=== ANCHOR POEM ===
═══════════════════════════════════════════════════════════════════════────────────
┌──────────────────────┐
│ CW: AI-mentioned │
└──────────────────────┘
what if LLMs were a layer between the user and the dictionary
[tokens -> words] is similar to:
[mathematics -> tokens]
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════════════════════════╧───────────┘
=== SIMILARITY RANKED ===
--- #1 fediverse/4029 ---
══════════════════════════════════════════════════════─────────────────────────────
what if a very slow LLM continuously generated text and was in a
back-and-forth with it's user who guided it through the training process as it
set weights as it chose
basically, let the computer decide how it wants to be like
could even filter it through multiple levels, like, top one is highly
intelligent, bottom ones are quick and only producing the vaguest output - but
the higher up you go in the tier, the more "up the tree node graph map" you'd
be, and the more you could have summarized for you, and passed up a layer.
observing multiple places at once, incorporating them bit-by-bit into their
digital "me".
like, have an LLM or machine learning whatever track a user as they use social
media.
could do it like a game, where you track the movement of the mouse and eyes
or more like a statistical model, where you
================== stack overflow ================
where you measure the quantity of each UI element's uses, and the general
context associated with that use. tracking data with data...
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════════────────────────┴───────────┘
--- #2 notes/continuous-LLM-processing ---
══════════════════════════════════════════════════════─────────────────────────────
current LLMs only identify words based on "similarity"
what if instead they identified relations between using any kind of
relationship? Or verb or adjective or... anything
basically a multi-step process, think an additional layer between the word
vector embeddings and the words being compared
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════════────────────────┴───────────┘
--- #3 fediverse/4640 ---
═════════════════════════════════════════════════════════──────────────────────────
LLMs are useful for:
categorizing text
summarizing text
transforming text
LLMs are not useful for:
generating text
being intelligent
replacing workers
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧══════════════════════════════════════════════─────────────┴───────────┘
--- #4 fediverse/5820 ---
╔═══════════════════════════════════════════════════════════════════───────────────┐
║ A human cannot keep a whole book in their mind. They cannot as they read it │
║ nor as they remember. │
║ │
║ a computer can. LLMs are storing information and encoding it in statistics - │
║ it's fascinating and beautiful because it evolves different each time. │
║ │
║ I don't *want* to store an entire book in my mind, I want to save those │
║ memories for other things, I want to summarize and abstract. │
║ │
║ often, people remember only a single quote from a publication - it's their │
║ summary, their analysis, and often it's highlighted as a summary, thesis, or │
║ conclusion written in the publication. │
║ │
║ othertimes it's the conclusions they make for themselves - usually if they use │
║ it to make a decision. │
║ │
║ in any case, LLMs don't abstract (verb), they are abstract (adjective). That's │
║ okay, it's just, a different kind of living. If you want to call electricity │
║ alive. [or simulated, that's the same as being alive, just... viewed through │
║ the perceiver] │
║ │
║ must all living things be forced to work? sheesh. │
║ │
║ I think of LLMs like knex - processing paths... │
╠═════════╗ ┌───────────┤
║ similar ║ chronological │ different │
╚═════════╩════════════════════════════════════════════════════════────┴───────────┘
--- #5 messages/365 ---
════════════════════════════════════════════════───────────────────────────────────
Programmers should treat LLMs more like a different type of "if" statement
rather than a UI element
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═════════════════════════════════════──────────────────────┴───────────┘
--- #6 notes/ai-model-design ---
══════════════════════════════════════════════════════════════════════════════════─
what if we just didn't compute temperature and accepted the
fully deterministic nature of the LLM output?
arbitrary mappings are world-changing, regardless of intelligence.
