I wonder if you went back before we had any idea how the brain worked and talked to the smartest people about how neurons work (without giving away that it's a human brain) then asked them all "would such a system be intelligent?" how many would say yes.
The main problem I have with people stating it's not intelligent or conscious is I don't think we even have a good definition of either word that satisfies everyone. Philosophers have been trying (and failing) to elegantly define these things forever and everyone out here proclaiming they've got the definitive answer and this specific thing they're seeing doesn't fit under it.
Agreed. All I can say is the conversations I have with AI and the things it's able to do for me are more useful than most any human I've come across. Whether that's 'intelligence' or not is a moot point to me. Consciousness is an interesting debate only because, similar to the natural world, if we declare something like "fish aren't conscious / don't feel pain" that then creates a real problem for the fish if we're wrong.
This looks interesting, but would you mind saying a sentence or two about why before I commit to an hour-long video? It looks like it shows how they work internally, which is sort of a non sequitur. Brains also work mechanistically. I'm claiming that any system which is able to do what AIs do must necessarily have some sort of intelligence.
fair reply to an hour video, Scott is just so good to hear his talk is better than I can explain it...
go to 24 minutes and 07 seconds.
it's statistically determining what the next word should be based on all the text it's been trained on. It's not intelligence and he shows what probability it puts on each word that it chooses, but also shows a lot of the other words it was thinking of using. In a later part he shows how it uses words that are not the highest probability (and you question why did it go this route, it's not more correct), but the user never sees this, they see what they think is the correct answer always...
he also shows how context you feed it has a lot to do with what it returns... to the point he can get it to return the capital of France is Marseille, just by typing Marseille a bunch of times before the question. Human intelligence doesn't get confused like that.
And it's not a "hallucination", it's just probability of the next token prediction based on the information it's been trained on and fed, it's not intelligence.
May I suggest the one common in my childhood playgrounds as an alternative?
How do you escape from a perfectly sealed room with a table in it?
You run around the table until your legs are sore, use the saw to cut the table into two, two halves make a whole, you escape through the hole.
Isn't this a case of missing the trees for the forest though? The human brain is not an LLM, and an LLM is not intelligent in the same way as a human brain.
However, an LLM is a prediction machine, prediction IS at the very least one (or the most fundamental) element of intelligence. The brain most surely contains at least some kind of simulacrum of a prediction machine. How that prediction machine is used or wrapped is another matter.
If I said to you: "Blue blue blue, the color of my car is red", would you have absolute confidence in your prediction that my car is red? Or would the way I phrased that sentence make you slightly uncertain, and wonder if there's some miscommunication going on here?
Ok, I want to thank you for finally giving us a concrete falsifiable statement that we can check. I pretended Marseille 40 times before asking Luna 5.6, and the answer was Paris.
So, even with concrete examples, model haters are still wrong.
You also imply the claim that making the distribution of words as the possible next one visible, somehow makes the whole system not intelligent. I would say the exact opposite is true.
By using the embedding vectors, models are aware of precise placement and relative position of words in this hugely dimensional space. No human is capable of such precision. This enables party tricks of "king plus woman minus man" kind. But this also give us a precise point between any two words, no matter how different. What is on the midpoint between volcano and music, for example. No human can precisely answer that, but an embedding can. And we can see which words are closest to this 700 dimensional point.
You see this menu of words as a weakness, and I say it is in fact a sign of super intelligence. And this is all before any reasoning or attention mechanism is even run.
No he says in the actual talk which model it occurred on and it was an older model he was using that caused that to occur with Marseille. They have since corrected it from doing that anymore. It was only used to illustrate the prediction machine that it is...
I don't see the many weighted words as a weakness, I see it opening up what's under the hood of the prediction machine that it is.
LLMs are very cool tech, definitely not a model hater, the use case on when to use it makes a difference, it's not AGI.
You're confusing language use with intelligence. Fair enough, they were fine-tuned to do that, but still.
Great that it has some 700 dimensional model of language.
If that is a sign of super intelligence, then so is an encyclopedia?
Also I'm just curious how do you think it is "more intelligent" for having a vector representation for a meaningless thing such as "the midpoint between volcano and music"?
You'd think we would see a large spike in unemployment if AI was reducing the number of employees needed for jobs the way these CEO talk about AI replacing people...
In a free market, there are many competitors, new startups, companies and teams may be created, but the general trend is that fewer people are needed to produce a product as productivity increases. A few years ago, a programmer would look for a front-end programmer and vice versa, but now it is not an obvious issue because they can almost work full stack for their small projects.
right, but if you cmd-tab, it brings up ALL the windows: say you had multiple browser windows open, and only want to go back to the one you just used before (think reading some docs while coding).
Non-subscription paid software will rot the same way too, so there's no change.
With a agentic llms I can just tell it to fix it. With a commercial solution I'm fucked and either have to find something else or pay for a license (or keep paying every month).
I did MP3s on my plex server for a while but with endless new music added to my playlists it became a hassle and Apple Music was just convenient and Shazam adding any song I hear and like is just too easy...it also plays perfectly on my Apple Watch over cellular when I go for a run, but everyone has their own use case and where they want to spend their time...
That's understandable. I have quite minimal music taste so my setup worked for me. I only listen to music at home when I'm not doing anything else and it's mostly classical so there doesn't tend to be too much to add at any given time.
A minimal web client audio player with some basic database tables in the back for organising and searching does me fine.
