reason why I think that report has confused Fable and Opus.
114.
by nl 2mo ago
It's interesting to note that the total compute forecast to be deployed in 2026-2027 is more than the amount that has been deployed during human history.
115.
by nl 2mo ago
Mythos/Fable are around 10T: > According to FT, industry estimates say Anthropic's most advanced Mythos 5 has about 8 trillion parameters and Fable 5 about 5 trillion https://www.reuters.com/technology/by
116.
by nl 2mo ago
Fable is strongly believed to be around 10T. The most conservative estimate I've seen is 8T. Eg: https://www.reuters.com/technology/bytedance-targets-mega-ai... That reports Mythos as 8T and Fable as 5T, but I thi
117.
by nl 2mo ago
Even if AI generated code isn't copyrightable (and in my other comment I point out that this probably isn't the case) the US economy still doesn't depend on that ruling. Even without being very imaginative, most non-open sour
118.
by nl 2mo ago
I think their capex will exceed their revenue, but $65B is a lot to spend on training runs! I suspect it's more likely they will spend the extra money on hardware and infrastructure (data centers) either directly or via suppliers.
119.
by nl 2mo ago
From their Twitter: "We hope to make this a permanent change to our plans, but strong demand for our models means that capacity may be tight over the coming weeks. We’ll keep you posted as things develop." https://xcanc
120.
by nl 2mo ago
What thought process gets you to that conclusion? It's clearly not right if you compare it to API prices. It's even less right if you compare to a SWE salary. Yes not using the full $200 means you aren't tokenmaxxing, but tha

More ›