Analyst Zephyr argues Chinese labs like DeepSeek and GLM no longer rely on knowledge distillation for capability gains

I would love for Zephyr to be correct but the issue is that the gap will come from aspects that aren't distillation availability. Americans have brought up really large clusters, at least Anthropic has an idea of what to do with them, some RSI has started, and there's a data moat

@zephyr_z9 That’s what makes a market 🫡

@zephyr_z9 Didn’t that guy claim that China was going to fold on the trade war? Why platform a clueless hack

Every model that even *maintains* the current gap from now on and until about Q3 2027 will be a heroic achievement by China.

@zephyr_z9 This guy has had the worst takes ever since tariffs started as well as throughout the entire Iranian conflict

@zephyr_z9 distillation as a moat feels like reading last quarter's playbook

the competition has already moved to scale and reasoning depth

@zephyr_z9 >U can try v4, K2.7 & GLM 5.2 to test this

with every update I try, and they fail at even the most basic workloads

@zephyr_z9 Distillation was never a thing but I am still bearish on open source models from China. Their most capable models will probably be closed source as they gain more mindshare in the marketplace.

Also the data industry in China is still pretty early stage.

@zephyr_z9 how would testing the models disprove that the distillation is important? I wouldn't be surprised if they had entire teams focused on distilling..

@zephyr_z9 That is one of the main gaps in his thesis.

@zephyr_z9 test what? clearly they are not on the level as mythos. it's clear they use distillation, lets stop pretending, but distillation is just one piece of the puzzle, it's not like you can just distill your way to a SOTA model

@BryanMcNamaraUS @zephyr_z9 对对，就像你妈，每隔2天就去卖淫来养你这个没人爱只会在家等死的废物

@zephyr_z9 Chinese open-source LLMs are unlocked and ready to speed-run self-iteration via RSI. The only thing holding them back right now is compute. But hey, at least open-source devs don't have to foot the massive compute bill for inference. 🤷‍♂️

@zephyr_z9 There is almost certainly a thumb drive filled with weights spirited from every American lab regularly. Or even a leaky process on a severer somewhere, dripping data to foreign nations.

@zephyr_z9 My market analysis is as follows:

👇Strategy as follows

👇

👇

👇

@maritimetank @zephyr_z9 He’s been right quite a bit fwiw

@zephyr_z9 Everyone thinks Claude will stay the best, yeah aight 😂

Don’t fade the chinese bros

@zephyr_z9 Agree. And the part that matters for the rest of us: these open models crossed "good enough for real products" a while ago. The distillation arms race is a frontier-lab story. The tier most companies actually build on is already here, and it can't be switched off.

@gaolisl139 @BryanMcNamaraUS @zephyr_z9 lol. I wonder if he will understand your message after translation.

@zephyr_z9 If you do any serious coding work GLM&co are extremely suboptimal when compared to the anthropic counterparts