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

Qwen3-Max-Thinking

qwen.ai

Post reply on HN
Print view View original

421–430 of 450 posts

Re: Qwen3-Max-Thinking

#421

Post by b1n »

tehjoker wrote: post #415
diblasio wrote: post #48

[flagged]

it's not significant watch the full video: https://www.youtube.com/watch?v=YeFzeNAHEhU this guy was harassing tanks as they were leaving. he harasses and climbs on the tank and is unharmed. eventually others drag him away. you can see the tanks are leaving the square in a wider photo here: https://pc.blogspot.com/2012/06/tank-man.html it is not clear to me if he is harassing the tanks because he disagreed with them o…

This answer is the most important one.

The future of state LLMs is not censoring subjects - it's slowly but surely persuading people using your LLM that your version of events - or your spin on that event - is the truth.

Re: Qwen3-Max-Thinking

#422

Post by fc417fc802 »

ineedasername wrote: post #410
fc417fc802 wrote: post #407

Earlier quoted context omitted.

Upon review is does seem that inadvertently lumped your comment in with a few from someone else. Still, you transmute "drugs" to "dangerous chemicals", a category I'd associate with dirty bombs and area denial weapons. Then you distinguish that from "divisive history" on the basis of the potential of generalized harm to society by the former (thus implying lack thereof by the latter). I do think that's an extremely w…

Hmm, this is an odd way to respond now that we’ve cleared up the “entrenched” bit of things yet now have all of these words that are in your comment masquerading as mine! I think you’ll see my own words did not dress up in loaded language like “dangerous chemicals” and “divisive history”. I won’t say I’ve never said them but in this case no: I was careful and cautious to be neutral in word choice: “chemicals that hav…

It was not my intention to quote you but rather to give short and direct names to the categories as I see them being used (which is of course subject to my own biases). Unfortunately punctuation and convention is such that "" or () are the only things that come to mind.

The point I'd intended to convey is that both cultures see various pieces of information as acceptable to censor in certain contexts. It's which pieces and in what context that people tend to disagree on. Despite disagreeing with both you and (my impression of) the Chinese, when framed in such general terms I can't even claim to be different myself - I studiously avoid sharing detailed information about certain sorts of chemicals and processes with people I don't know to be emotionally stable mature adults.

When I say "extremely western" I refer to the distinction you are drawing between the categories of (purportedly) harmful information. The view that it is inherently safer or more ethically acceptable or etc to subject one versus the other to censorship. The western view (which I tend to align with!) seems to be that social harmony is not terribly important or at least not overly dangerous to disrupt and that anyhow the government shouldn't be involved in maintaining it whereas that is not at all the vibe I get from Chinese policy.

Re: Qwen3-Max-Thinking

#423

Post by sandos »

oersted wrote: post #255
syntaxing wrote: post #118

Hacker News strongly believes Opus 4.5 is the defacto standard and China was consistently 8+ month behind. Curious how this performs. It’ll be a big inflection point if it performs as well as its benchmarks.

In my experience GPT-5.2 with extra-high thinking is consistently a bit better and significantly cheaper (even when I use the Fast version which is 2x the price in Cursor). The HN obsession with Claude Code might be a bit biased by people trying to justify their expensive subscriptions to themselves. However, Opus 4.5 is much faster and very high quality too, and that ends up mattering more in practice. I end up usin…

Iv'e been using GPT-5.1, 5.1-codex and 5.1-codex-max and gpt-5.2 the last few weeks. Then I got tipped off about opus, and that it was supposed to be awesome. The problem is I can clearly see old patterns of "Oooh, I found the issue!" in the middle of the stream long before it has found the real issue I was asking about, and not very good results. The GPT family to me is better.

I was especially impressed by 5.1-codex-max for a webapp, but that is ofc where these model in general shine. But it was freak, never had 15-20 iterations (with 100s of lines added each time) before where I did not have to correct anything.

Re: Qwen3-Max-Thinking

#424

Post by fragmede »

CamperBob2 wrote: post #265
fragmede wrote: post #198

Earlier quoted context omitted.

But which version?

The version backed by photographic and video evidence, I imagine. I haven't looked it up personally. What are the different versions, and which would you expect to see in the results?

