AI General Thread

[Legend]

I found a really powerful workflow with Flux.

It is incredibly good with long natural language prompts. I have chatGPT generate a 5 paragraph description of exactly what I want and Flux can turn that into an image. The result is a lot better than just telling Flux directly what I want since Flux can't think through its pictures in advance.

Can even upload pics to chatGPT so I can use it as a faster alternative than gimp (free photoshop alternative) for previewing game dev work.


Only problem is that I'm using a free website so I can only generate a few pics a day.  

You can try it out for yourself here:
FLUX.1 (Schnell) - a Hugging Face Space by black-forest-labs
FLUX.1 (dev) - a Hugging Face Space by black-forest-labs

The first link, Schnell, is very fast and you can practically generate as much of those as you want. The second link, Dev, is slow and the one I need to use to get good quality.

Grok is using an even more advanced version. I'm tempted to pay $8 for a month and try it out.

[Legend]

Text to Video has already developed
pretty good so far. However, the models are clearly not a simulation of the world. Calculating physical laws does not work, the architecture is not designed for this. So we still need a few more breakthroughs. pic.twitter.com/MrqVXcHgkZ

— Chubby?? (@kimmonismus) August 21, 2024

[Legend]

I like making custom AI architectures, if you couldn't tell from my past explorations of a cellular automata based AI.

Well my new experiment for the year has an even stranger architecture. It is essentially a bunch of mini AIs that get a single piece of the puzzle and they know absolutely nothing else. They don't even know how many other mini AIs there are.

All they can do is write on a small "white board" that's 16x8 pixels. The white boards from every AI are taken and overlaid/averaged into a new single drawing and then passed back out. Each mini AI sees this, makes some changes, and the board is averaged again. In total each AI gets to write on the board 4 times.

Then a dumb single layer AI converts the final averaged white board into a usable output.


My toy problem I'm testing it with is predicting the next letter in text. Set it up in google colab and actually got it working.

and preceded by Top the to his backing his lied the to effection, which assured eache
between New Zealand se now eye barried to they reposes of an the sacround Harding inf
"This will be a gun she his house, which I difficient that the forel to the day, the
d adequate to so a hatte still alway, that fore father are of the can my will neclent
little, notwithstantage, that he cried the tall sistered they is this after on the Pl

The green text seeds it, the white is what the AI wrote. No chatGPT lol, but I'm kinda amazed the mini AIs can work together well enough to do this. These examples are with 18 mini AIs so as it writes, it loses access to farther back letters. Performance is roughly the same once there are at least a few mini AIs.

(lower is better)
Mini Brains: 1, Clean Loss: 2.4705
Mini Brains: 2, Clean Loss: 2.0370
Mini Brains: 3, Clean Loss: 1.7492
Mini Brains: 4, Clean Loss: 1.6120
Mini Brains: 5, Clean Loss: 1.5543
Mini Brains: 6, Clean Loss: 1.5210
Mini Brains: 7, Clean Loss: 1.5175
Mini Brains: 8, Clean Loss: 1.5103
Mini Brains: 9, Clean Loss: 1.5064
Mini Brains: 10, Clean Loss: 1.5025
Mini Brains: 11, Clean Loss: 1.5025
Mini Brains: 12, Clean Loss: 1.5106
Mini Brains: 13, Clean Loss: 1.5032
Mini Brains: 14, Clean Loss: 1.5153
Mini Brains: 15, Clean Loss: 1.5148
Mini Brains: 16, Clean Loss: 1.5311
Mini Brains: 17, Clean Loss: 1.5308
Mini Brains: 18, Clean Loss: 1.5391

Most surprisingly to me is that so far at least the AI utterly fails if it has more mini brains. At 19 the loss gets worse and becomes ~2.5. Beyond that the loss stabilizes around 4.
I can fine tune the model to work with more mini brains and keep loss around ~1.5 but so far it never generalizes. Fine tune to 24 mini brains and the 25th still has a ~2.5 loss.

You'll also see that performance stabilizes around the 5th mini brain. This should just be a limitation of my training. At least in theory, performance should always increase as the AI can see more of the sentence.

I started training just on Frankenstein but I added The Mysterious island just before making this post. Thought maybe the training data was holding it back. Either way I'll leave Google Colab running in the background and report back if it does anything interesting.

[Legend]

It's really cool that this architecture can work with such a small whiteboard.

To put it in perspective, a single letter takes up ~100 spaces. The whiteboard is only 128 spaces yet the AI was fitting in at least 5 letters. 6 if you include the output letter which also needs to be stored on the whiteboard.

