That model stops working as soon as you try to distribute software at scale. You need some kind of standardisation and baseline. Otherwise your customisations conflict with someone else's customisations, with consequences that vary from mildly annoying to catastrophic. And the whole idea of shared software collapses.
It would work if you had strict access control to each feature in a common code base. But Git hadn't been invented yet.
The other issue is performance. Compiled ST isn't particularly slow, but you lose the modifiability. Interpreted ST has a mixed profile - some features are slow, some are fast - but generally it was slow compared to C.
Today that doesn't matter so much, but it was a drawback at the time.
It's a seductive model and I totally get the appeal. But it's not quite as straightforwardly superior as it might appear to be.
As far as performance, the Alto team described the performance of Smalltalk as between "majestic" and "glacial". From my experience with the Alto, Smalltalk seemed unusably slow, and I'm amazed that they could actually accomplish anything with it.
With the ParcPlace flavour teamwork with Envy version control was pretty slick. And in performance it was mostly compiled to bytecode with some cunning tricks in place to make interactive debugging appear to work at the source level. Truly great environment for exploratory development: concrete example, while building a comms server I could inspect why message handling failed, patch code, rerun from failure, much faster cycle time than most other langs/envs.
Git isn't the first or oldest DVCS. Smalltalk actually had multiple DVCSs, including ENVY (VA/Cincom), which dates back to the mid 90s, and Monticello (Squeak/Pharo), which dates to the early 2000s--both of which predating git.
Erlang systems don't do so well when multiple people are working on them with different understanding of what production is. (And especially if nobody actually knows).
I don't know that you need 'strict access control' as proposed, but you do need coordination, and a source code repository is a way to manage coordination.
Erlang doesn't have a different mode for fixed vs loadable code, although you may need to write code differently in order to be able to hotload changes, and that could affect performance; some code naturally fits the idioms needed though.
I find it amazingly valuable to be able to change the program without having to lose the program state; but it requires different thinking, and it is easy to mess things up in exciting ways that are different than many other systems. It's amazing to fix a bug and deploy it in seconds into a live system with millions of active connections; it's also amazing when you try to fix a bug and break millions of active connections in seconds, but in a negative way, because now you and your users have to pay the costs of reestablishing all of those connections.
OTOH, the key issue is interchange of state and messages between versions, and that's an issue that comes up in almost all distributed systems that update without downtime, so most people have to deal with that issue anyway, just usually not within the much smaller scope of a single task.
Icons make localisation much easier. In fact flat web design has evolved a fairly standard set of icons for basic operations. Most people know what a burger menu and x in the top corner of a window do. Same for copy, share, and so on.
The problem with Liquid Glass is that it's making the background style more important than the foreground content. No one cares if buttons ripple if they can't see what they do, because icons themselves are less clear and harder to read.
So I don't know what the point of this is.
Unifying the look with Apple's least successful, least popular, most niche product seems like a bizarre decision. I'm guessing the plan is to start adding VisionPro features in other products, but without 3D displays the difference between 3D and 2D metaphors is too huge to bridge.
I really liked Aqua. It was attractive and it was very usable.
This is... I don't know. It seems like style over substance for the sake of it, with significant damage to both.
Democratic voters want those things. It's not at all obvious the party establishment does.
The tell is that when Republicans push through their policies, Democratic opposition is weak and ineffectual. Instead of ferocious opposition the Dems send one of their famous sternly worded letters.
Since at least 2000 the party establishment has absolutely refused to do any of the things it could do to change this - including packing the Supreme Court, supporting and promoting grass roots activism between elections, using the filibuster, and so on.
Biden couldn't even get any of Trump's prosecutions over the line - including televised evidence of insurrection, and treasonous mishandling of official state secrets (!)
However it's spun, there is a very obvious reluctance to challenge the extremes of Republicanism.
The party is far more likely to censure one of its non-centrists than its centrists, while the opposite is true of the Republicans.
The Democrats operate as if they're controlled opposition. It's like their donors pay them to blunt their base. They haven't accomplished anything legislatively this century beyond pass the 1993 Republican healthcare plan under Obama's name. They couldn't even raise the minimum wage.
In my experience this is dead on. People have short attention spans but this has been happning the whole 21st century. In 2008 Obama won the primary despite the best efforts of leadership to nominate clinton. They even scrambled the "super delegates" (delegates who vote for the candidate chosen by senior leadership) hoping that even if Obama won more delegates, they could override the voters choice.
Of course, they failed, and democrats won 2 elections in a row running a candidate labeled a radical socialist. Obama became the only 21st century president to win the poplar vote twice, and the DNC has been trying to drag the party back in the 20th century ever since, blaming their own voters when it doesn't work.
It boggles my mind that they refused to even engage with the "undecided movement", which created a grass-roots get out to vote movement out of thin air. In swing states no less.
The starkest contrast between the two parties is womens rights and to a lesser extent LGBTQ rights. Although I'm not even sure how true this is anymore with so many politicians backing Cuomo, who resigned because an investigation found overwhelming evidence he sexually harrassed and assaulted female employees. And I'm pretty sure people like Chuck Schumer and other centrists view the LGBTQ community as a liability.
Both are true. The end of the Fairness Doctrine normalised the psychotic distortions and lies pumped out by Fox. But the same machine that uses Fox also runs bot farms, astroturfing operations, and curated social media algorithms to normalise even more extreme RW POVs.
With enough money, you absolutely could persuade Trump voters to vote for Democrats.
This isn't up for debate. PR and advertising exist because this is absolutely possible, and has been for a long time.
AI makes it much, much easier and more cost-effective.
This was how Cambridge Analytica and Facebook swung the Brexit vote. They didn't send out blanket "EU bad, vote Brexit" content. They created targeted ads that addressed hot-button fear points for individuals and niche demographics and used Facebook's ad targetting system to deliver them.
So some people were concerned about money for the UK's health system. They saw ads promising that Brexit would mean more money for the NHS. Others were concerned about potholes, or education, or - often - immigration.
Every group saw ads that triggered their personal hot-button issues and persuaded them to vote against their interests.
Ai can also nudge content choices towards autarch-sanctioned beliefs without the viewer being aware of it.
This has been happening for decades already. But AI can make it personal in a way that mass media can't.
Combine it with the kinds of psychological triggers and manipulations used in PR and advertising and you can convert almost anyone. You don't even need violence - just repetition.
This has already happened, btw. The Q phenomenon successfully radicalised entire demographics through careful use of emotional triggers and techniques to enhance suggestibility and addictiveness.
You don't need phenomenal consciousness. You need consistency.
LLMs are not consistent. This is unarguable. They will produce a string of text that says they have solved a problem and/or done a thing when neither is true.
And sometimes they will do it over and over, even when corrected.
Your last paragraph admits this.
Tokenisation on its own simply cannot represent reality accurately and reliably. It can be tweaked so that specific problems can appear solved, but true AI would be based on a reliable general strategy which solves entire classes of problems without needing this kind of tweaking.
This is a common category of error people commit when talking about LLMs.
"True, LLMs can't do X, but a lot of people don't do X well either!"
The problem is, when you say humans have trouble with X, what you mean is that human brains are fully capable of X, but sometimes they do, indeed, make mistakes. Or that some humans haven't trained their faculties for X very well, or whatever.
But LLMs are fundamentally, completely, incapable of X. It is not something that can be a result of their processes.
These things are not comparable.
So, to your specific point: When an LLM is inconsistent, it is because it is, at its root, a statistical engine generating plausible next tokens, with no semantic understanding of the underlying data. When a human is inconsistent, it is because they got distracted, didn't learn enough about this particular subject, or otherwise made a mistake that they can, if their attention is drawn to it, recognize and correct.
LLMs cannot. They can only be told they made a mistake, which prompts them to try again (because that's the pattern that has been trained into them for what happens when told they made a mistake). But their next try won't have any better odds of being correct than their previous one.
>But LLMs are fundamentally, completely, incapable of X. It is not something that can be a result of their processes.
This is the very point of contention. You don't get to just assume it.
> it is because it is, at its root, a statistical engine generating plausible next tokens, with no semantic understanding of the underlying data.
Another highly contentious point you are just outright assuming. LLMs are modelling the world, not just "predicting the next token". Some examples here[1][2][3]. Anyone claiming otherwise at this point is not arguing in good faith. It's interesting how the people with the strongest opinions about LLMs don't seem to understand them.
OK, sure; there is some evidence potentially showing that LLMs are constructing a world model of some sort.
This is, however, a distraction from the point, which is that you were trying to make claims that the described lack of consistency in LLMs shouldn't be considered a problem because "humans aren't very consistent either."