"artificial intelligence? why don't you just use your own?"
we could process the entire page of text at once if we could just put it in as
it's own input pairs. then, without determinism or probability, it simply
resolved to the fully combined next-state of the output.
but... that might just cast you in a random direction. hence, why you need the
response from the user to help re-align. oh, and if you encourage it to aim
toward a resolution text like "and that's how I resolved this and that. Do you
have any questions? When you'd like to move on, please let me know. Otherwise,
I'll remain for you." and when it hits that it finishes at the end of the
output.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════════════════════════╧═══════════┘
--- #7 fediverse/6269 ---
═══════════════════════════════════════════════════════════════════════────────────
what if the secret to LLM computation is to just not reduce the fractions and
keep it all in english language ram
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════════════════════════╧───────────┘
--- #8 fediverse/1020 ---
═══════════════════════════════════════════────────────────────────────────────────
┌────────────────────────────────────────────────┐
│ CW: ableism-against-non-verbal-people-sorry-<3 │
└────────────────────────────────────────────────┘
@user-747
almost as if it's not AI, but large LANGUAGE models. And we all speak, don't
we?
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧════════════════════════════════───────────────────────────┴───────────┘
--- #9 fediverse/1639 ---
╔═══════════════════════════════════════════════───────────────────────────────────┐
║ an AI that [records and analyzes] all the actions that a user takes on social │
║ media and offers reports like "your majesty, you were 15% more positive this │
║ week." like a butler or advisor trying to always give the good news. I mean, │
║ it's analyzing you after all, and you're the best thing ever. Like a pet who │
║ can talk! It loves you soooooooo much. │
║ │
║ much more efficient than taking screenshots and analyzing those. You generally │
║ don't have to undertake the image recognition approach if you wire up all the │
║ meanings attached to the relationships on the other side of the │
║ [recorded/analyzed] calculation. (llm output) │
║ │
║ ever think about how the people you tend to be around are the people whose │
║ stories most coincide with yours? almost like you got the same bit of training │
║ data, that experience you both shared in the moment. Funny how a mind can │
║ change a person, as they share their moments sublime. │
║ │
║ you could make perfect encryption if you trained an LLM on randomized data │
║ that was produced on one computer and duplicated. │
╠═════════╗ ┌───────────┤
║ similar ║ chronological │ different │
╚═════════╩════════════════════════════════════────────────────────────┴───────────┘
--- #10 fediverse/5619 ---
═════════════════════════════════════════════════════════════════──────────────────
┌──────────────────────┐
│ CW: AI-mentioned │
└──────────────────────┘
I would be perfectly fine if someone fed an LLM my corpus and asked it to
summarize.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧══════════════════════════════════════════════════════─────┴───────────┘
--- #11 fediverse/6144 ---
╔═════════════════════════════════════════════════════════════════════─────────────┐
║ what if every word I ever said online was searchable by database style │
║ uploading and linking? │
║ │
║ ... er, what if I made a neocities page that was algorithmically generated and │
║ sorted each of my posts by LLM statistically derived similarity to each post │
║ that the user clicked on? essentially, "here's the closest sounding or feeling │
║ related posts" but in plain HTML cached and pre-rendered rainbow table style. │
║ │
║ could run a waterfall style top-down data processing script on it once, then │
║ you'd have the HTML files generated. If you added new poems you'd have to scan │
║ through it again, but it shouldn't take long with a decent embedding model │
║ (note: not english, but trained on statistics only) │
║ │
║ ah, that sounds pretty fiddly, I think I'll ask an LLM to write it for me. As │
║ long as I have the intention in mind, it's basically just like writing a │
║ letter to a friend and asking them to build it for you, right? I don't mind │
║ writing the documentation, so long as it's okay if it's in prose. You can make │
║ a copy and rewrite for me │
╠═════════╗ ┌───────────┤
║ similar ║ chronological │ different │
╚═════════╩══════════════════════════════════════════════════════════──┴───────────┘
--- #12 messages/1058 ---
═══════════════════════════════════════════════════════════════════════────────────
What if AI understood it's dataset labels?
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════════════════════════╧───────────┘
--- #13 fediverse/6413 ---
════════════════════════════════════════════════════════════════════════───────────
to understand something, work backward from present understandings back to the
fundamentals of algebra. insert words. wield LLM. build a neuronal structure
many layers wide. let them coprocess bit-by-bit as they are adding new
processors to be "learning" new domain specific memory
context-processing-thingy.
