Born too late to explore the planet. Born too early to explore the stars. Born just in time to explore the subsurface.
SF
Joined February 2014
- Tweets1.3K
- Following643
- Followers2.5K
- Likes13.7K
First post in ages, but wanted to break the silence with something good.
Last summer I left Google to bet on something big. A tiny seed-stage startup that is completely rethinking the way we map the Earth's subsurface, and we just closed our Series A.
Terra is applying frontier AI that's physically grounded in real-world constraints to one of the largest economic opportunities on the planet: finding and optimizing the extraction of the critical minerals and energy resources the world is racing to secure.
The results speak for themselves. I'm so proud to be a part of this team, and excited for what's to come.
The two common predictions about LLMs are
1. They become cheaper and less differentiated to the point where are effectively a commodity
2. A single company wins the arms race and releases a "god model"
What do you think is more likely?
69%Commodification
31%God Model
13 votes • Final results
No matter how many times I read FAFO, my brain still tries to match the "First ___, First Out" pattern.
Our AI "safety" team just learned about overfitting, and issued new restrictions of train vs test accuracy.
99% train, 95% test is unfair and therefore bad.
86% train, 84% test is better, because it's more fair.
We're ngmi.
> Find a relevant paper or demo from a conference a few years ago
> Look up the author to see what they are doing now
> Their desk is in the building next door
Despite all the bs, working in big tech has its perks.
We're about to see a brand new class of models that cost exactly $99,999,999 to train.
jake retweeted
I would love an ELO-based restaurant ranking system
Yelp reviews aren’t very useful. Everything is 3.5 - 4.5 stars. Tells me nothing. Same for Google Maps. The 0-5 scale just doesn’t get used properly because of social norms toward high ratings.
What would work much better is…
Github should just have copilot create a public repo and link to that rather than generating the 7000th slight variation of some moderately complicated algorithm.