Humans are perfectly capable of being consistent when they choose to be. Human variability and fallibility cannot be used to handwave away lack of fundamental ability in LLMs. Especially when that lack of fundamental ability is on empirical display.
I still hold that LLMs cannot be consistent, just as TheOtherHobbes describes, and you have done nothing to refute that.
Address the actual point, or it becomes clear that you are the one arguing in bad faith.
You are misrepresenting the point of contention. The question is whether LLMs lack of consistency undermines the claim that they "understand" in some relevant sense. But arguing that lack of consistency is a defeater for understanding is itself undermined by noting that humans are inconsistent but do in fact understand things. It's as simple as that.
If you want to alter the argument by saying humans can engage in focused effort to reach some requisite level of consistency for understanding, you have to actually make that argument. It's not at all obvious that focused effort is required for understanding or that a lack of focused effort undermines understanding.
You also need to content with the fact that LLMs aren't really a single entity, but are a collection of personas, and what you get and its capabilities do depend on how you prompt it to a large degree. Even if the entity as a whole is inconsistent between prompts, the right subset might very well be reliably consistent. There's also the fact of the temperature setting that artificially injects randomness into the LLMs output. An LLM itself is entirely deterministic. It's not at all obvious how consistency relates to LLM understanding.
Feel free to do some conceptual work to make an argument; I'm happy to engage with it. What I'm tired of are these half-assed claims and incredulity that people don't take them as obviously true.
It's very localised and Californian. There were really two big tech scenes - one around MIT and Mass, and one around CalTech/Stanford and adjacent areas - with some also-rans in other areas that were mostly gov mil/aerospace spinoffs.
The Mass scene sort of fizzled in the 90s for various reasons - not dead, but not dominant - and the centre of gravity moved to the West Coast.
So if you were born in CA and studied there - and Atkinson did both - your odds of hitching your wagon to a success story were higher than if you were born in Montana or Dublin.
This is sold as a major efficiency of US capitalism, but in fact it's a major inefficiency because it's a severe physical and cultural constraint on opportunity. It's not that other places lack talented people, it's that the networks are highly localised, the culture is very standardised - far less creative than it used to be, and still pretends to be - and diverse ideas and talent are wasted on an industrial scale.
FWIW CalTech is in southern California and far away (both geographically and socially) from Stanford. Its strengths also tend to be primarily in physics, rocketry, and astronomy, rather than in CS - its primary ties are with JPL and NASA. The Bay Area tech scene is anchored by Stanford and UC Berkeley, though most Stanford alums would probably say it's just Stanford.
There's probably a book in there. The CA axis was probably Stanford/Berkeley with Caltech relatively small and in another part of the state and probably much more theoretical in focus.
Don't really buy Levy's thesis of the migration from east to west and Stallman as "the last hacker" hasn't aged well.
But Boston/Cambridge (really Massachusetts generally) did sort of empty out of a lot of tech for a time as minicomputer companies declined and Silicon Valley became the scene. I actually decided not to go that direction because, at the time in the nineties, it would have been a relative cost of living downgrade.
You said it yourself - universities are the major hubs that bring talented driven people together and provide access to some of the greatest teachers and researchers and other resources. MIT and Stanford are special, somehow, in this regard.
You see this as inefficient and maybe you’re right. I think about how little it has cost to run these schools compared to the wealth (financial, cultural, technological) they spin off and to me it looks very efficient.
> This is sold as a major efficiency of US capitalism, but in fact it's a major inefficiency because it's a severe physical and cultural constraint on opportunity.
I don't think social relationships and their geography are a particular characteristic of capitalism - let alone US-specific capitalism.
They - and the resulting hub/centralization effects - predate it by millennia. There is no shortage of historical cities or state that became major hubs for certain industries or research. How much of the effort in those places is "wasted" seems hard to quantify in an objective way, but again, the pattern of low-hanging fruit being more available to the first wave and then a lot of smart, hard-working people in the future generations working more around the edges is not capitalism-exclusive.
It would work if you had strict access control to each feature in a common code base. But Git hadn't been invented yet.
The other issue is performance. Compiled ST isn't particularly slow, but you lose the modifiability. Interpreted ST has a mixed profile - some features are slow, some are fast - but generally it was slow compared to C.
Today that doesn't matter so much, but it was a drawback at the time.
It's a seductive model and I totally get the appeal. But it's not quite as straightforwardly superior as it might appear to be.
reply