My dog pretended not to hear me calling and honestly same. I don't know who needs to hear this but sometimes the most productive thing you can do is rest
somebody's entire business model is sitting in here for free
a maths problem four separate groups of mathematicians could not close since 1973 fell in under
an hour, and the two pages of instructions that did it are a free pdf on openai's own server
that split is the product. a whole tier of startups takes your job, cuts it across agents, checks
the answers and bills you monthly for the arrangement. cutting the job is the only hard part of
it, and it is now written down
it goes by Graph Engineering: you stop asking one model harder questions and start deciding how
many run, on what, and who is allowed to check them
five rules carry the whole thing, and they work on the agents you already pay for:
→ never hand out "N agents for strategy X": fixed assignment is banned in writing, the split moves as the run learns what is dead
→ most agents are never told which approach is favoured, so they cannot all quietly land on the same one
→ every approach family goes into a shared registry, so the fleet can see which route is already packed
→ a dead route gets tagged blocked and stays blocked until someone brings a genuinely new mechanism
→ nothing counts until a separate agent attacks it against a named list of ways it could be wrong
that last rule is where most home setups quietly die: the agent that wrote the answer is the one
asked whether the answer is good, and it says yes every time
they also wrote the escape routes shut: assume a complete answer exists, no internet, "this is
unsolved" is not an accepted reply, eight hours of thinking before giving up is even on the
table. it finished in under one
width has a ceiling though. four agents score 91.9% against a single one's 88.8% on the same
test: three points for four times the spend. it pays where answers are expensive to find and
cheap to check, and on ordinary work the extra agents mostly agree with each other
the 64 that closed the proof were a one-off. the five rules are not, and they cost nothing
nineteen days public, and the Lean rewrite of that proof still sits at 119 stars
bookmark this before someone sells the five rules back to you ↓
Grok can give you a probability for any prediction market.
That does not mean it found an edge.
The useful workflow is simpler:
- collect fresh, verifiable sources
- separate facts from crowd sentiment
- write down what would break your thesis
- compare your estimate with the market price
- skip the trade when you cannot explain the gap
AI can process public information faster.
It cannot remove uncertainty, bad data, or risk.
The edge is not a secret prompt.
It is having a rule that stops you from betting when the evidence is weak.
A 27-YEAR-OLD FOUNDER IS CHANGING HOW APPS ARE BUILT
No coding.
No developers.
Just a conversation with AI.
Here’s how Anton Osika built one of the fastest-growing AI startups 👇
Anton noticed something simple.
Millions of people have ideas…
But very few know how to code.
So he asked:
What if building software was as easy as chatting with AI?
That idea became Lovable.
Describe your app in plain English…
AI writes the code.
Need changes?
Just keep chatting.
The growth was incredible.
• Launched in 2023
• Reached $50M+ ARR in about 8 months
• Grew to 2M+ users
• Became one of Europe’s fastest-growing AI startups
Today, founders can build MVPs in hours…
Instead of spending months hiring developers.
That’s changing how startups are created.
The biggest lesson?
The next billion-dollar company may not be built by the best programmer…
It may be built by the person with the best idea.
AI is removing the biggest barrier: coding.
Follow @Dexonfxf for more real AI founder stories.