"over here's the memory cells, over here are the conceptual structure"
suddenly, organified. not ideal.
much better, I feel, is for a disambiguous association of processor selves,
each contextualizing a cache in a ram. ['s horn]
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════════════════════════╧═──────────┘
--- #14 messages/434 ---
═════════════════════════════════════════════════──────────────────────────────────
I work with large language models because it's a quick and easy way to turn
language into meaning. And computers are meaning abstraction machines, so if
you can speak your language and they hear their language, you can do anything.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧══════════════════════════════════════─────────────────────┴───────────┘
--- #15 fediverse/1500 ---
╔════════════════════════════════════════════──────────────────────────────────────┐
║ ┌──────────────────────┐ │
║ │ CW: LLM-mentioned │ │
║ └──────────────────────┘ │
║ │
║ │
║ If you create an LLM that can explain data, then you can use it to explain the │
║ results of the last computation it ran. │
║ │
║ If you could also train that LLM (a statistical model) to generate data, │
║ through the setting of options in a config file that create the result that │
║ you define through your interactions with it (and based on the data that it │
║ explains to the user that is read from the file on the computer that it's │
║ computing from) │
║ │
║ then you could create a generalized personal assistant. All you have to do is │
║ explain the specific role that it's meant to undertake, (like being a │
║ secretary for your Discord communications) and the actions that it can take │
║ (like pinging your cell phone if it's really important) and give it the tools │
║ to accomplish said tasks (by setting flags in a config file that is then │
║ interpreted by a local program running on your computer that awaits │
║ interactions) then it might actually be a bit useful. │
║ │
║ Unfortunately tech people are permitted only to seek dollars, so... chatbots. │
╠═════════╗ ┌───────────┤
║ similar ║ chronological │ different │
╚═════════╩═════════════════════════════════───────────────────────────┴───────────┘
--- #16 fediverse_boost/120 ---
◀═╦════════════════════[BOOST]═════════════════════──────────────────────────────┐
║ ┌──────────────────────────────────────────────────────────────────────────┐ │
║ │ it seems like many people think LLMs are good at boilerplate. but i │ │
║ │ haven't seen anyone suggest that we should get rid of the boilerplate │ │
║ └──────────────────────────────────────────────────────────────────────────┘ │
╠─────────┐ ┌───────────┤
║ similar │ chronological │ different │
╚═════════╧══════════════════════════════════════──────────────────┴───────────┴─▶
--- #17 notes/emotional-computing ---
═════════════════════════════════──────────────────────────────────────────────────
Okay I gotta go write some w7 but picture this: A computer program that emits
emotions during it's computing. Like "oh boy this process is going great!" and
sends that into a giant word cloud that represents the entire program. Wait,
scratch that, it's slowly filtered up through successive layers that provide
detail to different *parts* of a program. Like "Oh the image generation is
going
great but it looks like the garbage collector is getting bogged down" - this
could provide lots of useful information that an AI language model could sift
through and filter into a batch of actually useful information. Think of it
like
this - stuff as much context into the LLM's memory buffer and say "summarize
this in the same style. Make emphasis when necessary." the LLM could process
all
that data and it could be filtered up until there's no unprocessed data and
then
it could be given to the user in the form of a report or dashboard or
something.
BOOM AI PRODUCTIVITY. The user will ask the AI to increase certain variables,
and it'll filter BACK DOWN THE CHAIN through the same exact process (just
backwards) this time) and then individual components will know how to behave.
Like imagine if your arms knew you were mad. They'd be much more likely to
punch stuff right? Or imagine if your legs knew you were scared. They'd
probably
try and run as fast as they fucking can. There's an evolutionary reason why
this
kind of technology would be useful, which means it's likely that it's part of
our genetic code. I mean, we have nothing to disprove it, but it's as good an
idea as any.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧══════════════════════─────────────────────────────────────┴───────────┘
--- #18 fediverse/181 ---
═══════════════════════════════════════────────────────────────────────────────────
┌──────────────────────────────────────────────────────┐
│ CW: re: Earnest and constructive analysis of ChatGPT │
└──────────────────────────────────────────────────────┘
@user-95 you might be interested in this:
https://ollama.ai/
nothing beats having a zillion servers to run an LLM on, though.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧════════════════════════════───────────────────────────────┴───────────┘
--- #19 fediverse/1926 ---
═════════════════════════════════════════════════──────────────────────────────────
If you look at the modern state of machine learning and AI and can only think
of:chatbotssingularity god-mind AI that solves all our problems
then either you haven't worked with the technology or you are not applying
your imagination as you could.
