25 ms·
Ezra Klein at the New York Times has been running a series of excellent podcast episodes and columns on the topics of both generative AI[1]. He and his guests
by dredmorbius 4y ago
Ezra Klein at the New York Times has been running a series of excellent podcast episodes and columns on the topics of both generative AI[1]. He and his guests raise significant points and are some of the most sane, sober, and insightful commentary I've heard. It's well worth reflection and consideration.
There's the very poorly titled column "This Changes Everything" ("this" is "generative AI"): <https://www.nytimes.com/2023/03/12/opinion/chatbots-artificial-intelligence-future-weirdness.html https://www.nytimes.com/2023/03/12/opinion/chatbots-artifici...>
And two podcast episodes in particular:
- "The Imminent Danger of A.I. Is One We're Not Talking About": <https://www.nytimes.com/2023/02/26/opinion/microsoft-bing-sydney-artificial-intelligence.html https://www.nytimes.com/2023/02/26/opinion/microsoft-bing-sy...> TL;DR: "Who will these machines serve?"
- "A.I. Is About to Get Much Weirder. Here’s What to Watch For." <https://www.nytimes.com/2023/03/21/opinion/ezra-klein-podcast-kelsey-piper.html https://www.nytimes.com/2023/03/21/opinion/ezra-klein-podcas...>
- And on a different but closely-related theme: "How the $500 Billion Attention Industry Really Works" <https://www.nytimes.com/2023/02/14/opinion/ezra-klein-podcast-tim-hwang.html https://www.nytimes.com/2023/02/14/opinion/ezra-klein-podcas...>
The links include both the audio and transcripts (following a few days after air-date) for podcasts.
I expect Klein to cover both aspects throughout the next year.
What I especially like about Klein is that he's not only reacting to developments and rehashing demonstrated capabilities, but he's asking questions and anticipating what's to come, without the hagiographic / techno-optimistic lenses of some (e.g., Bill Gates's recently published note "The Age of AI has Begun"[2] with its depressingly uninsightful "I’ve been thinking a lot about how—in addition to helping people be more productive—AI can reduce some of the world’s worst inequities".
What we've learnt about technology is that what it does depends greatly on who it serves. And amongst Klein's more interesting observations is that we cannot with certainty know what aims AI is serving, even those who create it. Klein repeatedly notes that many of those directly engaged in creating the technology itself have little idea where it is headed or what it will be able to do:
Since moving to the Bay Area in 2018, I have tried to spend time regularly with the people working on A.I. I don’t know that I can convey just how weird that culture is. And I don’t mean that dismissively; I mean it descriptively. It is a community that is living with an altered sense of time and consequence. They are creating a power that they do not understand at a pace they often cannot believe.
<https://www.nytimes.com/2023/03/12/opinion/chatbots-artificial-intelligence-future-weirdness.html https://www.nytimes.com/2023/03/12/opinion/chatbots-artifici...>
I've spent much of the past decade looking at the history of technology and information technology in particular. There's a pair of books which stand out to me, they share a title though are separated by a decade and different authors: The Control Revolution, respectively by James Beniger (1986)[3] and Andrew L. Shapiro (1999)[4].
Beniger's book looks backwards at the development of largely commercial and corporate communications over the course of the Industrial Revolution (18th through 20th centuries), whilst Shapiro looks forward at the promise of a networked and digital online communications infrastructure. Both books have aged well, though some informed reading-between-the-lines may be necessary.
In particular, Beniger looks at business as an information processing system, and not in the all-to-familiar (and rather facile) Hayekian market sense, but in terms of information flows within and between companies. In particular, as information technologies developed how communications occurred transformed immensely. Many of those changes to me seem to revolve around issues of trust.
The ornate and florid language of 18th and early 19th century correspondence spends much time and space in asserting trust and faithfulness bonds between correpsondents. (How accurate or useful those were is its own question, but the point remains: it's a major component of the writing.) Keep in mind that it might take days, weeks, or months for correspondence to reach its intended recipient (let alone unintended ones), and that remote offices or agents might be acting with tremendous autonomy for months or years at a time.
With the development of the telegraph, two things occurred:
- Communications became instantaneous, with multiple round-trip messages within the course of a day or an hour possible.
- Words got expensive.
Language became telegraphic.
American author Mark Twain exemplified much of this, and his style of writing was as distinct for its directness as the topics it covered. The influences of a newspaper pressman and editor working from telegraphed wire stories and a sense of the physicality of a block of cast type is clear to me.
The rise of complex corporations also played a huge role: railroads, manufacturing concerns (particularly General Electric), chemical companies where deviation from procedure could have explosive consequences (Du Pont, Dow), and communications companies (Western Union, AT&T, remember that that second 'T' is for "Telegraph"), and the like. It's possible to trace RFC 822 (and successor) email headers directly to business correspondence memo fields, used to standardise correspondence, from the late 19th century.
