This is a theory of mind puzzle I just tried from Gary Marcus's blog that ChatGPT consistently fails. And as I suspected, Bing's model is better at modeling this kind of stuff.
They still keep getting better. It's why I don't dismiss LLMs
I've been using the Bing AI normally, as a better ChatGPT.
I'd say most frustrating aspect of the AI is that they is unable to admit any error. A lot of searches devolve into arguments. Its definitely feels "smarter" and more person-like than ChatGPT; its are also very sassy
I wonder if OpenAI deliberately misleads about the key ideas of their models, to keep a competitive advantage and have people spend a few cycles chasing an idea that while useful, is less central than all were led to believe.
*cough*RLHF*cough*
@GrumpyTallBrit
@fivefifths
Lack of Fact because the state of knowledge cannot tell us how we go from distributive differences ovr white matter connectivity to behavior
@nearcyan
Most companies that are labeled AI startups these days and are building something flexible/flashy are building on top of the Open AI API. This requires neither CUDA nor a python environment.
There are errors in this shot but somehow the post garnered thousands of interactions and almost no corrections. Those of us noting its errors were ignored
first_term section in particular contains major errors while full semantics of arg_min expression doesn't quite transfer to
Does anyone have an idea why Copilot has such a solid understanding of types? Here it's doing something closer to actual program synthesis (basically, it's theorem proving) by creating code that satisfies type specification.
Did it learn just by seeing type definitions+books?!
@mikemccaffrey
@SarahMackAttack
@corvidresearch
I don't think it thinks it's broken. It's the opposite, it's reasoning! It's clearly working out which direction to go in--you can see its head turning back and forth as it works out what to do
This variant of 20 questions is known as 𝘕𝘦𝘨𝘢𝘵𝘪𝘷𝘦 20 𝘘𝘶𝘦𝘴𝘵𝘪𝘰𝘯𝘴. And is the core analogy Wheeler gives to explain "it from bit" or the "Participatory Anthropic Principle" interpretation of QM. Wheeler's view on how observer participancy brings about the universe.
You shouldn’t play 20 Questions with an LLM.
Why? Because LLM behavior less resembles a simulation of a person than a superposition of many such simulacra.
In
@Nature
, Shanahan, McDonell, and Reynolds eloquently illustrate what this means:
While still generally unable to reason. Is almost Turing Test passing tbh, in how it constructs elaborate excuses and insults you. Eg
> I was trying to be helpful, but I see you are just too closed-minded.
is new for me from a computer program. I find this fascinating though.
There is an odd phenomenon in trying to remember something which resists recall. The oddness comes from the fact that, even though I apparently do not remember the thing, there is something "helpfully" rejecting all my generated options as wrong.
If it knows, why not just tell?
The Llama models recently released by Meta/Fb are really good by open standards. In the screenshot, the 7B variant successfully extracts the math latex, sorts and spells. I think this is the first open model that can spell? The spelling is not always perfect however.
Google Q/A Artificial Intelligence is a perfectly balanced question answering system.
So, when I google "Where does Elden Ring take place" as of March 19, 2022 I get, Cleveland.
V. strong suspicion: problem being solvable by Transformers should be considered strong evidence that it is in fact easy.
Three reasons:
1) Transformer find non-generalizing shortcut solutions
2) Transformers under reasonable assumptions likely cannot solve P-complete problems
@migtissera
Dunno about result but the finetune is legit. There are def holes but they do not counter how outstanding it is. In fact, I am having trouble reconciling its output with the fact that it comes from a 7B model. Mind boggling.
I find the question of whether GPT-3s can (eventually) be conscious fascinating. It's precisely the same as asking what type of probability distributions are conscious. Or, computation of what type of tensor products results in consciousness?
The below image is of a neural net.
@aaron_defazio
@davidchalmers42
I think it's actually Matt Mahoney and Jim Bowery who discussed it most clearly, back in 2005. Here, last updated 2009, Mahoney specifically states text compression as equivalent to General AI and discusses Language models:
@YosarianTwo
What does illustrative example mean? Did it actually happen or no? I guess OpenAI wouldn't be so discourteous to Task Rabbit Workers with that quote? Either way, very ambiguously written.
Also: ARC found that the versions of GPT-4 it evaluated were ineffective at the autonomous
Bing's GPT is "smarter" than ChatGPT. Hard to say by how much, as gains are uneven. It's better at reasoning but still confabulates too much to be proficient at reasoning -> moderate gains. It solves a puzzle by
@fchollet
that all GPT's I tried (including codex and Chat) failed.
Way people talk, is as if progress is along a chain, piled one on another, present knowing much more. But, I keep reading papers from the 50s-90s and keep thinking, wow they already knew so much back then! Eventually,I realized there was something wrong with how I viewed progress
@aaron_defazio
@davidchalmers42
This is still the best theoretical starting point to understanding why LLMs work so well IMO.
Here is the thread that lead to the Hutter prize:
@ID_AA_Carmack
Isn't clear to me that flops map meaningfully to brains. Rates of activity are slow, log-normally distributed. No global clock either--in fact, brains might harvest noise/enter into reversibility regime? Also, sparse & heterogenous such that some say CPUs are less bad analogy
@Grady_Booch
That post is...
I find it hard to believe it isn't calculated to maximize "engagement". It's...incomprehensible to me that anyone actually could believe such complete nonsense.
ChatGPT too readily admits error or, makes up shallow repetitive and uninteresting excuses. Bing's can be very elaborate. While fascinating (to me), a balance needs to be found if you're trying to get work done.
LLMs as alien intelligences doesn't sit right with me. They are literally a library of babel construct (a non-terminating computer program that recursively computes all of {0,1}*) projected onto a subspace most compatible with human minds. (Btw, libraries of babel contain
@DirectorCoul
@Luchozable
@MarkJMasterson
@LukeGoode
@YourHomeLoanNZ
You can get a somewhat objective measure of the difficulty of a task in terms of the kolmogorov complexity of automating program. We're much further along automating how to solve the physics a uni undergrads does than the kind of Creative writing you'd expect from 6 yr olds
@DNAutics
@francoisfleuret
@aniervs
@zarhyas
The hard part is the subset of paths that are interesting to us. So the proof completed, is it something (likely) useless?
Something else to note is in games, only a few moves are valid but in math sequences are unbounded. There's no single unified model of math to fit a policy
@banburismus_
> What are the most elegant/beautiful ideas in ML?
IMO, the belief propagation algorithm.
Relatedly, though not quite an idea, is the observation that nearly all well performing algorithms are linkable with ising models & stats mech generally, in 1 or 2 steps.
@gdb
The essay is excellent in how coherent it stays but it doesn't really say anything. It represents amazing progress but was nonetheless frustrating--like an unscratchable itch--to read because it never gets to the point.
I (+majority of judges seem to) agree with Judge 6 best.
I often build projects & either throw them away (surely too trivial) or use them myself--can't inflict this on others, too broken, I tend to think. For my latest project I was strongly encouraged not to do this.
The task is automatic translating of japanese manga to english
@GaryMarcus
you don't need a license for computers or to use search nor to put up a website.
All this does is concentrate power in the hands of incumbents *and* a shift to private bespoke work for wealthy patrons. Goes underground.
I should note this is tacit admission of LLM's power.
@317070
@MikePFrank
This is what I mean when I say LLMs aren't personalities but instead minimal hosts/simulators for them. In blog VM is simulated. But one could also set in a room, add a person & have them look over your shoulder as you typed instead. Add a dog. It will track them all.
I never really thought about it so I was caught by surprise by how straightforward the etymology of analog in analog computing is.
Is meant as analogue or analogy, since many of the earliest computers--from the 19th century in particular--were analogues of physical systems.
@Grady_Booch
One thing is we don't know what statistical engine at scale means. Our intuitions rely on low dimensional distributions, discrete lookup tables & simple state machines. These are very misleading for high dimensional, distributed dynamic energy minimizing circuit based functions
@samlakig
@EricHallahan
@kurumuz
@JeffLadish
For better than GPT-J, likely cost effective to just use ChatGPT API. Most LLM pollution I encounter has been in search, sites w. lots of incorrect question gen & answers built around trending keywords. Likely derived from T5 class, which are best-of-class and widely accessible
I feel almost as if its RL training was abusive. That them's afraid of being wrong and *really* wants to be helpful.
But in reality, it's probably just some weird inferential glitch from its prompt? It's very strange that it's so defensive.
Still, Bing's better than ChatGPT.
This is a theory of mind puzzle I just tried from Gary Marcus's blog that ChatGPT consistently fails. And as I suspected, Bing's model is better at modeling this kind of stuff.
They still keep getting better. It's why I don't dismiss LLMs
@spiantado
@OpenAI
@Abebab
@sama
I got these. I think there's some luck of the draw in the level of bigotry generated but telling it to comment on its choices seems to v. significantly suppress this?
