It's strange to me that the author doesn't mention the primary cause of all homelessness, which is housing insecurity. By the numbers in the article, something like 99% of people with "severe mental illness" are not homeless, suggesting there are workable options for people with mental illness that don't require mass institutionalization.
I'm not sure if it's an objective of the "Manhattan Institute" to spin up these institutions for indefinite detention, but I will point out that providing housing universally is definitionally cheaper than institutionalization, which is housing with paid orderlies, guards, service workers, etc. If providing housing is too big-government, eviction protections were seen to curb homelessness during the pandemic. And for these right-wing think-tankers, housing affordability also has an available "market" solution which just requires the loosening of zoning restrictions with little other public investment.
We can't develop a universally coherent data set because what we understand as "truth" is so intensely contextual that we can't hope to cover the amount of context needed to make the things work how we want, not to mention the numerous social situations where writing factual statements would be awkward or disastrous.
Here are a few examples of statements that are not "factual" in the sense of being derivable from a universally coherent data set, and that nevertheless we would expect a useful intelligence to be able to generate:
"There is a region called Hobbiton where someone named Frodo Baggins lives."
"We'd like to announce that Mr. Ousted is transitioning from his role as CEO to an advisory position while he looks for a new challenge. We are grateful to Mr. Ousted for his contributions and will be sad to see him go."
Is there a streaming-native way of encoding multiple camera angles simultaneously in a video? The ability to toggle between different angles is a feature on DVDs and Blu-rays that IMO is underused but awesome for educational videos where different aspects or angles of a demonstration might be useful.
This is a great example of the type of question-begging and junk science "deep learning models" enable. The methodology is to take time series MRI of hundreds of regions of the brain and, labeling each data point as "male" or "female", train a model to differentiate between the two, then pick out the most determinative bits. The issue, of course, is that you are searching for what you want to find. With so much data, of course there is going to be some feature that has maximal differentiating power, so it's totally unsurprising to see a result showing maximal differentiation, since that was what the model was being trained for.
However, nothing in this points towards any intelligent conclusion… the highest differentiating region is called "R DLPFC, Subdivision A9l", with "R DLPFC, Subdivision A9/46d" coming close behind. No doubt though this will be passed around as evidence that gender disparities in professional fields and economic outcomes are biologically determined.
I think it's difficult to conceptualize a program's behavior as "buggy" when the program has no specified behavior. The vexing thing about LLMs and image generators is that they ultimately have no intended purpose or principled construction as artifacts, being mostly discovered through trial-and-error and pressed into service due to various misconceptions about what they are — human imitators, fact retrievers, search engines, etc. But they are really just statistical regurgitators over whatever the training dataset is, and while that type of an artifact seems to sometimes prove useful, it's not something we have yet developed any kind of principled theory around.
One example is in DALL-E initially going viral due to its generated image of an astronaut riding a unicorn. Is this a "bug" because unicorns don't exist and astronauts don't ride them? One user wants facts and another wants fancy. The decision about what results are useful for which cases is still highly social and situational, so AIs should never be put in a fire-and-forget scenario or we will see the biases Dan discusses. AIs are more properly used as sources of statistical potential that can be reviewed and discarded if they aren't useful. This isn't to say that the training sets are not biased, or that work shouldn't be done to rectify that distribution in the interest of a better society. But a lot of the problem is the idea that we can or should trust the result of an LLM or image generator as some source of truth.
It's odd to use dock workers as a contrast, since their heavy unionization has led to very high pay rates (which, for the record, I think are deserved):
"Full-time dock foremen earned 24% more than the average CEO’s base salary in 2022 and 20% more than neurologists. ILWU clerks earned just $5,600 per year less than airline pilots. Full-time dockworkers came in 21% higher than lawyers and 9% above dentists.
Average earnings in 2022 for all full-time ILWU registrants — $211,000 — were 3.4 times higher than the average wage for all professions calculated by the Bureau of Labor Statistics."[1]
Meanwhile the "Google employees" referenced in this story are contractors who are paid ~$19/hr, or ~$40k/year at full time pay.
This may or may not be real, but there has certainly been a lot of discussion about results that are similar to this from real models. My sense though is that nobody really has a solid way to fix these kinds of issues. You can basically just train with different goals, or regex out certain responses, but otherwise it seems like there's no agreed-upon method that gives people what they want here while also letting them train with business goals like "safety". Is that incorrect? Is there some kind of trick that people use to make everything respond "correctly" or are older models just more unobjectionable because they've had a longer time to manually smooth over the bad responses?
