oh_no

5.6 Luna costs far less and benchmarks far better, have you compared for this task?

buddy, they're on enterprise plans paying per token

i'm a little confused by this, should most piles of company documentation look pretty similar? what are we getting by tuning at the org level?

what if you have a bunch of teams or apps that have different documentation patterns?

how much does this degrade over time, it beats leading models with that static data set but clearly this edge will degrade with data drift, how quickly does that happen?

also, any "this is 100x cheaper" blogpost means nothing if not discussing TCO (I know your team didn't write this.) I don't care what inference costs are if I don't know training/overhead costs. what's the breakeven point. and again, how long is this RAG stack going to be worth keeping, you beat 5.6 Luna but at some point un-tuned models will beat you, so this is a temporary solution that needs to be re-upped at some point. benchmarks against data drift would help there

few days old but want to flag that there are zero apples:apples comparisons on this press release. set aside the benchmark vs older models they're testing their model with access to their internal legal research data vs open internet search on other models (and reasoning turned off on 5.5 for some reason?)

a bad product and an embarrassing press release

where are you seeing cheap Kimi? pricing I've seen is the same across the board (presumably due to licensing terms) and is in the Terra range.

I pretty strongly disagree about comparing this to Kimi and GLM, 5.2 was a big price hike for Chinese models, and Kimi K3 was a big price hike to that. K3 was within spitting distance of OpenAI pricing (more expensive than short context Terra, less than long context). And that's after months of OpenAI/Anthropic prices going up.

Now we have an American lab drastically cutting a price, feels like this is the opposite of that trend.

reading some of the comments i was expecting something really high end and polished, but yeah, this is something you could with a headset 1/10th the cost

based on pricing I think it's safe to say they're different. why would they charge half price when fable has been very popular?

seeing a jump this big is not a great sign for the continuing value of a benchmark

I have also thought about this a lot but have less developed views on fiction (short version: I don't currently want to read an LLM-written novel, but I expect that will change in the future and there will be some I enjoy)

But I want to say that non-fiction books are a lot more resistant to LLMs than you might think. I'm reading two non-fiction books currently. And both (https://www.goodreads.com/en/book/show/17801.Underground, https://www.goodreads.com/en/book/show/44824581-a-k-pop-live) could not have been written by an LLM. One is based on in-person interviews that were hard to set up, and the is based on a log of sources a number of which are going to live events.

Not every book fits this pattern of course. But even a more straight forward book (https://www.goodreads.com/series/139929-designers-dragons) will often rely on sources that are not easily accessed by an LLM, they may not be digitized, or they may not be accessible in your common digital collections.

I 100% think LLMs can do good synthesis work, don't really think we have the tooling/checking in place for them to write a competent non-fiction book yet, but that will certainly happen, but I still think the best and most interesting non-fiction remains firmly in the realm of humans.

(Also, a lot of non-fiction is memoirs, travelogues, etc. which LLM can't really replace unless you're dictating to it.)

GLM 5.2 and Kimi 3 both had huge API pricing jumps (GLM most expensive chinese model, by a lot, then then Kimi 3 a lot higher than that). the cost advantage is rapidly decaying. the oft-cited cost per task makes Kimi look good vs Anthropic but even, or slightly vs OpenAI

it has kimi 2.5, which isn't to say this will show up but who knows

i don't think so, i think it's 50% what work people are doing, 50% vibes. my experience with 5.5 is i like it more and get better results than 4.8/fable. which isn't to say i think it's a strictly better model, just been working better for me.

top of the line SSDs now eclipse DDR3 throughput, but DDR3 should retain a large edge in latency of orders of magnitude

absolutely no idea how useful any of that would be and what kind of latency degradation going through whatever adapter would cause

Foundation didn't create the dataset, just the framework for volunteers to do the work.

so i think you're a bit off. it's s/g but g is legit accounts who want to buy the steam machine.

we could say it's 5000 scalper accounts, and 50000000 gamer accounts. but it's not 5000/50000000, it's like 4500/20000. which isn't bad! but scalpers will still be way over-represented, because they'll be trying to buy it when most steam accounts won't.

now one fuzz factor is the queue system, as you're not putting down money to get in line i expect a lot of people who wouldn't otherwise sign up will, in case they decide to buy one when given the chance. so we might have 40000 gamer sign ups, but only 50% will pull the trigger. this also gives scalpers an out should the resale not be worth it.

(obviously all numbers made up)

except google does respect robots.txt so you do have a choice?

interesting that GPT Image-2 managed to 2-shot this with thinking turned on, I didn't save a copy and it disappeared from my window but I first got a failure very similar to the one in the article, but it saw the issue and said it was going to use a reference image, after which it came out with https://i.imgur.com/hlWpQNT.jpeg

1- Getting OpenAI's models in Azure with no license fee is pretty nice. 2- Microsoft owns ~15-27% of OpenAI, if the agreement was hurting OpenAI more than it was helping Microsoft, seems reasonable to change the terms.

yes but as far as i know gpt tokenizer is about the same as opus 4.6's, where 4.7 is seeing something in the ballpark of a 30% increase. this should still be cheaper even disregarding the concerns around 4.7 thinking burning tokens

oops, thanks. i had just been looking at their api docs

what's the source on that?

D&D is not a board game, nor is Warhammer 40k.

Tabletop games is already used as the broader category (covering board games, role playing games, miniatures games, and whatever I'm forgetting).

So I think this article is a bit misleading, he did not invent the polyhedral dice, he just made them better. "He was the first to create polyhedral dice for the U.S. market" is a weird sentence, I'm not sure what it's going for, but I think it's referring to the fact that early D&D dice were I believe imported, but I forget the details.

One bit I love from the early history of Gamescience is he didn't have the capital to make a full D&D set off the bat, so he'd get one dice mold made, release that one, then take the profits to make the next mold. Forget which was first but I think the d4 was early.