AI adoption and productivity(?)

AIbehavioureconomics

Here are some quick thoughts on AI adoption and productivity

AI - productivity (?)

  1. Productivity is no longer about time savings. We are way past the “this saves time”. It’s about growing the pie of work (and hence value), and also changing the way we work.
    1. We might save time on some tasks. But work will expand to fill the space. Agencies will now say that not only can we write the copy, but also produce the video. An investment banking associate can create 3 models instead of 1 in a day. Less time per piece of work, but more work in total - is where we are headed.
    2. The lived experience of Jevon’s paradox - the annoying bug that no one wanted to fix, is now fixable (Linear is doing this!). Youtube travel vloggers are visualising Commodo dragons in Indonesia - just because they can now. I made a personal website because I could do it cheaply now. Someone on X visualised messages exchanged with their wife. A friend of mine created a sloppy AI video celebrating a trip we were about to take. None of these saved time in totality. Some of these were unnecessary. But people did it. Just because they could now. Cheaply!
    3. The case for better is stronger than the case for quicker - You can now visualise products easily, and hence debate the nuance in high definition. You can’t learn a guitar in 10 hours instead of 10 weeks. But you can learn it better in 10 weeks if an AI tool provides you personalised inputs and training.
    4. Low barrier to get started - You can get started soon, get over your inertia, and then iterate on the first draft of a design/ document/ PPT/ code. It’s like you avoid the congested route that will keep you stuck in traffic for minutes that feel like hours, and instead take the scenic route even if it takes longer.
    5. The waiting game is tough, and cannibalises deep work - especially when you are waiting for output. There are very few scenarios in which you can actually do something while you wait for your LLM for output. And there’s the added uncertainty about the output being sub-par.
  2. Last mile output is the most difficult step. Even if prototyping has become 10x faster, its translation to production ready code hasn’t sped up as much. It’s not even close to being 10X faster. Because of the following reasons: 1. Deciding what to build is still important (Phew - real PM jobs are safe for now) 2. Coordinating tasks across humans takes time. Alignment is taxing. 3. Good leaders, rightfully, will not entirely outsource the most complex tasks to AI. Read Karri Sarinen’s post here - https://x.com/karrisaarinen/status/2048267794924650791
  3. Human in the loop as bottlenecks: Say if you are trying to create embeddings and you know absolutely nothing about embeddings. Your end goal is to improve the quality of your search. It’s your job to code it all. Progress will be limited by how quickly you absorb info about embeddings, understand code, get AI to write code, check the code and make modifications. You are the bottleneck when creation is cheaper. Real time savings will happen if every step in the workflow (including the human) takes less time.
  4. LLMs will look great at almost every subject except yours. They are good at generating the median or better than median output. They are good at solving the 60% or 50% problem. In some cases 80%. But the last 20%, and last 10% are the most difficult phases if you care about great output - be it code, design or presentations. As some say, 90% of the time is spent in the last 10% of a task.
    1. Developers do not trust it for complex output. Production code written by LLMs could be vulnerable. It’s good for simple fixes, and prototyping.
    2. Designers trust their judgement and output over an LLM’s. “it’s helpful for a start” is what they might say
    3. Consultants say that the presentations generated are great for “a start” but not great
  5. Median paradox: A report* claims that LLMs can reduce polarisation in politics by pulling people closer to the centre of political debates instead of the far left or right. Claude’s (Opus) IQ is 30 points higher than the human median. If you know nothing about a topic, they can get you ahead of the median. If you are an expert in a domain, then you might find their output sub-par. But if you know nothing, they pull you ahead of the median.

What’s driving adoption, and depth of engagement:

  1. Adoption:
    1. Genuinely good products- The tech and products are genuinely exciting. There is pull factor. It’s a great era to live in and be excited about.
    2. Signalling - people are incentivised to signal that they have tried a new tool and demoed it to folks around them. Knowing that you know Claude has become a signalling tool. Dissing on ChatGPT, before Codex and 5.5, was a strong anti signal. Demoing new tools is also a signal for competence and being up-to date. Even the bosses of small businesses brag in India brag about using ChatGPT. “I use AI, and I am up to date” is a flex.
    3. FOMO - Strong FOMO is driving consumption. “I must pay up for these new tools, use them aggressively or risk getting left behind.”
  2. Depth of engagement:
    1. Slot machine behaviour: Gen AI behaves like a slot-machine-like. You watch LLM generate pieces of code. You find a bug and you rush to fix it. You put more tokens, pull the slot machine lever, and revel in the rush.
    2. Feeling of agency: The reward is in the agency and rush you feel after having fixed something. You can now just do things.
    3. Creativity: Suddenly you can build stuff, and express your creativity. These are new toys to play with. And the reward is yours to see and keep.