What's amazing to me is that this is the first time I've seen the frame problem explicitly mentioned in online discussion of LLMs, despite it being an incredibly relevant and historically difficult problem in the exact field of AI. Nobody is directly grappling with it or seems to be familiar with the history of the field. The problems the frame problem causes are currently being attacked with efforts to radically expand the context window, but even infinite context doesn't answer the question of what is relevant to an action and what is not.
The 2nd episode of Stewart Brand’s BBC series “How Buildings Learn” is called “The Low Road” and also addresses the use of inexpensive or undesirable space.
Ultimately part of making room for innovation means allowing people to take risks, and that mechanism has in recent years shifted from wide availability of low-cost resources to wide availability of speculative capital to obtain high-cost resources.
What isn't really discussed here (and I wish was!) is the corollary to the price scaling of successive generations: the outcome isn't proportionate to the increased cost. GPT-4 cost thirty times as much as GPT-3 to train, but it isn't thirty times better. It's not even twice as good! I suspect the majority of people using an OpenAI product are still using GPT-3.5 in the free tier and are mostly satisfied with it, because the basic proposition — plausibly fluent English text — is still there, and GPT-4's improvements aren't needed for most of the use cases an LLM satisfies.
Following this out, you have the prospect of this company burning multiple trillion dollars to create a foundation model that is slower than GPT-4 and possibly only 5-10% better. The only way this makes any sense financially is if you think that extra 5-10% takes us to some kind of AGI that solves all of the reasons GPT-4 hasn't put everyone out of a job. I suspect OpenAI doesn't think this, but maybe "training another model" is easier to pitch than whatever the process would be to find product-market fit for their current model.
If the search terms are "competitive" what's the moat that keeps your customers' competitors from leveraging the same foundation models until that term is flooded with marginally-optimized copy? High-quality copy used to be a signal that the retailer or brand had enough success and therefore capital to afford to invest in the copy. Now anyone can do it, even retailers and brands that are terrible or maybe even fake. So what's the value of Google search if all it's surfacing is the highest-quality AI authoring tool? Maybe you can play the spread between now and wide adoption but it seems like eventually you'll be shooting yourself in the foot.
This is almost the opposite of what I envisioned when reading the title "folk computer", which makes me think of a computing experience that's fundamentally simple, convivial, and transparent to users, something like what's captured by the concept of "permacomputing". While the ideas explored here are no doubt interesting and probably important to consider, they seem like they require an enormous amount of abstraction and complexity to implement, while at the same time remaining somewhat impenetrable to the user. For example, using QR codes guarantees that the semantics of any given symbol are impossible to determine at a glance. Is it really an improvement over the mouse-and-screen paradigm to paste QR codes to your hands (the printed hands in the demo are actually two left hands), and set up a rig that requires camera mounts, projectors, and a large flat surface dedicated to object manipulation with extremely specific, non-intuitive semantics? I can't help but wonder how this would ever generalize to the number of uses or amount of convenience that people were able to wring out of even simple text-based terminals.
Though it seems like there is some sensitivity around the comparison of LLMs to Markov chains, and certainly an LLM is not generated in the same way, it is pretty accurate that an LLM could be represented by a sufficiently (ie very) complex Markov chain. The states of the chain would not be the tokens themselves, as in this example, but the total context window vector of N input tokens, which would fan out to states consisting of the new context N[1:] + A, where A is any of the tokens sampled from the resulting probability distribution, and the transition probabilities are just drawn from that same distribution according to the temperature settings.
You could even do some very hand-wavy math on how staggeringly complex the resulting Markov chain would get: BERT for example has a token vocabulary of 30,000 and a context window of 512 tokens. So the number of possible states would be 30,000^512, or ~1.9 x 10^2292, with each of those having a max fan out to 30,000 other states. So clearly the LLM is a more compact representation of the concept.
One way of answering this question would be to consider why, for example, Native American children were punished at boarding schools for speaking their language instead of English, why US colonial rule of Hawai'i included bans on Hawai'ian language usage, why Britain published missives against the Welsh language, or why nationalists in Taiwan or communists in China have worked to suppress languages like Taiwanese or Cantonese in favor of Mandarin. Languages are often considered delimiting factors for whether a group should be considered autonomous, a marker of shared ethnic heritage, and as a bulwark against cultural assimilation. Because they are part of a community's history, the eradication or revitalization of a language is often linked to the eradication or revitalization of that group of people as a cultural and political entity, as in Welsh revitalization efforts or protests to retain Cantonese broadcasts. So there is a connection to whether a group can be considered as a distinct culture, with its own distinct political rights, that goes beyond the utilitarian aspect of language as a means of simply transmitting information.
