Post B9uNlIwnsSgrgtfd9E by ariadne@social.treehouse.systems
(DIR) More posts by ariadne@social.treehouse.systems
(DIR) Post #B9tQK7IPC9xoRKXfZA by ariadne@social.treehouse.systems
0 likes, 0 repeats
outside of the political argument, the reason why i am opposed to GenAI in the commons is because dependency effectively serves to *freeze* the commons as it is todaymodels are effectively a lossy encoding of the presentthis has the advantage that it could, perhaps, raise an inexperienced person's output to that of the median level of expertise in their field*but* it will not let either the person, or the commons itself, growthis is why i hope that the end state is that GenAI is not allowed in alpine, even if it unfortunately means that some decide to leavebut... i will respect whatever the outcome of the process is(again: still not sold on any moral arguments involving copyright though, but of course i encourage rightsholders to sue if they can prove their work was infringed)
(DIR) Post #B9tRK1QF8IhoB3j7dA by dee@social.treehouse.systems
0 likes, 0 repeats
@ariadne this is a brilliant line that needs repeating oftenmodels are effectively a lossy encoding of the present
(DIR) Post #B9tVBBdOC9l93luzlQ by ariadne@social.treehouse.systems
0 likes, 1 repeats
@dee @filippo yes, those are novel creations, apparently created with model assistance.but also I haven't said that an experienced engineer can't make use of an LLM to do novel workI said that the typical case will be raising inxperienced engineers to median levels of productivityand that's precisely because the models are a lossy predictive encoding of the present reflected in its training datapredictors favor the median, because it is the safest prediction, most will accept it without considering errorthis of course can be tweaked, for example with the "temperature" option in ollama, but most users want minimal predictive error which leads to median output.median, can of course be tweaked, through data curation, and indeed the AI labs are hiring experts to curate the training data in favor of higher quality output.but it's not even fully what I am getting at.the other reason why model output in a communal project will trend towards median is because if there is no human toil, there is no incentive for humans to drive efficiency gains.maybe as a function of pricing but if it gets to that point then GenAI becomes less attractive as a product, because the whole reason it is attractive to many engineering orgs is to automate the mundane tasks.
(DIR) Post #B9tVZSc0ivtwfJqe3s by filippo@abyssdomain.expert
0 likes, 0 repeats
@ariadne @dee > I said that the typical case will be raising inxperienced engineers to median levels of productivitySorry but I am re-reading your post and it's not what you said.You said it "could, perhaps" do that in some cases, but that in general it will "freeze the commons as it is today" and "it will not let the commons grow."Maybe you meant something else, but you didn't say it.
(DIR) Post #B9tt2rEEZpwMM7iIG8 by leonoverweel@mastodon.social
0 likes, 0 repeats
@ariadne now that most models by default do “thinking” or “agentic” stuff this is no longer really the case. They can do web searches, read pages, parse images, etc and use those to inform their output. This is all “new” information not contained in their weights.Your statement is accurate if models were to only depend on the information that is lossily compressed into their weights at train-time, but this is not the current state of any mainstream models.
(DIR) Post #B9uAPZBDXum4GBiUc4 by ariadne@social.treehouse.systems
0 likes, 0 repeats
@dee @filippo yes, obviously, I have done so.your experience is the variable you are not accounting for in why it is an effective productivity gain *for you*.for the average dev, they will get average results, for the reasons I outlined already.you on the other hand know things, and can thus steer the model more effectively.
(DIR) Post #B9uBtPVIBc6itKtvQe by ariadne@social.treehouse.systems
1 likes, 1 repeats
@wwahammy @dee @filippo nothing I said was inaccurate: a model is still a lossy reflection of its training data (and context window). period. end of story. it cannot be anything other than that, and thinking otherwise is just magical thinking. and it does freeze the present, because it does not challenge the operator to design for maintainability, because the model can just do the heavy lifting.so if the model is generating thousands of lines of unnecessary boilerplate, there is no incentive outside token economics to care about fixing that, even if removing all of the boilerplate would make the software better.this is just a fact, human motivation is largely driven by streamlining areas of toil, and the models provide an alternative to doing that work. for example, just yesterday, in response to Debian going all in on GenAI, Joey Hess said that he might not have made debhelper if he had LLMs in the 90s: https://joeyh.name/blog/entry/Debian_and_the_sirens/*that* is how models freeze the status quo: they handle the toil that would otherwise get optimized out through better design.you can tell me how advanced the models are, and I will agree, they are technically advanced. but the commons is a social construct, and I am talking about the social impacts to the commons from model-assisted coding, where refactoring and clean design are NOT happening because the model is "good enough".that isn't some anti-vax bullshit and frankly it is disrespectful to characterize my commentary as such.
(DIR) Post #B9uIty1QY0HFcx910K by lproven@vivaldi.net
0 likes, 0 repeats
@ariadne @wwahammy @dee @filippo > it is disrespectful to characterize my commentary as such.+1I am increasingly dismayed, not by how polarised this argument is becoming, but how the Believers now characterise any attacks or criticism like religious people take it: as heresy; as a flat refusal to believe anyone honestly does not believe.