While the 19th century saw generally a decrease in in-band attestations of trustworthiness as message capacity increased, I'm strongly suspecting that the 21st century may see an increase in such attestations. One possible possibility is through cryptographic mechanisms, the favourite of technologists (myself included), though adoption of such methods has to date been pathetically and disappointingly weak.
Another is that multiple independent verifications of information will be required and increasingly common. This is already used in fields such as journalism and human-rights investigations.[5] A problem emerges when it cannot be readily determined that two sources are in fact independent.
For business increasingly dependent on remote interactions, the risks of impersonation and fraud (is that really the CEO calling on a scratchy phone line, or an AI bot?) is a huge and growing problem, along with invoice and billing frauds and the like.
How we're going to address this, and how the notions of "something you are, something you know, and something you have" as multiple forms of remote attestation will evolve ... is going to be an interesting set of questions.
________________________________
Notes:
1. "Generative AI" seems to be the general term describing collectively ChatGPT, Sydney, Bard, and other current-generation large language model (LLM) AI chatbots.
2. Note: <https://www.gatesnotes.com/The-Age-of-AI-Has-Begun https://www.gatesnotes.com/The-Age-of-AI-Has-Begun> HN discussion: <https://news.ycombinator.com/item?id=35250564 https://news.ycombinator.com/item?id=35250564>
3. <https://archive.org/details/controlrevolutio0000unse https://archive.org/details/controlrevolutio0000unse>
4. <https://archive.org/details/controlrevolutio0000shap https://archive.org/details/controlrevolutio0000shap>
5. See for example: <https://www.humanrightscareers.com/skills/beginners-guide-how-to-investigate-human-rights-violations/ https://www.humanrightscareers.com/skills/beginners-guide-ho...>
- e1g 4y ago> What we've learnt about technology is that what it does depends greatly on who it serves. You made many good points, but if this one resonates with anyone in the context of AI advancements, I recommend this long-form essay, "Meditations On Moloch" [1]. It's a thought-provoking exploration of the unrelenting drive for competition and self-interest, often leading to suboptimal outcomes for individuals and society in politics, education, economics, and the environment. [1] https://slatestarcodex.com/2014/07/30/meditations-on-moloch/ https://slatestarcodex.com/2014/07/30/meditations-on-moloch/ or the audio version at https://sscpodcast.libsyn.com/meditations-on-moloch https://sscpodcast.libsyn.com/meditations-on-moloch
- dredmorbius 4y agoCare to summarise what you think the core applicable message here of SSC's Moloch is? If it's that behaviours are emergent ... yes, I get that. (I've skimmed bits of the piece numerous times. Each time I attempt to dig in, the length puts me off. Formatted for print, the main text runs > 50 pages, and with comments to nearly 300.)
- e1g 4y agoI drew parallels here because Ezra Klein reached the same conclusions as SSC did at the end of those ~50 pages a decade earlier: when building a god machine, we must slow down forcefully if necessary. It's practically impossible, but the stakes are too high. Both Klein and SSC discuss four inescapable problems stemming from such complex systems: 1. Uncontrolled arms race: The competitive development of AI technologies is already an arms race where BigTech and VCs rush to create increasingly powerful AI systems without taking adequate safety precautions. 2. Misaligned Goals: systems are designed to optimize specific objectives, but if those objectives are not perfectly aligned with human values, the system will continue to take actions that are harmful to humans (the "alignment problem" in AI safety). 3. Economic Disruption: The widespread deployment will lead to significant economic and social disruption, including job displacement and inequality. Those people will not be getting better jobs. 4. Loss of Control: Highly autonomous systems could become problematic for humans to understand or control, especially if they are capable of self-improvement. The central thesis of Moloch is that in a competitive environment, or any environment that requires tradeoffs or optimization, the inevitable terminal state of the system is that everything that is not _the objective_ will be sacrificed to pursue _the objective_. When the desired objective is tightly coupled to human values, this is good because society will sacrifice unimportant things to pursue those important things. But those objectives can get decoupled from our values ("when a measure becomes a target..."), and that unavoidably creates a "race to the bottom," where individuals or groups are incentivized to engage in behavior that is ultimately harmful to themselves and others to gain a competitive advantage towards a Pyrrhic victory. SSC uses the metaphor of "Moloch" to describe this phenomenon, suggesting that Moloch represents the destructive force that arises from the uncontrolled competition and the pursuit of self-interest. edit: It took me two years to consume Meditations On Moloch. I blame the opening poem as it was incomprehensible and barely relevant to the rest of the piece. Ultimately, I found the audio version more approachable - just like listening to a lecture on YouTube. After that, I returned to the text version to reflect on specific chapters and follow up on references.