> All those Erewhon, whole foods and trader joes will turn into boarded window crack houses and garrisons for gang warfare.
This seems a little hysterical. Cities functioned extremely well, perhaps even better, before the advent of lengthy car commutes, office work, and mass commercial real estate. The businesses you mention - grocery stores - aren't even usually patronized by commuters, and their recent appearance in city centers has been precipitated by downtowns shifting from commercial-only blight to round-the-clock mixed use, something that will only accelerate as downtowns are liberated from massive buildings that sit empty 2/3 of the day, small businesses that have to shut down at 6pm and all weekend, and peaky congestion that requires all facilities to overinvest in capacity to handle a couple of hours on weekdays with much lower usage otherwise. Anyone who has walked around the ghost town of SF's financial district on a Sunday should know CRE is not conducive to vibrant city life.
"I have a worry about X that I can't justify in any way" isn't a good basis for forcing people into a decision that (as the GP mentions) has dozens of obvious, immediate, and measurable benefits in the other direction. The simplest explanation for why "leaders" are fighting for RTO is not that office work was, improbably, a perfect global maximum for productivity, innovation, esprit de corps, or whatever, but merely that office work is how things used to be done and a lot of people got used to it both organizationally and economically. From that perspective, I think the burden of proof is completely on those who are fighting for a return to status quo to argue why workers (and society at large!) should be willing to trade all of the obvious benefits of remote work for these supposed intangibles.
This law is true in some cases, but I think is too often used as the tech industry's version of "first they ignore you, then they laugh at you, then they fight you, then you win". The argument is that, sure, a technology's promise may have not panned out, but under Amara's law, we can expect a commensurately massive impact later.
As Carl Sagan pointed out, "…the fact that some geniuses were laughed at does not imply that all who are laughed at are geniuses." The same applies to technologies. The fact that a technology has shown slower-than-expected progress is not a foolproof indication that it will eventually have massive impact. It may just not be as useful as it initially seemed, or may even be supplanted by a completely different technology.
I’m perplexed by this article’s argument, which seems to be that Graeber is somehow betraying “anarchism” by theorizing the state as ill-defined or non-monolithic, or questioning the concept of “inequality”. All of these arguments seem totally compatible with anarchism, especially the type of modernism-questioning anarchism that would stem from anthropology. To me it feels like the author is coming from an overly rigid or subcultural definition of the term that takes any straying outside its bounds as indication of closet liberal behavior or a radical shift in thought. While it’s possible that his late works were cryptically hinting at a new angle, the inconsistencies in his writing and behavior are also adequately explained by Graeber being merely human and not the ideal paragon of anarchist ideology.
It feels like journalistic malpractice for this article to assign some kind of personhood to "AI George Carlin" and talk about this entity "dropping" a comedy special or "throwing shade". This is the production of another person, without Carlin's consent or his family's, using his voice and attempting to imitate him. It reminds me of an older article that claimed an "AI" was tricking people into solving captchas. What was actually happening was that a researcher was tricking people into solving captchas using an AI. When journalists credulously give agency to AIs they are papering over the harms being caused by actual people who have the decision whether to do something morally suspect or not.
I agree with your perspective overall but I find when something like this is posted the responses are usually along the lines of "how do we know that statistical prediction on a large enough scale is not akin to understanding?" Which to me is a somewhat lazy retort but highlights that one huge issue in this field is not really having definitions or an understanding of our own about what we are debating, and that we can't rely on saying "X is not Y" to dispel hype because at our current level of rigor around definitions, someone can just as easily say "nuh uh, X is Y".
For me the biggest indicator that something is still missing, and a reason not mentioned here, is the lack of any killer app besides general chatbots and code completion, even more than a year out from their explosion onto the scene (at which point both of those use cases already existed). It was assumed that the LLMs would quickly change everything but it appears their seemingly-minor limitations are pretty fundamental for many use cases.
I was interested to see a paper from a philosophical perspective but it seems this one fails to grapple with some of the more foundational critiques, primarily around reference where it claims that externalism militates for a view where reference is purely situated. I read the cited paper (Mandelkern & Linzen) and the arguments seem very weak to me that an intelligent agent "refers" in the absence of any grounding whatsoever beyond relative semantics between text. What it seems to argue is that because we sometimes are able to refer without absolute grounding (due to lack of direct experience of the referent), then reference overall does not require the referrer to have any world-grounding whatsoever. To me that's a leap. I would like to see more working against phenomenology here, especially Brian Cantwell Smith's so-far-unanswered critique of machine learning as intelligence.
