I wrote this a few months back and I'm even more confident now that the light of LLMs will have faded substantially in 5 years and we will wonder why we were so excited.
However I regret saying that they will have bad/random opinions as they are simply not capable of having them
I think in 5 years we'll look back on LLMs as a tool, one that is less helpful than we thought it was going to be.
It's still going to be spitting out wrong answers quite often and have bad/random opinions. And we'll still be figuring out when it is useful and mitigating damage
🚨Open PhD position🚨: I am looking for a PhD candidate for next year. The topic would be related to attention and working memory, using models, behavior and EEG. Reach out here or to bwyble
@gmail
.com if you're interested.
@Jabaluck
@IrisVanRooij
Try asking a group of people who primarily study intelligence and cognition instead of machine learning. You'll get a different answer because engineers rarely understand how complex our minds are.
New paper! People often do not remember information they just searched for ().
Subjects were asked to draw the contents of their memory immediately after searching for a target in a natural scene🧵
I'm going to put together an open letter on the existential risks imposed by visual attention and working memory. We must work to mitigate these threats and I'm willing to lead the effort, especially if it is well funded.
@MetacogniShane
@StuartJRitchie
@hisotalus
I think the point Stuart is making (correct me if I'm wrong) is that you and he might be having the same internal experience, but are describing it differently.
In peer review, so much time is spent dancing around the power of the reviewer.
You often can't say what you really think and you have to go to extraordinary lengths to appease hypothetical opinions, which are often inaccurate anyway. I support peer review but not like this
When I tweeted the below in June, people got very angry at me and I received misogynist attacks (will link thread below). But if you are interested in really understanding why AGI is not coming anytime soon, please check out our new preprint:
Being an expert AI researcher (I have been studying AI since the 1990's) I guess that I don't need to explain why I don't agree that AGI is coming soon.
If you don't agree that AGI is coming soon, you need to explain why your views are more informed than expert AI researchers. The experts might be wrong -- but it's irrational for you to assert with confidence that you know better than them.
Lots of discussion about this survey lately. It's important to note that the people responding to this survey are engineers with no training in understanding human minds.
Because of this lack of training, these people have no basis for making this evaluation. Instead >
@IrisVanRooij
What do you think would be the median forecasted time to AGI among your peers? A great many people would be extremely interested in seeing the results of a well-done survey showing something very different from:
While Katie recovers from this mental blowout, I'll add some context that I think she was distressed by, which is the fact that people as supposedly educated as Nate do not realize that language is rich in implicit information.
Many people study this, it's a whole thing.
@lukesjulson
This is turning into a blockbuster plot. I see Nicolas Cage as a washed up neuroscientist flying on a helicopter to DC to tell congress that we have to fix the DMNs before it's too late.
If you replace "GPT says that" with "GPT produces the following text", it becomes a lot easier to understand what it is doing.
The word "says" is simple but it brings much baggage.
Does anyone remember the story of the neuroscientist who always found amazing theta in hippocampal slices and it was found that this might have been due to all the cigar smoke in the lab? I'm trying to track this story down
Remember when back in the 1990's we would put 3d bar graphs in papers for absolutely no reason? We were spending our time to obscure information just because it was a new, fancy feature of recently developed tools.
It's good that we don't do that kind of thing anymore!
In the wake of the IIT letter (which I co-signed) there’s a lot of discussion about whether theories need to be falsifiable and I’ve had some thoughts that may or may not be useful. 1/n
I agree with this tweet.
It could be argued that chatbot therapy is better than nothing (not wanting to defend that, just setting it aside)
But the overwhelming reality is that tech solutions to social problems typically attempt to replace real solutions with false ones.
i can't put into words how angry it makes me to see someone ask "isn't chatbot therapy better than no therapy?"
people deserve real care, and the whole "this or nothing" turd is especially threatening against a backdrop of a trillion-dollar tech company's cheap facsimile of care
I hope that pundits and journalists use yesterday's colossal infowar failure as an opportunity to reflect on how effectively we are manipulated by Hamas. We must do better at holding them accountable, just as we are with the IDF.
I gave StableDiffusion the following prompt:
"Image of an old psychology laboratory from the early 1900's with chronometers, clocks, headphones, with a single female professor off to the side talking to several male subordinates"
wompwomp.mp3
Me: I need to say no to reviews to make time for my other work right now. Time to take control of my life!