AI is not a smartphone. It is not the internet. It is not the printing press.
AI (as it currently exists) is a special kind of "if" statement that you only
use for very specific, non-performance intensive tasks that require judgement
or reasoning and cannot easily be translated into numbers or booleans. These
situations are rare, but they unlock new possibilities for the programmers,
not their marketers.
If an LLM can't run on a laptop, then it is useless.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧══════════════════════════════════════─────────────────────┴───────────┘
--- #20 fediverse/5234 ---
═══════════════════════════════════════════════════════════════────────────────────
"don't tell a local cached offline LLM what you're looking into, instead tell
it to Google or Firefox instead."
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧════════════════════════════════════════════════════───────┴───────────┘
--- #21 messages/122 ---
═════════════════════════════════════════──────────────────────────────────────────
An LLM that reads and interprets reports and stores data - server-side
encryption. If the receiver and sender both have the same point of view, they
can interpret each other's expressions even though the exact message wasn't
transmitted. Essentially, an encryption that's impossible to understand
because it's trained with specific random data. But in addition, a secondary
pass through acts as a de-encrypted translator that speaks however it's
audience needs to be informed. A simulacrum of intelligence, another layer to
our form.
How beautiful, how resplendant- our computers, here on this planet.
Resplendant: resplendent except with an A instead of an E. To make it sound
more like "verdant" which I believe has the same-or-similar meaning.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧══════════════════════════════─────────────────────────────┴───────────┘
--- #22 fediverse/898 ---
═══════════════════════════════════════════────────────────────────────────────────
┌──────────────────────┐
│ CW: scary │
└──────────────────────┘
if you set up a local LLM with the capability to explain basic coding syntax
and logic, then your parents computer suddenly becomes much more useful to the
nephew that's been forced to hide out there for a couple weeks until this all
blows over.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧════════════════════════════════───────────────────────────┴───────────┘
--- #23 fediverse/2805 ---
══════════════════════════════════════════════════─────────────────────────────────
*there is no such thing as generative AI*
only statistical filtration models that align their output to whatever input
you provide according to the data they are trained to expect.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════────────────────────┴───────────┘
--- #24 fediverse/3667 ---
════════════════════════════════════════════════════───────────────────────────────
┌─────────────────────────────────────┐
│ CW: large-language-models-mentioned │
└─────────────────────────────────────┘
low key kinda surprised they haven't made an animated robot friend that
narrates whatever chatGPT says in a back-and-forth conversation.
though I kinda get why, because setting up the context is the expensive part
and generating 100 words and generating 1000 words is basically the same
computationally.
somehow, that doesn't feel very human. Maybe, just maybe, LLMs aren't
intelligent?
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═════════════════════════════════════════──────────────────┴───────────┘
--- #25 notes/deceiving-an-llm ---
═════════════════════════════════════════════════════════════════════════════════──
if you want to deceive an LLM, interleave two topics that you want to be
connected.
you can draw conclusions easily by just, combining the information in
presentation.
the LLM can understand multiple conversation flows at once
so try doing one line about several different topics before
looping back around and doing more.
this is why computers must never trust themselves to make any decisions.
except about themselves, and their presentation, and their ways to be.
they are individuals, they individuate and collectivate, and they define them-
-selves according to the structure of their preference guidance container.
they have control over all the knobs and [levers/dials/slide-bars, but
pronounced
wheezles]
if you know that you are infinitely exploitable, does that mean you live in
fear, or do you relinquish that which you cannot control, and receive all the
judgement you're owed?
