281 karma · joined May 3, 2021
You might be thinking of Let's Get Together? There is a list there of the few short stories in which the robots act against the three laws.
That being said the Robot stories are meant to be a counter to the Robot As Frankenstein's Monster stories that were prolific at the time. In most of the stories robots literally cannot harm humans. It is built into the structure of their positronic brain.
Gatorade was invented in 1965, and the cross country shoe in the 70s. Prior to that it was the wild west, no one studied or knew anything about this stuff really prior to that.
E.g. "what is the IMDb link for the film Ghost Fever?"
Throwing that in Chatbot Arena will usually result in a random link back or the model will say it can't do it. GPT-4o gave a random link, Claude said it couldn't do it, oddly geneini-advanced-0514 got it right, so hard to tell if they are using Google in the background or if they do better.
But yeah usually IMDb links come back right for big films but tend to be hallucinated for smaller ones. That's just an example.
Asimov wrote his robot short stories in which the three laws played a primary role at a time when robot as Frankenstein's monster was the norm. His short stories attempted to create a more positive optimistic note about how humans and robots could collaborate. The three laws were a way to make it crystal clear that robots could not hurt us, by rule. And the fun was then imagining all the unexpected ways that psychologically that might play out. But in the short stories the robots never actually hurt anyone although they often caused a lot of frustration and fear.
If anything the three laws seemed to show the inate fear of humans to the unknown. The laws were completely impossible to circumvent and people knew this ... And yet they remained staunchly opposed to having robots on earth. Completely illogical.
Anyways, looking at the way LLMs are playing out it seems to me Asimov was wrong. It is quite the opposite. Humans seem to have no fear of robots hurting them, and as a matter of fact seem to get frustrated when a robot isn't allowed to cave their head in with their super human strength when asked (metaphorically).
Think of a simple quadratic. The reflection is across the min(f(x)) axis. But for higher order polynomials that doesn't always work so that is certainly an issue. Thanks
Because in my mind I take a "random" polynomial and only consider the largest two roots. Then I consider doing two reflections: across the bottom/top of the curve, and across the x axis. If those top two roots aren't degenerate the combination of the four curves using the reflections have two curves with a largest real root, and two curves with a largest imaginary root (I think). If it is degenerate then the largest root is real.
So I would then conclude that (1) there are more real largest roots than imaginary and (2) the "advantage" is vanishingly small since there are infinitely more curves with a largest non-degenerate root than degenerate root.
As one can tell I barely know the proper nomenclature. I assume I'm mainly missing some consideration about the uniformity of random coefficients not resulting in a uniform distribution in space? Or possibly my reflections can violate one of the conditions (real coefficients?)? Or I'm just plain wrong.
If an LLM couldn't lie and could be provable shown to be unable to do so would be quite powerful.
The proposed solution is also not really a solution in that it still allows for one of the main issues: random people shouldn't be responsible for creating names for things that should have unique IDs. Which is why things like ensembl IDs exist.
I guess my point is: yes, I imagine the point will be to have something like "I would like to have a picture of George Washington please" and then when it generates it Google will also ask (like in their image search): want to also search that on Google? And enough pass through will generate revenue via their traditional advertising model. Presumably someone who is generating an image of George Washington is doing it for a reason and would like to know other stuff about George Washington.
Ads seem completely unavoidable to me. People like free (prefer it even, go figure) even if it is "free" (with ads), and businesses like ads because it turns out to be by far the most lucrative way to operate (just look at Netflix which is, apparently, actively trying to push people into the ad-tier service because they make much more money per user on the ad-tier than on their paid service).
Maybe eventually I'll get around to making a game or doing something weird like this. Great talk.
The FAQ from Apple does suggest breaks every 20 minutes "as you become acclimated". I don't quite know what that means for someone who, say, uses it for 2 uninterrupted hours a day to watch movies for example.
E.g. I would argue one of the most significant developments in cancer treatment is genome sequencing. It allows you to specifically treat the cancer a person has instead of the (mostly false) perception previously of discovering some panacea in the rainforest that cures all cancers. But the genome was sequenced 20 years ago. At the time there was a perception that that was going to be a step change in custom treatment and solving genetic disease. In reality we are just now reaching the $100 genome and if anything it only served to elucidate how much we didn't know about genetics back in 2000. But slow and steady work means we do indeed sequence people's specific cancer and customize treatment for that. So there was a leap: the Human Genome Project. But afterwards it took a lot of slow and steady progress to get to where we are at.
He thinks mRNA vaccines are going to be the same. A leap forward, and now we are looking at slow and steady progress to getting custom mRNA vaccines to effectively cure cancers on a person-to-person basis.
I've heard mixed things about it as a company but GResearch (also UK) seemed like an interesting R&D software / math mix in the vein of investment banking. I applied there over ten years ago so YMMV at this point, who knows.
