Artificial Intelligence thread

M

MortyandRick

Senior Member
Registered Member
E

Eventine

Senior Member
Registered Member
Please, Log in or Register to view URLs content!

In an unsurprising turn of events, Open AI has declared code red in an internal memo to the company, citing the challenge posed by Google and Anthropic as being existential to the company. To understand why they would say this, it has to be understood that GPT 5.1 was Open AI's preemptive answer to Gemini 3 and Opus 4.5. It was released in the same month and was meant to steal the thunder from their up & coming releases.

Yet, it did nothing of the sort, as Gemini 3 blew past them in the benchmarks, and Opus 4.5 was at least able to match (and in classic Anthropic fashion, beat them in agentic coding). Sam is right to be afraid - since falling behind in the US race isn't the same as falling behind vs. China. You can't get the US government to bail you out in this situation, as the government would just sit back and pick the obvious winner - ie Google, as predicted almost a year ago. And so Sam's dreams of winning the AI race are dwindling by the day.

1764661831385.png
Now, internal competition within the West aside, Gemini 3's release initially marked a water shed moment for the AI race, as for a period of time it looked as though Google has truly delivered something revolutionary and the West had pulled two generations ahead of China. However, after the dust's cleared a bit and looking past the marketing and benchmarks in more detail, Gemini 3 is really just a reemphasis of Google's strengths and weaknesses as an AI provider - it excels in context length, multi-modal capabilities (hence the high score on ARC-AGI 2), and wealth of knowledge (probably via tapping into Google's vast search engine data base). But when it comes to raw reasoning capabilities, it's just an incremental upgrade - something that DeepSeek 3.2 matched within two weeks of Gemini 3's release. It also tends to hallucinate quite a bit, similar to Google's previous models.

On the topic of DeepSeek 3.2, I think it's quite smart of them to release a model right now. I was getting a bit depressed, as for a while it seemed like Gemini 3 and Opus 4.5 had pulled a full generation ahead of Chinese models and we'd need at least 3-6 months to catch up again. But DeepSeek once again led the way for China's response to the Western hype machine - releasing an IMO Gold winning model to the public before any of the Western providers.

Of course, we should not celebrate too early. Google is still the company to beat, and they still retain significant advantages as stated above - context length, multi-modal capabilities, knowledge wealth. And Nano Banana 2 and Veo 3.1 continue to top charts in the Image & Video arenas. But the race there is very close - on image especially, I feel like China is starting to overtake the West especially since closed source providers keep shooting themselves in the foot with censorship. On video, I think the West retains an edge with a number of strong companies (Runway ML recently took the lead in the Text to Video leader boards, though they have no audio capabilities; and Alibaba slowed their momentum by keeping Wan 2.5 closed source). But I'm pretty confident we'll see a number of releases in the coming days by Chinese companies that will be able to match up.

Overall, China seems to be keeping up, if not closing the gap.
 
bsdnf

bsdnf

Senior Member
Registered Member
Eventine said:
Please, Log in or Register to view URLs content!

In an unsurprising turn of events, Open AI has declared code red in an internal memo to the company, citing the challenge posed by Google and Anthropic as being existential to the company. To understand why they would say this, it has to be understood that GPT 5.1 was Open AI's preemptive answer to Gemini 3 and Opus 4.5. It was released in the same month and was meant to steal the thunder from their up & coming releases.

Yet, it did nothing of the sort, as Gemini 3 blew past them in the benchmarks, and Opus 4.5 was at least able to match (and in classic Anthropic fashion, beat them in agentic coding). Sam is right to be afraid - since falling behind in the US race isn't the same as falling behind vs. China. You can't get the US government to bail you out in this situation, as the government would just sit back and pick the obvious winner - ie Google, as predicted almost a year ago. And so Sam's dreams of winning the AI race are dwindling by the day.

View attachment 165579
Now, internal competition within the West aside, Gemini 3's release initially marked a water shed moment for the AI race, as for a period of time it looked as though Google has truly delivered something revolutionary and the West had pulled two generations ahead of China. However, after the dust's cleared a bit and looking past the marketing and benchmarks in more detail, Gemini 3 is really just a reemphasis of Google's strengths and weaknesses as an AI provider - it excels in context length, multi-modal capabilities (hence the high score on ARC-AGI 2), and wealth of knowledge (probably via tapping into Google's vast search engine data base). But when it comes to raw reasoning capabilities, it's just an incremental upgrade - something that DeepSeek 3.2 matched within two weeks of Gemini 3's release. It also tends to hallucinate quite a bit, similar to Google's previous models.

On the topic of DeepSeek 3.2, I think it's quite smart of them to release a model right now. I was getting a bit depressed, as for a while it seemed like Gemini 3 and Opus 4.5 had pulled a full generation ahead of Chinese models and we'd need at least 3-6 months to catch up again. But DeepSeek once again led the way for China's response to the Western hype machine - releasing an IMO Gold winning model to the public before any of the Western providers.

Of course, we should not celebrate too early. Google is still the company to beat, and they still retain significant advantages as stated above - context length, multi-modal capabilities, knowledge wealth. And Nano Banana 2 and Veo 3.1 continue to top charts in the Image & Video arenas. But the race there is very close - on image especially, I feel like China is starting to overtake the West especially since closed source providers keep shooting themselves in the foot with censorship. On video, I think the West retains an edge with a number of strong companies (Runway ML recently took the lead in the Text to Video leader boards, though they have no audio capabilities; and Alibaba slowed their momentum by keeping Wan 2.5 closed source). But I'm pretty confident we'll see a number of releases in the coming days by Chinese companies that will be able to match up.

Overall, China seems to be keeping up, if not closing the gap.
Click to expand...
DeepSeek 3.2 is still being trained on the outdated, inefficient, and comparatively very small H800 cluster. If DeepSeek's Ascend cluster becomes available, V4 should be able to truly catch up to Forinter's performance.
 
You must log in or register to reply here.
Top