DeepSeek says that closed source models are accelerating at a faster rate than open sourced models. Thus, the gap instead of narrowing, is actually widening
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
The Intelligence Index score on DeepSeek V3.2 Speciale is limited due to its tool call support, which is why it is currently lower than the regular thinking variant. Dev expect the final score with tool call to be ~68 but we shall see.
Anthropic reportedly preparing for one of the largest IPOs ever in race with OpenAI: FT
- Anthropic is weighing a massive IPO while also exploring fresh private funding over the $300 billion mark, per the FT.
- The AI firm has reportedly engaged Wilson Sonsini and major banks as the startup races OpenAI for a public listing.
- The potential listings would test investor appetite for high-burn AI firms amid bubble fears and surging valuations.
Can't believe these guys are serious about this.
no, DeepSeek V3.2 is actually quite a bit better than MoonshotAI. The regular thinking version uses much fewer token than Kimi and is already at the same level in intelligence. The special version is SOTA. On the same level as Gemini 3. They just need to do some work still to get the tool calling worked out. There will be another major release before Chinese New Year, I'm sure of that.
But here you are comparing apples to oranges. These small AI labs in China are quite AGI pilled also, whereas Qwen team and Doubao team are more about putting AI in products. That's why these big tech have to develop a full suite of AI models and products.
Qwen basically has full control of small model fine tuning world at this point.
just posted this today. ByteDance's Doubao phone just came out and the AI-native feature here is quite impressive.
Alibaba is also out with their Quark AI glass and browser.
you go to IPO at this point only if you cannot raise enough private funding. These guys are just all doing media tour at this point to raise the value of their asset.
But make no mistake, Gemini release was crushing for these two.
Chinese firms are famous for humble brag. Where they understate the performance of their own product. At this point, it's quite clear the Chinese open source models are far closer to the American closed source models than at this point last year. V3.2 Speciale is SOTA and Kimi K2 was SOTA at its release. You couldn't have said anything like that a year ago.
This is so cool, we have AI glasses for visually impaired. Improved AI + sensors allow people to correctly find their target destination and avoid getting hit by eBikes.
Gemini 3 Pro and even its Deep Think version already feel nerfed just weeks later.... Google is watering down, and its no where close to the perf/power/intelligence of the benchmarks...
What Deepseek isnt admitting to is that with Kimi K2 Thinking, Z-Image, Qwen models and its own V3.2, all open-weights and open license, it establishes a permanent cognitive floor... that will hit white collar knowledge workers in the US the hardest while making the US AI tech companies ability to extract AI API tax from the rest of the world, including US itself, much harder...
OpenAI, Anthropihic, will never turn a profit. AGI is no where in site, LLMs rate are plateuing, and US AI companies are not advancing fast enough to outrun its cash spend vs gpu obselence vs Chinese open weights pressure
High HA EUV is end of road, China will soon catch up with SMEE/SMIC/Huawei...
LLMs /Transformers will never get us to full AGI...
In the beginning of 2025 folks were hoping China might eventually come out with a reasoning model on par with o1 by end of year... I think we are well past that now.
We have seen how fragile US really is if China presses hard on REE cards...
Deepseek is being disingenious on level of Gordan Chang here, its the West that is running out of time, a dollar short and day late... earlier this year I was legit worried it waa going to be the other way around...
China has already won
I think so, China's economy and society offer a much larger surface area where AI can operate on whereas in the West, the only use AI can be put to is to churn out a flood of AI slop on Youtube and social media, not really helping better a person's life.
Genuine use of AI in real life application esp in enterprise use is boring as hell. Your average person doesn't wanna hear about it. There are a lot of genuine AI companies that people will never hear about because they work on the backend that's frankly too complicated and boring for them. Even though ironically That's where the real change happens. That's why you have people like Sam who are doing all this hyping nonsense about stuff like AGI and yet it is him who they listen to.
Last edited: