Scratchpad
2025-06-21
The ugly: - monetization warps incentives, memetics and abrasive slop, ragebait to get max impressions -> verified replies shifts the nature of the game - actual grifting normies eg AI influences, programming “influencers” (unhinged devs) - parasocial, delusion, gossipy in nature, bootlicking - unfortunately exposed to the literal bottom of the barrel in thinking capacity and their thoughts with no control
The good: - Lots of talented kids, ML researchers, indie game guys and a small % of very interesting CS, hardware and math folks - Crazy reach potential with companies, industry folks - For me, for example: - Got to talk to frontier lab folks, countless amazing people working at great companies like Tenstorrent, TigerBeetle, Exa etc… - Got to help tons of people land jobs by just connecting talent to companies for fun - Got to positively impact a bunch of people’s lives by answering DMs or just posting and having some people for whom it clicked
Maybe the only way to use X is to simply post and ignore TL? TL mutes, blocks do not actually work, nature of feed changes too often, so it’s too dangerous and frustrating to actually read it
Lists work, notifications work, but then you get FOMO and/or less exposure to potentially great content from cool folks
2025-06-13
MORE SQLITES, SQLITES EVERYWHERE save every single new tab save parent tab build tree structures LLM innards (self-hosted models? CF? deepseek API?)
-> visualizations -> research trails -> publish hem?
2025-06-13
One thing I notice about tpot and a lot of the friendships between the more “normie” folks is how they believe being equally kind and like some kind of mop to everyone is a virtue. Asking questions about each other’s personal lives when they revive their account a break, like “how was your sister’s wedding” etc… when there is almost no chance they actually meaningfully care about this, so they just RP in this very large and sad charade. Because it’s preached as a virtue, but it can’t be virtuous if it is applied to all with no standard. But really, not everyone deserves it.
breadth first empathy is an incredible sign of a human battery
2025-05-13
there is some research (that is moderately contested) that brain sizes have became smaller over the last 15kyrs, and it relates to the collective brain hypothesis (which im not 100% or even 60% convinced by but has some really tangible hypothesis that explain many phenomenas well and are back by a mixture of ok to great scientific studies and processes) part of that is due to the increasing interconnection between populations and the capacity to specialize more and more, offloading some skill needs to the institutions and structures you live in type shit I think that’s moderately interesting in that one day the downstream impacts of hyperconnected populations by the internet might be able to be observed or studied and will be able to explain a lot of cultural evolutions it caused
2025-05-10
aight here’s another plane thought, I think that was from Henrich who’s some cultural evolution / collective brain hypothesis guy (and many other things in the past including an aerospace engineer) my notes/takes: Iq reflects institutional goals IQ is a set of cognitive abilities that leads to auccess in 20th century contemporary institutions (Flynn effect) -> wouldn’t correlate to success -> more iq more science more abstract thoughts -> no indicator that IQ is good for future -> most creative ppl rarely biggest iq
minds that might lead us into a new world (read fundamental changes to our institutions) might not have the highest IQ because said new world might not require what current IQ values
I think the main point of the thought came from all the x “iq arcs” that i saw with ppl sharing iqs and making memes about it
and the thing that i claim is that IQ reflects the institutions of our times and has less “intangible quantifiers of intelligence” than we think
and the capacity to like, remember long chains of numbers in reverse is not going to be immensely relevant in the future for example, so ways that we quantify intelligence in huamsn will change
2025-05-08
Benchmarks ideas #2
- Tinygrad in context -> implement Arxiv paper / HF repo
- RL idea for GUI use
- Generate synthetic dataset of various HTML/CSS UIs
- RL the model on button interactions
- Generate synthetic dataset of various HTML/CSS UIs
2025-04-17
- symbolic reasoning
- language in reasoning
- tokens vs latent space reasoning
- rotate cube is it reason?
- training embeds some products of reasoning already
- circuits
2025-04-17
- symbolic reasoning
- language in reasoning
- tokens vs latent space reasoning
- rotate cube is it reason?
- training embeds some products of reasoning already
- circuits
2025-04-19
deutsch/friston/wheeler/random stuff
Structure == patterns - structure is basically regularity or patterns that can be discovered, described, leveraged, compressed
Compression == prediction - kolmogorov complexity - compression information inherently discards redundant details, keeps essetnail features that predict unsee or future data - better compression == better predictive capabilities
Building predictive models building ~= learning - active inference, free energy principle, Friston - biological/cognitive systems operate by minimizing “free energy,” effectively reducing surprise (prediction error) about incoming sensory data
Intelligence is the ability to approximate patterns (structure) within reality through efficient compression (predictive representation) - intelligence measures how efficiently a system uses observations to build generalizable models, adding a quantitative dimension to our qualitative discussions
Learning and cognition are emergent from this, as processes continuously refining these approximations (building predictive models) by effectively integrating and leveraging noise