Skip to content

Knowledge Work in the AI Era

Knowledge Work in the AI Era

Reducing bottlenecks in AI-human collaboration + responding to uncertainty

2025-03-25 (updated: 2025-07-22)

0. Speaker

Worked as a developer, researcher, data engineer, planner, etc. at a portal site, a game company, a UX agency, and so on.

Currently planning/developing AI-infused services at various companies.

1. Before We Begin

Problem: No one has lived very long in the AI era.

❌ “This is how to live in the AI era.”

✅ “I thought about this and tried that, and this is what it was like for me.”

“Anyone who says he knows the truth exactly is probably someone who doesn’t know very much.” —1928, Bertrand Russell

Topic

Sharing experiences and thoughts from things I’ve worried about and tried.

Subjective opinions on small habits, lessons, tips, and attitudes.

It will be a success if this makes people want to put together a healthy level of alertness + a concrete execution plan.

Problem: Progress is Too Fast (last 3 years)

(Reference: LLMflation)

Problem: Progress is Too Fast (last 7 days)

July 14, 2025: Moonshot unveiled Kimi K2 (1000B-A32B). It surpasses Opus 4 (non-thinking) on a considerable number of benchmarks.

July 21, 2025: Qwen3-235B-A22B-2507 released. It’s one-quarter the size of Kimi K2 (two-thirds by activation), yet better on nearly every benchmark.

235B?

  • 235B 16FP: requires 512GB RAM (about 20 million KRW for a Mac Studio)
  • 235B 4bit quantized: requires 128GB RAM (about 5 million KRW for a Mac Studio)

“For paid subscribers, I send the summary version.” —Simon Willison, Pay me to send you less

Problem: Prediction is Difficult

Academia: LLMs alone have limits (July 1, Thinking Beyond Tokens)

OpenAI: Oh, but they can? (July 20, Gold medal-level performance on IMO)

Gary Marcus: Impressive, but it’s unofficial, isn’t it? (same day, a post left on Twitter)

Google: Ours is official😎 (the next day, Advanced Gemini officially achieves gold-medal)

What Do the Experts Say?

Caution: argument from expert opinion

  • “AI is too stupid to be dangerous. Used well, it is beneficial.” —Y. LeCun (Meta)
  • “AI is already too smart and dangerous.” —G. Hinton
  • “AI is a stupid stochastic parrot, but dangerous because it is discriminatory.” —E. M. Bender
  • “It’s not AI that’s dangerous, it’s the greed of corporations using AI.” –-T. Harris (a remark aimed at Y. LeCun)
  • “From the perspective of Chomskyan linguistics, it means absolutely nothing.” —N. Chomsky (founder of Chomskyan linguistics)
  • “Stupid until you mix in symbolic AI” —G. Marcus (roughly a disciple’s disciple of Chomsky)
  • “Quantum computers? At least 20 years away (January). Oops, I misspoke (March).” —J. Huang (after the Willow announcement vs. after Quantum Day)

A quasi-religion? Either worshipping it (e/acc) or fearing it (longtermism).

What About Other Fields

There aren’t many fields using AI more aggressively than IT. An exception? War:

The singularity in war: the point in HOTL situations where AI’s OODA speed surpasses the speed of human understanding, and humans in effect lose control.

It is useful to think not in terms of a generalized concept of singularity, but concepts of singularity specialized by field. Another example:

“Ray Kurzweil’s singularity is the point when computers surpass human strengths. That time has not yet come, but the point when computers surpass human weaknesses has already arrived.” —T. Harris, Social Dilemma

What Situation Am I In?

