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

GLM-5.2 is a step change for open agents

interconnects.ai

Post reply on HN
Print view View original

211–220 of 240 posts

Re: GLM-5.2 is a step change for open agents

#211

Post by cameldrv »

jerojero wrote: post #3

Open weight models from Chinese labs tend to be significantly cheaper. I think theyre absolutely needed. I can't afford 200 USD a month for personal use of coding AI, and I don't think such prices are reasonable for most of the world economy anyway. Not to mention US firms might be giving their employees a lot more than that. It's increasingly feeling, to me, that theres a gap building up between haves and have nots.…

Yes, but you’re paying with your data unless you’re hosting with a provider you trust or self-hosting.

Re: GLM-5.2 is a step change for open agents

#212

Post by pizzafeelsright »

nunodonato wrote: post #132
christophilus wrote: post #47

I've been working with Deepseek V4 Flash (with opencode as the harness). It's been almost indistinguishable from Codex / Claude Code for me. I'm sure I'll run into problems when I get to a stickier ticket to tackle. But so far, it's been quite good, and I find it writes straightforward code. I do think the Chinese models are good enough for an 80/20 rule use case.

it would be a really great option if it didn't lack vision

this is mcp or custom call to lowest cost model

someone did a webcam + agentic + capture of other computer bios/boot -> upload to image model -> back to agent

Re: GLM-5.2 is a step change for open agents

#213

Post by geye1234 »

Curious to hear if anyone has tried running the 2-bit or 3-bit quantization of this. With a bit of investment I may just be able to swing it locally. I already have 96GB VRAM, so with 192GB RAM, which seems to be the most one can find these days with a 4-slot motherboard, I may be in with a shot. Yes, it'd be slow, but I could give it overnight jobs. But I don't know if running at such a low quantization would make it hallucinate with only a small context.

Qwen and Gemma are great, but they need babysitting every 30 mins, which is quite a cognitive load.

Re: GLM-5.2 is a step change for open agents

#214

Post by mtlynch »

dotancohen wrote: post #206
mtlynch wrote: post #167

Earlier quoted context omitted.

Not GP, but just being smaller makes it easier to achieve reliability. Like if you're a git forge with 100 similar customers, you can likely achieve an order of magnitude better reliability than GitHub, who is trying to serve millions of customers with wildly different needs.

The problems for services such as GitHub with scaling are reducing cost per customer. That's even more pertinent when discussing inference at scale.

> The problems for services such as GitHub with scaling are reducing cost per customer. That's even more pertinent when discussing inference at scale.

I don't think that's true. When I look at GitHub's incident history,[0] it doesn't read to me like a company that's struggling to cut costs. It looks like a company that's trying to do a million things to serve a million use cases, and the growing interconnections between all those distinct services and workflows cause unexpected failures.

[0] https://www.githubstatus.com/history

Re: GLM-5.2 is a step change for open agents

#215

Post by timcobb »

59nadir wrote: post #134
timcobb wrote: post #21

Can people share their GLM and open model setups in general please? What provider do you use. Why do you trust it with serving full quality? What harness do you use? Why do you trust it not to have malware (most harnessed are TS apps). I am just trying GLM 5.1 from Nvidia build in open code would love to hear how you all do it, thanks.

> What provider do you use? 1. My own harness + Local (which usually means Qwen3.6-35B-A3B), I use this fairly often for research gathering on topics, info gathering on code bases, etc. 2. My own harness + DeepSeek v4 Flash served by DeepSeek, I added $20 quite some time ago and somehow still have $18.77 in there after I don't know how many prompts. I use this pretty often, slightly less than my local setup, it's gre…

Do you write /maintain evals? This is something I want to get into more. Otherwise I feel really blind and feel compelled to just drop money on frontier.

Re: GLM-5.2 is a step change for open agents

#216

Post by andriy_koval »

metobehonest wrote: post #205
dotancohen wrote: post #131

Earlier quoted context omitted.

What kind of off topic political ideology spam is this? Do you not think that the Chinese kill their enemies? The Chinese are genociding Uyghurs as we speak, purely for being Muslim, in numbers that dwarf any harm the US has done.

The US funds Israel and it is only those funds and military aid that keep it from collapsing unto itself. That's the state that orchestrates the largest scale genocide by a "first world" power since WW2, as recognized by the United Nations and independent organizations like the Amnesty International. https://amnesty.ca/wp-content/uploads/2024/12/Amnesty-Intern... Nothing China did comes close to this.

> as recognized by the United Nations

its not, this would require voted resolution to declare genocide. It was some report on inquiry by individuals with unknown bias.

Re: GLM-5.2 is a step change for open agents

#217

Post by matheusmoreira »

ttoinou wrote: post #196
matheusmoreira wrote: post #60

Earlier quoted context omitted.

For the record, 200 USD is around 60% of the brazilian minimum wage.

How about brazilian median software developer wage ?

According to Glassdoor statistics, brazilian developers make between 600-1600 USD per month on average. Seniors might rise above 2000 USD.

So a 200 USD subscription falls between 10% and 33% of an average brazilian developer's salary.

Re: GLM-5.2 is a step change for open agents

#218

Post by arikrahman »

tsss wrote: post #130
arikrahman wrote: post #40

Earlier quoted context omitted.

Someone else on this forum put it well, U.S. is trying to achieve AGI at all costs, while Chinese models are seeking widespread adoption.

Everyone wants widespread adoption, of course. I'm sure that China is also working on more expensive frontier intelligence models behind doors, but they're lagging behind America on that front. Going for cost-optimized open weight models is their bet to stay relevant in a market where they can't compete for the "luxury" segment. It is important for them to get a foot in the door and maintain a presence in the press t…

I didn't think of the European angle, that's something I can update in my synthesis.

Re: GLM-5.2 is a step change for open agents

#219

Post by Farmadupe »

geye1234 wrote: post #213

Curious to hear if anyone has tried running the 2-bit or 3-bit quantization of this. With a bit of investment I may just be able to swing it locally. I already have 96GB VRAM, so with 192GB RAM, which seems to be the most one can find these days with a 4-slot motherboard, I may be in with a shot. Yes, it'd be slow, but I could give it overnight jobs. But I don't know if running at such a low quantization would make i…

2 and 3 bit quants are often closer to gibberish than hallucination, and that'll happen regardless of context.

I shouldn't claim too much, I haven't tried GLM5.2 at 2/3 bit quantization, but if I were a betting man I'd put money on "useless even as a chatbot"

Re: GLM-5.2 is a step change for open agents

#220

Post by 59nadir »

timcobb wrote: post #215
59nadir wrote: post #134

Earlier quoted context omitted.

> What provider do you use? 1. My own harness + Local (which usually means Qwen3.6-35B-A3B), I use this fairly often for research gathering on topics, info gathering on code bases, etc. 2. My own harness + DeepSeek v4 Flash served by DeepSeek, I added $20 quite some time ago and somehow still have $18.77 in there after I don't know how many prompts. I use this pretty often, slightly less than my local setup, it's gre…

Do you write /maintain evals? This is something I want to get into more. Otherwise I feel really blind and feel compelled to just drop money on frontier.

Not really. I have one I made for fun where I let LLMs control a text editor called Kakoune, and then give them no other way to do things, to see how they deal with it, but that's not really a scenario I expect them to do well at.

So far most of them have done very poorly on that one, because they are all overtrained on just executing shell commands.

A former colleague of mine and I made a simple test for some baseline "Everything worth using should be able to do this pretty easily and swiftly" but that's some very minor code generation with a very straight forward, boilerplate-type pattern.

Post reply on HN