Rant on AI in the modern workplace

The speaker expresses frustration with the direction large language models (LLMs) and artificial intelligence (AI) are taking, particularly in the workplace. While acknowledging the immense utility of LLMs – citing examples like recovering a corrupt 20-year-old voice recording or efficiently sorting through thousands of customer service tickets – the core issue lies in the desire for AI to entirely replace human work rather than augmenting it. The problematic sentiment is captured in a statement from an article about a company wanting to capture employee mouse movements and keystrokes for AI training: "The vision we are building towards is one where our agents primarily do the work. And our role is to direct review and help them improve." This pursuit of full automation, where AI agents become the primary workforce, is viewed as driven by greed and a misunderstanding of current AI capabilities. The speaker highlights a critical flaw in current LLMs: while they perform exceptionally well 90% of the time, the remaining 10% can be catastrophically bad, producing results so nonsensical they are "stupider than what the stupidest person imaginable would do." This inconsistency makes full automation risky and counterproductive. Two approaches to using AI are presented: 1. **Augmentation:** AI presents information or suggestions for human review, significantly speeding up work. The speaker is in favor of this, as it leverages AI's strengths while mitigating its weaknesses. For example, an AI could analyze thousands of customer tickets to identify the most urgent ones and explain why, allowing a human to prioritize their responses. Even if the AI is wrong 10% of the time, it still saves substantial time by flagging high-priority items for human review. 2. **Full Automation:** AI performs the work autonomously without human intervention. This is where the speaker sees the "clusterf***" occurring. The desire for AI to "do all my work for me," despite the "massively higher risk of error," is deemed greedy. An illustrative example involves a call recording system used in a store since 2011 to resolve disputes about pricing. An LLM could transcribe calls and then analyze the communication to determine who is telling the truth and why. This is a valuable augmentation. However, instructing the LLM to "respond based on your findings" crosses into greedy territory because it risks the 10% error rate directly impacting customer interactions, potentially leading to significant problems. The speaker emphasizes that current LLMs are essentially "spicy autocomplete." While impressive, they are not true artificial intelligence and are not yet capable of reliably performing all tasks without human oversight. The push for surveillance – monitoring every mouse click and keystroke – to train models for full automation is seen as a misguided attempt to overcome this fundamental limitation. A personal anecdote further illustrates the danger of full automation. After using Gemini to optimize website content for Google, the speaker got "a little greedy" and automated the process for every page. The result was a bizarre suggestion for customers to ship MacBooks in anti-static bags inside their boxes, a nonsensical instruction that a human reviewer would have immediately caught. This highlights the "10% of the time that it does something massively stupid." The concept of "auditory exclusion" (or a similar psychological phenomenon) is introduced to explain why relying too heavily on AI for full automation is problematic. Just as the brain tunes out constant, irrelevant stimuli (like the feeling of sitting in a chair), it can also tune out the need for active attention when a system (like self-driving cars or AI agents) is mostly reliable. If an AI is correct 90% or 99% of the time, a human overseeing it might become complacent and miss the critical 1% or 10% error, with potentially disastrous consequences. The speaker concludes by reiterating that companies should be content with AI's ability to augment human productivity and reduce time spent on mundane tasks, rather than striving for complete replacement. The current state of AI leads to "utterly useless, garbage, sloppy, shitty" results when pushed for full automation. The speaker questions at what point employees would draw the line if their every digital action was monitored to train AI, or if AI cameras watched and listened to them continuously, asking what their "breaking point" would be in such a dystopian work environment.