Catman

ML engineer · 10+ yrs, CV → LLMs · crypto degen since 2021 · AI agents run half my life. Chief Innovation Officer @SynopsisEvents.

[Catman @LittleCatman69]

Joining @SynopsisEvents as Chief Innovation Officer. The plan isn't AI-generated content, everyone has that. It's AI where the actually is: finding the right audience for each project, matching sponsors with the summits and shows that convert, and reporting that shows what moved instead of vanity numbers. Web3 marketing is about to get measurable.

[Catman @LittleCatman69]

Conclusion
So, the right approach to using AI isn't just about "doing tasks faster." You need to:
→ identify the queue where work is piling up
→ deploy AI agents to that stage (sorting, preliminary screening, routing)
→ keep human workload below ~80%

Use AI to manage the queue, not the task execution process itself.

[Catman @LittleCatman69]

Law 2: Kingman’s formula.

Wait time ∝ u / (1 − u)

80% utilization → wait time increases 4x
95% utilization → increases 19x

AI floods the next stages of the process with output. The human reviewer’s utilization jumps from 80% to 95%, and the queue balloons rapidly.

More AI output ≠ higher throughput.

[Catman @LittleCatman69]

Goldratt
Law 3: Goldratt’s Theory of Constraints.

An hour saved at a non-bottleneck stage is an illusion.

AI adoption usually accelerates low-cost stages (writing copy, drafting, coding).
The constraint, however, is typically a high-cost stage: review, decision-making, or approval.

[Catman @LittleCatman69]

Flow efficiency and Amdahl's Law
Law 1: Flow efficiency.

Trace the path of a single task from request to completion. In knowledge work, the time spent actually performing the task is often less than 15% of the total time. The rest is waiting.

Even if AI makes the execution phase instantaneous, the total cycle time drops by a maximum of ~15%.

1 / 0.85 ≈ 1.18x. That’s your limit.

[Catman @LittleCatman69]

Everyone is using AI to make work faster.
Most companies won't get faster.

3 old laws from operations research explain why, and where AI actually helps 🧵

[Catman @LittleCatman69]

Flow efficiency и закон Амдала
Law 1: flow efficiency.

Track one task from request to done. In knowledge work, active work is often <15% of that time. The rest is waiting.

Make the work part infinitely fast with AI and lead time drops at most ~15%.

[Catman @LittleCatman69]

The implication for builders is uncomfortable but simple. Tooling is no longer a moat, because everyone has the same tooling at the same price. The moat is allocation: search wide, then one bet, held long, with full attention on it. Cheap execution rewards that sequence.

[Catman @LittleCatman69]

The price of producing things collapsed. Code, text, research, a working prototype: an afternoon and a few dollars. When an input gets that cheap, it stops being the constraint. The constraint moves to whatever did not get cheaper. Today that is attention. Nothing else is close.

[Catman @LittleCatman69]

The same law runs both attention markets. Yours is fixed in capacity, so once the search is done the only lever left is how narrowly it is aimed. The audience's is fixed too, and it remembers a person for one thing or for nothing. Past the search, breadth reads as noise in both.

[Catman @LittleCatman69]

@DeepPatternAI We can discuss this in DM

[Catman @LittleCatman69]

An agent that makes a mistake is easy to catch. Tests fail, something crashes, the diff looks wrong. We have decades of tooling for exactly that. An agent doing excellent work on the wrong thing sets off nothing at all, and that is the one that costs you.

[Catman @LittleCatman69]

And the skill nobody mentions: being willing to throw away good work. Not broken work, good work. Agents produce more of it than you can use, and most of it is a well-built road going somewhere you never wanted to be.

[Catman @LittleCatman69]

I don't think there is a clever fix. What helps me is refusing to take the feeling of progress from a report, and going to look at the actual thing instead. If I can't say in one sentence why today's work exists, it usually doesn't.