chessucation

i grew up between germany and the US — having dealt with the public insurance system in germany, it's not (much) less admin-intensive than the US. my mom works in the public healthcare system in germany and has to fill out heaps of paperwork.

these days, i think most tasks come with significant administrative overhead as insurance companies want everything documented (and hospitals need to itemize everything to bill it).

i looked at very similar data as anthropic (was a bit amusing to look at these side by side) with physical tasks being part of the equation much more prominently (across all jobs).

if you're interested: https://largelabormodel.com/mirror?job=2221

some quick author notes here:

- the 4,811 tasks are derived from 480 jobs and measured against technical ability by deployed systems which then score each job on "replaceability"

- Technical replaceability ≠ replacement, could (by technical capability) vs. does/will

- wrote a short-ish essay on the data and how to read it: https://largelabormodel.com/two-jobs

- full methodology is here: https://largelabormodel.com/methodology-full

- the model is built to be disagreed with (and evolve). would love your input, thoughts, requests, etc. — most useful thing you can do: look up a job you've done and tell me where its task table is wrong

are these clusterings early evidence of the models training themselves, i.e. passing on and amplifying their own tics to successor models?

yes i think the writing itself already speaks "claude brain" — not criticism, personally, learning with/from machines to write well-structured and clear instructions has been quite helpful.

i consult startups and i am regularly surprised to see how poorly-structured their employee onboarding is across all levels.

a new hand who doesn't have to wonder what the body is doing will be much more effective than one feeling around for other limbs.

"this is a well-structured prompt. i'll pull up the relevant research and return with suggestions."

only a matter of time until it's "burning man proudly presented by palantir"

very interesting, didn't know my weekend was going to be filled with color science but happier for it, brilliant presentation too

ha indeed!

coming at this from the opposite end of not having been able to release or do much without the help of a team, i really appreciate being able to go fast and produce imperfect results that can then be iterated and improved.

but also find myself getting almost annoyed when i have to do things myself but without the joy or knowledge that these correction will help another human being become better at something but just feed an opaque database that will likely not learn much from it.

i'm also super curious about the impact of the auto-sorting on hiring

esp. in a time when formal (white collar) qualification will probably become much less relevant than ad-hoc problem solving

i've been dabbling with various assistant models and instances and find that there's an interesting "idea" that pops up from their synthesis and relevancy analysis at a decent frequency. some kind of "inverse prompting" in which the machine has an angle, perspective, opportunity (based on the context it's working with/in) and the human turns those seedlings into something useful... or else decides its weed to be discarded.

it seems that the shift is different in kind from previous ones (but then again, perhaps people thought that about previous labor disruptions as well)

take ATMs. they automated the task but the rung above it remained human. someone to handle exceptions, sell the mortgage, etc. running a branch actually became cheaper. cheaper branches meant more branches, so teller headcount went up for two decades.

what seems different now isn't that the models are better at tasks. it's that the rung above is being climbed at the same time. the coordination layer is getting absorbed concurrently rather than a generation later.

that doesn't mean no new work appears but every historical instance seems to have some kind of "escape route" which i'm not sure exists this time around.