Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

GLM-5.1: Towards Long-Horizon Tasks

z.ai

Post reply on HN
Print view View original

271–280 of 285 posts

Re: GLM-5.1: Towards Long-Horizon Tasks

#271

Post by usef- »

usef- wrote: post #251
ReptileMan wrote: post #237

Earlier quoted context omitted.

It is easy to get 30B when you resell something you buy for 50B

The proverbial "50B" is investment in next year's model. The current model cost under "30B", and therefore "is profitable". It is a bet on scaling, yes, but that's been common throughout the industry (see, eg, Amazon not being profitable for many years but building infrastructure)

Also see the Dario interview with Dwarkesh:

> If every year we predict exactly what the demand is going to be, we’ll be profitable every year. Because spending 50% of your compute on research, roughly, plus a gross margin that’s higher than 50% and correct demand prediction leads to profit. That’s the profitable business model that I think is kind of there, but obscured by these building ahead and prediction errors.

(a lot more at the link)

https://www.dwarkesh.com/p/dario-amodei-2?open=false#%C2%A70...

Re: GLM-5.1: Towards Long-Horizon Tasks

#273

Post by BoorishBears »

jaggs wrote: post #268
BoorishBears wrote: post #266

Earlier quoted context omitted.

This one's even more interesting https://aibenchy.com/compare/anthropic-claude-opus-4-6-mediu... Who knew Anthropic was this far behind???

Yeah, but actually that's not a good look. Anyone who's used Gemini will know how random it is in terms of getting anything serious done, compared to the rock solid opus experience.

Their benchmark is chock-full of things like that: It's deeply flawed and is essentially rating how LLMs perform if you exert yourself trying to hold them entirely the wrong way.

Re: GLM-5.1: Towards Long-Horizon Tasks

#274

Post by adrian_b »

uvu wrote: post #197
Ms-J wrote: post #196

Z.ai and their GLM models are pretty low quality. I've been testing it for awhile now since it seemed to have potential as a local model. With this new update it still cannot parse simple, test PDFs correctly. It inconsistently tells me that the value in the name field in the document is incorrect, and has the name reversed to put the last name first. Or that a date is wrong as it's in the past/future, when it is not…

Completely agree with this statement "Z.ai and their GLM models are pretty low quality." I have been trying out and it's kind of useless compare to SOTA models.

I do not doubt your experience, but such statements should always be qualified by specifying the kind of tasks for which you have tried the models.

For all existing models, including for all SOTA models, you can find contradictory statements, that they suck and that they are great.

It is very likely that all these statements are true simultaneously, because each model may succeed for some tasks and fail for others, so without specifying the tested tasks any claim that a model was good or bad is worthless.

Re: GLM-5.1: Towards Long-Horizon Tasks

#275

Post by adrian_b »

fireant wrote: post #272
adrian_b wrote: post #219

Earlier quoted context omitted.

The point of doing local inference with huge models stored on an SSD is to do it free, even if slow.

You are just trading opex for capex. Local GPUs aren't free.

True, but this is not only a trade-off between opex and capex.

Local inference using open weight models provides guaranteed performance which will remain stable over time, and be available at any moment.

As many current HN threads show, depending on external AI inference providers is extremely risky, as their performance can be degraded unpredictably at any time or their prices can be raised at any time, equally unpredictably.

Being dependent on a subscription for your programming workflow is a huge bet, that you will gain more from a slightly higher quality of the proprietary models than you will lose if the service will be degraded in the future.

As the recent history has shown, many have already lost this bet.

I am not a gambler, so I have made my choice, which is local AI inference, using a variety of models depending on the task, i.e. both small models completely executable on relatively cheap GPUs (like the new Intel GPUs), medium models that need e.g. 128 GB on a CPU, and huge models that must be stored on fast SSDs (e.g. interleaved on multiple PCIe 5.0 SSDs).

Such a strategy is achievable with a modest capex, in the lower half of the 4-digit range.

Re: GLM-5.1: Towards Long-Horizon Tasks

#276

Post by monooso »

nl wrote: post #253
monooso wrote: post #232

Earlier quoted context omitted.

> Anthropic is up to $30B annual recurring revenue. I wish I had failing business models like that. And profit? A company can have $300B annual revenue, and still be a failing business if it's making a loss. Somewhere along the line we seem to have forgotten this basic fact. Eventually there will be no more rounds of funding to feed the fire.

Anthropic has raised $64B in total since they were founded. Even if you say we are going to measure profit in the very special hacker news way of looking at money taken in from customer revenue against money invested and we say they can't do things like counting building data centers or buying GPUs as capital expenses and instead have to count them against profit then in 2 years time they will have made more money th…

The numbers are indeed extraordinary. But I still don't agree with this statement.

> ...in 2 years time they will have made more money than they have taken in investment.

In two years' time they will have generated more revenue, that is not the same as "making money."

Re: GLM-5.1: Towards Long-Horizon Tasks

#277

Post by airstrike »

plandis wrote: post #72
airstrike wrote: post #29

Earlier quoted context omitted.

100k tokens it's basically nothing these days. Claude Opus 4.6M with 1M context windows is just a different ball game

Claude Opus can use a 1M context window but I’ve found it to degrade significantly past 250k in practice.

250k is still massively more than 100k and 1M prevents it from having to compact

Re: GLM-5.1: Towards Long-Horizon Tasks

#279

Post by Ms-J »

CamperBob2 wrote: post #257
orbital-decay wrote: post #223

Earlier quoted context omitted.

>by the way, which is the best uncensored model at the moment? There are no such models, depending on your definition of censorship. If you're referring to abliteration and similar automated techniques, they're snake oil.

That is absolutely not the case. Try HauHauCS's Qwen 3.5 models. They don't refuse anything, and they don't lose a noticeable amount of capability.

Thanks for the tip.

That user has produced other interesting models, including aggressively uncensored variants which claim 0 refusals.

I will definitely try it as there can never be too many uncensored AI implementations.

Re: GLM-5.1: Towards Long-Horizon Tasks

#280

Post by orbital-decay »

CamperBob2 wrote: post #257
orbital-decay wrote: post #223

Earlier quoted context omitted.

>by the way, which is the best uncensored model at the moment? There are no such models, depending on your definition of censorship. If you're referring to abliteration and similar automated techniques, they're snake oil.

That is absolutely not the case. Try HauHauCS's Qwen 3.5 models. They don't refuse anything, and they don't lose a noticeable amount of capability.

Refusal training is only one part of censorship (hence "depending on your definition"). Most permanent biases baked in by the devs are impossible to correct automatically.
Post reply on HN