A few thoughts on the emergence of DeepSeek and the future of expertise

The third decade of the twenty-first century has brought about the start of a radical change to a central paradigm that was dominant throughout human history. From the beginning of the internet age, knowledge has been slowly becoming a commodity. No longer are the days when you had to journey to another corner of the country to attend an elitist educational institution, admission to which cost years’ worth of an average man’s salary. People still follow this practice, either as a force of habit or as a kind of maturation ritual. Nevertheless, the infrastructure required to obtain knowledge that was only accessible to a mere few started to become almost freely available to everyone with internet access. Put simply, all the knowledge required to obtain a graduate-level understanding in many fields - from computer science to chemistry - is available today for everyone to consume without setting a foot in a university or a college. MIT OpenCourseWare was one of the first harbingers of this era, opening its courses in the form of lecture recordings and printed exercises to the public. Since then, many universities followed suit and hundreds are currently offering free online courses. Organizations such as the Open Source Society University have even published online syllabi for people who want to study completely on their own from online courses only, by mixing and matching free online courses from different sources.

The radical change that has started to take shape in the third decade of the 21st century is the commoditization of expertise in many fields, for which the already commoditized knowledge is a requirement. No, people have not grasped the general availability of knowledge with both hands and started to earn PhDs en masse - that was not the case. We handed off this task to machines, which, in contrast to Homo sapiens, happily devoured the information and became super intelligent. As a consequence, a computer programming task that requires years of training—and is therefore priced accordingly—can be accomplished in a few minutes by a person with access to one of the publicly available AI systems and a very basic familiarity with programming to turn the code into a finished product. Products like Devin and its subsequent open source implementation “OpenHands”, are able to develop a POC in a matter of minutes. Critics will say that the quality doesn’t rival the quality coming off the hands of experienced programmers, which may be true at the moment, but I suspect that only a few are willing to gamble that AI won’t catch up eventually.

As exponential progress usually plays out, the next round of transformation hasn’t waited 10–20 years, but just a little over a year. The expertise of developing such an O1-level of intelligence, which was reserved to a mere few with budgets measured in billions, has also become widely available, and soon, we will probably all be able to run it on our laptops or even handheld devices.

The principal in my opinion is this: knowledge - and, very recently, intelligence, can be easily encoded as bits (zeros and ones) and transmitted across the world in a fraction of a second through fiber optic cables at a laughably low cost. What we can’t encode into ones and zeros and transfer easily across the ocean has mostly kept its value through time (not taking into account 3D printing, which may also change that). That’s why lemons and potatoes still have more or less the same prices as they had before the internet emerged. For this reason, companies that focus on expertise—knowledge that can be easily replicated with bits and bytes—will face ferocious competition from open-source or open-access alternatives.

The field of protein design—my bread and butter—is an excellent microcosm of this phenomenon. Proteins are physical objects, and some generate billions of dollars in annual sales (the charts from Novo Nordisk illustrate tis perfectly). Meanwhile, several companies claiming to develop AI for designing novel proteins have been successfully raising millions, promising to capitalize on their algorithms. Isn’t that what Google did, making billions off its algorithm? The problem is, those days are gone. You can no longer profit merely by touting an “AI breakthrough.” Open-source models and algorithms (OpenFold, Boltz, RosettaDiffusion), developed by the community, are just as good as commercial ones—those commercial solutions now struggle to demonstrate additional value over public alternatives. I predict that companies in the realm of protein engineering will have to pivot to physical products (pharma, food, bioproduction, and more) or vanish.