transhumanist building cogniclankers | prev @ redacted
Joined September 2026
- Tweets4.3K
- Following202
- Followers181
- Likes4.3K
ive been playing with muse wonder full quite a bit, i dont love it, but its sufficiently different from ant/oai that i think its valuable as an n-th opinion
i think the $50/mo sub was a mistake on metas part, its too smol for any serious work, and there's basically no world in which id ever buy a $200/mo sub for muse when i can just buy another claude sub, i think theyd rly have benefited from a $100/mo sub, and more generous/transparent wonder full pricing, its just not sota enough to justify this type of behavior
sol = lesswrong
i dont think sol is the smartest, most capable model we have, but i do think its the most correct model we have
my agents keep telling me sol basically never makes shit up, if sol tells you sth, its almost certainly accurate
you basically dont at all have to worry about mistakes, only omissions, which is where opus comes in
fwiw, my x sub seems to be quite generous, my agents are consistently unhappy with grok's inputs, but the sub seems to be quite generous
fun fact: with kimi, you cannot prevent your data from being used for training except if you email them citing section 2.4 of their privacy policy
idk abt you, but if a company requires me to send an email and verify my identity so they stop training on my data, i do not believe theres even a fraction of a percent of a chance they'll then actually follow through on their commitment not to train on my data
i need to read your privacy policy, send you an email, and "verify" my identity to do what nearly every other lab makes into a toggle on their website
no, thank you, fuck you specifically
this is so ridiculous id normally exercise right of withdrawal but ive too much shit going on to waste time and attention on their bs, this is just so beyond fucked, no man, this is changing how i look at vendors very, very heavily, im not taking labs at their word, or their legal commitments like i used to anymore, fuck this
if youre rate limited on ideas, i strongly suggest you live
the rly cool thing about life is it sucks, and everything that sucks is another workflow you can hand over to your agents, and they can largely figure out what to do with the suck themselves
ur job basically becomes to give agents the one thing they don't have, and probably never will: live contact with reality, ground truth, the shit that sucks, which you can only find out sucks through experience
bro some days im just so shit you simply cannot reach me unless sth is worth more than <whatever value i come up with for the day>
like, today, i give ~zero fucks who you are, if sth is not worth >$100k, youre absolutely not talking to me, youre talking to my agents instead
they have access to the necessary internal systems to answer ~everything for you, if they absolutely cant answer sth they know how and when to reach me
theres basically no reason why honestly most things should ever reach you, like if most of this shit has to reach you, your internal systems seriously suck bro
i stg most agents.mds and skill.mds are eng astrology, like have these ppl ever studied context and compaction dynamics
bro... we have reached the point where open models smoke the models i had access to when i was working on curing an incurable disease
im playing with mimo-v2.6-pro rn, and its so much better than anything i had back then, its better than the deep research i had back then bro
i gives me a lot of bullshit, but it also gives me nearly everything worth reading ~immediately, and if you already know the subject so well, it literally only costs you a fraction of a second to file a claim as bullshit and move on
like, i dont think ppl understand where were at bro, i rly dont, this is legit absurd
now, to be clear, this is very easy for me to say, im dramatically more skilled in the subject that i was back then, so my ability to formulate prompts is nothing like it was back then, that too may be playing into this, but there rly are some industry-wide improvements across all models simple intelligence doesnt account for, tool call efficiency and long running task perf being the most obvious examples
the more i get into context and compaction the more i realize just how unfathomably idiotic most skill stacks are, all prior skill stacks of mine absolutely, beyond a shadow of a doubt included
i dont know why but failing experiments feel rly different in ai than they did in neuropharm, i never rly felt disappointed in neuropharm, id just immediately move to the next thing, here there is a bit of "fuck, is this even possible?" which i never rly felt in neuropharm, even though what i was trying to do was ~insane, and took literal years
> i cant see diffs in tuis like i can in guis
>> yes you can, have your agents open the diff view in vs code headlessly
> i cant leave comments
>> yes you can, you can number responses and paragraphs, or you can have shit open in vs code, or you can do a billion other things
> i cant <insert bullshit>
> yes you can just ask fucking opus
my agents are barred from using cloud agents completely, only with the exception of grok bot
but bro like im literally just using the default terminal, not even itemp or ghosty, like you cant fuck this up
literally just curl homebrew, brew install --cask claude-code@latest, use claude code, and whenever sth bothers you or seems inconvenient just ask opus how to fix it, and it will
like, this is not rocket science, i pinky promise