Every time it’s like
“City dwellers, how do you live a suburban lifestyle given the constraints of the city”
>we don’t live a suburban lifestyle, we live an urban one, fitting our preferences
“That doesn’t make sense because it doesn’t match my preferences”
Despite claims to the contrary, with the smartest models, you really do have to watch constantly. Mainly if you're working on new things. If there's a pattern in the codebase, it's usually pretty good, but man, otherwise.
This. It feels a bit unsettling in systems academia. Worse yet, the problems many papers try to optimize are already well-represented in the scientific literature, and even a good (albeit agent psychosis-induced) result often stems from the LLM having been trained on the solution.
I am absolutely in favor of pushing AI to solve hard problems, but I worry we are missing the plot. Maybe the current trend of writing papers using AI to (re)discover systems algorithms is on the path to AI truly discovering novel systems optimizations. If not, I hope we correct course soon.
The models get smarter with each release. Proportional to that increase in intelligence is an even greater increase in their ability to bullshit you into believing that they know what they're talking about.
Man goes to doctor. Says he's depressed about AI. He fears the permanent underclass.
Doctor says, "Treatment is simple. Read Gary Marcus. LLMs are stochastic parrots—they can't reason out of distribution."
Man bursts into tears. "But doctor..." he says, "I am in distribution!"
on memory/continual learning, he's very clear that that will be an app level thing, not a model thing, because it's not practical, economical or even possible to fine tune per user models.
i think we'll find a lot of other limitations that can never be solved in the model layer, simply because we've kinda painted ourselves into an architectural corner with transformers. which is why i think transformers will be a primitive within AGI systems, but are very unlikely to be the sole thing giving us AGI.
on a more positive note, that means there's ample opportunity for folks outside the big labs to innovate and create systems that are more than just a chatbox hooked up to a model.
it also means, that a lot of this work will basically be relearning all the things from information retrieval ca. 2000s, and combining that with new methods. same for ensemble techniques, etc.
a certain CEO at chroma is gonna like this tweet :)
@usr_bin_roygbiv lol new york is a nightmare in many special ways but I do appreciate not ever having to check my fucking shoes or under a chair before I sit down. miss my old newtown terrace tho
There is a common pejorative attack on rwers, hinted at here, which boils down to “they’re just bitter that their liberal friends think they’re scum.”
And sure, there is something very funny and self-defeating about “owning the libs to get them to like you”. Moreover, the typical response pattern is often full of vice-signaling, commons-plundering, and all manner of “I don’t care if this hurts me, just so long as it hurts you more” behavior that simply validates those same negative labels to which they so strongly object.
It’s childish. It’s grotesque. It’s counterproductive and negative-sum. It is ignobility at its most ignoble. And perhaps most of all, it’s just plain embarrassing.
HOWEVER. I think that the widespread appeal of this reactionary behavior ought to serve as a damning indictment of the overall “theory of change” that the 2010s woke movement represented. The theory that, if you leverage cultural power to shun, shame, and banish wrongthink from the domain of acceptable discourse, that this will somehow engender a form of lasting moral and ideological victory—that by making certain viewpoints unacceptable, you would make them unthinkable.
This theory of change, misguided as it was to begin with, culminated in a mass cultural backlash, the first and second term of Donald Trump, and a reactionary movement so hellbent on revenge that they have willingly destroyed an immense amount of valuable institutional capital just for the chance to inflict a similar degree of anguish upon what they see as their oppressors. By any reasonable metric, this must be considered an abject failure.
This failure should have been foreseeable. In a political system in which everyone’s vote is equal and private, and a new media landscape that eschews top-down enforcement of norms, instead rewarding self-radicalizing echo chambers and simclusters, this strategy is predictably disastrous. Ideologies which are not allowed to participate the light of open debate will merely fester in the dark.
Ancient societies that committed wholesale massacre of the defeated, brutal as they were, at least understood this simple principle: if you want lasting victory, your enemies must either be captured, convinced (i.e., integrated), or killed. Silence is not an option. They will not go quietly into that good night. Slaves without shackles will always revolt once they locate the power to do so, and in doing so they will rarely take care to maintain the prudent constructs of their former masters. They will simply burn it all to the ground.
My current theory of politics is that the #1 problem with our system today is the hollowing out of the conservative movement. In a way, the memetic power of wokeness did succeed in vanquishing traditional conservative ideas from the domain of elite discourse. But if conservative ideas are universally treated as vile, stupid, and evil, then the only men who will remain to champion them will be those who embrace such labels. And in the absence of a Reagan or Friedman or Romney to root for, the shunned masses will search for salvation in men like Trump. If they cannot see themselves in Captain America, they will simply embrace Homelander.
This sort of crazy deference to vagrants is how we end up with a north american gypsy culture. Between Canada and the west coast it'll happen inside a couple generations unless something changes. Seems bad.
Another day, another insane judicial ruling, and a good response from Premier Ford. An Ontario court has ruled that the Region of Waterloo cannot clear a 30-person tent encampment from a parking lot it owns to build a major transit hub.
Buried in the decision is that the court declared homelessness an analogous ground under s.15 of the Charter. A "constructively immutable characteristic." A "discrete and insular minority." This argument has been tried before and failed, including in this very case, three years ago.
If this holds, every municipal bylaw that differentially affects homeless people faces Charter equality scrutiny. The court went further, ruling the region cannot use its own land unless it first provides an alternative legal encampment or "tenting protocol" with equivalent services.
Elected officials passed a bylaw, amended it, dropped fines and offered housing plans. None of it mattered. A single judge overrode all of it and made himself the region's chief housing policy-maker. The Charter has become not a shield against state overreach but a sword by which courts dictate municipal governance on questions that belong to elected governments.
this is what i mean when i say european cities are rubber rooms. turbo nerfed theme parks. but in America, simply walking down the street, any manner of looney tunes fate could befall you. swept away by rapids, or a piano could fall on you, your head emerging with keys for teeth
@usr_bin_roygbiv Ah, I see. I reckon I can get it closer with more scaffolding. I'm trying to swap it out for a subagent evidence gathering workflow in our research app, not necessarily for harness util stuff.
@captgouda24 The prostitute comment is quite apt. God made men differ in strength, but Sam Colt made them equal. Likewise, God made men differ in intelligence, but Sam Altman made them equal. We’re now competing on beauty. The concern is not if scaling has hit a wall. It’s if you have.