“Even users on the OpenAI Plus plan can now use Deep Research 15 times a month. I should max it out and use every single one.” — from my diary on March 4, 2025

“It turned out I had fewer questions than I expected, and I can’t even finish reading the reports that did get generated.” — from my diary on March 13, 2025

  • “Everybody has a plausible plan. Until they get punched in the face.”
  • “I’m gonna go wreck Deep Research” / “But the one who always got wrecked was me”

AI capability overhang and generative indigestion

Interim Summary

Change is fast, prediction is difficult, expert opinions all differ, and interests are entangled → responding to uncertainty

In leading fields, it is obvious that humans are the bottleneck → reducing human bottlenecks

Amid all this uncertainty, one relatively certain thing is that the human bottleneck problem will gradually become more severe.

Someday(?) AI and robots may do all labor, distributive justice may be realized, and humans may receive universal basic income and be completely liberated from alienated labor, but for the time being(?) we have to adapt and live through it.

How long “for the time being” means—1 year, 5 years, or 10 years—I don’t know.

2. Human Bottlenecks

In fields where AI-human collaboration has developed, humans are already the bottleneck resource, and this phenomenon will likely accelerate for the time being (one of the few predictions I feel confident about).

The performance (throughput) of the overall process is constrained by the bottleneck (queueing theory).

Analyzing Human Bottlenecks

“What you write down doesn’t mean exactly what you think it means. And when it does, it doesn’t have the consequences you expected.” —Daniel Jackson, Software Abstractions

The AI-human collaboration loop: (excluding proactive and/or ambient agents)

  1. The human overcomes laziness
  2. The human makes their intention concrete
  3. The human conveys their intention to the AI
  4. The AI generates a response ← watch out for hallucinations 😇
  5. The AI conveys the response to the human
  6. The human understands the AI’s response
  7. GOTO 1

Task: improve human capability + improve the communication channel

Habit: Externalize Thoughts in Advance

One way to reduce latency is pre-loading.

If latency arises from having to convey thoughts to AI? Externalize the thoughts in advance. (Reference: wiki gardening rules)

Difference from a cache:

  • Cache: if the hit ratio is low, it’s wasteful
  • Externalizing thoughts: even if the hit ratio is low, the ongoing act of managing thoughts itself is beneficial (reference: value embedded in the process)

Try externalizing tacit knowledge too (reference: Knowledge elicitation)

Habit: Automation

Eliminate duplication (duplication; meaningless repetition) (reference: Hammerspoon LLM script)

Habit: Manualization

Introduce redundancy (redundancy; meaningful repetition) (reference: value embedded in the process)

Habit: Squaring Capability

Use AI to increase capability → use AI to get better at “using AI to increase capability”

Examples:

Habit: Refining Thought

Get friendly with formal languages (reference: limits of specifications written in natural language)

  • You can express intentions clearly
  • You can verify the output of LLMs (a highly effective remedy for hallucinations)
  • You can think more clearly

Dilemma: Does Using AI Reduce Capability?

To collaborate well with AI, you need to improve your capability, but if you collaborate with AI, your capability declines?

  • The group that studied using generative AI performed better on the test, but when they were later prevented from using AI, they actually scored lower than the group that had never used it in the first place (Generative AI can harm learning)
  • The group that wrote essays using LLMs showed the least brain activation (Your Brain on ChatGPT)

Question 1. What Even Is an Idiot Box? 🤔

“Those who acquire writing will cease to use their memory and so will forget more.” — what King Thamus of Egypt said to Theuth, inventor of writing (or so Socrates said (or so Plato wrote down in Phaedrus))

“A balanced approach is advisable, one that might leverage AI for routine assistance but still challenges individuals to perform core cognitive operations themselves.” —Your Brain on ChatGPT

“Is it safe to say that LLMs are, in essence, making us dumber? No!” —from the FAQ written by the author of Your Brain on ChatGPT

Even Plato did not criticize writing as a whole. Reference: Plato’s critique of writing

Question 2. Is Every Capability Precious? 🤔

“The group that studied using a scientific calculator performed better on the test, but when they were later prevented from using a scientific calculator, they actually scored lower than the group that had never used one in the first place” (is that really so surprising?)

