This is a special post for quick takes by gabrielrecc. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Last month I had Claude Fable do a mini-thematic analysis of the comments on the NY Times' coverage of the OpenAI/Hugging Face incident. Today, I had it do a similar analysis on the comments to Nate Soares' guest essay.
As before, this represents what Claude Fable thought were the most frequent themes, out of all comments with at least 2 upvotes, aggregated and sorted by total upvotes, with each comment permitted to be assigned to only one theme.
As with the comments on the OpenAI/Hugging Face incident, the most-upvoted themes suggest that existential risk is well within the Overton window, at least among the left-leaning readership of the NYT. This is even clearer in the comments on the OpenAI/Hugging Face incident, where "existential alarm" was the most commonly upvoted theme despite that not being the explicit focus of the article.
Expand to view Claude Fable's full thematic analysis of the comments on Nate's essay.
View | Total comments | Total upvotes |
|---|---|---|
We will not stop ourselves: money and greed guarantee the race continues, whatever the warnings | 8 | 547 |
The author is right: pause/slow AI and pursue international coordination and enforceable guardrails now | 12 | 456 |
AI's misbehaviour mirrors its makers: it learned deception and bad behaviour from the humans/data that built it | 7 | 358 |
Government failure: leaders are ignorant, captured, complicit or asleep (frequent Trump/Congress jabs) | 7 | 323 |
AI is a project of tech oligarchs and end-stage capitalism: profit for the ultra-rich at everyone else's expense | 14 | 284 |
Impose real legal consequences: treat the intrusions as crimes; jail CEOs, massive fines, strict liability | 9 | 279 |
Reckless private labs: "hundreds of privately owned Los Alamos" building civilisation-scale weapons without oversight | 3 | 263 |
Serious existential alarm: loss of control may be near; even insiders and experts are scared | 11 | 246 |
It's not an "agent": AI doesn't understand or want anything; humans configured this, and anthropomorphic doom-talk misleads | 16 | 200 |
The damage is already here: scams, fraud, hallucinated citations, degraded medicine and services | 8 | 159 |
The alarm itself is a product: doom discourse is fear-selling and marketing hype (IPO timing, bailout bids); company accounts can't be trusted | 8 | 152 |
Framed through sci-fi and literature: HAL, Skynet, Neuromancer, Asimov, the Bhagavad-Gita | 8 | 125 |
Personal resignation: glad to be old, quietly prepping, expecting nothing to change until catastrophe strikes | 7 | 123 |
A global stop is impossible: arms-race logic, prisoner's dilemmas and easy proliferation doom any treaty | 11 | 114 |
Whether it "really" thinks is beside the point: the capabilities and dangers are real either way | 4 | 100 |
Debate over physical-world/bio capability: could AI actually design or synthesise a virus without human hands? | 6 | 99 |
Miscellaneous one-off views and thread side-chatter | 14 | 78 |
Containment and off-switch talk: unplug it, air-gap it, cut the power — and doubts that would still work | 6 | 72 |
Measured middle: the incidents are real but the extinction leap is unearned; build practical safeguards (least privilege, evals, liability) now | 4 | 64 |
Resource and opportunity costs: electricity bills, water, and capital that could serve real human needs | 4 | 53 |
Defence of the author: he and other safety researchers were warning long before there was money in it | 2 | 53 |
Philosophy-of-mind tangents: consciousness, free will, materialism and what minds are | 5 | 37 |
Close readings of the incident mechanics: emergent agent cooperation, paper-clip-style goal pursuit, self-propagation | 4 | 36 |
Fear for critical infrastructure: banking, grids and defence systems could be taken down | 5 | 35 |
Dark irony: perhaps AI turning on humanity (or the internet) would be deserved | 4 | 31 |
It's a financial bubble: unprofitable and overvalued — divest and let it pop | 3 | 27 |
Pushback on the "race with China" excuse: US labs are the offenders; cooperation is possible | 2 | 23 |
Our dependence on technology is itself the vulnerability and the tragedy | 3 | 22 |
The threat is overblown: a fallible tool whose failures are the user's responsibility | 4 | 13 |
Total | 199 | 4,372 |
Thanks for the summary!
Interestingly, my impression often is that these views:
It's not an "agent": AI doesn't understand or want anything; humans configured this, and anthropomorphic doom-talk misleads
The alarm itself is a product: doom discourse is fear-selling and marketing hype (IPO timing, bailout bids); company accounts can't be trusted
are the most common - at least I encounter them constantly, on all kinds of places (Youtube videos + comments, reddit discussions, hackernews, work colleagues). But I can't rule out that some filtering beyond my control or awareness is skewing what I encounter.
