Charley Cunningham

Sailing On

[Charley Cunningham @charley3c]

@michael_chomsky https://t.co/IUwmWchsJY

[Matt Gapp @mattgapp]

Today we're launching Dollars 💰

Most AI startups spend so much time and effort building an AI product with negative gross margins. We skipped straight to the end.

We sell US dollar bills for $0.80.

TAM: $1T+ in circulation. We’re going after all of it.

Let’s scale to the moon together.

dollars for less dot com

[Charley Cunningham @charley3c]

@d3layd @michael_chomsky https://t.co/QVUP3Gplb3 > https://t.co/1dexAylfB0

[Charley Cunningham @charley3c]

@benswerd @NirvikBaruah @ArjanChaudharyy > there are only so many ocean themed sandbox cos

just wait until the next YC batch, Garry Tan will bless us with an unlimited supply of newly funded hetzner wrappers (and some of them are bound to be ocean themed)

[Kavin Anand @kavin_anand]

🤔

[Charley Cunningham @charley3c]

@zack_overflow You forgot to include Sailboxes! Autosleep and usage based pricing means $0 for a forgetful weekend. https://t.co/AtDoCHLlPs

[Charley Cunningham @charley3c]

Current cost of this method is $1 quadrillion per ounce.

Would RL in a nuclear physics simulation engine environment be worth the compute? Need a superintelligent model to evaluate how to optimally allocate compute between this and other possible experiments

Alternatively we could capture an asteroid

[Dr. Dating @drdating007]

His Flirting Strategy Was Crazy.

[Charley Cunningham @charley3c]

@jediahkatz 🐐

[Nirvik Baruah @NirvikBaruah]

Active memory/disk billing is essential for hosting background agents at scale.

Our users' agents only use a tiny fraction of their reserved memory/disk but most sandbox providers make them pay for their full allocation. It seems incredibly obvious that you should do active-everything billing if you want to encourage ambitious long-horizon work, so I'm always surprised that sailboxes are basically the only sandbox that supports it.

[Charley Cunningham @charley3c]

@peywalt Winter is coming

[Charley Cunningham @charley3c]

@Aaroth @aryehazan Called it

[Charley Cunningham @charley3c]

@Aaroth Wish labs would allocate focus/compute to making this labor intensive step easier, rather than gain of function cyber RL

[Charley Cunningham @charley3c]

@Aaroth @majatrebacz might have some interesting thoughts here

[Charley Cunningham @charley3c]

@willdepue Interesting, maybe they distilled Astra onto the new model super quickly/efficiently (if that counts as pretraining)

[Charley Cunningham @charley3c]

@mehtaab_sawhney What did the drowning number theorist say?

log log log log…

[Charley Cunningham @charley3c]

@thsottiaux @thsottiaux Maybe people prefer using oai models in Codex or OSS tools via their oai subscription

But cursor defaulting subagents and new users / auto users to composer prob also contributes to this

[Charley Cunningham @charley3c]

@appsicle_ @thsottiaux @karpathy @OfficialLoganK @poteto cursor: @jediahkatz

[Charley Cunningham @charley3c]

@cwamidon Hmm maybe they need more datacenters

[Charley Cunningham @charley3c]

@kunchenguid @cerebras @sailresearchco 👀