One of their other arguments, which is under the umbrella of the "re-description fallacy" hinges on the observation that describing (e.g.) a piano as "hammers hitting strings" doesn't preclude it from more complex behavior like harmony, and so a simple description of LLMs as autocomplete etc doesn't preclude more advanced understanding. While it's true that any complex process can be inappropriately dismissed by simplifying its description, I think this too-neatly sidesteps more complex critiques which advance from simple descriptions of LLM behavior as contrasted with complex behavior from sentient intelligence that cannot be implemented by those simple mechanisms.
It’s interesting to me that these criticisms come mostly from the right when they’re essentially the result of a corporate attitude towards universities, where rather than being so heavily subsidized that they are essentially free for the populace, and sheltered from popular opinion, universities are constantly competing for student attention, investment resources, prestige, and cooperation from industry. As a result the complained-about behaviors are essentially things we see in competitive industry — insincere lip service to diversity as advertisement, cutthroat behavior and outright cheating to attain status, self-congratulation for unremarkable or fake accomplishments, fiefdom-building, and general focus on profits above any semblance of a mission. But an effort to restore an academic environment would probably be rejected as leaving academics unaccountable to the market as they instruct the next generation.
It's kind of fascinating that we've taken the lowest-bandwidth way of proving agency and effort - writing - and made it tremendously more labor-intensive to deduce whether any agency or effort were involved at all. I think the knock-on effects are probably huge - not only are the reporter and the maintainer both wasting time here that could be better spent on something useful, but the whole bug-bounty-crowdsourced-CVE process is now compromised by a lower signal-to-noise ratio and the probable outcome of raising barriers to submission as a means of dealing with the spam, meaning fewer bugs discovered and fixed, more security vulnerabilities, and everything that results from that.
The same dynamic is at play in other sectors, meaning we find ourselves increasingly unable to assign any trust to product reviews, court filings, recipes, how-to guides, medical advice, and so forth. One of the promises of the internet was the rapid expansion of content through democratization of publishing, but I think we're witnessing the gutting of any benefit that had left.
I wonder if we'll eventually realize, similar to Solow's productivity paradox, that whatever efficiency "gains" we get from AI are just cancelled out by the increased need for fact-checking, some increased incidence of major errors being blindly trusted, and suboptimal outcomes (e.g. being bamboozled by good copy or fake reviews into paying for an inferior product). All this in addition to the opportunity cost of the brainpower and energy currently being poured into multiple largely-comparable models.
The article is sourced from The Information, which says "Google plans to reorganize a big part of its 30,000-person ad sales unit, an executive told some staff last week, prompting anxiety that some departments will face job cuts."
That doesn't sound like laying off 30,000 people. That sounds like a exec said there would be a reorganization of a "big part" of a unit of 30,000 people total, and people are worried it will include some layoffs, which is not the same thing.
Maybe I wasn't clear, but my point of view is that efficiency is — rightly — not an important feature of farmers markets (at least to me or people I know), so I'm put off by this paper that is trying to somehow optimize them using POS data. The argument is that this way of looking at farmers markets seems to miss why people patronize them, which is closer to the points you bring up and much less about efficiency. Though as others point out there are probably cases where the FM is more or less efficient than a supermarket, or more well-stocked, my point still stands that the paper feels like it's aiming at something most people aren't asking for.
What is it about text generation that makes us so credulous about its relationship to reality in a way that we aren't with other generative tech? I assume that a lawyer would implicitly understand that they can't, for example, ask MidJourney to generate a photo of a certain crime's perpetrator and then submit that as evidence in court. The phenomena of what Bard or GPT-4 does with text is almost exactly analogous to what MidJourney does with pixels insofar as its relation to reality, and yet when it comes to text, even the creators of a generated work don't understand it's not real.
I never really thought of farmers markets as a location for efficiency in the economy. Normally you are shutting down a street, hauling product in small trucks directly from an upstate farm, and having 1-2 staffers per product type make direct customer sales in a non-refrigerated tent. It's pretty objectively a waste of energy, money, and time if all you are looking for is sustenance. But I think people with the means go to farmers markets to make the utterly joyless but necessary trip to the grocery into a kind of fun and interesting simulacrum of an old-time open-air market, even at increased cost or less selection. While of course the vendors are trying to make money and probably want to make more of it, I can't help but intuitively recoil at the attempt to calculate this out. In some way it seems tasteless or missing the point, and if every vendor were to standardize on some product mix or strategy for sales, I think it would reduce the eclectic appeal of the market to most of its regular customers.