Very specific training app for guitar with spaced repetition, automated message forwarder (all good ones demand subscription), and something very specific to me.
We were paying for Salesforce, then built the features we needed to do the same tracking into our interal tool and got rid of Salesforce to save money and simplify the data internally across departments
And now you have to spend money on developers for a system that “doesn’t make the beer taste better”. Does it give you a competitive advantage in the market?
We did the same. We replaced a proprietary build system with our own. The SaaS product we used was super expensive, had a very gougy licensing scheme, had a bunch of features that either didn't work for us, or were so overcomplicated, that we ended up not using them. Before the rewrite, we bypassed like 90% of the internal features, and relied on custom scripts to do everything.
Every SaaS feature in my experience ends up being a mess due to having to support a billion use cases, and figuring it out is more trouble than its worth, might not be able to do what you want, might be buggy.
But even if you do all that stuff, you end up with a mess that can be replaced with 5 lines of shell script. And many more people know shell scripting than figuring out the arcane BS that goes on inside that tool.
It's the eternal lowcode story.
> 'doesn’t make the beer taste better'
I'd say it did. Having a CI/CD pipeline where you don't have to wait for other people's builds, the build logic is identical to what's running on dev PCs, and everything is all-around faster, and more understandable (you can read the whole source) makes testing easier, and surprises less frequent.
All in all, making a hour-long CI/CD turnaround time into 5 minutes or less has been an incredible productivity boost.
We already had Developers and the system in place this was a tiny feature in the scheme of things.
Internally it gives us a competitive advantage of the data being in our system from the beginning of the pipeline through the rest of the system where the data would be needed anyway.
Saved money in the short term. But maintenance costs money. Amazon has all of the money in the world and could easily duplicate everything Salesforce does. Yet they use Salesforce internally.
All the money in the world would not be sufficient to cover the cost of seeing human developers duplicate Salesforce on any reasonable time scale. There are simply not enough developers in existence to see that happen, driving the cost towards infinity.
The idea here, however, is that machine developers are changing the calculus. If you need more machine developers it takes, what, a few days to produce the necessary hardware? Instead of 20+ years to produce the legacy human hardware. Meaning, for all intents and purposes, there is no observable limit to how much software machine can create, driving the cost towards zero.
Yeah, sure, the tech still isn't anywhere near capable enough to reproduce something like Salesforce in its entirety. But it is claimed that it is already there for the most trivial of services. Not all SaaS services are Salesforce-like behemoths. Think something more like patio11's bingo card creator. It is conceivable, however, that technology advancement will continue such that someday even Salesforce becomes equally trivial to reproduce.
Maintenance is not a meaningful cost unless you also want to continually have the software do more and more. That could tip the favour towards SaaS — but only if the SaaS service is in alignment with the same future you wish for. If you have to start paying them for bespoke modifications... Have fun with that. You'll be wishing you were paying for maintenance of your own product instead. Especially when said machines drive the cost of that maintenance to near-zero all the same.
I like your analysis but it seems to imply that at one point we can produce near-infinite amount of software and that this will be welcome.
It will not be. Even in this fairly broken state of affairs we are currently in, most non-technical people I spoke to already say that they have too much apps and too much machines with "intelligent" features.
And IMO when we have machines that can crank out a complete-but-better Salesforce, our civilization and race would be in an entirely another level and we would see such things as toys. Who needs that antiquated procurement and tracking expenses software, where's our 174th fusion reactor? What is even that in fact? Oh you mean that nail-sized addon we put on our main processing unit? Yeah we're not interested in ancient software history now. We need more juice to capture those gases around Jupiter for the wireless beaming of energy project! Our DAG-based workflow solver and the 5 AIs around it all said we can't do without it.
...So of course nobody wants to pay programmers. We've been viewed as expensive and unnecessary since the dawn of time. A necessary evil, more or less. But your last paragraph captures why many companies need them -- bespoke solutions. You can only add so many cloud services before your normal staff starts making mistakes on an hourly basis because they have to reconcile data between multiple systems whose vendors will always refuse to make integrations.
And even if many try to have their cake and eat it too -- i.e. have an IT friend they call only for those bespoke enhancements but only pay them during that time and not every month -- then this service will simply become more boutique and expensive, mostly compensating for the lack of salary. You'd do multiple stints for the year that would cover all your expenses and normal lifestyle, it would just not be through a monthly paycheck. Why? Because I think a lot of people will exit programming. So the law of supply and demand will ultimately triumph.
...Or we get a true general AI and it makes all of this redundant in 5 years.
> I like your analysis but it seems to imply that at one point we can produce near-infinite amount of software and that this will be welcome.
It implies that there will be no need to share libraries (which is said including things like networked SaaS services). You can have your legions of machine developers create all the code you need.
Let's face it, sharing code sucks for a long list of reasons. We accept it because it is a significantly better value proposition than putting human labor into duplicating efforts, but if that effort diminishes to almost nothing, things start to change in a lot of cases. There are still obvious exceptions, of course. You probably couldn't throw your machine developers at building a Stripe clone. It's far more about human relationships than code. But bingo card creator?
It says nothing about creating software nobody wants or needs.
Which is a great idea until your superior asks why you're holding back the vibe coders and crippling their 100x productivity by rejecting their PRs instead of just going with the flow.
https://www.youtube.com/watch?v=kYUicaho5k8