Pablo Lopez

Exploring how humans and AI can collaborate to make everyday life a little less painful. Shaping AI at @duckbillai. All opinions are my own.

[duckbill.ai @duckbillai]

been getting a lot of texts like this lately!

[Pablo Lopez @PabloLopezDotAI]

@duckbillai @meetgranola amazing combination, your most omniscient notetaker now with unlimited real world access

[Meghan Verena Joyce @meghanvjoyce]

this turns meeting notes into results

[Pablo Lopez @PabloLopezDotAI]

@karpathy In @duckbillai we had this case where our Agent will bring up a user’s nut allergy continuously, for a dentist appointment, for getting a car maintenance appointment (making sure the place had no nut fragance)

[Pablo Lopez @PabloLopezDotAI]

@svpino you can do this with @duckbillai and a human assistant will do the payment step ;) can be from your imessages or via MCP, as you prefer

[Pablo Lopez @PabloLopezDotAI]

@Dropbox Why not finetuning if you have labeled data?

[Pablo Lopez @PabloLopezDotAI]

@luisgonzaleznf wtf, that's my PR. want to cover this in our next 1:1

[Pablo Lopez @PabloLopezDotAI]

@luisgonzaleznf that's my PR, @luisgonzaleznf we will discuss this in our next 1:1

[Pablo Lopez @PabloLopezDotAI]

@qadir5000 @dotta Hey, our @duckbillai MCP allows your agent to task humans with any remote task, for when you need more reliability or your agent is stuck with sth etc even for purchases is a better option.

[Pablo Lopez @PabloLopezDotAI]

@_kaitodev @DanielleFong @karpathy Where is AI engineers?

[Pablo Lopez @PabloLopezDotAI]

@luisgonzaleznf @NickADobos Agreed, “Tell agent to” is a pose

[Pablo Lopez @PabloLopezDotAI]

@LeverCRO @mattshumer_ LOL

[Pablo Lopez @PabloLopezDotAI]

@Polymarket @duckbillai is the real rentahuman that works, though it doesnt have any associated scammy shitcoin yet

[Pablo Lopez @PabloLopezDotAI]

@luisgonzaleznf @mattpocockuk 2 weeks of traditional ML research in 4 hours. That's when it hit me.

[duckbill.ai @duckbillai]

@Polymarket you can literally just text us

[Pablo Lopez @PabloLopezDotAI]

@jimprosser @goyal__pramod Hey @jimprosser we built an MCP that connects claude code with human personal assistants for phone calls, purchases, travel. I think you would love how it works in your setup: @duckbillai

[Pablo Lopez @PabloLopezDotAI]

@fchollet Just by looking at mathematicians playing sports and viceversa, I would say not much

[Pablo Lopez @PabloLopezDotAI]

This seems real

[John Rush @johnrushx]

I think they've revealed this themselves many times.

You can switch between low/mid/high/exhigh.

These levels are just enums with a coefficient multiplied by the base coefficient:

baseCoeff = 1
userCoeff = low: 1; mid: 2; high: 10; xhigh: 50
finalCoeff = baseCoeff × userCoeff

They adjust the base coefficient in real time based on load and their needs (they increase it during release week and gradually reduce it afterward).

[Pablo Lopez @PabloLopezDotAI]

@emollick I think the point is missed, code first, then research, then AI self improves. Then robotics and everything else follows. At least that the theory! IMO the agents will be perfect but the real world still fragmented, that’s @duckbillai bet

[Pablo Lopez @PabloLopezDotAI]

@hwchase17 Langsmith?