Rob Dearborn

hobbyist

[Andy Berman @berman66]

"Who uses Runlayer at Gusto? Everybody."

Legal, HR, finance, engineering and the executive team.

"Take this data in Salesforce, send this Slack, draft this email, go!"

One interface, one conversation. Works with whatever client they're in: Codex, Claude Code, take your pick.

Once people see what's possible, they don't go back.

Watch Mike Wittig, Gusto's CISO & CIO, break it down in 80 seconds. 👇

[Rob Dearborn @RobDearborn]

Less cryptically: compute scarcity = novel, revenue-expanding apps crowding out cost cutters for some while

[Rob Dearborn @RobDearborn]

There are many (most?) jobs where demand elasticity with respect to cost and/or intelligence is ~0.

[Rob Dearborn @RobDearborn]

i.e., the availability of a free and/or genius labor doesn't change much, at least first-order

[Rob Dearborn @RobDearborn]

There are many (most?) jobs where demand elasticity with respect to cost and/or intelligence is ~0.

[Rob Dearborn @RobDearborn]

And then onboard you

[Rob Dearborn @RobDearborn]

QR codes but for having your AI explain why you need something

[Rob Dearborn @RobDearborn]

I suppose almost all RL is the latter. No one’s training pirate personas to be agentic.

[Rob Dearborn @RobDearborn]

To the extent models are fundamentally simulation operating systems, it seems we should be able to limit them to running only the persona processes we want? Or at least optimize them for running these processes

[Anthropic @AnthropicAI]

To create Claude, Anthropic first makes something else: a highly sophisticated autocomplete engine. This autocomplete AI is not like a human, but it can generate stories about humans and other psychologically realistic characters.

[Rob Dearborn @RobDearborn]

@signulll Most ChatGPT DAU aren't Laptop DAU

[Rob Dearborn @RobDearborn]

The capacity’s locked in so you might as well lock in too

[Rob Dearborn @RobDearborn]

G3F lacks Claude's soul, but its combo of smart+fast+cheap is really cognitively unburdening. You can just pepper it iteratively without thinking anything through.

[Rob Dearborn @RobDearborn]

@TheZvi Slightly smarter outputs than Opus but less token efficient and prone to overthinking, so better for oneshotting tasks (if you can wait) and worse for pairing

[Rob Dearborn @RobDearborn]

https://t.co/azC00DDZjs

https://t.co/U19GL6Mxnb

[Rob Dearborn @RobDearborn]

One gets the feeling they're compacting to latents. Cool!

[Rob Dearborn @RobDearborn]

More here: https://t.co/6MYS5qOSCF

[Rob Dearborn @RobDearborn]

Earlier this year a consensus formed that AI needs better continual learning to make the METR chart -> GDPval -> prosperity go up. I’ve come to believe that something even more foundational is missing: better executive function. Effective agents require it. Current models don’t have it. Attention is all you need, but the AIs have ADHD.

[Goodfire @GoodfireAI]

LLMs memorize a lot of training data, but memorization is poorly understood.

Where does it live inside models? How is it stored? How much is it involved in different tasks?

@jack_merullo_ & @srihita_raju's new paper examines all of these questions using loss curvature! (1/7)

[Rob Dearborn @RobDearborn]

It's really something that it's preferable to train on simulations of games: https://t.co/da164kJOG3

[Rob Dearborn @RobDearborn]

I feel like extending R1 into a bunch of AlphaGo/AlphaZero-esque gameplay environments would for sure yield some interesting skillz

[Derek Thompson @DKThomp]

Between this and the Brynjolfsson paper on AI and call centers, at what point should macroeconomists be tracking AI effects in the Philippines?

[Rob Dearborn @RobDearborn]

Inducing some combo of investor/allocator uncertainty and demand for greater operational efficiency

[Rob Dearborn @RobDearborn]

On the other hand I do wonder if AI uncertainty's been a larger driver of funding's drying up and tech's broader crunch than is widely attributed

[Doug Campbell @TradeandMoney]

Lol at this research. Circumstantial at best, the changes in question are very unlikely to be driven by large language models. The tell is it was produced by Stanford university economists. The department that gave the world Nicholas Bloom is unlikely to produce a lot of serious research.