GPT-5 Disappoints as China Advances and DeepSeek R2 Emerges

GPT5 has only been released for a couple of days, and I have finally replaced Grok and Deep Seek to try it out. Honestly, I find the transition smoother than when I tried Gemini, but in comparison to earlier GPT versions, I don't have a difference. I find GPT5 worse than Grok, and Gemini is pretty much useless here.

Deepseek works best for writing, but Grok is good at translation, with GPT5 also doing fairly well. Honestly, GPT5 has been a complete letdown. I even feel that the name GPT5 should be put on the backburner for the time being, and something like GPT4.5 or GPT4.9 would be better. Prior to GPT4, every version had obvious improvements, providing remarkable improvements and increasing work efficiency. But with GPT5, I don't feel any major improvement in any area for my work.

I like using text-to-speech to listen to my drafts—it helps me catch mistakes I’d miss while reading. I’ve tried both Microsoft and iFlytek, and while they sound pretty similar, Microsoft is just less of a hassle. iFlytek needed extra checks for “sensitive” stuff, and honestly, I don’t like anyone or anything snooping through my writing.

I also found a bunch of open-source AI speech tools you can run on your own computer, some of the best are made by Chinese developers. They’re really cool. I was blown away by how realistic some of these voices are. Some almost sound like real people talking. Honestly, I wouldn’t be surprised if AI speech is one of the first areas where we see huge jumps toward AGI.

That aside, the AI image and video production facilities I was waiting for most eagerly have not seen any significant progress. They remain in the toy-like proof-of-concept stage and I believe significant breakthroughs are not feasible in the near future. Today for websites, more of their features rely on AI. From server setup to modifying source code, changing looks, and developing plugins, Claude is certainly the most professional in that respect, but it is far from substituting human effort and there is no expectation of that in the short run. It is essentially a high-end toy. GPT, DeepSeek, and Gemini all have their strengths and weaknesses, but they are capable of only fixing little things or producing well-defined blocks of code, not producing whole projects.

AI agents can also replace some simple work environments, but their cost savings are very small. It would probably take a sudden reduction in the cost of computing power in society and at least a tenfold increase in computing capacity to make them cost-effective. This also encompasses energy costs like electricity, which I am not going to discuss here. My opinion is that the current AI software environment can only replace simple work in the short term.

Countries like India, which have been called the world's office for call centers, low-level clerical tasks, and low-level programming outsourcing, could experience severe impacts. But few countries have anything to worry about. Current AI is far from a true intelligent agent. The term artificial intelligence is more fiction than reality. Individual workers who master AI can achieve the productivity of a small team and significantly enhance their competitiveness.

But to use AI, you must be extremely skilled in your profession. For these powerful AI models, my recommendation would be to stay busy, chat with them every now and then, and use them for light jobs to stay updated without getting replaced.

Lastly, let us touch upon the China–US tech competition. The US administration has relaxed restrictions a bit more on chip sales from AMD and Nvidia to enable them to once again sell some chips to China. This means Chinese companies are now able to produce chips with the same performance. Huawei keeps pushing forward with innovation in chip and memory technology, e.g., potentially bringing a local HBM memory solution to prevent reliance on foreign competitors. Huawei's dominance in connectivity technology also makes up for its lag in chip process manufacturing to some degree.

China will struggle to penetrate chip manufacturing limits in the near future, its horizon being 5nm-equivalent processes for the foreseeable future. American firms, boosted by TSMC's 2nm process, will enjoy their computing muscle edge.

China's dominance of rare earths provides it with strategic clout, with the possible impact of compelling concessions from the US on its techno-containment policy. China can harm the US tech sector seriously. US sanctions will not impact it, but Deepseek is silent with nothing new reported on R2.

I think the three-month-old Deep Seek V3 upgrade was actually the R2 version, but its performance upgrade was not strong enough to match the new versions like Gemini or GPT40, so it failed to earn the R2 designation. Moreover, as most Chinese government agencies and large-scale companies have utilized the Deep Seek platform, its priority has shifted to serving these large-scale customers.

The R2 release has to be a breakthrough version to be able to make some contribution; otherwise, it will be futile. I think Deepseek is waiting to assess the technical level of GPT5 before determining the release schedule of R2. Now that GPT5 has been released, we should anticipate R2 in the fourth quarter, or at most by the 2026 Spring Festival in March. We look forward to it eagerly.

So, where does this leave us? GPT-5 may have underwhelmed, but AI isn’t slowing down. China is clearly pushing ahead, and DeepSeek R2 looks like it could shake things up. If you’re following AI, keep experimenting, stay curious, and don’t blink, because the next leap could come sooner than you think.