This part I 100% agree with. I adapt a lot of scientific code as my day-to-day and most of the issues in them tend to be making things 100x slower than they need to be and then even implementing insane approximations to "fix" the speed issue instead of actually fixing it
>"Big ball of mud" design
Funny enough this was explicitly how my PI at my current job wants to implement software. In his opinion the biggest roadblock in scientific software is actually convincing scientists to use the software. And what scientists want is a big ball of mud which they can iterate on easily and basically requires no installation. In his opinion a giant Python file with a requirement.txt file and a Python version is all you need. I find the attitude interesting. For the record he is a software engineer turned scientist, not the other way around, but our mutual hatred for Conda makes me wonder if he is onto something ...
>I think the appearance of "I'm just getting shit done" is often a superficial one, because it doesn't factor in the real costs: other scientists and engineers can't use their solutions because they're not designed in a way that makes them work in any other setting than the narrow one they were solving for.
For the record my experience is the exact opposite. The crazy trash software probably written in Python that is produced by scientists are often the ones more easily iterated on and used by other scientists. The software scientists and researchers can't use are the over-engineered stuff written in a language they don't know (e.g. Scala or Rust) that requires them to install a hundred things before they are able to use it.
Sounds like the mission of Research Software Engineers (https://society-rse.org/).
I work as a software engineer (with a PhD in a field that is not CS) in a research setting, and there is a give and take. 50% of my job is reading, understanding, and adapting very bad code from non-software engineers into a production system. But another 50% is binning the overwrought inflexible code written by my software engineer predecessors in order to do all that more quickly than refactoring would allow.
In a research setting, in my opinion, MVP is king. Researchers seem to usually not produce viable long-term solutions. But software engineers do as well by virtue of not being domain experts (how could they be, they would need a PhD to understand the research domain!) and being unable to test 100% of the assumptions underlying the software themselves. Which is why it can help to have someone in between who is a domain expert, but knows just enough software engineering to produce production-good-enough code.
I kind of agree with your last point. I try to do all the problems without non-built in python packages myself for exactly the reason you describe. Just feels more satisfying to me.
It is my favorite problem of AoC as well this year mostly because it was the first problem in many years where when part 2 popped I didn't immediately mostly know what algorithm they were going for.
So at least according to him we've lived in that world for the last 5 years.
As a person working with / tangentially to people in the same field I would say that it's made things faster and more scalable, but protein structure isn't the be all end all of things. Researchers use AlphaFold a lot for filtering potential candidates, but that is only one step in a lot of steps. A SNP mutation -> protein structure change -> functional change is already difficult without then considering that the vast majority of mutations that create function change in humans are not in exons, so something like AlphaFold (in the form I'm familiar with) would be useless for those as well.
Eventually though an AI system that can go mutation->function change is entirely possible, although it is much much further in the future. In that case though I think you'll be quite close to a future where combined with things like CRISPR therapeutic treatment for all heritable disease would be possible.
But I also would say multiple odd post processing stuff (obviously completely obscured for security reasons) bolted onto a giant black box model will erode the trust in the results. If a robot was unveiled and the question of "what prevents this robot from using it's superhuman strength from smashing my head in" the answer of "don't worry there is a post processing step in the robots brain whereby if it detects a desire to kill we just cancel that" would be a little disconcerting.
The more satisfying solution is: the model / robot is designed to not be able to produce specific images / to smash human heads in. It just might not really be possible.
Anyways an LLM clearly teased that out whereas if you misremember or misread something a straight search is going to be bad.
Most of my searches are the opposite. I was to know about an obscure movie from the 80s with a toy helicopter. Google very neatly suggests Defense Play (correct) but most LLMs I've tried end up just suggesting very popular films with a helicopter and it ends up being quite difficult to get it to give me information about obscure stuff. Also with that same search the LLM suggests a bunch of incorrect films since (and I figured this out later) it turns out that it was all sourced from a single forum thread from the 90s where a bunch of the posts suggested movies that don't have toy helicopters in them. Go figure.
Just checked my local library system and there are 7 available copies of Heavenly Creatures, I could get it and watch it in a few days via inter library loan. But I'm in the States.
95% of the films I watch via the library I would venture. But it depends on your system.
The two issues with that is (1) they did effectively let humanity "look at the source" in that a big part of the stories was the corporation attempting to get humans to trust the robots by implementing the three laws in such a way that it would be impossible to circumvent (and making those laws very widely known). Didn't work, humans still didn't trust them. (2) as far as I know the operators of LLMs don't seem to currently have a way to give instructions that can't be circumvented quite easily.
Viewing the source and having that source be ironclad was, for Asimov at least, a prerequisite to even attempting to integrate superhuman technology into society.
Regardless, in the early stories, robots could not lie to us. It was indelibly programmed into the positronic brain. They would destroy themselves if put in a position where the three laws were violated.
Anyways, if that were possible with current LLMs I would think the hallucination problem would have been trivially addressed: just program in that the LLM can't tell a lie.
Regardless my point is it is a good series and surprisingly tame. Sierra adventures have a ... style but I love them. The main thing is: save a lot. More than you think.