Divine
Region: GB
Thursday 09 July 2026 06:33:22 GMT
138163
6240
195
741
Music
Download
Comments
Hamoon :
GLM 5.2 was the first model to shock me. I just wanted to test out its reasoning across disciplines and formulate innovative ideas out of component parts, so I gave it a complex philosophical essay and it was so precise in its logic and the way it used the sources... really brilliant stuff at a fraction of fable's crazy cost
2026-07-11 08:56:41
70
Kenny :
Distillation is merely the final step in training large language models (LLMs); it is not a core part of the process. In both China and the US, the truly critical elements of model training are pre-training, data curation, data engineering, and model architecture. To understand who leads in these areas, you only need to look at the quality, quantity, and citation counts of research papers. You will find that many researchers, even those in the US, are of Chinese descent. DeepSeek's innovative open-source papers are a perfect example of this foundational work. In contrast, distillation is considered a lower-tier technique. Its purpose is twofold: 1. To create a series of "small but beautiful" models based on a primary, powerful model. 2. To align a model with other high-performing models on the market. Chinese models utilize distillation from American models, but the reverse is also true. For instance, when Claude Opus was first released, it would sometimes identify itself as GPT or DeepSeek when prompted about its identity, which shows how prevalent distillation is across the board.
2026-07-11 02:25:46
77
︎ :
antrophic is ass. openai and grok are wayy cheaper
2026-07-09 18:32:36
33
hdk :
This guy is highly biased.
2026-07-11 01:17:40
5
Robinwood🧸 :
7.5x cheaper? What does that mean? This is trump math
2026-07-10 21:06:34
31
Saad Rabia :
You are oversimplifying what is really happening and the comparison is absolutely wrong. It is the same as the car production industry comparison. It is flawed because things in China are heavily funded by the government at a huge loss.
2026-07-12 00:20:51
2
user2948448304 :
but why don’t american companies distill?
2026-07-11 18:37:28
2
UncleRoost :
The question.. how long till the US AI bubble pops?
2026-07-11 13:43:05
6
NTHM :
Distillation is a common practice by all AI models, including American ones. It’s the use of efficient architecture, better algorithms, lower parameter counts, cheaper compute, cheaper power costs, among others, that lower the overall cost of Chinese models
2026-07-09 22:55:23
40
stevejnemes :
any business giving their data to an open Ai system is a fool. closed systems only
2026-07-09 17:33:57
6
lorat25 :
Where is Europe?
2026-07-10 08:25:53
0
rdrctn :
open weights NOT open source
2026-07-12 22:40:57
1
AW :
interesting. can we compare all the CEOs' total compensation packages of all the AI companies? maybe that is also part of the reason the cost is different?
2026-07-10 23:02:40
9
rollingtoons :
China published 70% of research papers on A.I as well patents, wonder who copied who? knowing all Math Olympic winners are 99,9% Chinese.
2026-07-10 10:17:28
0
Or Cohen307 :
do I need the fastest ai model to manage my business the answer is yes for all tech companies your argument is senseless.
2026-07-11 05:33:58
0
tommy :
can’t see Z.ai in my appstore though
2026-07-12 01:14:49
0
vancouver4pal 🍉 :
It’s a bubble
2026-07-11 04:59:38
0
Ashtray :
为什么美国AI不自己蒸馏自己的模型来降低成本？
2026-07-11 06:46:39
1
hentaiNo.1 :
Do i need the most expensive car to get me from a to b?
2026-07-13 05:14:19
1
Leonal Messi :
i peraonally like deepseek , stop making everything for money
2026-07-11 09:47:24
2
Hi :
Competition is good for consumers 👍
2026-07-11 01:08:54
2
LK :
The bubble's popping
2026-07-11 17:18:24
1
soorajldisg :
How to do distillation? I ask questions but don't know how to distill models. I want to do one with my AI machine. What are all the steps?
2026-07-09 21:25:15
0
Ai.Celebs :
Its the mindset to make the most with the least amount of ressources
2026-07-11 08:16:11
1
Aziz Aboulmahasin :
Who’s that guy speaking ?
2026-07-11 16:36:27
1
To see more videos from user @divine251b, please go to the Tikwm
homepage.
