How was Qwen previously able to train on Claude? Prompt injection?

What's with the random Japanese-looking gibberish in the screenshot?

From what I've read, it seems like a lot of the companies monitor API calls and will posion responses for suspicious traffic. This basically means they'll modify the output to be harmful to the scrapper.

Outside of that they can block accounts/IPs, limit what is returned (just the final answer vs all reasoning steps, etc.). However it's a bit like all cyber security and an evolving game of cat and mouse.

I see. Is there some way to protect models from being trained on them aside from just removing it from the market?

Its believed (probably more just fact) that a lot of the open source models leverage larger models. Either directly through distiallation (meaning training a model directly off of another) or by using the larger models for knowledge distiallation (basically using the larger models as "ground truth" for training workflows).

Either way, this means the open source models would always be "a few months behind" as they rely on the progress of the closed source models to improve themselves.

Really? What do you mean?

Tbf, qwen is a distilled version of claude :')

It's just the new GLM 5.2

Deepseek and Qwen are catching up fast and at this point in the race it doesn’t matter so much if it takes weeks or even months to catch up to whatever the US releases, but cost is definitely starting to play a much bigger role. Neither country is close enough to AGI yet.

Its the poor grammar and speling mistaks that give us away.

China was like, Bezos is starting Prometheus. Ok, us too.

unfortunately, most people cant tell.

Yeah this is an interesting way to compose the ensemble though. Rather than averaging over sub-networks they're synthesizing the panel of responses:

A judge model reads every panel response and produces structured analysis: consensus points, contradictions, partial coverage, unique insights, blind spots

Whereas beam search injects variance by trialling candidate sentences and mixture of experts has competing sub-models here we're reconciling different ideas.

Notice though that they're always using Opus 4.8 as the judge so I think the claim (of surpassing frontier models) is over-inflated. I'd instead characterise it as outsourcing legwork to cheaper models.

I'm still optimistic about ensembles of smaller models though. You could imagine a specialist synthesis model or advantages from combining latent activations instead of text responses

As an ML Acolyte, Even I can tell that it's a well known phenomenon that ensemble methods are more resilient to overfitting issues. What's fascinating is humanity's penchant to forget the past and rediscover stuff.

The writers street doppelganger. https://youtu.be/LTYZNr1mF1M

“Significant productivity in the economy is going to raise the standard of living,” he said. “People who today have two-earner households, they’ll become one-earner households. Maybe some people who are working overtime will stop working overtime.”

Yea this will never happen with you greedy fucks trying to minimize labor cost as much as possible. You are more like to use AI as leverage to pay people less than to allow them to work less for the same income.

Theyre just making words up at this point

I wonder from what angle they are coming at this. Is it the reasonable "a team of experts with sufficient experimental capacity can now prototype potent biological weapons much faster, even though they still have to test them" or is it the wackadoo "you can ask an LLM how to creat a bioweapon and it will list you all the steps".

The first one, yeah, sure but bad actors (including US) will always try to use any technology for such leverage. If second one, then it is a weak attempt at securing more funding.

This is a shitty attempt to make congress think it needs to throw more money at AI to keep up with China.

For LLMs: For the most part, yes. However, an agentic loop, a sufficient knowledge base about chemicals & reactions, and some simulation tools to help judge, can come up with a top 5 promising solutions/approaches, which can then be tested in a lab. Faking what scientists might try themselves, just faster.

Then there are the non-LLM specialty AI models, which are increasingly being developed. Probably developed and trained by a small team of domain experts with the help of a coding agent, and seeded with domain-specific datasets.

The times ARE changing, but I would not expect Little Johnny to be whipping up bioweapons with chatbots anytime soon.

AI doesn’t mean ChatGPT, copilot etc. There are all kinds of models. There are definitely various medical models, I’d imagine there are some kind of protein folding and chemistry models as well.

What? It's not out here reasoning about chemistry. If the chatbot can tell you how to manufacture anthrax, it's because some Geocities page has the same information for similar searches.

gasps!

I'd better buy as much of it as possible before someone else does!

You see, the problem is our product is just so good. It’s too damn good, oh god it scares me how good it is.

Why do I feel it is always a responsibility of Linux foundation and other non-profits, to optimise and standardise things, instead of the actual for profit companies doing these things?