Joe Hakim

Semi illiterate rock herder
--
ex PhD candidate at Harvard-MIT's HST

[Joe Hakim @JoeBHakim]

@tenobrus Are you 50 50 over under 25% year over year?? Would you wager $1000 (now-bucks), 1:1, >10% (year over year, so cum its like 61% (!))? Idk how rich you are so lmk your price, maybe $1000 today-bucks?

[Joe Hakim @JoeBHakim]

@ZekeEmanuel @nytimes Please when you post about x -> health, include which confounders they control for! Otherwise glib "isolated people ALSO are poorer/more rural/less active in going to doctor/..." could undermine validity

[Joe Hakim @JoeBHakim]

@sincethestudy $2k ~?~ most homes? Also look at how revolutionary techs diffuse slowly (vaclav smil wrote extensively on this), even among rich nations

[Joe Hakim @JoeBHakim]

@nagpalchirag I miss the era of your old survival phenotyping work, don't like but understand that EVERYONE kinda has to work on kernels or rl these days

[Joe Hakim @JoeBHakim]

Get it at https://t.co/9HEz7T0gFR

Works with acronyms, terms like "vramlet", "readme dot em dee" -> readme.md, "five point four millimeters" -> 5.4mm, etc. I don't enunciate IRL so that this works for me is pretty neat!

[Joe Hakim @JoeBHakim]

Last night tried a friends speech-to-text-to-claude stack on a mac, REALLY TRASH-TIER STT (!?)

Did a local whisper alternative

(note: linux+X11 only, but you can fork this for your system)

[Joe Hakim @JoeBHakim]

>- don’t train models for very high-risk capabilities without strong evals, justification, and containment

Agreed, dont even bother training biology or biomedicine or law stuff for that matter, or only contain it to Very Trusted Individuals, we need to prevent Evil™️ things TODAY at whatever cost to tomorrow

[Joe Hakim @JoeBHakim]

Make it legible enough that YOU can read it and should be fine, for local small models! DM me for details if you want, standard vibe-coded (but manually checked (but I'm not that assiduous)) disclaimers.

*I don't believe in elbows.

[Joe Hakim @JoeBHakim]

"should I not make typos in coding agents?"

Maybe! Did some local modeling:

[Joe Hakim @JoeBHakim]

...that uses a word regex to independently scramble characters (in words that are long enough, either swaps to transpose adjacent letters, drops leters, duplicates , or does keyboard adjacent keys.) CRUXEval adapted from Gu et al 2024.

Might try this on more or more expensive models, see if drop "elbow"* is later. My take:

[Joe Hakim @JoeBHakim]

It might look more like a "routing to whatever model doesn't excessively refuse, along the pareto-fronteir".

There seem to be 2 routing universes, open routing to best specialized OW model, and big-lab opaque routing to worse models for "safety". 4.8 works for pathogens, for now

[Joe Hakim @JoeBHakim]

@zakkohane Yep! Like the DARPA bioatribution challenge they just hosted to process ~ 1 petabyte in < 24 hours. Depending on method can be sort of embarrassingly paralell, think doable on 4 H100s.
Lots of other logistics GETTING that petabyte on a cluster though!

[Joe Hakim @JoeBHakim]

@difficultyang Id have more patience to actually dry run, do in several test dirs, tweak commands etc. with a bunch of cc candidates

[Joe Hakim @JoeBHakim]

Recently churned chatgpt -> claude $200/mo plans, and baffled there's no `pulse` equivalent, miss it!
...might work on a FOSS/ pay-me-to run-it-for-you kind of alternative.

Will happily fry it to oblivion if/when anthropic does this feature. EVEN GEMINI HAS SOMETHING LIKE THIS!

[Joe Hakim @JoeBHakim]

@norpadon Or just have the machines write it! But yeah, Typst >>, but don't think they have templates for most journals yet?

[Joe Hakim @JoeBHakim]

@agupta Actually I found entirely the opposite w.r.t bio work, relative to 4.7, which is allowing VERY cool hopefully useful stuff w.r.t. not thinking I'm a bioterrorist. Actually made me churn off 5.5

[Joe Hakim @JoeBHakim]

standard disclaimers with all my tools, but hope this is helpful for people learning/wanting to experiment. Currently working on some examples from hernan-robins' what if book. as usual, let me know what features youd like here! happy DAGing!

[Joe Hakim @JoeBHakim]

Epi, econ ppl, you ever use dagitty?
Decided to superset it with
`https://t.co/Ah9a52H0RR`!

A thread, to be updated later:

[Joe Hakim @JoeBHakim]

the way I see this being used is mostly internet-argument-winning.

A: chess elo is OBVIOUSLY correlated with IQ
B: actually, when you select for elite samples, there are cases where its negatively correlated!
A: bullshit!
B: (note the selection on elite sample = 1 out of {0,1})