Original video in Chinese.
Key Takeaway
- The updates from LLM giants (OpenAI, Google) are “eating away” at the market share of smaller vendors and startups.
- OpenAI’s GPT-4o integrated image generation, delivering a brand-new text-and-image interaction experience and widening the gap with competitors.
- Google’s Gemini 2.5 Pro has seen major improvements in coding and reasoning, and it comes with an ultra-large context window, showing strong all-around capability.
- Top models are moving toward a “model as application” direction, rolling out multimodality, code generation, tool use, and more across the board.
- I’m pessimistic about startups in the AI era, believing AI’s power and centralization will compress the room for entrepreneurship, and I emphasize that relationships between people are something AI can’t replace.
I’ve recently had this feeling that LLMs are no longer something small vendors can really play with. Every update from the big companies eats into the share of smaller players and also takes away opportunities from a batch of founders. Just look at OpenAI and Google these past couple of days—if I were in this industry, I’d definitely feel exhausted and hopeless.
First, OpenAI. They updated their GPT models and integrated the most advanced image generation capabilities into GPT-4o. As a result, overnight, Twitter was flooded with Ghibli-style images generated by ChatGPT. It wasn’t just users making memes; a lot of big names started joining in too.
Honestly, it’s been a while since I’ve seen this level of hype in the AI space. Altman really knows distribution. Ghibli’s art style already has a huge mass appeal. When you turn real photos into that style, the contrast is especially suited to social media spread. It’s hard not to go viral.
And this OpenAI tech isn’t just image generation. It should also be able to understand the background information in an image. Because one user found that in the lower-left corner of this image, there was a “ceasefire agreement” on the table, which suggests GPT knew what the original image meant.
This is exactly what I said in the previous episode, “The Rise of Gemini”:
Now AI can answer your questions in a way that combines text and images.
Whatever images you want to generate or modify, AI can make it happen instantly.
This kind of completely new experience never existed before. With this update, OpenAI wiped out half of ComfyUI’s territory and once again widened the gap with other vendors.
Actually, it’s not just founders and smaller model companies that are frustrated—Google probably isn’t too happy either. They released Gemini 2.5 Pro at the same time, but all the attention got stolen.
That said, this model is really, really impressive.
First, Gemini 2.5 Pro’s coding ability has improved significantly and is now close to Claude. Look, I asked it to write a script for 100 small balls bouncing inside a sphere, and it handled it very easily.
Second, Gemini 2.5 Pro’s reasoning ability has improved significantly. Once reasoning gets stronger, combined with the ultra-large context window, it gave me the pleasant surprise of “global understanding.” Whether I use it to analyze scripts or translate PDFs, I feel Gemini 2.5 Pro works better than other models.
You can see that this is what a top global model should look like today. This industry has long since moved beyond simply competing on text generation.
You use reinforcement learning, and I do too. You have chain of thought, and I do too. On top of that, I have a larger context window, native multimodality, the ability to generate and edit images, write code, call tools, and even real-time voice and video with users.
So many capabilities have now rolled out across the board. They all have one goal: to turn the model into a complete application.
That’s why I’ve actually always been pessimistic about startups in the AI era. AI is too powerful, and too centralized. The room for survival for founders will be much smaller than it was in the internet era.
So what is it that AI can’t replace? I think the final answer can only be people. Because only people can’t be replaced by AI; and only the relationships between people are something AI cannot generate.
OK, that’s it for this episode. If you want to learn about AI, come join our Newtype community. See you next time!