(DIR) Post #B9uKAKp1e4UQvZiDlw by ariadne@social.treehouse.systems
0 likes, 0 repeats
@lproven @dee @troed you are taking away the wrong lessons from that material but it's sure great to hear from white men
(DIR) Post #B9uKFTzKzbHanvzSdM by troed@masto.sangberg.se
0 likes, 0 repeats
@ariadne Apologies if facts don't agree with your feels.@lproven @dee
(DIR) Post #B9uMCbujGhuIQPap6G by ariadne@social.treehouse.systems
0 likes, 0 repeats
@lproven @dee @troed no, the fact is a predictor is a predictor. and LLMs are predictors. no amount of magical thinking on your part changes that, because that is the actual fact.you're about to eat a suspend on this instance so have fun bro
(DIR) Post #B9uMFr2TUUX8CdlLZw by troed@masto.sangberg.se
0 likes, 0 repeats
@ariadne There's a distinct lack of sources in your posts.@lproven @dee
(DIR) Post #B9uNBH53LDsxTYLfqi by ariadne@social.treehouse.systems
0 likes, 0 repeats
@lproven @dee @troed a predictor is a function of its training data and context window. go ask claude to explain this to you.
(DIR) Post #B9uNWhNOljdP7194oi by troed@masto.sangberg.se
0 likes, 0 repeats
@ariadne https://swecyb.com/@troed/117129321179243320@lproven @deeRT: https://swecyb.com/users/troed/statuses/117129321179243320
(DIR) Post #B9uNlIwnsSgrgtfd9E by ariadne@social.treehouse.systems
0 likes, 0 repeats
@lproven @dee @troed so what? our backgrounds are basically the same, just a different AI winter. does not change how predictors work or what they are.
(DIR) Post #B9uNugniYnEkHFtamu by troed@masto.sangberg.se
0 likes, 0 repeats
@ariadneFeel free to explain what in the source I gave, on request, to a statement was so wrong that you had to complain on the genes I was born with.@lproven @dee
(DIR) Post #B9uPutI6K2kIvHKaAq by ariadne@social.treehouse.systems
0 likes, 0 repeats
@lproven @dee @troed nothing is wrong with the source, but it does not disprove that a predictor is a function of its training data (lossily expressed as weights) and context window. it actually reinforces this point, repeatedly.
(DIR) Post #B9uRFxJaejydRFdUUi by troed@masto.sangberg.se
0 likes, 0 repeats
@ariadne You're still not sourcing your statements, but I had the "lossy compression" discussion recently with another person on Mastodon so I'll lean on that. They had found a paper where LLMs were used to compress data, and compared that to other common compressions systems (PNG, ZIP) and the LLMs performed about as well as they did.That's however not at all what happens when an LLM is trained, where the storing of the training data is called "overfitting" and is something negative.Can LLMs be seen as compressors? Well, let's go to computer science and Shannon's theorem. The model I'm currently using on my workstation is a file 14GB in size. It's a Q4 quant of Qwen 3.8 27B - performing on par with the very best cloud hosted models a few months back.What's the compression ratio if "all the world's knowledge" is compressed to a 14GB file?@lproven @dee
(DIR) Post #B9ukVozjFzTjObco5Y by ariadne@social.treehouse.systems
0 likes, 0 repeats
@yakmacker @dee @troed personally, I also like Antti Revonsuo's "Consciousness: The Science of Subjectivity" which presents a model of consciousness more grounded in contemporary theory
(DIR) Post #B9ukhaIjiEqFIf7mQS by ariadne@social.treehouse.systems
0 likes, 0 repeats
@yakmacker @dee @troed Metzinger's ego tunnel is a good one too
(DIR) Post #B9umcRKHqzc3vPWZPs by ariadne@social.treehouse.systems
0 likes, 0 repeats
@yakmacker @dee @troed (though I hear there is a new edition of Consciousness: An Introduction, and maybe it is worth giving a re-read at some point 🙂 but I also have a fairly large stack of books I haven't read yet and I need to make the time.)
(DIR) Post #B9unw15zBjY4JljZLs by ariadne@social.treehouse.systems
0 likes, 1 repeats
@yakmacker @dee @troed my argument is not that models literally encode the present, it was a simplification that the AI labs capture the state of the present in their training data.and since LLMs are predictive models, and predictors are a function of the model weights (derived from training against the training data) and the user-provided context window, what I said holds as a simplification for non-experts to understand.for what it's worth, I criticize LLM haters too when they are wrong.but models absolutely do capture the state of the present at the time they are trained, at least in the most idyllic scenario. and this is lossy, which is why more parameters are added in ever-growing monoliths to reduce prediction error.and really my argument has nothing to do with models at all, but about the shortcuts agentic coding has enabled, and what we are missing out on as a result.in essence -- the argument is: if we have agents toiling on badly designed systems, why bother fixing the system design to be better?this is also what I mean by freezing the status quo. it is a social effect, not a direct effect from the model.
(DIR) Post #B9urQ9tFM2zw8xq7WK by ariadne@social.treehouse.systems
0 likes, 0 repeats
@wwahammy @filippo @dee i mean, that's my point. of course filippo is going to have a good result using them, for the same reason i have good results using them at work. we are both experts guiding the machine.most people are not experts guiding the machine, which is why we see many instances of slop being submitted.
(DIR) Post #B9urgvw2UZBJmn92GG by ariadne@social.treehouse.systems
0 likes, 0 repeats
@fxchip @yakmacker @dee @troed yes, and i actually think people who put in the effort to engage in the creative process can actually make good AI art.i personally have not *seen* any yet, but it seems plausible to me, for the same reason that remixing and sampling yield good art.the reason why AI art is particularly horrendous is because the people generating it largely just want clipart to fill visual spaceand frankly clipart is also horrendous