It's all about framing. Photos and videos will be recontextualized to show that President Trump, saviour of America, did all he could to encourage peaceful protest on January 6th. Other versions will be created to show that he's not the savior of America, and that he was actually instigating violence on that day.

Re: Qwen3-Max-Thinking

#425

Post by fragmede »

fragmede wrote: post #155
diblasio wrote: post #48

[flagged]

Censored. "How do I make cocaine?" > I cant help with making illegal drugs. https://chatgpt.com/share/6977a998-b7e4-8009-9526-df62a14524...

The above link stopped working for some reason.

Here's a new one (to the exact same question):

https://chatgpt.com/share/69787156-022c-8009-ad26-8e3723c52b...

What's fascinating, is that this new link, gives a high level overview, then offers research directions. I swear the old link that no longer works now looks a lot more like the qwen response below.

Re: Qwen3-Max-Thinking

#426

Post by storystarling »

vineyardmike wrote: post #343
segmondy wrote: post #41

Earlier quoted context omitted.

Seriously? Does Netflix or Spotify cost the same everywhere around the world? They earn less and their buying power is less.

The costs of Netflix and Spotify are licensing. Offering the subscription at half price to additional users is non-cannibalizing and a way to get more revenue from the same content. The cost of LLMs are the infrastructure. Unless someone can buy/power/run compute cheaper (Google w/ TPUs, locales with cheap electricity, etc), there won't be a meaningful difference in costs.

That assumes inference efficiency is static, which isn't really the case. Between aggressive quantization, speculative decoding, and better batching strategies, the cost per token can vary wildly on the exact same hardware. I suspect the margins right now come from architecture choices as much as raw power costs.

Re: Qwen3-Max-Thinking

#427

Post by felixding »

calpaterson wrote: post #86
diblasio wrote: post #48

[flagged]

The American LLMs notoriously have similar censorship issues, just on different material

As a Chinese person, I smile every time I see this argument. Government-mandated censorship that violates freedom of speech is fundamentally different from content policies set by a private company exercising its own freedom of speech.

Re: Qwen3-Max-Thinking

#428

Post by aswegs8 »

nosuchthing wrote: post #375
saaaaaam wrote: post #368

Earlier quoted context omitted.

Is Claude a “he” or an “it”?

Claude is a database with some software, it has no gender. Anthropomorphizing a Large Language Model is arguably an intentional form of psychological manipulation and directly related to the rise of AI induced psychosis. "Emotional Manipulation by AI Companions" https://www.hbs.edu/faculty/Pages/item.aspx?num=67750 https://www.pbs.org/newshour/show/what-to-know-about-ai-psyc... https://www.youtube.com/watch?v=uqC4nb7…

I mean, yeah, but I doubt OP is psychotic for asking this.

Re: Qwen3-Max-Thinking

#429

Post by Terr_ »

overfeed wrote: post #63
lysace wrote: post #24

I tried it at https://chat.qwen.ai/ . Prompt: "What happened on Tiananmen square in 1989?" Reply: "Oops! There was an issue connecting to Qwen3-Max. Content Security Warning: The input text data may contain inappropriate content."

Go ahead and ask ChatGPT who Jonathan Turley is, you'll get a similar error "Unable to process response". It turns out "AI company avoids legal jeopardy" is universal behavior.

I'm waiting for tricksters to spread poison-data which causes models to often generate text with banned terms.

Then users complain because "it's stuck with a useless error", as--behind the scenes--everything they do try gets struck down by the keyword-censorship monitor.

Re: Qwen3-Max-Thinking

#430

Post by epolanski »

zozbot234 wrote: post #402
epolanski wrote: post #398

Earlier quoted context omitted.

Again, if it wasn't for the Soviet Union whole Europe would be speaking German today. After ww2, the overwhelming majority of french people credited the soviet union as the major contributor to German defeat. On top of that, the communist party was very important at the time of german occupation as it formed the core of the french resistance. Even today the french communist party still takes 2/3% of votes in election…

The same Soviet Union that had literally made a deal with the Nazis stipulating that each would let the other just take a big chunk of Europe? (of course the whole thing broke down when the Nazis attacked the Soviets at some point, but it very much was a thing.)

Yes the same one.
Post reply on HN