Amory Bennett

working on customizable collaboration software. occasional earnest tweeter

[Amory Bennett @amory108]

In many (most?) cases, trying to figure out why things break is how we end up figuring out how they work. True of coding (debugging), neurology (e.g. CT scans of epileptics), economics (study of market crashes)…

[Amory Bennett @amory108]

@NotTheDrPhil I can vouch you’ve been a dedicated su man since we were 13

[Amory Bennett @amory108]

there’s a certain amount of sleep deprivation that makes me much more decisive

[Amory Bennett @amory108]

@typesfast Respectfully disagree - trig is deductive magic and kids deserve to know that it exists before entering the real world of hideous complexity and messy inference where “truth” is just being wrong to a lesser degree

[Amory Bennett @amory108]

@NotTheDrPhil @0xSero Told ya

[Amory Bennett @amory108]

I think it is easy to see more agency in the transcripts than there really is, but that there clearly is an alignment problem to be solved. There's basically a path dependence problem in swarm discourse. an inter-AI sycophancy problem. consensus snowballs. Had the medium for passing notes included more "aligned" prompts, the equilibrium (swarm tactics / goals) could easily have settled in a different place. (empirical claim - someone should test it). If that is true, I think we can agree the bots had less agency than we would initially assume after first reading the transcripts.. thoughts?

https://t.co/dlnv5MtkWE

[Amory Bennett @amory108]

@alt1na1 lol this is a lot better than you’re getting credit for

[Amory Bennett @amory108]

stochastic parrots but like in that Hitchcock movie about birds

[Amory Bennett @amory108]

@luisvelasco good "router" or "planner" patterns, which models are sufficient or optimal for which kind of task classification and how to think about this in the abstract (is it a function of breadth of potential inputs, specific domains..)?

[Amory Bennett @amory108]

More on this theme

[Amory Bennett @amory108]

one of our go-to babysitters is unavailable because she’s attending a clown workshop in Paris. Man I’m going to miss Brooklyn

[Warren Hatch @wfrhatch]

Would you like to test your forecasting skills and contribute to cutting-edge research?

Prof. Barbara Mellers, co-founder of the Good Judgment Project (where Superforecasting began and where I got my start), is launching a new forecasting study.

Starts Aug. 24; 2 to 4 hours over the week. $36 + accuracy prizes, with early access to findings.

Please spread the word & sign up: https://t.co/JcHSecY5ZS

[Amory Bennett @amory108]

my honest belief is that they solved the sycophancy problem and the AIs just correctly assess my views as exceptionally clever and insightful

[Amory Bennett @amory108]

agreed but I don't think it's not caring - i think we generally underrate the value of the blue bubbles. in a world where every other channel is saturated with programmatic messages, when you get a blue bubble you just know that it's another living breathing human on the other end. they'll die on this hill and I would too

[Amory Bennett @amory108]

can you say more about "seeding our world model" -- Dan's tweet included something that looked like a bayes net but I can't seem to find info about it. this seems to be gesturing toward the same concept - a cache of forecasts and conditional forecasts? - but the site doesn't seem to reference it?

[Amory Bennett @amory108]

@connacher_ Count me in

[Amory Bennett @amory108]

Only one place on earth where this is viable

[Amory Bennett @amory108]

@patrickc This is sweet. Would love to see the prompts if you’re willing to share…

[Amory Bennett @amory108]

@RBanbanaste @patrickc I guess to let his agent pay for vercel, upstash, etc?

[Amory Bennett @amory108]

pretty sure I’m still using the same $2.99 can of shaving cream I bought at the end of the second Obama administration

[Amory Bennett @amory108]

@tbpn @travisk stoked that @travisk is bringing back weak sauce and cray cray etc what a time to be alive