And this was before the AI era lmfao. Just imagine the number of exploits exist today.
Reality is 90% of modern “engineers” don’t even know what it means to be an engineer.
This is serious.
I just saw with my very own eyes that even non-technical people can exploit CVEs like PHP2Shell using certain AI models and I am actually frightened.
Once you figure it out, it is downright trivial to do. The digital infrastructure of the world is at risk.
This is probably one of the highest IRR VC investments of all time
Martin / A16Z led the series A ~2 years ago at a $400M valuation… now Cursor is acquired for $60B.
~120x return in 2 years after accounting for dilution…
~1000% IRR, and a 10 digit gain for LP’s!
MVP is often mistaken for slop and that’s a huge issue.
An mvp is meant to be a dumped down version of your app. The least number of features that deliver your value proposition is your MVP. That is the definition of mvp
It’s okay to have some slop, it’s not okay to only have slop
LOL tiktok has a huge exploit where you can boost a video to make it trendy within 1-2 hours XD debating whether I report it or no.
I am not talking tricking algorithm, I AM TALKING EXPLOIT IN CLIENT SIDE CODE!
Design your problems in BFS style.
Every problem you want ai to solve, be it coding, research or whatever, break it into a tree and apply BFS on it!
Each node should be executed by a small LLM not a large llm. Small LLMs produce better output. The more tokens an LLM support the worst it gets over time.
Large reasoning AI should be designing your tree and essentially taking the results of all smaller AIs and put them together. In other words big LLMs should be the orchestrator in a way.
Orchestration takes place only after a leaf set with common parent are done.
I absolutely agree. More abstraction doesn't automatically mean a better system.
The distinction I am making is between implementation complexity and search complexity.
Recursion isn't introducing more work; but rather, it's changing how we search for the actual solution. The goal is to reduce ambiguity by recursively discovering smaller, well defined building blocks before trying to generate them.
AI doesn't need better loops or better graphs.
It needs another abstraction.
I think that abstraction is recursion.
If you are building AI coding agents or even using one, I think you'll find this model and algorithm worth thinking about. https://t.co/ffQzx9EBWU
They still hallucinate but I agree with what most said, we are used to it.
But I think most importantly is that most models now do “thinking” aka reason with itself by going back and forth which helped decrease hallucinating by a lot. In other words, it’s holding itself accountable.
P.s. I want to add, models are not learning. This is a huge misconception, someone said that in your comments but it’s absolutely not true. They do not learn a this stage.