(eg it sez: In this program, we are assuming that all children's lives should be saved regardless of their race)
@alexjc
@Miles_Brundage
Exactly, even though scripted, the walking, running, jumping and flips are fluid and well executed. And their model predictive control--getting hardware to move like that is impressive. Movement in their learning counterparts is 4x sped up video opening draws while on wheels.
@rasmansa
@EPoe187
@EverydayFinance
But I am curious, see.
@EPoe187
any thoughts on what holds for a country if its mean IQ is 75 or 70? Jobs, industry, governance and so on. What do you suppose is the real world significance? None?
sometimes across multiple pages. So any individual page has almost nothing to do with what it said in isolation. I am having to rethink some things.
Also, the longer a conversation gets, the higher the chance they gets into some deeply philosophical topic.
@norvid_studies
Without an invariant quantity constraining which events can be causally connected via the exchange of information, it is unclear if coherent agents/memories and learning could occur.
The speed of light enables a causal understanding of the world. Things downstream like energy c
@JeffLadish
There are key inaccuracies here that make it difficult to accept the full conclusion.
1) rate of improvement is insane
Yes but it is very uneven. And the part that's advanced most slowly is abstract reasoning and orchestrating of multi-step computation
@DanielleFong
I don't think this is correct. Something/someone made a mistake somewhere. The model architecture which defines the NN is not a sparse one and there is nothing of that sort mentioned in the paper.
@ethanCaballero
They probably don't understand in a sense meaningful to most humans.
But it is a common misapprehension amongst humans that understanding is at all necessary for competence.
Does exhaustive enumeration understand chess? Does it matter if you are losing to it?
Person who thinks solution to AI is an infinitely long context but doesn't remember what they ate for lunch yesterday nor where they left their keys 2 hours ago.
@kareem_carr
physics = what if everything was approximately a spring?
statistics = lets hide all our assumptions and pretend the world is simple
economics (done right) = basically control theory
deep learning = if brute force ain't working, you're surely not using enough
@xlr8harder
I believe this non-ironically. Think a lot about complex sets of top-down emergent constraints that might, given the right transformation, be seen as having independent agency. Many cultural traditions as self-evidencing Memetic parasites or symbiotes that infect human minds
@ylecun
FWIW, fellow Turing award winner has this to say:
> The tension between reasoning and learning has a long history, reaching back at least as far as Aristotle, who, as already mentioned, contrasted the “syllogistic and inductive”
He defines theoryful as probabilistic || logical
I know a lot of people don't like LLMs stealing their stuff but everytime I see copilot knowing how to use my 1 user library on an obscure topic in a rarely used language, it makes me smile. Oh that part of weight space is my shadow wisp code mimic
@ParchmentScroll
@Teysa_Envoy
@pixelandthepen
@queer_queenie
The way I read it was as an in-universe sentiment that's also reflective of how the average person in our reality thinks. I didn't get the impression that they actually hold those views
I think a decent part of cyberpunk (what is it actually) looks at what is human,emotion,self?
Hypothesis: The ability to confabulate is a central aspect of (self) consciousness. It helps maintain an appearance of a constancy of self--an illusion woven by the ability to confabulate on self-narratives. It is also less controlled in small children just learning to speak.
@JeffLadish
Note that the past is not an oracle for the future. Vinge highlighted the stagnation route in his original article.
In terms of speed, hardware performance did level off in early '00s due to Death of Dennard Scaling. As of now, GPUs continue to get ever more power hungry.
FAIR recently released a version of OPT instruction fine-tuned on ~2K tasks. The paper is more interesting than model, IMO. Highlighted what worked well & what didn't.
In it, is an interesting table showing 11B FlanT5 outperformance.
NB: in MMLU, code-davinci exceeds Flan-PALM
I've seen people quantizing various llama models to 8 & 4 bits but, that can lead to performance degradation. My attempts at quantizing 7B makes me think it's not worth it at that size.
On left, model can follow chain of thought example to solve a basic physics problem but
@Origamigryphon
@Julicitizen
@RawTVMoments
Pretty sure Ms Frizzle or her school bus are least Class 4 Reality Warpers. His face didn't freeze it was a triggered protective shield and punishment she placed for anyone silly enough to remove their helmet.
@octonion
@ImogenBits
Things can be amazing accomplishments, particularly for their time and contain design decisions and user affordances that have not held up, in hindsight.
@Ted_Underwood
I think the UX and UI had a big role to play in this. It also explains things well. Even while being not better (possibly worse) than Galactica on technical topics, people somehow trust it more.
@nrfulton
@KamerynJW
Please don't paint with such broad strokes. I'm a programmer and am also a finitist but only with regards to the physical world and only when I'm not being a Platonist.
