I'm a Doctor/YouTuber that loves anime ✨♪( ´▽`)✨ but I also code, so building AI tools for myself and sharing what I learn in public. YouTube strategy @dabidoyt
I think most people who are decently proficient at AI would hopefully be aware that if you want deterministic outputs, you need a deterministic way to handle inputs. This is almost certainly with code, not AI.
There’s ways to increase determinism of an LLM of course: for example, if determinism matters, then you have an evaluator agent that focuses on improving it towards some set number of outputs.
More simply put, a second AI agent should nudge the output of the first one towards whatever you define as “good”.
The way I’ve been doing it for YouTube strategy/business agents, for instance, is to get it to use rubrics and score against those rubrics.
Most importantly, you also need to have a “critic” agent, because AI agents tend to find its own responses the strongest.
That’s it! Overall conclusions are as below.
If you enjoyed this thread, please feel free to give it a like and retweet.
This profile is all about AI stuff.
Leopold Aschenbrenner is an insanely smart person:
> 24 years old
> Was at OpenAI, fired from OpenAI
> Turned $1B into $4.5B by investing in AI
> Became well known for a thesis about AI and AGI
> Turned that into an investing thesis where he’s betting EVERYTHING on the things that contribute to AI — its infrastructure
When everyone else is investing in NVIDIA and TSMC, his biggest investments are hundreds of millions of dollars into companies people have never heard of:
- CoreWeave. About $775M + $437M invested in this alone.
- Bloom Energy, which makes fuel cells that can be placed right next to AI centers. He invested more than $876M here:
- A bunch of data centers once used for mining bitcoin, but because they have real estate and GPUs ready to go, are perfect to pivot to AI compute
It’s so interesting. So, I broke down the full portfolio:
- Real estate + AI compute are essentially AI infra plays
- He’s essentially done the equivalent of shorting by putting a PUT on Infosys. This is a company that outsources cheap labor, but it’s the labor that’s all automatable by AI.
This might look like YouTube...
But nope! I coded a YouTube thumbnail visualiser from scratch in NextJS and TailwindCSS, using Cursor AI.
It's a 1:1 copy of the YouTube channel page. You can see from the URL that below is localhost, not YouTube.
Here's how I did it: 🧵
It's interesting that there's a kind of ironic implication that AI fluency is one of the most important skills needed in this world...
which is that it requires the ability to observe patterns rather than being told them explicitly.
i.e. definitely hard agree with the tinkering component of things
Can you multiply 97 x 94?
...in less than ten seconds?
I couldn't.
So I asked ChatGPT how this maths trick actually works, and how you can do it too.
🧵/2