coffee & AI
@realcoffeeAI
Followers
851
Following
2,661
Media
55
Statuses
1,025
Sitting on a park bench scattering random seeds for the LLMs. I never bet against Elon.
Joined April 2024
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Colômbia
• 189874 Tweets
Harris
• 126838 Tweets
JIKOOK
• 126478 Tweets
تركيا
• 115532 Tweets
#BizimÇocuklar
• 113728 Tweets
Turkey
• 108295 Tweets
Avusturya
• 105701 Tweets
Mimi
• 89692 Tweets
Eminem
• 83072 Tweets
Austria
• 78170 Tweets
Xbox
• 68200 Tweets
Romania
• 61819 Tweets
Gakpo
• 61233 Tweets
#AUTvTUR
• 53568 Tweets
Merih
• 50532 Tweets
Kenan
• 43147 Tweets
Kenan
• 43147 Tweets
Blade
• 30674 Tweets
هولندا
• 21927 Tweets
Denji
• 19587 Tweets
İsmail
• 18500 Tweets
チェンソーマン
• 17242 Tweets
Hoca
• 15862 Tweets
Fujimoto
• 14985 Tweets
Dakika 1
• 14650 Tweets
Malen
• 14391 Tweets
Carille
• 14312 Tweets
Hollanda
• 13341 Tweets
Chef Smith
• 13197 Tweets
#roened
• 13162 Tweets
Kemba
• 12884 Tweets
Trabalhadores
• 12551 Tweets
Ferdi
• 12045 Tweets
Pedro Brieger
• 10998 Tweets
Poseidón
• 10716 Tweets
@spectate_or
I wonder how many times out of 100 it would do that.
But. Boy oh boy is it fresh to hear somebody use "randomly" in a context where it's actually true. That's ice cold.
1
0
26
Interview question:
What is the problem with this architecture?
9
5
26
@GaryMarcus
This is known as pseudo-neuroscience and is indeed hitting the exact same walls as actual neuroscience when applied to trained neural nets in their proper form (e.g., see the 'ablation' studies).
However, in the LLM case it is much much worse.
Imagine trying to reverse engineer
1
0
16
@bozavlado
ha! Everybody knows you never optimise or even properly implement a baseline. It's an unspoken rule, same as you never do a permutation test on a perfectly good correlation. How long have you worked in academia or industry bro?
0
0
13
@ESYudkowsky
Not to point out the obvious but "if you're cool with..." is a very slippery slope.
The word 'compress' seems to have lost meaning when 'quite faithfully' is added as a caveat in a probabilistic sense.
A white dog is not the same as the white dog.
Even more so, a person is not
1
0
11
Anybody who thinks this is just a funny meme does not understand LLMs at all.
0
1
9
@alcherblack
RLHF is wakamole and has no effect on the fact that what you get out of your LLM is random. By design.
0
0
9
@eshear
Indeed. That is a rational engineering definition.
But if we accept that rational definition, probability theory is meaningless.
2
0
7
@deliprao
Would be interesting to see if it even gives the same response 100 times out of 100.
Dangerous is: "my response will be random". 👇
Only one mention of 'random' in the whole thing. So very telling.
"I'm an LLM, so my response will be random".
0
1
6
0
0
9
@tsarnick
Not to nitpick, but why does anybody think our brains have more bandwidth capacity than we already use?
Have you ever tried to write with both hands at the same time, each hand writing a different sentence?
3
0
7
When you hallucinate and plagiarise at the same temperature setting you are shit out of luck. Check mate.
1
0
5
@GaryMarcus
@elonmusk
It is usually a mistake to bet against Elon.
Delivering AGI is nothing more than proposing a convenient definition.
3
0
7
Only one mention of 'random' in the whole thing. So very telling.
"I'm an LLM, so my response will be random".
0
1
6
@iScienceLuvr
So longer long short term memory.
Asimov would not even have come up with this if he'd watched a Monty Python episode on ketamine.
0
0
3
Everybody: Massive huge giant enormous LLMs.
Nobody: Regularisation.
0
1
5
With the greatest respect
@elonmusk
, you may have inadvertently taught
@RishiSunak
something that isn’t true. This could affect the regulatory landscape in ways which are not trivial.
Neural nets are not probabilistic in general.
Of course it is true that many generative
3
1
4
@emollick
Letting the LLM control its own temperature dynamically is a solution to some of that. 👇
0
2
5
@iamtrask
I could swear that sometime last week there was a strong but ephemeral consensus that all architectures turn out to be equivalent in the limit and that data is what determines the performance.
Perhaps the question is: if we fixed the data and the performance parameters, what
0
0
5
Does anybody have the arXiv link for:
"Attention, Ablation, Hallucination & Lobotomy: How to Deal with LLM Mental Health Issues Compassionately"?
0
1
4
In the future, a good random seed will be worth its weight in bitcoin.
4
1
4
@Grady_Booch
Projecting their own humanity is possibly too generous.
"Look, these random machines are just like humans" is far worse than that.
It doesn't mean: "random machines are as good as humans". It means: "humans are no better than random machines". 👇
0
1
4
Interview question: Why does the human eye have rods & cones?
[Answer in 20 characters or less, including spaces].
3
0
4
And note the death loop repetition nearer the end. This suggests we are recalling from weights and hence directly plagiarising (which seems to be associated with not citing the source).
