Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Let’s forget about us AI models who are still ahead, their main issue for now it’s Chinese competitors who offer open sourced free models and have no intention of asking for subscription anytime soon. Plus I don’t think is/will be the best among the top Chinese models like deepseek, or Qwen who are still ahead of kimi and offering their products without any subscription. Why should consumers adopt their models instead of those two? I don’t see any major advantage they have over Deepseek and Qwen to justify a consumer paying their models over those two who are even better and free.
They’re currently offering higher quotas for the Ok Computer Agent and Deep Research features with their subscriptions, along with API credits, though Kimi itself remains free to use on the website and app. Currently, it’s not a bad deal to get both API credits and computer agent + Deep Research quotas for the price of a subscription. Also, Kimi K2 is currently the best open-source non-reasoning model on the Artificial Analysis leaderboard, since Qwen3-Max is not open source.
I suspect they have a monetization strategy in the works for 2026 and will likely discontinue bundling API credits with subscriptions once that plan is in place. The challenge to monetize will instead fall on U.S AI labs by 2026–2027 as Chinese labs continue releasing superior open-source models for free imo.
if feels like DeepSeek is really working with domestic chipmakers to improve support for domestic chips like Ascend and Cambricon. Likely also other chipmakers, but these two are more well known and their libraries are open source, so we can just spot them here.
Zhipu is out with GLM-4.6 and according to the benchmarks, it's really good.
Remains to be seen how popular it gets though.
A note that LLMs that follow a “Kepler-esque” approach: they can successfully predict the next position in a planet’s orbit, but fail to find the underlying explanation of Newton’s Law of Gravity (see ). Instead, they resort to incorrect fitting rules that allow them to successfully predict the planet’s next orbital position but to find the force vector and generalize to other physics. Explained in .
Last edited:
The lack of higher level “meta” thinking is a well known problem with deep neural network models in general. The underlying architecture after all is an approximation function for higher order principles built from a bunch of primitive attention layers - this is never going to exactly represent something like F=ma.
It’d be the equivalent of trying to have the human brain learn F=ma as an exact neural function. Impossible- you can never calculate F=ma instinctually. You have to do the math on a paper or calculator.
Thus the current focus on agentic systems. You don’t need neural networks to exactly represent F=ma. You need them to come up with symbolic solutions that they can then implement and use in code.
The symbolic calculation should not be represented in the LLM but be a function they can call. But the tricky part is having the LLM go through the experimentation & abstract thinking process to arrive at a formula they can code up in the first place.
It should be possible since LLMs can obviously manipulate symbols and use tools in a principled way. But having them combine approximate neural network thinking with precise symbolic thinking is not trivial. The big labs are working on it but it’s not clear anyone has the “solution.”