Journal X: This paper says the key finding that underlies all of your research doesn't replicate, do you want to review it?
Yes this. No part of the objective function for LLMs even attempts to discern truth. The hope is that truth is absorbed statistically given a dataset that reflects truth.
Humans have metacognition that forces us to care about truth, even though we're not perfect at it.
Re: AI systems and whether they can reason:
Humans train for our whole lives on what is true and what is not true.
(The whole of academia is about working to build new truths.)
LLMs are not trained on truth. They have no representation of truth. We can't say they 'think'.
Even if you keep the test set locked away, you can still get overfitting if you repeatedly re-use it. We provide an example analysis here that generates spurious signal despite cross validation.
Machine learning can easily produce false positives when the test set is wrongly used. Just et al in
@NatureHumBehav
suggested that ML can identify suicidal ideation extremely well from fMRI and we were skeptical. Today retraction and our analysis of what went wrong came out.
As a cognitive scientist I would argue that the puzzle was already solved mentally before it was picked up. The 3.13s is just the motor execution time.
Here are four examples. Again, the lawsuit includes one hundred of them. You get the point. I find this exhibit to be an incredibly powerful illustration for a lawsuit that will go before a jury of Americans. Again, it's impossible to argue with this. /14
@AndreTI
@IrisVanRooij
@Jabaluck
Writing a poem isn't the same thing as conquering the world. Neither is playing chess, or Go, nor solving mathematical proofs....
@xriskology
@GaryMarcus
Yes, also it is absolutely relevant *who* is pushing these worries. Because if they control the narrative, they will also be in control of how the solutions are developed, and who those those solutions will benefit.
Pay attention to red flags, that's what they are for.
There is going to be a literal bodycount from the widespread acceptance of the Hamas statement about who bombed the hospital without evidence.
This has to be a turning point in how we digest and interpret information, especially now that Twitter has lost its rudder.
Scholarly publishing doesn't have to be stuck in a world of outdated document technologies. It's due for an upgrade.
This is why
@neuromatch
via our spinoff organization, Scholar Nexus, is partnering with
@curvenote
on our open publishing project! 🎉🙌
Open PhD positions: Drs. Janet van Hell and Michele Diaz are accepting students for Fall 2024 to explore the neural and cognitive mechanisms of language learning and bilingual language processing (van Hell's lab) >
@ylecun
@SpriteAttack
@survivor_343
This is equating human interests to AI interests, which you should know is not valid. There's no reason to extend the same rights to an AI as a person. (and the analogy is also technically inaccurate as well, but that's a separate point)
What they are giving us is an echo of what they have heard from the AI talking heads and futurists. The fact that it is ~700 voices is irrelevant because they are mostly drawing from the same small pool of sources. The number could be raised to 70k with no increase in validity>
@StuartJRitchie
Have you read the substance of the points that are described elsewhere? The point is emphatically not mere disagreement (which is standard in the field), but rather that IIT often doesn't subscribe to the standards of testability that we typically expect of theories.
>
@pwlot
@emollick
@GaryMarcus
I think people just have rose tinted glasses. They remember that sense of wonder that they first experienced back in the day, and now the flaws are more apparent.
The audacity of for-profit publishing is amazing. If an article is retracted, it is still viewable, and still for $$$. The paper that initiated the retraction is ALSO for $$$.
So they MAKE MONEY WHEN A PAPER IS RETRACTED. Let that sink in when you think about incentives.
This whole thread exemplifies nicely that GPT-4 is not as good at grammar as people think. It overrides grammaticality when it conflicts with its built in sexist stereotypes about occupations.
I cannot imagine what it is like watching people on your own campus essentially calling for your entire culture history to be expunged from the world. We should be able to criticize Israel's conduct in the war but this rally is appalling and noone should support it.
Students at
@Princeton
today were shouting “globalize the intifada” and “there is only one [final?] solution: intifada revolution!”
Hamas showed us what “intifada” means: mass murder, beheading children, gang rape, taking hostage babies, women, and the elderly.
#TheWestIsNext
Recent talk at
@CentreforVisio1
: The awesome
@bradpwyble
spoke to us about "Building compositional memories from latent spaces in a generative model". Wonderful clear and engaging presentation, great intro by
@J_D_Crawford
, and fantastic audience questions throughout.
