Ok. It’s strange, then, that wikipedia does not say this. On the contrary, it says:
The notion that bilateral trade deficits are per se detrimental to the respective national economies is overwhelmingly rejected by trade experts and economists.[2][3][4][5]
(This doesn’t necessarily contradict your claim, but it would be misleading for the article to say this but not mention a consensus view that trade surpluses are beneficial.)
Do you believe running a trade surplus causes a country to be wealthier? If so, how do we know that?
And so, like OpenAI and Anthropic, Google DeepMind wants the United States’ AI to be stronger than China’s AI. And like OpenAI, it intends to make weapons for the US government.
One might think that in dropping its commitments not to cause net harm and not to violate international law and human rights, Google is signalling its intent to violate human rights. On the contrary, I believe it’s merely allowing itself to threaten human rights — or rather, build weapons that will enable the US government to threaten human rights in order to achieve its goals.
(That’s the purpose of a military, after all. We usually don’t spell this out because it’s ugly.)
This move is an escalation of the AI race that makes AI war more likely. And even if war is averted, it will further shift the balance of power from individuals to already-powerful institutions. And in the meantime, the AIs themselves may become autonomous actors with their own purposes.
Google’s AI principles used to say:
In addition to the above objectives, we will not design or deploy AI in the following application areas:
Technologies that cause or are likely to cause overall harm. Where there is a material risk of harm, we will proceed only where we believe that the benefits substantially outweigh the risks, and will incorporate appropriate safety constraints.
Weapons or other technologies whose principal purpose or implementation is to cause or directly facilitate injury to people.
Technologies that gather or use information for surveillance violating internationally accepted norms.
Technologies whose purpose contravenes widely accepted principles of international law and human rights.
As our experience in this space deepens, this list may evolve.
On 2025-02-04, Google removed these four commitments. The updated principles seem consistent with making weapons, causing net harm, violating human rights, etc. As justification, James Manyika and Demis Hassabis said:
There’s a global competition taking place for AI leadership within an increasingly complex geopolitical landscape. We believe democracies should lead in AI development, guided by core values like freedom, equality, and respect for human rights. And we believe that companies, governments, and organizations sharing these values should work together to create AI that protects people, promotes global growth, and supports national security.
Update: It’s even better than that. Not only will they make a lab order for you, but they will also pay for the test itself, at a steep discount to the consumer price.
I didn’t know about ownyourlabs, thanks! While patients can order a small number of tests directly from Labcorp and Quest Diagnostics, it seems ownyourlabs will sell you a lab order for many tests that you can’t get that way.
Exhibit 13 is a sort of Oppenheimer-meets-Truman email thread in which Ilya Sutskever says:
Yesterday while we were considering making our final commitment given the non-solicit agreement, we realized we’d made a mistake.
Today, OpenAI republished that email (along with others) on its website (archived). But the above sentence is different in OpenAI’s version of the email:
Yesterday while we were considering making our final commitment (even the non-solicit agreement), we realized we’d made a mistake.
I wonder which sentence is the one Ilya actually wrote.
check out exhibit 13...
Section 3.3(f)(iii):
Within 120 days of the date of this memorandum, DOE, acting primarily through the National Nuclear Security Administration (NNSA) and in close coordination with AISI and NSA, shall seek to develop the capability to perform rapid systematic testing of AI models’ capacity to generate or exacerbate nuclear and radiological risks. This initiative shall involve the development and maintenance of infrastructure capable of running classified and unclassified tests, including using restricted data and relevant classified threat information. This initiative shall also feature the creation and regular updating of automated evaluations, the development of an interface for enabling human-led red-teaming, and the establishment of technical and legal tooling necessary for facilitating the rapid and secure transfer of United States Government, open-weight, and proprietary models to these facilities.
It sounds like the plan is for AI labs to transmit models to government datacenters for testing. I anticipate at least one government agency will quietly keep a copy for internal use.
So was the launch code really 000000?
A fun lesson on the scientific method: A few days ago, OpenAI announced a rollout of “Advanced Voice Mode”. Confused by the messaging, I ran a test to see if their phone app could respond to an interruption:
Nisan: Can you count to 10, and then I’ll interrupt you in the middle and see how that works.
ChatGPT 4o: Sure! I’ll start counting now: One, two, three, four, fiv—
Nisan: Stop. Interruption.
ChatGPT 4o: How’d the interruption go?
Nisan: Oh, that went great. You stopped at five.
ChatGPT 4o: Awesome! Glad it worked smoothly. Do you want to try another test or experiment?