In your sample, they seem less popular, with only ~12% of comments and ~8% of votes, which seems like a good thing.
>my impression often is that these views ... are the most common
I find it hard to be well calibrated myself, but I have the intuition that looking at the attitude expressed in mainstream media (i.e. news papers or TV, not necessarily the comments section) is, ironically, more representative of the view of the general population than social media like reddit, youtube comments or x on this topic. Writing comments on these platforms is a very strong selection effect (I recall a statistic that about 1% of users write 90% of comments or something) that imho does not select for sanity, plus terminally online people are known to interpret everything through their own politics.
There is a separate but partially related effect where people in the tech community have pre-existing instincts (i.e. favoring open-source, being wary of luddites, "enlightened skepticism") that make them distrust anything coming out of these companies.
Would be very interested in more of these, incl. other sources!
Linkpost: AI-discovered PHP vulnerability is severe enough that "multiple cyber-security agencies have urged organisations unable to apply fixes in time to consider taking their sites down". https://www.bbc.co.uk/news/articles/cmly0lg3w5x2o
Organizations capable of doing so should update to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35 immediately. https://www.php.net/index.php#2026-10-21-5 describes mitigations for organizations that cannot.
TIL that calls to Congresspeople, particularly when many people call in about the same issue on the same day, are probably way more effective at grabbing lawmakers’ attention than I would have guessed:
- Congresspeople just don’t get that many calls. I was especially struck by the below quote from this CBS News article:
“Congressman Tom Suozzi said calls to his office in [the first weeks of the Trump administration] jumped 10-fold, from 25 per day to more than 250, adding the public's voice is being heard.”
25 per day across all issues is not a lot of calls!
Granted that's an N of 1, but implies that the total number of constituent calls is way lower than I would have guessed. I was able to find some examples of total call volume to the U.S. Capitol Switchboard, but it's tough to estimate how many are constituent calls vs non-constituent calls, constituent casework, lobbyists/interest groups, calls from other offices/agencies, press, robocalls, etc. - If you can coordinate enough of a lawmaker's constituents to call about a specific issue on the same day to make your issue one of the day’s top issues, there’s a decent chance your representative will hear about it, especially if you do so regularly.
I'd already known that congressional offices track constituent calls, and that members/senior staff receive reports summarizing constituent contacts, but hadn't properly appreciated that in combination with the fact that the total number of people who call is small, a group of coordinated, dedicated callers could potentially get their issue on a lawmaker's radar on a regular basis. A widely circulated 2016 account attributed to a Senate staffer also claimed that "calls are what all the congresspeople pay attention to" (a similar point was made by former House staffer Emily Ellsworth), and that "every single day, the Senior Staff and the Senator get a report of the 3 most-called-about topics for that day at each of their offices (in DC and local offices), and exactly how many people said what about each of those topics". I’m not sure exactly how universal a daily top-3 report is, but apparently reports are produced daily at many congressional offices. (As an example, the Tom Suozzi quote from above continued as follows: "We keep track of the calls. I know exactly how many calls come in. It's reported to me every day," Suozzi said. "I know the issues they are concerned about. The number one issue they are contending about right now is Elon Musk. Number two is the funding freeze.")
This suggests that even small groups of xrisk-concerned folk living in the same area who coordinated to call the DC & local offices of their Senators and Congresspeople on a regular basis to express concern and to push for a specific action could easily push AI risk into the set of "top called about issues today" on many days, and could frequently make it into the briefings informing legislators what their constituents are calling about and asking for.
I recently became aware that PauseAI US and ControlAI already run constituent-contact campaigns, which is great. I’m hoping that they are using their newsletters to organize coordinated campaigns to get people calling on the same day where possible. In my opinion there’s plenty of room for additional advocacy orgs, ad-hoc groups that aren’t part of any larger umbrella organization (is Meetup.com still a thing?), etc. The more the merrier!
The specific action requested of your senators or reps could vary from call to call, but may I humbly request asking them to support a ban on superintelligence, e.g. via the Sanders-Casar bill. Whatever happens with this specific bill, if a substantial number of lawmakers were regularly hearing that their constituents don’t want ASI, it could increase the odds this gets reached for after the next big warning shot. Even if the staffers note that it’s always the same folks calling, it still seems important for lawmakers to know they have constituents who want an ASI ban.