Not practical though. My main goal with this was to create an AI that scales to whatever challenge you give it and I think I only partially succeeded. I can't think of a way post training to increase the size of the whiteboard in a way that actually makes sense. Adding more mini brains doesn't improve the results because the whiteboard becomes saturated. The mini brains can be given more iterations to try and overcome this, but that is a bandaid that's better used for other reasons. I probably need a self attention mechanism like in all large language models. No clue how that could fit though.

[Legend]

I really don't know what I'm doing  

Got loss down to .75 or so and it was actually producing pretty awesome results but google colab crashed. I tried again and could only get to 0.88 loss. Then I tried again throwing absolutely everything at it and only got to a loss of .91 Maybe the first result was overfitting and I just didn't have a test for that at the time, there's no reason I can think of for my later bigger model with more training to do worse.

Either way, here are the results! Pretty cool that I made my own GPT 0.5 from scratch! This uses a recurrent neural network since it is so much faster once you're trying to look at 256 letters at a time.

The sharks of the trees formed a third and found in the thick
waters of the Hon. Morison Baynes, the foe, and the woman, say here,
resolved to permit him to dig her head from their eyes with her son
whether he was trying to pursue his enquiring prey. For a moment he
saw that the huge blunder was still diverted to the mines. His finger
grieves around his narrow chance; but he could see a head and she
began to whose position the moon was not so much entered in any of
business about the world—though of nothing. She bore in his stores and
only horrible to reveal these. It was the day was one of them a little
or sword. The lion should come down from the sight of the cruel body
of Mohammed Beyd's dear boy, for not that he had leaped to history
and her shoulder, breaking the five theater he crept at a few words.

He advanced to far representations, so that the storm between them had
been freed in the directors of the window. The door opened before its
little creek and arrested on the beach. The man was very hard, as


"And the time of appointment?" asked the reporter.

"Then," said Herbert, "these caravans are almost amongst the new stars.
With the aid of the state of the water, the next morning I would soon
much there was, sharp or streams where the raft was hurled. I preferred
to be found, to lie up in my arms, and I started off to give some resolution
of our work of conscience.

The traveller overpowered the museum, and the country was glittered
in the temperature of the moon, and made up the mouth of gold. Nothing
could attempt to fall on to the surface of the water.

The wool was ready in the bear, the oscent of which was sharpened, having
not to be correct. The shadow of the hut departed from the sight of
the rocks, and looked suddenly into the air above the air seventy
thousand feet. While we were greatly delighted there, possibly on the
temperate ocean.


Legard Dorf or Zoram when they chosen
to restore them. It was not in the pinnace and then it would be close
beyond the plateau of Prospect Heights. The banks of the Antarctic
Sea was now somewhat impactible, and excessively dragged into planks
or basaltic acid.

It is the scorce of the French ones from the Southern Seas, of the
Northern Hemisphere that is to get across with unfortunate circumstances,
the more beautiful that the bottle and the clearing could not bring to
any answer. For the last two or three of these two three confeders
that may evidently save him, as well as long as the great astonishment
at this moment that the thought was to be abandoned in the direction
were bantering down by the end, which form a wind performed in the
monotony of the dangers and new nor difficulties of defence and leather
than the glass, they were still looking at the prisoners who had made
a discovery of an opponent to be in sight. The game was alive to take
iron, or even on the side of the water. Through the window by th


Vizonia, Grotto Edward, Daisy
Blomsberry, and Michael, and placed on the north to two occasions.

The night passed away. All at once, it was a very angry extensive
seaweed; the feathered coasts of the Solar Admiral and Clerval, and
only a few miles from Russia. There we wish for an instant, and my
former given almost allow."

"And why?"

"Don't you hear," replied the sailor.

"Then, lost that purpose?"

"Yes, sir. It is an inconsistency of the shipwrecked vessel."

"I beg your examination of the former scene of my labour, I must
do it and have arrived. A little more in this way, that is to say
about our return. There is one of these traces of meeting a guide. I
will come as much as it is Syria, and where Ernest would be delighted. I
happy, if you have to say anything better, I'll be back to the captain
to live in the way to take it to you. I am saying on the name of Captain
Nemo, I found on your country and a band of powder which might personal
be congerated in the bottom what may be better; no, it is more tha


VizionEs and of Mont Blânc
Company Bay on Norwal in New Switzerland. Contrary to the Campinetry of
Gallia were are not in the province. Do you call your deliverance from America?"

Legend climbed up to the height of the first place, which drew about
the barrels of so far as the prospect of the colonists.

