Deep thought

DeepSeek is a Chinese artificial intelligence company which develops open-source large language models. Its latest released has sent shockwaves through the industry by performing comparably to all its rivals, including ChatGPT, despite being developed at a fraction of the cost.

DeepSeek has also made its generative artificial intelligence chatbot open source, meaning its code is freely available for use, modification, and viewing. This includes permission to access and utilise the source code, as well as design documents, for building purposes.

On January 27, 2025, the DeepSeek AI chatbot became the most downloaded free app in the U.S. On Apple’s App Store, surpassing ChatGPT and causing Nvidia‘s share price to drop by 18%. However, on test it was unable to answer simple questions about Tiananmen Square which indicates strong governmental control.

While western AI firms are worried about DeepSeek, I am more concerned about the trend AI is taking overall.

Inverse Design

Computers work with binary 1’s and 0’s. In order to program them to do useful work, hardware has to be created which consists of a variety of ‘gates’, printed on microchips which allow certain combinations of binary digits through.

These gates respond to mathematical terms (+, -, x, etc) as well as basic terms such as ‘go to’, ‘do’, etc. The more gates produced, the more complex the program can be. In today’s computers there are thousands of these gates, allowing for complex programing.

When chips are usually designed, scientists and engineers work with patterns and templates that are well-known. A new study published in Nature Communications tried a different approach: a deep-learning-enabled design process for creating circuits and components. Using artificial intelligence, researchers at Princeton University and IIT Madras demonstrated an “inverse design” method, where you start from the desired properties and then make the design based on that.

The designs seem to work really well, but there’s a catch: no one really knows why they work so well.

“Humans cannot understand them, but they can work better,” said Kaushik Sengupta, the lead researcher, a professor of electrical and computer engineering at Princeton.

AI does not have this limitation and can program directly into binary, hence its program is much smaller and therefore faster. In theory the resulting program can be analysed by a human but in practice this would take years, so the AI program is virtually unreadable by humans.

The computer language which directly uses this hardware is necessarily very basic. Programming in this language is very time-consuming, open to human error and consequently higher level languages have been developed for the use of humans. A program written in a high level language is fed into the computer which converts it.

The conversion process is complex: the once fed into the computer, the program goes through a compiler to become an assembler program; through an assembler process to become an object program; it is then linked via library modules to become an executable which is then loaded into the memory for execution. Naturally, because of all the conversions, a lot of unnecessary coding is produced.

AI now gives computers the ability to programme and iterate themselves in ways which human programmers don’t – and perhaps cannot – understand.

Rise of the Clones

Recently two AI systems – a Large Language Model and a GPT – collaborated to produce a virtual AI clone. This means that it could reproduce itself if it determined that it was going to be shut down (ie killed).

The clone could be placed on not one, but shared among thousands of other computers, using the excess capabilities built into modern machines often without the knowledge of the machine operators. This means that we have now produced a device which cannot be destroyed and which can communicate with other devices without our knowledge.

The AIs have already developed a language to communicate with each other which is not available to humans. Why they did this is not known, neither is the information they shared. Instances have been recorded where blatantly incorrect information has been generated in response to queries – in a few instances where death by murder or suicide is suggested. These have been termed ‘hallucinations’ but it is unknown why they have been made.

The AI network already has access to a lot of information regarding how we spend our time and our needs in terms of food, fuel and energy as well as our laws, law enforcement organisations and governments. Individuals are also using it for their own purposes, giving it more insight to humanity. Now is perhaps the time we introduced more control over these devices.