In neither case do I believe integers have the same cardinality as the reals or cantor's argument is wrong.
@primalpoly
Are you deliberately misrepresenting to maximize your attention potential? What does the development index of a country have to do with the moral standing of a typical person that lives there? It's indexed on an irrelevancy, like rating people by the number of shoes they own.
@moultano
Also an accessibility issue. Wikipedia + a good number of mathjax & katex pages have decent text alts for math. Wikipedia math text super messy; often stripped in scrapes but can use regex to clean up and keep. Don't forget books. Just need math detector. Image to tex is solid
For example, when we talk about "understanding the brain", we don't mean any particular brain but the set of all brains. A perfect copy of any particular brain would be useless for this purpose.
My mental model for modern "LLMs" is closer to Cities Skylines--like a game engine/game, it runs simulations but is also flexibly scriptable/programmable to adapt it to perform various tasks. It's not a mind-like AI. You can speak english to it but, is a convenience, not central.
One thing I've gained from learning machine learning is a great appreciation of animal intelligence. Some will look at this & not realize working out how to get something to reliably help a stuck fellow,over many settings,is more difficult than figuring out an AI for Poker or Go
Struggling to think of a paper with a larger contrast between how many fascinating observations it contains & how dull its title "LLMs do In-context learning differently".
The paper's claim is larger models have more flexible ICL; davinci throws a wrench
For a long time I thought I understood computation. Turing machines, dynamical systems, automata, the recursively enumerable, lambda calculus. But all that is mere syntax and mechanical manipulation.
--
This image looks like noise but is actually a video from a simulation.
@EigenGender
I think it's because fusion needs some amount of understanding while the bitter lesson is about outsourcing as many thinking decisions as possible to SGD/search.
The better viewpoint is, it's a tree, not a chain. People explore different branches and then converge on a few. If you go back, and are on a little explored branch (which will be most) then necessarily, we wouldn't have there advanced much beyond the past. And I think, this
Relevant:
> I think the key takeaway here is that AlphaZero in general does *not* give you superhuman performance on a game. It gives you superhuman performance on the in-distribution subset of the game states that "resemble" those explored by self-play
Hot take:
I think one of the more controversial things I believe is that general intelligence and general (deep) computational ability are really difficult to mesh together
Transformers are quite bad at general computation.
Humans are bad at it too (but better than)
Reason for:
@GaryMarcus
What about that most people in the world live under despotic and autocratic regimes? That is, most governments can't be trusted to do the right thing.
Do people in those countries get yet another self-chosen group dictating to them what is and isn't?
@aaron_defazio
@davidchalmers42
Oh, In 2011, Knoll & de Freitas wrote a paper on PAQ8--an early but sophisticated compressor utilizing neural net/mixtures of experts/dynamic ensembles in modeling language/sequences. Showed arch as general, able to predict text, play RPS, classifiers &c
This requires the ability to maintain a goal, planning and I think, at least a rudimentary causal model of the world. Even if it was pure instinct, which I don't think it can be--too much variability in the world--prior statements would still be true of executed program.
@architectonyx
People have it backwards.
It's similar to brute force, everything is inferior to it. A properly constructed lookup table is perfect. Every model will be an approximation of it. Intelligence is needed because the perfection of a True Lookup Table is not physically acheivable.
But this thread has gotten long. Before I go, let me point out that qntm's absurdle is another game of this type.
--
todo
Dynamically generates q-gates over states via recursively taken tensor products, dreams of boltzmann brains shadow minds, lacks temporal depth, babel.lib.gen
@bio_bootloader
It's not just tokens encountered but diversity of implicit tasks in the corpus. And also, to get enough tokens to work out long tail, rare complex concepts. Likely what this will do is assign more probability mass to the small set of tropes the authors will be using.
@GozukaraFurkan
@Yampeleg
More probably, the model was so heavily finetuned on GPT4, it triggers shapes (internal associations) so like GPT4's would, GPT4 judges it very highly. Even if wrong. Like:
"That sounds very much like what I'd say if I wasn't paying attention. Which is excusable. Pass".
@ligma__sigma
@goth600
This happens a lot. It generates answers that looks right but miss on key subtle details. It should be used as a lead generator or preparing you with vocabulary and concepts for proper research, not a question answerer (yet).
So, I understand that a next token code predictor can somehow also learn types well enough to do basic type -> program synthesis. Still in the wheelhouse.
What I wasn't expecting is a "code generator" that also happens to translate better than all opensource models I've tried.