Ethan Mollick’s Post

AI doesn't just handle the easy 80%. It often tackles the tedious 80%. The human 20% isn't just the hard part. It's the creative, insightful, and judgment-driven part. Expertise isn't about doing everything. It's about knowing where to look and what questions to ask. The real value of experts in the AI age? Discernment, not data processing. AI amplifies human intelligence. It doesn't replace it. Perhaps the future isn't AI or human, but AI-augmented humans outperforming both. The best AI users aren't those who blindly trust it, but those who skillfully verify and enhance its output. What if AI's greatest contribution isn't doing our work, but freeing us to do our best work? In a world of AI-generated content, human insight becomes the scarcest resource. The expert of the future isn't competing with AI. They're conducting it. AI is a tool. The human mind is the toolmaker. One evolves, the other innovates. Are we asking the right question? It's not "Can AI do my job?" but "How can AI help me do my job better?"

It’s a huge lever for experts, but you have to know what you’re doing. It’s easy to understand if we think of LLMs as cultural technology, an interaction between a senior developer and junior developer will result in high quality code, whereas an interaction between non coder and junior coder will result in something that may not work. This is consistent with what is in the training data 😀

AI does not remove the need for real expertise. In fact, from what I've seen in real client situations, those with experience / specialization can use AI better as they can have more precision in what they ask, how they break tasks into smaller parts, and how they review final outputs. #humansmatter

When we say AI does 80% of the work, we're assuming work is just... well, work. But in reality, projects aren't assembly lines where we can neatly divide tasks. Instead, think of it like cooking. AI might be great at chopping vegetables, measuring ingredients, and even following recipes. That could easily look like 80% of the work if you're counting steps or time spent. But ask any chef - the magic isn't in the prep. It's in knowing when to trust your instincts and go off-recipe. It's in that pinch of salt added at just the right moment, or the creative fusion of flavors that creates something entirely new. So maybe instead of percentages, we should be asking: How does AI enhance our ability to create? How does it free us up to take bigger risks, to innovate in ways we couldn't before? The real value isn't in AI doing 80% of the work. It's in how it might allow us to reimagine the other 20% - and perhaps the entire project itself.

Excellent point. The human-AI symbiosis is indeed more nuanced than often portrayed. In practice, this means AI might draft reports, analyze trends, or generate initial designs, while humans refine, validate, and add the crucial insights that stem from experience and intuition. This synergy potentially leads to outcomes surpassing what either could achieve alone. The challenge now lies in developing workflows and tools that optimize this collaboration, enhancing human capabilities rather than replacing them outright.

I would guess that achieving such grandiose levels of automation, given the field of custom software development, would favor only a few top organizations globally, leaving the rest with average generic co-pilots due to the upfront investment and expertise needed. This disparity arises because developing highly customized and optimized GenAI systems requires substantial financial resources and access to top-tier talent. Additionally, the complexity of creating tailored solutions necessitates a deep understanding of the specific industry and its unique challenges. Organizations that can afford these investments are more likely to innovate and push the boundaries of what automation can achieve, while those with limited budgets and expertise may rely on more generic, less effective solutions. Consequently, this creates a competitive edge for the leading companies, enabling them to deliver superior performance and efficiency.

In this case wouldn’t we call that more an agent? An AI is very broad, and here it’s implying that someone has already done the heavy work of building something that solves part of the problem (LLMs). In my experience even with these tools, for projects that are past basic level of complexity, there is more than 20% left. It’s shrinking, tho! But for instance if you want to split shipments into several warehouses based on the forecast demand, this would be using a specific model, hence the different between AI and agent or LLM. But maybe it’s considered the same now :)

80% of the work != 80% of the cost/effort. It might be only 5-10% of the cost/effort. Assumptions of linearity are what marketing people do.

I rather think that for many industrial caess it can solve/do up to 80% of the tasks (to completion) using the right CoT. It's a matter of breaking up problems in the right dimensions. For those it cannot solve (80% means it cannot solve) we still use AI but in that case we don't need the "smartest" model but rather fastest (llama3.1 70b in groq like) 1-2 sec reply in an iterative conversational way (almost as a smart transcriber). I rather get that speed with e.g. 65% right in iterative mode that a CoT taking some mins and resulting in a 80%. my two cents from users perspective.

Using A.I. to do the “annoying” work is actually under utilisation of it’s capabilities. The more we start exploring AIs capabilities from the “Collaborative” and “Aggregation” viewpoint, the more potent will be the outcomes. As with all software under utilisation to exploitation to catalytic potential takes a full cycle - with the human being being the slowest cog in the cycle 🫣😇😇