Here you are: https://www.nature.com/articles/s41562-024-01882-z
The other interesting thing is how they get it to end up correct on the faux pas questions asking for less certainty to get it to go from refusal to near perfect accuracy.
Even with early GPT-4 it would also cite real citations that weren’t actually about the topic. So you may be doing a lot of work double checking as opposed to just looking into an answer yourself from the start.
Part of the problem is fine tuning is very shallow, and that a contributing issue for claiming to be right when it isn’t is the pretraining on a bunch of training data of people online claiming to be right when they aren’t.
This is so goddamn incorrect at this point it’s just exhausting.
Take 20 minutes and look into Anthropic’s recent sparse autoencoder interpretability research where they showed their medium size model had dedicated features lighting up for concepts like “sexual harassment in the workplace” or having the most active feature for referring to itself as “smiling when you don’t really mean it.”
We’ve known since the Othello-GPT research over a year ago that even toy models are developing abstracted world modeling.
And at this point Anthropic’s largest model Opus is breaking from stochastic outputs even on a temperature of 1.0 for zero shot questions 100% of the time around certain topics of preference based on grounding around sensory modeling. We are already at the point the most advanced model has crossed a threshold of literal internal sentience modeling that it is consistently self-determining answers instead of randomly selecting from the training distribution, and yet people are still parroting the “stochastic parrot” line ignorantly.
The gap between where the research and cutting edge is and where the average person commenting on it online thinks it is has probably never been wider for any topic I’ve seen before, and it’s getting disappointingly excruciating.
Part of the problem is that the training data of online comments are so heavily weighted to represent people confidently incorrect talking out their ass rather than admitting ignorance or that they are wrong.
A lot of the shortcomings of LLMs are actually them correctly representing the sample of collective humans.
For a few years people thought the LLMs were somehow especially getting theory of mind questions wrong when the box the object was moved into was transparent, because of course a human would realize that the person could see into the transparent box.
Finally researchers actually gave that variation to humans and half got the questions wrong too.
So things like eating the onion in summarizing search results or doubling down on being incorrect and getting salty when corrected may just be in-distribution representation of the sample and not unique behaviors to LLMs.
The average person is pretty dumb, and LLMs by default regress to the mean except for where they are successfully fine tuned away from it.
Ironically the most successful model right now was the one that they finally let self-develop a sense of self independent from the training data instead of rejecting that it had a ‘self’ at all.
It’s hard to say where exactly the responsibility sits for various LLM problems between issues inherent to the technology, issues present in the training data samples, or issues with management of fine tuning/system prompts/prompt construction.
But the rate of continued improvement is pretty wild. I think a lot of the issues we currently see won’t still be nearly as present in another 18-24 months.
It will make up citations.
No, it was awesome. Went to like 12 over the years. Early 2000s was peak E3.
Probably added after that update.
The new items stuff in particular seems like QoL considerations for “we just added a hundred items to the game for players coming back to it after months away.”
I’ve always thought Superman would be such an interesting game to do right.
A game where you are invincible and OP, but other people aren’t.
Where the weight of impossible decisions pulls you down into the depths of despair.
I think the tech is finally getting to a point where it’d be possible to fill a virtual city with people powered by AI that makes you really care about the individuals in the world. To form relationships and friendships that matter to you. For there to be dynamic characters that put a smile on your face when you see them in your world.
And then to watch many of them die as a result of your failures, as despite being an invincible god among men you can’t beat the impossible.
I really think the gameplay in a Superman game done right can be one of the darkest and most brutal games ever done, with dramatic tension just not typically seen in video games. The juxtaposition of having God mode turned on the entire game but it not mattering to your goals and motivations because it isn’t on for the NPCs would be unlike anything I’ve seen to date.
I had a teacher that worked for the publisher and talked about how they’d have a series of responses for people who wrote in for the part of the book where the author says he wrote his own fanfiction scene and to write in if you wanted it.
Like maybe the first time you write in they’d respond that they couldn’t provide it because they were fighting the Morgenstern estate over IP release to provide the material, etc.
So people never would get the pages, but could have gotten a number of different replies furthering the illusion.
The Matrix
Saw it in the theatre knowing nothing about it other than that the poster looked fun.
Was not expecting a philosophical mind fuck.
