Will AI Take Your Job

4th June 2025

Will AI Take Your Job?

The short answer: Not yet, if ever.

Today, Australia’s Federal Science and Industry Minister, Tim Ayres commented on the increasing role of unions in protecting workers in the advent of AI in our workplaces. This comes on the heels of Anthropic CEO Dario Amodei’s comments that “AI could wipe out half of all entry-level white-collar jobs — and spike unemployment to 10-20% in the next one to five years…”. This makes for good forward planning, however such a threat is still undetermined.

Right now, the question: Is the technology good enough? If not, how and when?

We have all experienced hype cycles, and as a scientist, I have experienced these first hand through assertions such as: Can carbon nanotubes build a space elevator so that we no longer need to build space rockets? Can fusion energy achieve net zero and solve all our energy problems? Can quantum algorithms crack RSA encryption protocols in seconds? One more familiar is the prediction a decade ago that we will all be in self-driving cars by 2025. As with all these- hypothetically, yes, they could happen, but not without at least one more “eureka” moment. The current hype around AI is the same. We’re just not there, yet.

The 2017 paper “Attention is All You Need” was one eureka moment that led to ChatGPT. These large language models (LLMs) are nothing short of incredible. Fast-forward to the current day, they outperform the average human in exams for most areas of expertise. However, you need to be looking at robust sets of Evals that have questions and answers hidden, or questions that require answers so complex, training on them is not feasible. Examples of these are “Humanities last exam”, “SWE-bench”, “FrontierMath” where the top models score 20%, 34% and 2%, respectively.

How does this translate to white collar work?

Agentic systems are like Rube Goldberg machines and many of the agents only have accuracies of 70-80% for their given tasks. This creates a compounding probability that the system will fail with increasing complexity. So far, the best examples of agentic systems are coding agents, and even that is falling short. Just Google “watching AI slowly drive Microsoft employees insane” to understand how that’s going!

In summary, the grand challenge is this overreliance on engineering due to the imperfect LLM accuracies. I think the key thing that’s missing is a way to easily create persistent memory. Much like humans, AI needs to remember their mistakes and learn from them. It’s currently the elephant in the room that tech leaders and most of Silicon Valley AI experts are trying desperately to address with trillions of dollars, GPUs, and significant brain power.

Until another breakthrough is made, white collar workers are safe. For now, I'll continue to build AI systems that augment and assist as per their capabilities.