These stories of settler violence in the West Bank are heartbreaking. The Israeli gov't must stop the settling, which harms people and damages any prospects for peace
On unjust killings in the West Bank, on murders by settlers and IDF raids that kill civilians, I asked some of my Israeli peace movement friends and they recommend
@Yesh_Din
if you'd like to pursue justice
Note that in this example all that GPT does is list the definitions of the terms in the figure. This is not what an "expert tutor" does. You could instead just look up all of these definitions in a textbook and also get contextual/narrative support, which is lacking here
It is hard to overstate how important Ella's competent oversight of NMA operations has been. Behind the scenes, she has deftly managed innumerable mechanisms and without her capable hand on the tiller it is hard to imagine how NMA would be so successful.
The time has come for me to step down as executive director of
@neuromatch
Academy. It’s been an absolute privilege to serve in this capacity for the past two years, primarily because it allowed me to contribute to and work with such an inspiring and dedicated community. (1/N)
Hey everyone, we're doing aphantasia again. It's on page 417 of your beginner's guide to the Discourse
(it keeps coming up because it's both interesting and hard to resolve)
My hot take: this isn't real. Nobody can actually "see" stuff that doesn't exist in the way implied here unless they're actively hallucinating. This is just an issue of these things being extremely hard to explain using language, and people using words in subtly different ways.
@Dick_retired
I think the argument was buttressed by other data, such as spontaneous emergence of grammar in children inventing languages, and the general similarity of grammars for all regions of the world.
Can a large language model be used as a "cognitive model" - meaning, a scientific artifact that helps us reason about the emergence of complex behavior and abstract representations in the human mind? My answer is YES.
Why and under what conditions? 🧵
@KordingLab
No, they make money because they find ways to optimize money which in many cases harms people rather than giving them what they want.
Examples: Monopolies, child labor, lobbying to change the rules of competition, destroying the environment.
The history of >
I agree that who gets credited vs penalized for interdisciplinary work is a heavily biased function. Some academics, typically men, are praised for working in many areas, while others (often women) are marginalized as not serious if they have multiple areas of expertise.
A pattern I have noticed is that even tho I have worked in baby, decision-making, cog/neuro/psych labs and even a year out of academia, my interdisciplinary expertise is subtractive. Men's is multiplicative. I'm not even seen as a "real AI" person, even tho I work in an AI dept.
@o_guest
Given the frequency of "oops LLMs fail here too" and the lack of new "omg look at this" papers, it seems like we are in the hype walkback phase. Watch for the ranks of sparks defenders to back away
@EikoFried
I'm not sure if leave-one-out is particularly worse than k-fold wrt overfitting. I suspect the bigger problem is repeatedly optimizing hyperparameters on the same data set and then still trusting the output of the cross validation.
more here:
It's possible to use DL models within the context of cognitive modelling. But the DL model is NOT the cognitive model. You have to put in the hard work of incorporating, testing, developing the theory. There are other aspects.
Having always worked on artificial neural networks (ANNs) — my PhD a decade ago was on critiquing ANN semantic memory models — I'm kinda grossed out (but not surprised) to see academics totally misunderstand ANNs as cognitive models, likely because of motivated reasoning. 1/
A good thread about the decision faced by
@VSSMtg
. I'm not endorsing all of this, but I agree that Flight is a better choice than Fight. We should also make massive efforts to protect the more vulnerable attendees in Tampa over the next two years (e.g. free rides from TPA)
I've had some time to reflect upon the
#VSS2023
business meeting and thought I'd share some thoughts about
#moveVSS
. The tl;dr - in my opinion, the only ethical option is to try something different than what is being done, and that would require moving VSS out of Florida. 1/n
@TaliaRinger
They absolutely do believe it. My feed is currently full of people who are dead certain that AIs are *already* better than practicing doctors.
@criticalneuro
Yes exactly. AutoGPT is a great example of this. They are going to need to re-invent cognitive psychology to get this thing to be useful, but they have no idea how much work lies before them (and how much we already know about this)
@chazfirestone
What are the odds that someone like SBF building a crypto bank in Silicon Valley would be a huge scammer?
Wait... I'm being told that the Bayesian priors are hugely favorable.
@simonw
@patrickdward
And of course a prompt injection can be quite extensive, covering lots of possible queries.
