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

GLM-4.5: Reasoning, Coding, and Agentic Abililties

z.ai

Post reply on HN
Print view View original

141–150 of 153 posts

Re: GLM-4.5: Reasoning, Coding, and Agentic Abililties

#141

Post by gojomo »

ed wrote: post #95
vineyardmike wrote: post #81

Earlier quoted context omitted.

There was a recent paper that showed you can spread model’s behavior through training on outputs, even if you don’t directly include obvious markers of the behavior. It’s totally plausible that training off Claude’s outputs subtly affected GLM into mentioning “Claude” even if they don’t include the direct tokens very often. https://alignment.anthropic.com/2025/subliminal-learning/

Subliminal learning happens when the teacher and student models share a common base model, which is unlikely to be the case here

Given that other work shows that models often converge on similar internal representations, I'd not be surprised if there were close analogues of 'subliminal learning' that don't require shared-ancestor-base-model, just enough overlap in training material.

Further, "enough" training from another model's outputs – de facto 'distillation' – is likely to have similar effects as starting from a common base model, just "from thge other direction".

(Finally: some of the more nationalistic-paranoid observers seem to think Chinese labs have relied on exfiltrated weights from US entities. I don't personally think that'd be a likely or necessary contributor to Z.ai & others' successes, the mere appearance of this occasional "I am Claude" answer is sure to fuel further armchair belief in those theories.)

Re: GLM-4.5: Reasoning, Coding, and Agentic Abililties

#143

Post by polotics »

cheema33 wrote: post #77
polotics wrote: post #73

so sad, i had a few questions about the CCP, China's constitution, and very quickly on a very factual non challenging question:... Question: what proportion of the NPC's seats is held by the CCP? GLM-4.5 Uh-oh! There was an issue with the response. Content Security Warning: The content may contain inappropriate content. I could see some decent thinking in gray, but bam, all hidden in red. How many of the NPC's seats…

> so sad, i had a few questions about the CCP, China's constitution, and very quickly on a very factual non challenging question:... It is a Chinese model. It censors. Next you will tell us that Pope is religious. And water is wet. Did you make any effort to test its coding abilities and compare it with other models in the same price range?

Thank you so much for the amazing and reusable example of whataboutism. And please do not insult the Chinese by equating Chinese with censorship, I have too much respect for the intelligence and worth of this civilised people to not hold them to the same standards as the best of the best. Saying Chinese in no way implies any censorship as a necessary condition in my mind.

Re: GLM-4.5: Reasoning, Coding, and Agentic Abililties

#144

Post by SkyPuncher »

princealiiiii wrote: post #37
gwd wrote: post #35

Earlier quoted context omitted.

I wonder if they're falling back to the Claude API when they're over capacity?

I asked why it said it was Claude, and it said it made a mistake, it's actually GLM. I don't think it's a routing issue.

LLMs don’t know who they are.

This comes up all the time on Cursor forums. People gripe that their premium Sonnet 4 Max requests say they’re 3.5.

Realistically, the LLMs just don’t know who they are.

Re: GLM-4.5: Reasoning, Coding, and Agentic Abililties

#145

Post by stx5 »

oc1 wrote: post #138
stx5 wrote: post #112

Earlier quoted context omitted.

oh really？ China considers itself a democracy with its own characteristics. According to China's White Paper "China: Democracy That Works" published by the State Council Information Office, China practices "whole-process people's democracy," which integrates process and results democracy, procedural and substantive democracy, direct and indirect democracy, and people's democracy and state will. China's democracy is d…

Thanks for clarifying. Actually Chinas democracy sounds really nice now. Would love to try it out where people are actually the masters of the country :) (Btw this chinese ai-propaganda-bot parent is the best example why we need to keep china out of our digital infrastructure and ban apps like tiktok. They are already brainwashing western kids with shit like this fake user above. )

Telling you a joke that's very popular in China, we call it a 'Soviet Joke'. --- American: "Why did you come to America?" Soviet: "I came to learn your brainwashing techniques." American: "Nonsense! We're a free and democratic country—no brainwashing here!" Soviet: "Exactly! That's precisely what I want to learn!"

Re: GLM-4.5: Reasoning, Coding, and Agentic Abililties

#146

Post by majormajor »

littlestymaar wrote: post #140
majormajor wrote: post #114

Earlier quoted context omitted.

Take a look at the API calls you'd use to build your own chatbot on top of any of the available models. Like https://docs.anthropic.com/en/api/messages or https://platform.openai.com/docs/api-reference/chat - you send the message history each time. You can even lie about that message history! You can utilize caching like https://platform.openai.com/docs/guides/prompt-caching and note that "Cache hits are only possibl…

Exactly, but caching doesn't work if you switch between providers in the middle of the conversation, which is my entire point.

If you're selectively faking things you don't care. You may not even be aware because the caching is transparent to you and you send the whole set of messages to the system each time either way. From the perspective of the person faking their model to look better than it is, it requires no special implementation changes.

And if you're faking your model to look better than it is, you probably aren't sending every call out to the paid 3rd party, you're more likely intentionally only using it to guide your model periodically.

Re: GLM-4.5: Reasoning, Coding, and Agentic Abililties

#147

Post by littlestymaar »

majormajor wrote: post #146
littlestymaar wrote: post #140

Earlier quoted context omitted.

Exactly, but caching doesn't work if you switch between providers in the middle of the conversation, which is my entire point.

If you're selectively faking things you don't care. You may not even be aware because the caching is transparent to you and you send the whole set of messages to the system each time either way. From the perspective of the person faking their model to look better than it is, it requires no special implementation changes. And if you're faking your model to look better than it is, you probably aren't sending every call…

> because the caching is transparent to you

It isn't when you look at your invoices though.

> aren't sending every call out to the paid 3rd party, you're more likely intentionally only using it to guide your model periodically.

I'd you do that, you're going to have to pay each token multiple times: both as inferred token on your model, and as input tokens on the third party and your model.

If the conversation are long enough (I didn't do the math but I suspect they don't even need to be that long) it's going to be costlier than just using the paid model with caching.

Re: GLM-4.5: Reasoning, Coding, and Agentic Abililties

#148

Post by dash2 »

MaxPock wrote: post #115
dash2 wrote: post #113

Earlier quoted context omitted.

Slightly OT but when I hear a set of short comments like this, all agreeing with each other, I become very suspicious. It just doesn't seem like a natural conversation.

Do you think two comments agreeing that the Chinese government has mandated the teaching of AI and robotics in high schools is suspicious?

There's four, and they read like a single pro-Chinese thesis broken up into bite-size chunks.

Re: GLM-4.5: Reasoning, Coding, and Agentic Abililties

#149

Post by sunaookami »

emerald_rat903 wrote: post #13

I typed "Hello" in their chat [1] and it replied back with "Hello! I'm Claude, an AI assistant created by Anthropic. How can I help you today?" Hmmm.... [1] https://chat.z.ai/

I thought people on Hacker News knew how LLMs worked but we still get these posts every time. No, it's not "actually routed to Claude behind the scenes", it's just polluted training data.
Post reply on HN