Are capabilities like these still as important as they used to be?

  • Matching parentheses by eye without syntax highlighting
  • Memorizing and typing APIs without autocomplete
  • Making good PowerPoint slides

Think about which capabilities are important now but will become less important in the future, or vice versa.

Dilemma: Exploration Over Exploitation

A exploration-exploitation dilemma situation. A good MAB algorithm increases the frequency of “exploration” in the following cases:

  • When many attempts remain
  • When uncertainty is high (because there has not yet been enough exploration, or because circumstances are changing quickly)

But humans?

  • Path dependence: you have to keep using your brain to prevent or delay the cognitive decline that begins around your 40s (the less you explore, the less able you become to explore)
  • Increase in conscientiousness: as people age, Big-5 conscientiousness tends to increase (exploration decreases)
  • Increase in life expectancy: life expectancy keeps rising. If you take good care of yourself, healthy life expectancy also increases (maybe 140 years old is possible?)

Let’s put more weight on exploration over exploitation. Especially these days, when uncertainty is high, exploration is all the more important.

3. Responding to Uncertainty

The basic approach of agile methodology: “If prediction is difficult, respond nimbly.”

What does it mean for the person themself, not software, to become nimble?

The Fieldstone Method

If I discover a tool or practice that seems worth trying, I write it down first even if I don’t need it right away, and then try it when the opportunity arises (borrowed from The fieldstone method of writing)

“Optionality: The right, but not the obligation, to take action.” —Optionality: How to Survive and Thrive in a Volatile World

Biased Gullibility: Be More Biased Toward “It Can Be Done” Than Toward “It Can’t”

If things are uncertain anyway, rather than assuming it won’t work and therefore not doing it, try believing that it will work and doing it.

“특출난 노학자가 무언가가 가능하다고 말하면 거의 옳을 것이다. 하지만 그가 무언가가 불가능하다고 말하면 아마도 틀릴 가능성이 높다. (When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong.)” —Arthur C. Clarke #

This maxim is beneficial in two directions.

  1. Be less affected when someone tells me something can’t be done
  2. Not casually tell someone else that something can’t be done

Managing Information Channels

From the perspective of system dynamics, if a wiki is Stock, then information channels are Flow.

Maintain channels I know because I need to know them: Feedly, Twitter, GitHub release notes, newsletters, Reddit, Hacker News, podcasts, YouTube, etc.

Discover channels I need to know but didn’t know about: Google Alerts, Google Scholar Alerts, and if I happen to find a post interesting, figure out how to encounter posts like that more often (follow, subscribe via RSS, etc.)

Reduce information flowing in that I don’t need to know: (I’m still not very good at this. Maybe use things like regular expressions and free LLMs to put filters on RSS feeds with a high noise ratio?)

Doubt Yourself

No matter how important a signal is, it’s useless if you are not prepared to accept it.

Am I prepared to overturn my own thinking? (The motto of XP is “Embrace change”)

Delay forming opinions (Lean software development says something similar with Decide as late as possible)

This is especially important the older you get, the more you know, and the higher your social status.

Don’t Doubt Yourself

If you are in a situation where there are already lots of people(😇) doubting you for you, then don’t doubt yourself.

Example: if you’re a frontend developer or a designer, don’t be swayed by talk like premature UI design extinctionism.

(Of course, keep reassessing your social position.)

Update Knowledge

Update outdated knowledge (a kind of refactoring)

  • Even if there is a familiar way, search to see whether there is a better one, or try talking with an LLM
  • Use Deep research to find out the latest trends related to things you studied in the past
  • Wake up in the morning, kick off one Deep research, and then take a shower

Consider the shelf life or depreciation of knowledge.

This is especially important the older you get, the more you know, and the higher your social status.

End

Thank you

By
ak
Created
Updated

Machine-translated from the Korean original.

2026 © ak | markdown | 한국어