How will AI effect specific tech roles?

The previous article about the Effective use of AI stimulated several good conversations. One of them was about what AI means, in a practical sense for people's work today.

One question I would have for you is what is your view on the implication of work in general — let's say for the data scientist, for the software developer, for the typical analyst at McKinsey.

I think in the short term there is, and will be, ongoing disruption to work in general. But the hype is too thick, and the disruption is as much, or more, from the hype then from the actual capabilities.

The tune I'm hearing in tech circles is like This is as big as the industrial revolution, We won't need programmers in 5 years, A complete change to the nature of intellectual work.

Or maybe my favorite:

Oh, you are a programmer? I'm sorry, what do you plan on doing then? You know, as your next career?

I think that people don't understand what those things mean, at least on a quantitative basis. The industrial revolution effected something north of 80% of all labor, and took at least 80 years. AI just isn't at that level, and definitely isn't going to make the change 16x faster.

What I've seen so far is enthusiastic experimentation, and then what looks like premature action from executives. I hope these articles may help provide business leaders with some additional grounding.

Currently there is little to no evidence of broad based increases in productivity from AI. The IMF estimates that only 40% of jobs are exposed to risk of change from AI an EU parliamentary study suggests a (positive) impact of up to 14% on global GDP over the next five years.

These are significant impacts, but not the "end of work as we know it".

Everything that I've seen so far indicates that the generic application of AI results in a zero to net negative productivity gain. The benefits of AI appear to accrue almost exclusively to using specialized AIs and the narrower the task the greater the return.

I think a great example is radiologist. Their job is much more "mechanistic" than most in the medical profession. They don't interact directly with patients, but rather almost exclusively with other professionals. Their job can be modeled (almost, and with apologies) as a black box "image in" → "diagnostic write up out". A specialized medical AI, trained in medical image recognition can perform the same function, cheaper, and faster, and with greater accuracy. This only seems to work because of the specificity: image recognition, and text generation, for a specific domain. This is a great return on investment. I expect that this won't "replace" radiologists. Instead I think it more likely to result in increased quantity of work, additional human attention available for the more challenging/interesting cases, and an expansion of the duties of radiologists.

Similarly for the specific roles mentioned in the quote at the start of this article, predictions vary depending on how general or specific the role is, and external factors other than AI.

For data science, I think AI will have few direct effects. LLM techniques generally don't work as well as their existing algorithms and techniques, in the cases where LLMs based methods do work better, they will be adopted as specific tools. There may be increased data science work to produce AI training datasets.

I do think that a specialized LLM could do wonders for ETL (extract, transfer, and load) processes. ETL is typically a data engineer's job, rather than data scientist, but it is a closely related field. I've done a lot of ETL with government and private sector IT modernization projects over the last six years, and would love to create specialized LLM based data cleaning tools. And am currently looking for partners to help fund and commercialize that work. This is another example of specialized AI, you can't feed the data into Grok and get good answers, but a specialized LLM based tool that could identity error patterns and automatically correct them could save huge amounts of manual effort. It would not replace data engineers, but rather leverage them. There is a lot of data engineering work that could be done if it was cost effective, and AI's need for training data actually increases that demand. So I expect to not see a decrease in the number of such engineers, but an increase in the work they would handle.

I'm am personally underwhelmed with AI for software development. AI is moderately useful for writing code, especially in more verbose languages and frameworks. It can help someone who can't code, or can only barely code, to get code out the door. It is much less useful in editing existing code, debugging, gathering requirements, performance testing, and everything else. It would also be more useful for developers to reduce boilerplate, than to generate more of it with AI. It does seem useful for throwaway / one off scripts, and for overcoming "dumb things". For example Salesforce requires 80% code coverage (across an entire project) to deploy a change to production. I've encountered multiple times when old projects no longer pass the 80% metric, but the business does not want to maintain them, they just want to ship new work. Having an AI generate ugly and "useless" tests that exercise the code paths negates Salesforce's restriction with marginal cost to the business.

It looks like people at the frontier of trying to replace software developers with AI are experiencing terminal setbacks. There is also a rising frustration with open source AI based contributions.

In summary I think there will be churn around many roles as companies reduce hiring, and or engage in layoffs based on the AI hype. I think it will be a short term effect, followed by benefits to those companies that use specialized AIs to augment appropriate roles.