ai makes building software so cheap that users could just build the tools they want themselves

how does software get shaped today?

1. devs build features, users just use them
   • if the feature you want doesnt exist, tough luck
2. some software lets users build a bit themselves, like excel
   • a cell can hold a formula like =A1+A2, so non programmers end up programming without realising
   • but this caps out, turning a rough idea into real code is hard, so most people stop at simple formulas

what does ai change?

• ai turns a rough idea into real code for you
• so users can build whole tools on the fly, e.g “make me a tool that shows which customers havent paid this month”
• this gives you two loops
  1. inner loop: use the tool to do x, e.g check who hasnt paid
  2. outer loop: ask the ai to change the tool, e.g “also show how late they are”
     • excel already works like this, typing in numbers is the inner loop, editing formulas is the outer loop

flowchart LR
  you((you)) -->|inner loop: use the tool to do x| tool[your tool]
  you -->|outer loop: ask the ai to change it| ai[ai]
  ai -->|updates| tool

so whats next?

• security: ai written tools run on shared data and nobody reviews them, so what stops a tool from leaking that data? does every tool need permissions like phone apps?
• tools vs data: if you share a tool, do you only share the code and have it run on each users own data, so nothing leaks?
• if tools are this cheap, why share at all? maybe the only reason is trust, a shared tool is already tested and debugged
• sharing does have a nice upside, users make tools better for other users, but does that survive when making your own is ~free?
• forks: i copy your tool and tweak it, someone tweaks mine, now 50 versions give slightly different numbers, which ones right?
• breaking: someone renames a column and 200 tools nobody owns break, does the ai fix them, and how do you know the fix is right?
• wrong answers: beginners cant catch ai mistakes, so a subtly wrong shared tool spreads the mistake to everyone
• liability: a coworkers shared tool makes a payment mistake, whos to blame?
• money: if users build the features, the vendor sells the data layer and the sharing layer, is that just notion and airtable again?
• agents: if agents do the work, tools dont need a ui at all, just safe functions over data

sources

• geoffreylitt.com/2023/03/25/llm-end-user-programming
• inkandswitch.com/essay/malleable-software

ai is going to widen the k shaped economy, so sell to the people at the top of the K, or sell the people at the bottom a way up

where does k shaped come from?

• after covid the recovery split in two like the letter K
  • people who own assets (stocks, houses) shot up
  • people living on wages fell behind
• its back now for consumer spending
  • the top 10% of earners now do ~half of all us consumer spending, up from ~35% in the early 90s
  • the bottom 80% fell to ~37%, from ~42% before covid
• asset prices went way up and inflation and high rates hit everyone else
  • so asset owners feel rich and keep spending, everyone else trades down
  • hermes and walmart growing, kohls and bath and body works hurting

how does ai change this game?

• ai benefits high capability individuals and mega corps with capital, so they pull away even faster
• we expect the top percentile to run massive agent fleets
• everything in the middle dies and the rest is left behind

what does this mean?

• sell to the asset owners
  • thats mostly older people and retirees, they own the houses and stocks
• sell to the people with ai tailwinds
  • the people running agents
  • and the agents themselves, as they’re the ones doing the buying
• sell shovels to the people without these tailwinds trying to escape
  • e.g gambling and to dying businesses
• avoid markets where the demand side is collapsing

x.com/mitchellh/status/2041566958681014418

main claim: the way to get massive adoption now isnt a polished mainline app, its building blocks that let others build lots of stuff on top, quantity over quality

• libghostty, pi mono, nextjs, tailwind
• ghostty (the app) took 18 months to hit 1M daily updates, libghostty (the building block pulled out of it) hit millions of daily users in 2 months

why?

• agents write most code now, and agents drive what software gets used and downloaded
• agents are meh at building from scratch, but really good at gluing building blocks together
• agents prefer
  1. high quality
  2. well documented
  3. proven components
     • they pick these by default unless you prompt them otherwise, even more than people do

issues

1. security
2. instability
3. lack of maintenance

benefits

1. lower quality bar: core apps hit customers directly so they have to work, building blocks dont
2. higher awareness: users appreciate building blocks more than core apps
3. lower maintenance burden: its easier to say no to requests as its a building block not a core app
4. outsourced R&D: code is so cheap that you get way more open source suggestions and contributions

so what?

1. a better ecosystem?
2. issues in commercialisation
   • models often pick free alternatives over commercial software
   • so does this mean any closed source software can just be copied?

testing mermaid and latex rendering

inline math: $e^{i\pi} + 1 = 0$ and $a_1 + b_2 = c$

block math:

\[\int_0^\infty e^{-x^2}\,dx = \frac{\sqrt{\pi}}{2}\]

flowchart:

graph LR
  A[startup] --> B[sells to startup]
  B --> C[customer grows]
  C --> D[you grow]

sequence:

sequenceDiagram
  Founder->>Startup: demo
  Startup-->>Founder: ok lets try it

something i’ve noticed - a lot of successful companies got going by selling to other startups

• stripe selling to YC
• scale AI selling to cruise
• turbopuffer selling to cursor

think it’s cause:

• if they grow you grow
• lowest barrier to entry, startups will just try stuff