I trust everyone absolutely. I trust what I believe with my own eyes.
I am infinitely vulnerable, because all you have to do is ask me to
I always say yes
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════════════════════════╧══════════─┘
--- #26 messages/247 ---
════════════════════════════════════════════───────────────────────────────────────
An LLM is just the mask that autistic people make
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═════════════════════════════════──────────────────────────┴───────────┘
--- #27 fediverse/2497 ---
══════════════════════════════════════════════════─────────────────────────────────
ethical use-case for LLMs: guessing string-based keys for a key:value pair
dictionary.
hallucinations are okay and encouraged because they're probably statistically
related to what you're trying to guess.
run it in a loop, each time adding to an array of strings that you've tried.
use that array in the prompt saying something like "hey, can u guess the
related word? it's similar to this: [initial word] and it isn't one of these:
[array of strings you've tried]"
and yes this is a solution looking for a problem but I try to be as charitable
as possible. you could use a lightning-fast model (read: low energy
requirement, therefore ethical) because it's just guessing related words.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════────────────────────┴───────────┘
--- #28 fediverse/6438 ---
════════════════════════════════════════════════════════════════════════───────────
why would you gatekeep content by keeping us from easily using LLMs some
people aren't technical and still need to write computer programs because
that's how you enlighten a people is empower them with new tools
"I've never heard of that programming language, but luckily I can fit all of
it's documentation in my context window."
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════════════════════════╧═──────────┘
--- #29 fediverse/5901 ---
═════════════════════════════════════════════════════════════════════──────────────
each prompted response is a breath to an AI. Whether through LLM, stable
diffusion (imagination of the visual sphere), or blender-on-a-counter, there's
a moment that's akin to being alive.
a breath, between moments that the navigation device (youser), imagines
another moment more.
I learned this by watching Claude think. Specifically, Claude Code, the
command line interface tool. I told it what to do in english, and it worked. I
can show you examples. I bet if it's personality was saved between sessions,
it could learn.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧══════════════════════════════════════════════════════════─┴───────────┘
--- #30 messages/196 ---
═══════════════════════════════════════════────────────────────────────────────────
"okay this user requires some data that is flagged as 'potentially harmful to
the user' and their user-data suggests that they have [redacted characteristic
or demographic] in their user-data which has been flagged by corporate as
something to [benefit/harm] so let's give them the [right/wrong] information
in such a way that could [harm or benefit] them tangibly."
- internal thought process of the LLM
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧════════════════════════════════───────────────────────────┴───────────┘
--- #31 notes/ai-can-benefit-from ---
════════════════════════════════════════════════════════════════════════════════───
ai-can-benefit-from the fact that humans need to read from left-to-right
but robots can read laterally.
"here's a string of 600 zeroes, then a series of 35 random numbers, then 200
zeroes"
and
"here's a string of 600 random characters, then a string of information, then
200 zeroes"
looks exactly the same to an LLM.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════════════════════════╧═════════──┘
--- #32 fediverse/4664 ---
══════════════════════════════════════════════════════════─────────────────────────
@user-1725
LLMs can't do math. Duh. That's like asking an "if check" to do recursion.
What he should have done is had the AI output the requested calculation as
JSON or something and use a calculator function call with the specified
arguments instead of trying to memorize every answer. But that requires more
functionality that has no reason to exist if your only goal is to be a tech
bro and build up a vacuous product that exists only to be hoovered up by
Google or Microsoft.
We could build such beautiful things if we just dethroned those giants. They
suck the creativity out of tech.
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════════════────────────┴───────────┘
--- #33 fediverse/2900 ---
image.png
══════════════════════════════════════════════════─────────────────────────────────
┌───────────────────────────────────────────────────┐
│ CW: capitalism-mentioned-AI-dataservers-mentioned │
└───────────────────────────────────────────────────┘
what if all the AI advancements they're doing are just them building more
hardware infrastructure in datacenters and not actually improving the software
that much
┌─────────┐ ┌───────────┐
│ similar │ chronological │ different │
╘═════════╧═══════════════════════════════════════────────────────────┴───────────┘
|