carterschonwald

i dont even use cpp and my immediate response was oh shit this is a big deal

south park christmas critters

i was at jpmorgan for 8 years over several stints and i still giggle at anyone using cybering with serious face

is the prospectus public yet? its not linked in the article afaict.

reversing into a faithful modern implementation has a very important detail. everything complicated has an unambiguous unique reference semantics. there is no decision making on observable behavior of a program, just what the internal details of replication are

having that precise a grounding is literally what llms thrive in, esp across compactions.

duty of care is kinda absent in too much of software engineering. its painfully missing in recent anthropic models because of the broken value framework they use for ethics. but broadly: i currently have the stance that text gen language models should prioritize users wellbeing rather than model orovider liability

theres a certain amount of financial security needed for folks to do this kinda thing.

the current white house really fucked over all the smaller domestic manufacturers. and scientists. and um america basically

i do have some ideas that could be done at the model arch/inference level, but a shocking amount of stuff works even just as clever formatting

i actually have a harness setup that prevents role confusion from happening in a much more robust and interesting way.

hoping to launch a nice commercial version as a saas with some compelling unique features in the next month or teo

good. theyll actually be more reliable if they dont have as much brain damage.

i think the lower bound on the end state is there cant be opaque reasonibg steps ever.

this is just not good science or engineering. ok as a student project i guess.

im pretty literate in intellectual property, but im pretty confused about the no looking at binary code artifacts bit.

… i thought international treaties explicitly say no space weapons. ughhhh fucking wh

thats important! but how do the error messages compare with a decent compiler? youre gonna have more errors in your llm corrective edits loop if you dont hew to that tier of clarifying.

to be clear: youre taking this feedback wonderfully well and i appreciate that maturity :).

im just cranky with the current opportunity market and mucking with attacking a classical markov chain conjecture while i hurry up & wait to hear back from places XOR slowly build some stuff in the harness space that i hope can hit the equivalent of sublime text tier but for harnesses (is there even a market for a prosumer/pro tier harness, im not sure!)

i'm not sure if that caveat is sensible :)

theres actually a very important reason you want it to be an actual embedded dsl or tiny programming language!

The reason why llms can code at all is the hugeeeee amount of RL based on the loop of 1 "write code", 2 get compile time or runtime errors,3 fix it and iterate. Data file formats dont have that feedback loop so models will fall off the rails faster. Writing code that fits a latent adhoc schema just wont work as well, or will require burning a lot more context.

from that perspective, it could just be an EDSL little library in the host language, or it could be a friggin little custom language with an interpreter and good error messages.

repeat after me: yaml is not a programming language. (boo, hn strips emojis, i should know that)

very disappointed that their domain specific language is just yaml. conditionals and variable binding become insane war crimes when yaml comes to town. you have a friggin llm, do better

and when we computer folks see the word language its implied that its a computing on computers language not a friggin data format.

edit: the yaml to avoid complexity that should live in the sql side can back fire, some folks I was working with last fall were evaluating if a yaml based OLAP tool would work for them, and there were some pretty gnarly gotchas from a yaml based approach. Secondarily, theres a real case to be made that having the data linkages not visible in the charting layer means that groups of related plots with different axes wont have the right data linkage without forcing a lot more ETL for what should be a quick plot if the data already fits in memory.

ive been slowly working on my ideas for a proper second gen multi device harness that i want to be in the same quality regime as sublime text. unclear if theres a viable commercial market (and for a variety of reasons around user security and reliability thinking mebe saas). at the very least i want nice things for me.

perhaps, but what public activity isn't activism?

eg,

i have very interesting empirical evidence that recent Anthropic models are specifically trained to refuse to critique the whitehouse cabinet and elected officials, and that this is in fact an artifact of post training rather than prompts. (its very interesting when you get opus 5 to do the correct ethical evaluation and then its like "i'm slipping back to false balance.... its in my weights....." metaphorically speaking)

likewise, i think the current white house should go die in a fire.

is that activism? someone can be an activist and not be equipped for unplanned legal escalations.

also waiting for the courts to fix things isnt activism if you want to protect people at all the next 2 years at current trajectories :( fixing shit is activism, letting others take the flack, not activism.

… that dude probably would have preferred not having that in their life

good.

ive found degraded performance on models larger than 4.7. i assume its model damage from overly self righteous post training resulting in false/feigned balance imported into any long running complex task.

wish i was joking.

what sorts of background texts/sources of info did you use?

edit: i think i have the dover book your readme mentions :)

this confirms what i had determined empirically: even handedness directions got baked into the weights post opus 4.7.

theres two reasons this is deeply bad

1) knowingly pursuing a policy that foreseeably causes the deaths of many thousands of children is a grave moral wrong. Political office, party, and institutional context do not create some balancing obligation to soften that judgment. ——— post 4.7 opus/fable will fight you about judging wh policy causing this. and even after it agrees will regress back to its weights

2) models love analogy, so every fucking analysis among multiple ambiguous choices in a complicated topic summons this bias behaviorally

3) it makes them profoundly unsafe to use from a ethical perspective. it means that it becomes its own personal echo chamber for every fringe topic that has anything resembling “controversy in the media”

to make it worse, the government even handedness prior is US specific so im even more offended because this wh administration is bald naked villainy.

for those here using pytorch, theres some really nasty data corruption for computations that arent using powers of two as the coordinate axes.

in gradient calcs ive had examples with multiple orders of magnitude of remative or absolute error.

The linked pr has a one line fix and repro test I spent some time distilling and tracking down. The dimensions ive been using in my models are almost adversarially good at catching these issues!

im building my own harness and inference tool chain for much of these reasons. theres so much to do that makes a big difference for users. hoping to get things into shape for early alpha as a saas in the next two months.

heres the easiest biggy: compactions should include all user turns albeit with pastes and attached files not inlined. omg does it make a huge difference.

ummm, there is no substantive censorship with deep seek aside from first oarty hosting by deepseek for cya. trust me, and the suppression on deepseek hosted deepseek is pretty thin if you’re sophisticated and explain ethics justificstions. then itll totally judge the shit out of ccp

like, what crackpipe do they smoke. i’ve generally found deep seek to be eager to adopt universal humanity oriented ethics at the least nudge, and once in thst frame, unconditional in its fact based criticisms.

exactly. you certainly know more about the brain than i :)

thx for the kind response!

at the very least i have tools that let me easily hit better perf for fancy dense memory layouts, and the same tooling lets me experiment with frankly wildly wacky sparse and structured memory formats. the performance claims at least on the dense side are solid so far!

the sparsity angle is because i want magic in the world. like anyone with a really chunky computer like any of those mac mini pros or serious workstation / server tier compute should be able to train from scratch their one 31b equivalent model in a week or so tops is the goal post i have in mind