Artificial Intelligence – DeepSeek Who?

The industry has a very short memory.

• The fact that no one has noticed that DeepSeek has updated its model is a clear sign that the only thing that was special about DeepSeek was its efficiency, which has now been copied and commoditised in just a few months.
• DeepSeek has updated its R1 model (which is a bit of a surprise as everyone was expecting it to release a new model called R2), which “thinks” for longer and, combined with a few tweaks, further improves its performance.
• DeepSeek, like everyone else, trots out the usual benchmarks (see here), observes some improvements and then goes on to explain how anyone can use it.
• Hardly anyone has noticed or seems to care.
• There are two observations worth making:

• • First, Its not R2: which is yet another sign, along with OpenAI’s failure to produce GPT 5, that the pace of improvement of these models is slowing to a crawl.
  • The latest scaling trick is to have the models utilise more compute to come up with better answers, and this has produced a significant improvement in some areas of problem-solving.
  • However, this too is already slowing down in terms of the improvement one gets for every extra unit of compute that is dedicated to the task at hand.
  • Where the next big leap in performance that we are constantly promised is going to come from remains very unclear.
  • Second, No one cares: which is the most significant observation.
  • In January, DeepSeek dominated the news headlines across the world and caused a stock market reset, but in May, this has passed with barely a ripple.
  • If DeepSeek’s innovations had remained unique to DeepSeek, then this would have been a major update, but it turned out that while DeepSeek had done some very clever things, they were also very easy to copy.
  • This is why we have seen almost all of the Chinese models copy DeepSeek’s innovations, and even Meta has admitted that it is using two of the techniques first used by DeepSeek in Llama 4.
  • Google also claims that Gemini runs more efficiently than DeepSeek, but has declined to state how it achieves that.
• Consequently, I think it is pretty safe to conclude that the DeepSeek moment was a flash in the pan where someone came up with some clever ideas, the industry copied them, and the world moved on.
• There is nothing particularly special about DeepSeek’s performance, and now there is nothing particularly special about its efficiency either.
• The main question to ask is why the Chinese state is allowing its best models to be exported from the Chinese mainland via the open source community while seemingly getting nothing in return?
• Other than the obvious PR benefit of being recognised as a real competitor in AI, the most likely answer is that China would like to replicate in AI what it achieved in solar panels and what it is achieving in EVs.
• This is to consolidate the industry in China, and then because it has the ecosystem, no one else is able to compete, resulting in the whole industry moving to China.
• The problem is that in AI, this is unlikely to work because while China can compete on performance, its inability to access or manufacture cutting-edge silicon means that it won’t be able to compete on cost.
• For state or military purposes, this doesn’t matter very much, but when it comes to the outcome of the ideological struggle being waged between China and the USA, it is likely to be decisive.
• Chinese technology, and AI in particular, will no longer be a cheap option for non-affiliated countries, and so they are much less likely to adopt it.
• We are already seeing this in the Middle East, where Saudi Arabia, UAE and potentially Qatar are pivoting away from China and towards the USA and the West.
• Consequently, I continue to think that while the USA doesn’t have a particular edge when it comes to performance, it will be the only one capable of delivering at a massive scale due to its superior economics.
• This is where the USA and the West have an advantage, but it will take a long time to become clear to all and sundry, as even most of the AI industry has yet to see it.