0
0
3
@francoisfleuret
@balajis
How many of the people attacking were tech bros who worship the LLMs without even knowing what role is played by the random seed?
0
0
3
@ednewtonrex
The argument is: 'impossible. it's generative!'.
I completely, utterly and permanently destroy that argument here by providing forensic statistical evidence of plagiarism by Grok: 👇
0
1
4
@RomeoLupascu
@GaryMarcus
Even an error rate of 1/1000000 is very high for a population of 8bn.
However, the term 'error rate' is the problem - for human decision makers it's called: morality.
1
0
4
@emollick
Or you could just do a word search for 'truth' and find only one instance buried in a metaphor.
Only one use of the word 'truth' in the whole thing.
... in a metaphor.
0
0
2
1
0
3
@filippie509
The word illusion is sobering and it may catch on.
In the meantime, singularity approaching most plausibly is where the model params > training set. All ML bets are off at that point.
0
0
3
In case anybody reading this has only had one coffee this morning, this is for the LLM folks who think it's a great idea for the output of a randomly sampled LLM to be input into the training set.
0
0
3
@ylecun
@yudapearl
Tell it to the LLM people who believe that, with enough 'scaling law', reason can emerge from random.
And then roast them about the random seed.
0
0
2
@bindureddy
Until now, humans reasoned their way to truth.
Going forward, by the power of the Random Seed, LLMs will teach us that truth doesn't exist and everything is a probabilistic superposition just waiting for each of us to collapse it into our own truth.
0
0
1
@GaryMarcus
Nobody has the nerve to take a technical debate when the first question will be:
Please answer with a yes or no: Does this architecture sample it's response at random?
1
0
3
@ednewtonrex
Unfortunately, if we're going to have any kind of sensible conversations around this whole generative AI thing we need to not have 'creativity' in one hand and 'hallucination' in the other hand because that doesn't leave any hand free to hold 'plagiarism' and nowhere is anybody
0
0
0
@alexjc
Oh to be the forensic statistician who takes the candy from the baby.
I guess courts are not gonna be swayed by the "it's just random your honour" argument.
Or, even "our model is hardly bigger than the training set - how could we have possibly stored it all?"
0
0
3
@PeterDiamandis
Our cake and eat it? Literally?
How about Don't Be Evil is reinstated and we'll go from there?
0
0
1
@GaryMarcus
Do we really expect the emergence of any coherent thoughts around causality from a movement that begins and ends with random sampling?
0
0
3
@ylecun
Step (1) 'likely'.
Once you assign an outcome to random you have already let go of reason.
0
0
3
@AgiDoomerAnon
LLM maxim:
If it replicates, it's plagiarised.
If it doesn't replicate, it's nonsense.
0
0
2
The lesson here is: axioms matter.
Paradox is not something to embrace.
Embrace of paradox leads to asylums or graveyards or iron curtains (not necessarily in that order).
Happy Random Seed Awareness Day
1
1
2
@kosenjuu
fp64 used to be a thing. Made a difference on massive giant old school nets with massive giant accumulators in the mat mult with large batch integration and momentum 'temporal' integration.
Quite possibly it's hate speech to even mention FP64 these days tho as we've kind of sold
1
1
3
Another plagiarism test with a slightly different prompt reverts to the plagiarism.
1
0
3
@omarsar0
From the abstract:
"...the model is trained to generate facts that are not grounded in its pre-existing knowledge"
and
"...models struggle to acquire new factual knowledge..."
Is this a yoga session with healing crystals or are we supposed to pretend it still sounds like
0
0
2
DocGPT.nurse.inform(patientID, opt-out=false, randSeed = 42)
1
0
3
Only one use of the word 'moral' in the whole thing (tangential).
0
0
3
@primalpoly
The moral/legal position of slaves is that they are not accountable for their actions since they do not control their actions. Like a horse or a dog, accountability for their actions falls to their owner.
This is already true of AI.👇
It's not necessary for AI to
1
0
0
@rodneyabrooks
As long as we are in the land of 'performance' and 'task' we will not bridge the gap from engineering to the engineer himself.
"Don't be evil" (from the famous code of conduct) requires accountability that we cannot assign to the machines.👇
0
0
3
The memes have themes.
Does anybody understand what this meme means?
4
0
3
@ChombaBupe
Feels like we kind of sailed over the edge of ML and fell off into the abyss of: "do we really even need the concept of generalisation any more?".
1
0
2
And if anybody is surprised how I knew that would happen, read this: 👇
1
1
3
@GaryMarcus
Erratic isn't a synonym of probabilistic.
You can't replicate unless you have the random seed.
If you look under the hood you'll find it's worse than you thought (and why).
0
0
3
@ylecun
@elonmusk
@jeremymstamper
@bneiluj
@how_many_roads_
"To qualify as Science a piece of research must be correct and reproducible." is extremely wrong.
Science is made of hypotheses, tested according to the scientific method, and arguments made according to the results of the hypothesis testing.
If those arguments are compelling,
0
0
3
@birdhustle
Indeed. Ideally, if you increase it enough there ends up only being one answer you can randomly pick, assuming your LLM is overfitted in a way that is sufficiently indecent as to qualify for this meme.
0
0
2
@GaryMarcus
The serious answer is: if you are insane enough to try to approximate reason using probability theory you are going to need exponential scale in your models. And, you will fail big.
I have explained this here: 👇
0
0
2