Share prices for SAAS companies has fallen dramatically of late, amid fears of AI competition. Why buy software, when you can ask an AI to create tailored solutions? Here is what Cofounder and Head of AI at Pingwire Carl Ringqvist thinks about the future for SAAS companies. Carl has a PhD in Mathematics from KTH, where he specialised in AI research.
# How much has your productivity increased from using AI in coding, in your personal experience?
I think it has easily doubled from the pre-chatGPT era. And rapidly increasing.
# Lately, there has been intense debate on on the future of human coding. Do you think coders will become obsolete?
I think the profession will change, profoundly. I do not think coders in their current form will exist for very much longer. But I think their skillset makes them well-positioned for performing tasks that will still be in demand.
# What are those skillsets then?
The ability to abstract and fundamentally understand. Aptitude in pragmatic theory and being able to oversee and build large scale designs and plans. Doing this in a coding context requires, or is at least greatly facilitated by, having classical coding skills.
You might make an analogy with the history of mental arithmetic. This skill used to have intrinsic value in the labor market, as it was the only way to get computations done. Then came the electronic calculator and the demand for the skill evaporated. Analysts, or people working with numbers, did however not disappear. They just started to spend all their time figuring out what should get computed, rather than doing computation. I think that, in the same way you can argue for that the analysts of today are similar in talent, skill and interests to the computational executioners of the past; you can argue for that the future system overseers and architects will be similar in talent, skill and interests to the coders of today.
Task composition of the profession change, maybe beyond recognition, but the same type of people is needed just one abstraction layer up.
# Do you think that the fact that the coding profession is so profoundly changing will change the market dynamics for SAAS?
I think the market will change, but not disappear. I think the market will move from delivering a system, to delivering the infrastructure and knowledge needed to build a certain type of system. But as long as humans are needed in the development cycle, I think the forces motivating the existence of SAAS, like the need for specialization and competition for the sharpest people, will remain.
# Heavy voices in the industry states that AGI could be here in just a few years. Will such an AI not be able to do all of the tasks you mention? And in fact build even complicated systems from scratch, tailored to each client?
Today, I think there is extreme uncertainty in any guess about the future of AI. The research and development is intense, and the resources being allocated are extreme. Yet I see a few reasons worth mentioning for why humans will probably be needed in the development cycle, even in the foreseeable future.
# Can you elaborate?
I generally think that the roadblocks from the present up to a human free development cycle consists of the following:
1. Not enough high quality data,
2. Not large enough compute (large enough context windows),
3. The problem with unverifiable problems, and
4. Hallucinations
I think the first two roadblocks (1) not enough high quality data + (2) Not large enough compute will be solved in the near or at least not very far future. The second two, I am not so sure about. These are potentially hard and fundamental problems to crack.
3) A verifiable problem is a problem for which the solution can be easily and automatically scored. Typical examples are exam questions of any type. Verifiability is what lends a problem class to reinforcement learning: a process whereby the model is trained through iterating the steps of solution generation, solution evaluation (with scoring) and adjustment. The scoring function is what allows the model to be trained towards the desired behavior here. Reinforcement learning, in turn, seems to really be a necessary ingredient in order to reach superhuman performance in generating solutions for a problem class. This is why AI systems today beat even the best humans in exams in nearly all fields of knowledge. Unverifiable problems on the other hand, are problems that are not verifiable. Which is to say a lot. In fact, I think most of the problem solutions generated by humans in their everyday working life are not so easily evaluated. And for such problems, the models are still not performing super well. And it just seems like a remedy is not just around the corner; instead the road for super human performance in unverifiable domains seem longer.
4) Hallucinations are errors, generated by the AI model. These constitute an even bigger problem. In fact, it might just be that there exist mathematical truths proving that the hallucinations is inherent to any high dimensional model. If this is the case, and if the AGI defininition is taken to include getting rid of hallucinations entirely, well then we are far from a solution I would say, as the AI universe of today entirely rests on building the necessary complexity out of very high dimensional mathematical spaces.
# So your best guess is that the SAAS business model will stay relevant given the current technology trajectory?
That is my guess! With a big amount of reservation for uncertainty and bias. But still!