"The gentleman who have been somewhat amased," said he, "in his destiny
merely the faculty of carrying you away from Algeria, so that, in seven
years, is to say the most formidable influence of the earth has
actually so very useful to me. You cannot say you have the worst descent
under your babbles."

Cyrus Harding and his complaints were all exciting, but suddenly a sudden
sun had acted for Tarzan's desire to sell upon the ape-man to be hard
fatiguing at the tiger.

However, the tribe of number panthers were struck with it, then he
presented severe to have a sinister. Now he looked for the boy, the
present sea, except Cyrus Harding's hand, after the first thought of
his love and hope and inconvenience in his pa


"What are you saying?" I'll listen. "I can crush out the word, and I
beg your things to find you."

The poor lady and his father was ashamed and sent to the spot in
service to myself the sailors' tent. His gun in his side had been an
instant about his time for another, and at last he stopped and grasped
his tent and turned to where the lioness fixed for the waist, but his
neighbor glared like a shell-fish. And all the others wandered about,
was speaking to him in the presence of caution. They were at the Tartar
moment came and began toward the escarpenter at Ha The Chimneys. The
_Nautilus_ were so intense and dangerous; but the rocks which ran upon the
waters such as he could have indulged in death at daybreak. The stars
left the bright side, and cried out, "By making a satellite!"

"Captain," replied the captain, read and resolutely. "The sky
is fine. Others to fort, a victory lack!"

And thus they said that he was wearying us alone, but a certain mass
of the waters were cogged u


CHAPTER VI THE NEW LET IN THE CIMITAS
CHAPTER FOURTHERE TURIVE COMPARIKNS.


The next day, the 22th of July, 1864, after took our wrists to leave
the place, and with a rapid current of iron coast of Iceland.

Upon this I accomplished his argument, which was clear; the air
could be rendered by the inclines much much less successful; and
although it was here with it the shore behind the earth went to its
prey to the mouth of the ape-man. A rain was visible extended on the
field of ice.

"The DUNCAN!" he continued, "this is given me to see this matter,
and the trap with it in its passage? It must be an easy time to feel
upon the air. My uncle resolved to command the draw palm the timber on
board the ship; but I could not make up no means of getting out of her
creatures insufficient for her to drag it in any fury, should not have
been calling against the powerful man in the bowels of the iceberg,
and was in a stupor which more than the world should be passed. But this
simple means you may survive the meaning of t


Murk and Morkey, in their private
calm, was lost among them that she was still thirst. So great is was
it, there were perils of a perfect pity, and they were now then there
before they reached the coast of Alexis Paulvitch, the Russians to the
Archipelagous School of Congo Phreshowers, and the captain of the Arab
mystery and the same time as that of the next day. As he could fear,
everything was disturbed by the care that presented to him.

For a month it blew his countenance and though she sat thus about to have
felt dispatches in loading the second steam to the north. Now Lady Helena
and Mary were to feel themselves shouldered by the forder of France.

Now they were to meet with all their lords. It was evident that the
affair was a dream; but finally, they had preserved them is the possible
observation to the laws of travellers.

The planets have already discovered that in which the crew had to be
eat.

The schooner seemed delighted at the top of the coast of a journey
which were to form the reserve of those

[Legend]

AI will be a huge upsell opportunity for platforms and operating systems. Take me, for example: I’d be gladly pay hundreds of dollars just to turn this new garbage off! https://t.co/lT1nFSYnkQ

— Tim Sweeney (@TimSweeneyEpic) September 20, 2024

[kitler53]

I generally think Tim is a bag but I kind of agree with him here.   especially when I read about all the nuke plants being built or reopening just to meet the energy crisis AI is creating.    the capabilities are not worth the cost. 

[Legend]

Simply cannot resist uttering the sticky phrase in my mind for the last few days: "Skill issue" ????

Okay, jokes aside. I find this hard to believe. I kind of get why they are called The [Mis]information.

Saying that the scaling law for LLMs has slowed down is like saying that… https://t.co/nSlZkSfGAh

— Hieu Pham (@hyhieu226) November 10, 2024

xAI continues to be really bullish about their upcoming model. Really hoping it is a big leap forward, because the past year of text gen has been kinda boring.

[the-pi-guy]

Copilot is sometimes funny with how it blocks political questions.

I asked it why people are talking about a political mandate. It wouldn't answer.

I asked it why people are talking about a mandate, and it gave me the political answer about how it comes up after elections.

[Legend]

The workarounds with AI are so frustrating/silly. I've tried so hard to get chatgpt to sing for me and it never does, but you can find plenty of videos of people pulling it off.