NitpickLawyer

6 ms·

61.
by NitpickLawyer 2mo ago
> A abusive software supplier would be fired and replaced. Oracle has joined the chat :) (and them joining the chat is an additional clause to the initial contract, billed at 300% hourly rate, with a minimum block of 900 man-hours)
62.
by NitpickLawyer 2mo ago
It depends on the actual implementation. There really isn't anything stopping them from including a "pre system prompt" or "root prompt" or whatever they want to call it, before your system prompt, even for API call
63.
by NitpickLawyer 2mo ago
> even if you require and get > 99.9% accuracy [...] there will be anecdotes of patients getting wrong results given that millions of people go to the doctor a year. This is known as the long tail problem in ML, and it's a realit
64.
by NitpickLawyer 2mo ago
Only relevant benchmarks are those you make yourself, targeted specifically for your workflows. Anything else is just number go up on a pretty graph, and every model out there is probably benchmaxxed to hell on the public ones anyway. Keep
65.
by NitpickLawyer 2mo ago
> The absolute worst market time to etch a model to a chip is right now Slightly disagree. It really depends on the price-point at which they can do that etching. ~1k usd / ~30B model in a hdd-sized case that fits on your desk? I&#x
66.
by NitpickLawyer 2mo ago
You're generally correct, but in this particular instance it does actually make sense to call it the dark side. Dark, as in "quiet" from RF interference. Just like some areas on Earth, near radio observatories, are declared &
67.
by NitpickLawyer 2mo ago
> is not the addictive algorithmic feeds. The Internet has been addictive since forever. Algorithmic feeds take the blame recently because simple explanation makes brain feel good, but they are at most a catalyst. We used to hog the phon
68.
by NitpickLawyer 2mo ago
I'm confused. You said "he doesn't use AI" and I took that as a general "he never used AI". If I was mistaken then ignore this whole thread, that's my bad.
69.
by NitpickLawyer 2mo ago
> Why? I really don't understand this. Why can't a tech writer take a deep but neutral intellectual interest in something? Isn't it important that some do? Do people have to be stakeholders or clearly on one given team, pr
70.
by NitpickLawyer 2mo ago
> He is not using AI That's ... even worse? So we're all here in the comments trying to figure out what the author means, and what their overall point is, while clearly they don't even use the damn thing? Oof... What a was
71.
by NitpickLawyer 2mo ago
Yesterday I tried both as well. I do a quick "explain this repo" + "any security issues" convo to do a "vibe check" on the models and make sure everything works w/ serving and harness. Both qwen and glimme
72.
by NitpickLawyer 2mo ago
> Does anyone know about this technique? How does that work? The term you are looking for is Curriculum Learning. There are several papers exploring this. From memory, it leads to faster initial loss drop on some experiments, it can be g
73.
by NitpickLawyer 2mo ago
The way they explain it implies they're using this at the sampler level and not trained into the weights themselves. So unless you're using an inference library that does this, the open models will not have this kind of a watermar
74.
by NitpickLawyer 2mo ago
I wonder if people still think that the hf incident was an OpenAI "marketing stunt". Out of everyone involved in this kerfuffle, hf and zai are riding the PR wave, while oAI is getting laughed at, having congressional letters thro
75.
by NitpickLawyer 2mo ago
> Beats Opus 4.7 Max I'm a huge open model fan, and have used them since forever, even have daily drivers for on-prem dev, but no. They do not beat opus on real-world usage. Qwen models are impressively good for what they are, are &
76.
by NitpickLawyer 2mo ago
Yeah, could be. The article linked here had very few details on the generators, besides calling it "hybrid electric" which led me to google a bit and found those ranges, but not additional info on how it'll be used. I guess w
77.
by NitpickLawyer 2mo ago
> Labs have already used up internet-scale data Despite this being the topic du jour of 2025, it was never true. Most of the "we've hit a wall with data" came from communicators / media and not researchers. It got po
78.
by NitpickLawyer 2mo ago
"I think it would be a good idea"
79.
by NitpickLawyer 2mo ago
> It uses electric motors for zero-emission short hops up to 124 miles and backup reserve generators for an extended hybrid range up to 300–500 miles. I was wondering how they're gonna deal with alternate airports and "fuel&quo
80.
by NitpickLawyer 2mo ago
There's a huge list of "low hanging fruit" that are at the same time niche enough or low-user enough that no one has taken a stab at it so far. Or, if they did, they were expensive / captive / terrible ux solutions.
81.
by NitpickLawyer 2mo ago
> 100 instances of the same LLM may converge on the same result as 10. Not in the highly verifiable domains. There you can take it from say 80-90% maj@x to 99% pass@n. Math, some parts of programming and cybersec are examples of highly v
82.
by NitpickLawyer 2mo ago
3.8-27B is confirmed for Friday. They didn't release their whole 4B-400B range of models since 3.5. And 3.6 only got 27B and 35B MoE. So yeah, slowing down, but not completely out of the small model game.
83.
by NitpickLawyer 2mo ago
unshloth started as a finetuning library with lots of optimisations so you could finetune on lower end hardware. Kind of OGs of the local community. Started by two brothers Michael and Daniel(?) a math wiz and a community builder/commu
84.
by NitpickLawyer 2mo ago
It really depends. It used to be easier to have a rule of thumb, but now it's not clear anymore. Now there are a lot of things to consider, such as a model's kv efficiency (how much context you can fit), MoE v. dense, QAT or not (
85.
by NitpickLawyer 2mo ago
Should have worded that better, my bad.
86.
by NitpickLawyer 2mo ago
Sure, but that's for "personal" serving. I meant for 3rd party providers. Usually we get a good indication on what it costs to host this, as the prices settle on open router. That's why I said it's tougher to serve
87.
by NitpickLawyer 2mo ago
No, Kimi k3 is 2.8T params. This is 2.4T params but ~5TB weights because it was released in bf16 and ~2.5TB for the fp8 version. Kimi k3 launched with QAT 4bit, so ~1.5TB weights.
88.
by NitpickLawyer 2mo ago
Supposedly this is a Kimi k3 rival. Bit of a chonker, especially since they only released bf16 and fp8. So at launch this will be harder to serve than k3. No QAT on q4 means that someone with deep pockets (nvda?) will have to quant it, with
89.
by NitpickLawyer 2mo ago
Breaking a platform's ToS is a civil contract violation, not a criminal offence. Stealing is. Potato, avocado.
90.
by NitpickLawyer 2mo ago
Sure, but the current one is charged per token in & token out. Not per completion / task / hour / whatever. You can't charge per token and then say "you stole that token". Again, they can unilaterally decid

More ›