This seems really basic. I don't understand why people don't get how bad this is likely to be.
It's like have a really dumb guard at a jail and you can just ask him to let you out.
@DrYohanJohn
By conflating these different notions of falsifiability, it is hard to have a clear discussion of what counts as a valid scientific theory so we should try to keep these various distinctions in mind when we throw that word around. /end
@ryxcommar
For me GPT-4 got it wrong, even at various temperature settings. The Original post also pointed out that GPT-4 got it wrong. I realize that sometimes it gets it right but that's still not great.
As I recall, when the movie ended the theater was just quiet. None of my friends or I knew how to react. We were all disappointed, but everyone was reluctant to be the first to say it, given our incredible enthusiasm.
Seeking instructors for Graduate Seminars in Cognitive Attention:
I want to update my syllabus/readings for a graduate seminar in attention, and I'm looking for instructors interested in sharing materials to build better courses
If you're interested: bwyble
@gmail
.com
pls retwt
I admire this kind of work and I agree with Will that WM theory can often fall into debates that really don't bear fruit because of the nature of WM. IMO the key feature of WM is its flexibility, which confounds efforts to measure it in a general sense using a given paradigm.
My ambitious solo paper reviewing our field is now online! "Mapping visual working memory models to a theoretical framework" published in Psychonomic Bulletin & Review (
@Psychonomic_Soc
):
@DrYohanJohn
Finally it may be unfalsifiable because the axioms of the theory and how they are applied make it something close to a tautology. It has been argued that IIT falls into this category, and this is where I think the notion of a theory breaks down. 4/n
We are now seeing a glitzy video that undermines the narrative of the music festival massacre in Israel.
For US folks, before you retweet, this is exactly how the "J6 didn't happen" controversy was manufactured. A few video snippets coupled with high production value. >
@mmitchell_ai
It seems that they are allowing the LLM to have multiple attempts to get the right answer selectively for the ones it go wrong. Huge problem obviously.
I love that Fetterman is an activist Senator. His outspoken voice here on social media is proof that someone can be an effective senator despite temporary damage to their oral communication abilities and validation that he was a great choice to lead the commonwealth
@LeoVasanko
@davidad
@GaryMarcus
I am correct, vast majority of developmentally normal children are generally able to make good choices in tic tac toe, Spitz & Winters (1977)
@AndyXAndersen
@henrymerrilees
@nicholdav
@IrisVanRooij
Of course people do not specify that they are questioning her views *because* she is a woman. These attitudes are typically obscured, but once you start noticing the pattern, it's everywhere.
@KordingLab
business regulation is a never ending battle to build guardrails to prevent their objective function from destroying society. It is ultimately the government that protects us (inefficiently!) from the harms of this objective function.
A theory can be unfalsifiable for different reasons, not all of them equivalent.
It may be currently unfalsifiable because the measurement tools don’t yet exist (e.g. we can’t measure enough neurons in the brain to answer a question). To me this is still a theory 2/n
I think as you get more senior you start to pay more attention to the intro/discussion because it gives you a sense of what people in the field are thinking, and that happens to matter quite a lot.
The longer people are in academia, the more they realize that when reading papers it's best to ignore Intro, Discussion etc. and just look at Methods and Results
This is an example of smart, influential people misunderstanding LLMs. Sydney almost certainly did not conclude that the man's wife didn't love him from data it had about the wife, rather it was confabulating, perhaps from a romance plot.
But this still isn't clear to many
It is imperative that we make sure that new technology like artificial intelligence works for us, and not just a few billionaires who already control everything else about our economy.
@jfrickuga
@NIH
@NSF
The people who push these anti-tenure agendas would be secretly thrilled at such an outcome. They want to see universities collapse.
& for fun we report an amusing effect in our practice trial at the start of the experiment. We asked subjects to simply draw a house to get used to the canvas and tools. We quite often got horses! Here are some of our favorites.
@Jabaluck
@IrisVanRooij
Do you understand that the vast majority of people at these conferences are solving engineering problems and know next to nothing about cognition?
The good news is that these approaches will keep faceplanting due to infeasibility so at least our thoughts are safe.
The bad news is that this won't prevent companies from pretending that they work, which can cause many other harms.
It's unfortunate that the "mind reading" power of brain imaging techniques is being overstated in service of an overreaching agenda that could lead to severe restrictions on data sharing for research.