Tencent's 'Hunyuan-T1'–The First Mamba-Powered Ultra-Large Model

> LLM labs are so focused on making those scores go up it’s becoming a bit of a perverse incentive.

This seems like an odd comment to post in response to this article.

This is about showing that a new architecture can match the results of more established architectures in a more efficient way. The benchmarks are there to show this. Of course they aren’t going to say “It’s just as good – trust us!”.

Being _perceived_ as having the best LLM/chatbot is a billion dollar game now. And it is an ongoing race, at breakneck speeds. These companies are likely gaming the metrics in any and all ways that they can. Of course there are probably many working on genuine improvements also. And at the frontier it can be very difficult to separate "hack" from "better generalized performance". But that is much harder, so might be the minority in terms of practical impact already.

It is a big problem for researchers at least that we/they do know what is in the training data and how that process works. Figuring out if there are (for example) data leaks or overeager preference tuning, that caused performance to get better for a given task is extremely difficult with these giganormous black boxes.

Intelligence is so vaguely defined and has so many dimensions that it is practically impossible to assess. The only approximation we have is the benchmarks we currently use. It is no surprise that model creators optimize their models for the best results in these benchmarks. Benchmarks have helped us drastically improve models, taking them from a mere gimmick to "write my PhD thesis." Currently, there is no other way to determine which model is better or to identify areas that need improvement.

That is to say, focusing on scores is a good thing. If we want our models to improve further, we simply need better benchmarks.

Ironic and delicious, since this is also how the public education system in the US is incentivized.

Goodhart's law - https://en.wikipedia.org/wiki/Goodhart%27s_law

Zero trust in benchmarks without opening model's training data. It's trivial to push results up with spoiled training data.

This is already a problem for years in AI.

If anything I feel like “Ok, so…” is wasted tokens so you’d think RL that incentivizes more concise thought chains would eliminate it. Maybe it’s actually useful in compelling the subsequent text to be more helpful or insightful.

Surprisingly, Gemini (Thinking) doesn't do that—it thinks very formally, as if it's already formed its response.

A large amount of work in the last few years has gone into building benchmarks because models have been going though and beating them at a fairly astonishing rate. It's generally accepted as true that passing any one of them does not constitute fully general intelligence but the difficult part has been finding things that they cannot do. They are giving them more and more difficult tasks. The ARC prize in particular was designed to be focused on reasoning more than knowledge. The 87.5% score achieved in such a short time by throwing lots of resources at conventional methods was quite a surprise.

You can at least have a degree of confidence that they will perform well in the areas covered by the benchmarks (as long as they weren't contaminated) and with enough benchmarks you get fairly broad coverage.

The trick is that the benchmarks must have a wide enough distribution so that a well scoring model is potentially useful for the widest span of users.

There also would need to be a guarantee (or checking of the model somehow) that model providers don't just train on the benchmarks. Solutions are dynamic components (random names, numbers, etc) or private parts of benchmarks.

I mean all optimization algorithms do is game a benchmark. That’s the whole point.

The hard part is making the benchmark meaningful in the first place.

Typically you train it on one set and test it on another set. If you see that the differences between the two sets are significant enough and yet it has maintained good performance on the test set, you claim that it has done something useful [alongside gaming the benchmark that is the train set]. That "side effect" is always the useful part in any ML process.

If the test set is extremely similar to the train set then yes, it's goodharts law all around. For modern LLMs, it's hard to make a test set that is different from what it has trained on, because of the sheer expanse of the training data used. Note that the two sets are different only if they are statistically different. It is not enough that they simply don't repeat verbatim.

We've been able to pass the Turing Test on text, audio, and short form video (think AI's on video passing coding tests). I think there's an important distinction now with AI streamers where people notice they are AI's eventually. Now there might pop up AI streamers where you don't know they're an AI. However, there's a ceiling on how far digital interactions on the Turing Test can go. The next big hurdle towards AGI is physical interactions, like entering a room.

When actual people start using it

You could ask the same question of a student who has just graduated after passing specific tests in school.

I try to generally keep up with the overall trends, but I’m an engineer at a resource-constrained startup, not a research scientist. I want to see real-world application, at least mid-term value, minimum lock-in, and strong supportability. Until then, I just don’t have time to think about it.

No-one really knows until the dust has settled. Look back 12+ months and the picture will be much clearer.

Trying to drink from the firehose of ML research is only valuable for extremely active research participants. Can be fun though :)

For me nothing has been groundbreaking nor significant. What we are seeing is the same in every new innovation, a suite of micro-innovations which improves efficiency and reduces cost.

But LLMs are still fundamentally a stochastic parrot that depends heavily on source data to produce useful results. So we will go through a lull until there is some new groundbreaking research which moves everything forward. And then the cycle repeats.

As someone who frequently thinks in both English and Chinese, I wonder if this "proves" that the Whorfian hypothesis is correct, or maybe at least more efficient?

Its system prompt says it should reply in Chinese. I saw it discussing its prompt in the thinking process.

To be fair, that's a pretty common human behavior in my experience. ;p

It also appears to be intentional:

> [Q:] Do you understand English?

> [A:] 您好！我是由腾讯开发的腾讯元宝（Tencent Yuanbao），当前基于混元大模型（Hunyuan-T1）为您服务。我主要使用中文进行交互，但也具备一定的英文理解能力。您可以用中文或英文随时与我交流，我会尽力为您提供帮助~ 若有特定需求，也可以随时告知我切换更适配的模型哦！

In relevant part:

> I mainly use Chinese to interact, but also have a certain ability to understand English. You can use Chinese or English to communicate with me at any time, [and] I will do my utmost to offer you assistance~

Do you know? Are most LLMs trained in a single or multiple languages? Just curious.

Thank you for sharing this riveting discussion with a chatbot to all of us.

I remember pushing the R1 distill of llama 8B to see what limits had been put in place. It wasn’t too happy to discuss the 1989 Tiananmen Square protests and massacre, but if I first primed it by asking about 9/11 it seemed to veer more towards a Wikipedia based response and then it would happily talk about Tiananmen Square.

Models tend towards the data they are trained on, but there is also a lot of reinforcement learning to force the model to follow certain «safety» guidelines. Be those to not discuss how to make a nuke, or not to discuss bad things that the government of particular countries have done to their own people.

I guess you are conflating "democracy" and "republic", as Jefferson (?) pointed out. The key thing is not democracy but the separation of powers, and the rule of law, which is more or less what a "republic" is meant to be.

Firstly, these things do not think but regurgitate data they are trained on.

But to call China simply a dictatorship is grossly inadequate. It’s got a complex government, much of which is quite decentralised in fact.

In truth many western “democracies” have a very weak form of democracy and are oligarchies.

[flagged]

> 1T

What's the best technique to train the model to stop responding? A bit of fine tuning on texts with EOS markers?

I think it's particularly egregious that they use such a lossy encoding. I can't read the hanzi, but at least "Hùn yuán" would have been more helpful, or even "Hu4n yua1n" would have enabled me to pronounce it or look it up without having the context to guess which characters it was representing.

Agreed. We all have a duty to respect languages and their official transcription. Pinyin with tones does not look much different from French with accents. In both cases, most people aren’t likely to pronounce it correctly, though.

The irony is not lost on me that Tencent themselves did that.

> The naming is similar to the popularity of greek mythology in western products. (e.g. all the products named "Apollo")

Popular? So you’re saying that all the VPs who have come up with the mind bendingly unique and creative name Prometheus didn’t do so out of level 10 vision?