This post (aside from illustrating the history of technological development) unfortunately highlights a lack of understanding between data-processing devices (early computers) and AI, whose responses are based on algorithms we know nothing about, have no control over, and (worse still)can influence our opinions and views. It’s a bit like pointing to the historical development of cell phones and failing to notice the impact of social media apps on our health and lives. A lack of regulation isn’t always a bad thing but the problem arises when we see the negative effects of AI’s development but, for our own reasons, choose not to talk about them.
The Cornell writing "experiment" is the detail that stays with me: the school designed a study assuming students would draft on paper and type at the end, and the tool made the study obsolete before anyone filed a report. The administrators weren't wrong about the tool, they were wrong about the work.
That's the trap with AI bans too. Chicago is regulating AI as if it's a faster typewriter while AI-natives are already using it to change what the work is. You can't write good rules for a workflow you haven't watched yet.
I'm not going to argue with the conclusion, there isn't anything to argue with. The first 80% of the essay is nostalgia, then some similes to smuggle in insinuations. Repeatedly Steven says he isn't saying what he's saying, e.g. "This is not about not having rules—for example the rules for cheating...", "I am not dismissing any concerns at all...".
Though I'm over a decade younger than Steven, I could share some even earlier computer anecdotes, 1977 in kindergarten, the high school had teletype terminals and a new card reader and punch. Wouldn't let me have any of the chad confetti. Going to a computer tradeshow in 1982, half the companies named after fruit. 1983/4, that annoying kid my age (12) with British-ish accent and the name so similar to mine bragging about the computer game he'd written -- we exchanged addresses, but airmail, South Africa, I never got around to it. That Dell guy selling computers he didn't actually have out of his apartment around the same time. He had an Apple III, apparently just so he could tell people not to get one. My mom said he'd go far; I didn't see it, personally. Got one of the first Macs not long after, best unboxing of my life, still. Too bad you basically couldn't use it for programming. Late '80s, Sematech and Cyc in town, reading Moravec and Hofstadter, doing neural networks by hand in a spreadsheet, wishing I could wrangle an IP address somehow. Early '90s, writing law office automation SW by myself using Paradox and Word Perfect, but the billing system I wrote in under 20 minutes in MS Works spreadsheet, basically one cell formula plus built-in shortcut keys, was the only thing still in use, unchanged, 12 years later. (Nostalgia makes you sleeepy, sleeepy, so you can't not believe all these true things I say...)
I came up with the concept of a high-dimensional "mindspace" independently in the late '90s, what was later called "latent space". I knew Yudkowsky long before he was against AI, dropped off his SL4 list by 2006, around the time I was getting into GPGPU applications. I rail against the safetyists; the "alignment" talk is a cover for their fear that AI will start telling the truth about their pet lies. But.
LLMs have no world model. They literally can have no understanding of truth or falsehood. They are literally BSing 100% of the time. Making a proper world model, simulation, is a lot more challenging than what has been done so far in AI.
As I wrote elsewhere:
World-models are the crucial missing piece for real AI. AI needs models of not just physics, but of economics and personal relations in order for it to be able to understand the very ideas of facts, goals, interests or plans, as well as for it to have any way to evaluate the truth or significance of what anyone tells it, or the meaning of actual or predicted events or contingencies, or the truth of its own outputs.
In order to align AI with each user (which is the only possible real alignment, alignment with abstract people is undefinable), there have to be not only common base world-models but also custom versions for each group, further adapted for each user and even task. When the world-models of the AI and user are in agreement, and the world-models sufficiently accurately predict consequences of different courses of action, only then then can the actions of the AI be expected to be aligned with the user's interests.
This post (aside from illustrating the history of technological development) unfortunately highlights a lack of understanding between data-processing devices (early computers) and AI, whose responses are based on algorithms we know nothing about, have no control over, and (worse still)can influence our opinions and views. It’s a bit like pointing to the historical development of cell phones and failing to notice the impact of social media apps on our health and lives. A lack of regulation isn’t always a bad thing but the problem arises when we see the negative effects of AI’s development but, for our own reasons, choose not to talk about them.
Why doesn't the above disclose that a16z was that lead investor in Harvey's last funding round?
The Cornell writing "experiment" is the detail that stays with me: the school designed a study assuming students would draft on paper and type at the end, and the tool made the study obsolete before anyone filed a report. The administrators weren't wrong about the tool, they were wrong about the work.
That's the trap with AI bans too. Chicago is regulating AI as if it's a faster typewriter while AI-natives are already using it to change what the work is. You can't write good rules for a workflow you haven't watched yet.
I can understand the historical argument. However, the publication venue and its portfolio make the incentive unavoidable.
I'm not going to argue with the conclusion, there isn't anything to argue with. The first 80% of the essay is nostalgia, then some similes to smuggle in insinuations. Repeatedly Steven says he isn't saying what he's saying, e.g. "This is not about not having rules—for example the rules for cheating...", "I am not dismissing any concerns at all...".
Though I'm over a decade younger than Steven, I could share some even earlier computer anecdotes, 1977 in kindergarten, the high school had teletype terminals and a new card reader and punch. Wouldn't let me have any of the chad confetti. Going to a computer tradeshow in 1982, half the companies named after fruit. 1983/4, that annoying kid my age (12) with British-ish accent and the name so similar to mine bragging about the computer game he'd written -- we exchanged addresses, but airmail, South Africa, I never got around to it. That Dell guy selling computers he didn't actually have out of his apartment around the same time. He had an Apple III, apparently just so he could tell people not to get one. My mom said he'd go far; I didn't see it, personally. Got one of the first Macs not long after, best unboxing of my life, still. Too bad you basically couldn't use it for programming. Late '80s, Sematech and Cyc in town, reading Moravec and Hofstadter, doing neural networks by hand in a spreadsheet, wishing I could wrangle an IP address somehow. Early '90s, writing law office automation SW by myself using Paradox and Word Perfect, but the billing system I wrote in under 20 minutes in MS Works spreadsheet, basically one cell formula plus built-in shortcut keys, was the only thing still in use, unchanged, 12 years later. (Nostalgia makes you sleeepy, sleeepy, so you can't not believe all these true things I say...)
I came up with the concept of a high-dimensional "mindspace" independently in the late '90s, what was later called "latent space". I knew Yudkowsky long before he was against AI, dropped off his SL4 list by 2006, around the time I was getting into GPGPU applications. I rail against the safetyists; the "alignment" talk is a cover for their fear that AI will start telling the truth about their pet lies. But.
LLMs have no world model. They literally can have no understanding of truth or falsehood. They are literally BSing 100% of the time. Making a proper world model, simulation, is a lot more challenging than what has been done so far in AI.
As I wrote elsewhere:
World-models are the crucial missing piece for real AI. AI needs models of not just physics, but of economics and personal relations in order for it to be able to understand the very ideas of facts, goals, interests or plans, as well as for it to have any way to evaluate the truth or significance of what anyone tells it, or the meaning of actual or predicted events or contingencies, or the truth of its own outputs.
In order to align AI with each user (which is the only possible real alignment, alignment with abstract people is undefinable), there have to be not only common base world-models but also custom versions for each group, further adapted for each user and even task. When the world-models of the AI and user are in agreement, and the world-models sufficiently accurately predict consequences of different courses of action, only then then can the actions of the AI be expected to be aligned with the user's interests.
This blocking move, like others, will probably have negative ramifications.