Expanding autonomy @variantfund | Alum @penn | Still tokenmaxxing
SF
Joined August 2021
- Tweets1.4K
- Following904
- Followers3.7K
- Likes10.9K
Caleb Shack retweeted
Robotics research meetup happening at SF tomorrow night!
Enjoy pizzas and papers :)
luma.com/ud78bwvv
Caleb Shack retweeted
Will gladly hire you and pay you a very competitive salary if you can prompt this!
Caleb Shack retweeted
If you're building in prediction markets, you’ve probably had a few conversations like this lately:
"Who's our regulator?"
"Where's your user?"
"New Jersey."
"CFTC."
"Nevada."
"Gambling regulator."
"New York."
"Both, somehow?"
We’ve been keeping track: predictionmarketstracker.com
Running Oasis takes one H100 per player. That's ~$3/hour of GPU for one person's world. Roblox monetizes an hour at about 6 cents. Inference costs have been falling ~10x a year. Do the math on when those lines cross and you know when the metaverse ships.
Caleb Shack retweeted
@stripe : "We think it is important that deployment of AI enhances human agency… and [to] protect economic autonomy."
@variantfund: "Autonomy is fundamentally about human agency: the degree to which users are in control of their lives, assets, and identities."
AUTONOMY != AUTOMATION
Started testing out some local AI apps through Pinokio(.)co
Pretty impressive what you can do on a macbook nowadays at 0 marginal cost
Seems like a good time to refresh this piece now that the “trenches” are back
The strategy extends to any new social app: acknowledge the hierarchy of users and build top-down
Huge props to @fomo for out-executing on the social trading vision
Spent some time today running a world model. My takeaways:
1) World Models are the unlock for the next great medium of entertainment
2) Long-horizon memory is an open frontier
3) We need A LOT more data
The path forward for consumer world models may require solving constraints their economics cannot incentivize on their own
The ability to easily, quickly, and correctly verify agent outputs determines what you can safely delegate to automation. An open question is what is hard to verify today that becomes easy in the future?
Passive judgement decomposition technology is the missing layer in Sovereign AI
By the end of 2027, automated intelligence will replace human labor on all rules-bound tasks and processes, but automation won't stop there
FDEs are actively deployed to fill the judgement gaps while expert data infrastructure providers are commoditizing them to the frontier labs
The process of judgement decomposition still requires active participation from those involved—why can't this be passive?
I love vibecoding but once I push something into prod I hate maintaining it. Anyone have good suggestions for vibemaintaing platforms or tactics?
Thinking out loud, it feels like the anthropomorphic wave of robotics is taking a big swing in the wrong direction.
Rationale for humanoid robots seems to be:
1. Humanoids fit environments we already built to serve ourselves
2. Egocentric data is theoretically abundant for training (e.g. everyone starts wearing meta glasses)
3. Humanoids are the gateway for human-robot trust
> Environments can rather be designed for application-specific robot capabilities.
> Training data can be gleaned from simulations on these environments.
> Human embodiment carries the baggage of emotional distrust humans have towards one another
These are weakly held thoughts, curious how others are thinking about this