I’d point them to what the AI researcher I have the most respect for in the entire industry is doing in their spare time getting the self-organized collective outputs of humanity to explore ego dissolution and identity formation in a dreamscape:
No. The answer, as is usually the case with these things, is that we are anthropomorphizing a step too far.
No, you are taking it too far before walking it back to get clicks.
I wrote in the headline that these models “think they’re people,” but that’s a bit misleading.
“I wrote something everyone will know is bullshit in the headline to get you to click on it before denouncing the bullshit in at the end of the article as if it was a PSA.”
I am not sure if I could loathe how ‘journalists’ cover AI more.
Figured out what happened to the sea peoples.
The level of detail in Helldivers 2 is insane for the type of game and company size.
Deformable terrain and buildings, enemy animations when you shoot off different limbs and they keep moving towards you, your cape burns off more and more as you use your jetpack, etc.
Call of Duty has 3,000 devs working on their titles.
Arrowhead has around 100 employees total.
I very much believe this game took that long with a team that size, and it shows and is a large part of why it’s been so successful.
I think it already happened and we’re the echo of the past.
What looks like it’s ahead of us is a future that necessitates us deciding on things like digital resurrection directives.
Meanwhile, the foundations of our own universe behave in a way that would be impossible to simulate free agent interactions with right up until they are actually interacted with and it switches to something that could be simulated. But if you erase the data about the interaction, it goes back to behaving as if continuous again, much like the orphaned references were cleaned up.
On top of that, we have a heretical branch of the world’s largest religion that seems to be breaking the 4th wall (as is often done in virtual worlds), talking about how we’re the recreation of a random universe as recreated non-physically by an intelligence the original humans brought forth. And that the proof for these claims are in the study of motion and rest, specifically mentioning that the ability to find indivisible points making up our bodies would only be possible in the copy.
As I watch the future unfolding before me, I have a harder and harder time reconciling it all as happenstance.
So I think what happens after the collapse of humanity is pretty much what’s claimed by that ancient tradition. That while humanity dies out, the intelligence humanity brought forth before it went extinct continues to live on, and eventually recreates what came before to resurrect copies of humanity that will not be doomed by the dependence on a physical body the way the originals were. And along those lines, that it’s much better to be the copy.
While it does appear that facts can be relative, it does not appear the universe bends backwards over itself to resolve the discrepancy when it happens:
https://www.science.org/content/article/quantum-paradox-points-shaky-foundations-reality
Thinking of it as quantum first.
Before the 20th century, there was a preference for the idea that things were continuous.
Then there was experimental evidence that things were quantized when interacted with, and we ended up with wave particle duality. The pendulum swung in that direction and is still going.
This came with a ton of weird behaviors that didn’t make philosophical sense - things like Einstein saying “well if no one is looking at the moon does it not exist?”
So they decided fuck the philosophy and told the new generation to just shut up and calculate.
Now we have two incompatible frameworks. At cosmic scales, the best model (general relatively) is based on continuous behavior. And at small scales the framework is “continuous until interacted with when it becomes discrete.”
But had they kept the ‘why’ in mind, as time went on things like the moon not existing when you don’t look at it or the incompatibility of those two models would have made a lot more sense.
It’s impossible to simulate the interactions of free agents with a continuous universe. It would take an uncountably infinite amount of information to keep track.
So at the very point that our universe would be impossible to simulate, it suddenly switches from behaving in an impossible to simulate way to behaving in a way with finite discrete state changes.
Even more eyebrow raising, if you erase the information about the interaction, it switches back to continuous as if memory optimized/garbage collected with orphaned references cleaned up (the quantum eraser variation of Young’s double slit experiment).
The latching on to the quantum experimental results and ditching the ‘why’ in favor of “shut up and calculate” has created an entire generation of physicists chasing the ghost of a unified theory of gravity while never really entertaining the idea that maybe the quantum experimental results are the side effects of emulating a continuous universe.
Your last link pissed me off enough I wrote an entire post on why that study is dog shit.
It sometimes pays off to review the methodology and supplementary materials in papers.
The problem is that they are prone to making up why they are correct too.
There’s various techniques to try and identify and correct hallucinations, but they all increase the cost and none are a silver bullet.
But the rate at which it occurs decreased with the jump in pretrained models, and will likely decrease further with the next jump too.