Nisan: Yeah, let’s do another interruption test. This time, please count to 20.
ChatGPT 4o: Got it! Starting now: One, two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirt—
Nisan: Stop. Interruption.
ChatGPT 4o: How was that?
Nisan: Wow. [Thinks: It seemed to stop slightly before I interrupted it!]
ChatGPT 4o: Glad that impressed you! Want to try something else?
Nisan: Uh, hold on.
ChatGPT 4o: Take your time!
Nisan: [Generates a random number.] Okay, please count to 30, and I’ll interrupt at a random point.
ChatGPT 4o: Alright, starting now: One, two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen—
[Nisan interrupts at some point, I forgot exactly where.]
ChatGPT 4o: —sixteen, seventeen, eighteen, nineteen, twenty, twenty-one, twenty-two, twenty-three, twenty-four, twenty-five, twenty-six, twenty-seven, twenty-eight, twenty-nine, thirty. Looks like I made it to the end! Did you mean to interrupt somewhere?
Nisan: Yeah, I did. It must not have worked. Let’s try that again. [Generates a random number.] Please count up to 30.
ChatGPT 4o: No problem, let’s try again! Starting now: One, two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen—
[Nisan interrupts at some point.]
ChatGPT 4o: —sixteen, seventeen, eighteen, nineteen, twenty, twenty-one, twenty-two, twenty-three, twenty-four, twenty-fiv— Did it work this time?
Nisan: Uh, no.
I believe ChatGPT can’t hear me when it’s talking. But it correctly guessed when I would interrupt on the first two trials!
The coin flip is a brilliant piece of technology for generating trustworthy random noise:
Making a two-headed coin is forgery, which is a crime.
Such trick coins can be foiled anyways by calling the toss in the air.
Thus when teaching the concept of a Bernoulli variable, we use the example of coin flips, because everyone already knows what they are. This is unfortunate because the very next concept we introduce is a biased Bernoulli variable, which corresponds to a “weighted” coin. But weighted coins don’t exist! If it were practical to manufacture trick coins with arbitrary biases, coin flipping wouldn’t be as popular as it is.
If there was a consensus among the 8 as to which tuning is better, that would be significant, right? Since the chance of that is 1⁄128 if they can’t tell the difference. You can even get p < 0.05 with one dissenter if you use a one-tailed test (which is maybe dubious). Of course we don’t know what the data look like, so I’m just being pedantic here.
Progress towards a robotic piano tuner: Entropy piano tuner attempts to accommodate “variations in string thickness, stretching, corrosion, dents, the harp flexing”, etc. by minimizing the entropy of the power spectrum. Using it should be better than mindlessly tuning to a digital guitar tuner.
According to the website, professional pianists still prefer a human-tuned piano, but no one else can tell the difference. And the general opinion on piano tuner message boards seems to be that it’s not quite good enough to replace a professional tuner’s judgment.
This post is wrong. Thanks to SymplecticMan for the thought experiment demonstrating that a mixture of ideal gases follows a law rather than my proposed law. (It’s also different from Newton’s law.)
I made a pretty but unjustified assumption — that a cooling baking sheet can be modeled as a dynamical system where each possible transition is equally likely and in which heat is transferred in fixed quanta, one at a time. This contradicted Newton’s law, and I got excited when I realized that Newton’s law was merely a first-order approximation.
My mistake was not noticing that Newton’s law is a first-order approximation to any model of cooling where heat transfer increases with temperature difference, so I had not observed any reason to favor my model over any other.
In penance I have acquired a copy of Non-Equilibrium Thermodynamics by de Groot and Mazur, with the intention of eventually reading it.
This is the perfect time to start an AI + education project. AI today is not quite reliable enough to be a trustworthy teacher; and in the near future generic AI assistants will likely be smart enough to teach anything well (if they want to).
In the meantime, Eureka Labs faces an interesting alignment problem: Can they ensure that their AI teachers teach only true things? It will be tempting to make teachers that only seem to teach well. I hope they figure out how to navigate that!
On 2018-04-09, OpenAI said[1]:
OpenAI’s mission is to ensure that artificial general intelligence (AGI) [...] benefits all of humanity.
In contrast, in 2023, OpenAI said[2]:
[...] OpenAI’s mission: to build artificial general intelligence (AGI) that is safe and benefits all of humanity.
This archived snapshot is from 2023-05-17, but the document didn’t get much attention until November that year.
Another example is risk compensation: You make an activity safer (yay) and participants compensate by taking more risks (oh no).
archive