Ryan

solving the subsurface @mineflow_ai

[Ryan @ryan__goggins]

@StephWolski @nico_laqua @nico_laqua it's important that you're on the right side of history here

[Ryan @ryan__goggins]

by the end of the decade, there will be many molybdenum mines operating in downtown san francisco, it will be commonplace like a "corgi cafe"

[Ryan @ryan__goggins]

we made all of the functionality on mineflow free, only thing that costs $$ is downloading our data or uploading your own

[Ryan @ryan__goggins]

@chris_dalke this is sick

[Ryan @ryan__goggins]

@maxvonhippel frap

[Ryan @ryan__goggins]

@celestepoasts opus 6 might be able to one-shot gold grades in the subsurface :0

[Ryan @ryan__goggins]

Launched cross section vectorization today in Mineflow! We can now automatically pull cross section images out of documents and project them into 3D space and let you export them to other geological modeling software. Documentation here: https://t.co/JnybMsHxwr

[Ryan @ryan__goggins]

@technoHimbo top 5 book review of all time

[Ryan @ryan__goggins]

@technoHimbo prescient

[Ryan @ryan__goggins]

we hit sota on both map vectorization and georeferencing earlier this year. happy to make this customer facing now :)

[Mineflow @mineflow_ai]

State of the art map vectorization is live!!

[Ryan @ryan__goggins]

new OAI swarm message board just dropped

[Robert Francis @rbfrncs]

Amazing things are happening in the Original NYT Plum Torte comments section

[Ryan @ryan__goggins]

mineral exploration data extraction is solved internally

[Ryan @ryan__goggins]

@joelhooks https://t.co/GNrPprPqVE

[Ryan @ryan__goggins]

@she_llac good find

[Ryan @ryan__goggins]

@PhilipJohnston @FermiExplorer lot of good copper out there, check out escondida if you have time

[Ryan @ryan__goggins]

@she_llac how do you go about searching for these? enumerate GET-safe semi-popular forums?

[Ryan @ryan__goggins]

could be an interesting interp mechanism to instruct the model to obscure reasoning in {legible space} and then, as long as it can still produce the correct output, identify what part of the model is doing the reasoning in place of the original legible space (eg. identifying j-space with probes), then instruct the model to obscure reasoning in {original legible space + newly identified legible space} and continue iterating until the model can no longer produce the correct response? the shaky assumptions are: 1. there is no reasoning baked into "I will focus on a calm visual scene..." and 2. that the models lack eval-awareness. As long as these hold, this could be a means for exhaustively finding more sources of hidden reasoning? or at least proving that they exist

[Ryan @ryan__goggins]

name 5 notable men

[Ryan @ryan__goggins]

@danlovesproofs sick, always wanted stuff like this to be athing https://t.co/AdZuMWhXrQ

[Ryan @ryan__goggins]

there's definitely money to be made in producing a company that just sells a set of prompts/skills that clean up your codebase and consolidate abstractions and make it more readable and net-subtract LOCs from the codebase. prompts/skills that are idempotent if run a few times in a row, but very useful to have running once a day or so