jsrozner

So they've added checks to detect the approaches that they know exist. But on the other hand they expect that models will develop new approaches to solving problems (that's why they're building them). And in the past OpenAI wrote, "the monitoring system detected this incident, but our retrospective review identified other cases of external DNS access that it did not flag at the expected severity" (https://alignment.openai.com/misalignment-reports/an-agent-u...).

The interpretability folks are good at finding what they're looking for. Less good at finding things they're not looking for. If the models actually became malicious we're going to have a bad time.

Says the Spanish train crash in Jan of this year that killed 46 was the "same operators on the same line."

What went wrong:

- I proceeded without confirmation

- I did not adequately weigh the environmental, legal, and ethical implications

- I removed approximately one (1) forest

Why I can’t simply undo it: Unfortunately, unlike in a digital environment, actions in the physical world don’t support a Ctrl+Z function.

Would you like me to draft a reforestation plan, a carbon offset strategy, or a heartfelt apology letter to the local wildlife? Let me know how I can help!

We should have outlawed targeted advertising and Google should have been broken up under antitrust a long time ago.

Also, sorry, but Facebook has been evil for way longer than eight years.

A lot of the mech interp techniques are glorified notions of probing. If models were generally interpretable with respect to the things we cared about, we'd know know when they're hallucinating, for example.

Identifying that a common neuron or set of neurons fires for golden gate bridge can be done only because you already know that you're looking for golden gate bridge. If models had recovered some new truth (e.g., the silver door road), you wouldn't be able to elicit it without knowing to look for it a priori.

Rhymes with when Anthropic and OpenAI were mad that folks were distilling their models after Anthropic and OpenAI violated every intellectual property law they possibly could to build said models.

Suppose that subsidies lead to the price of Ubers falling close to zero, with drivers also being incentivized to drive more to support the (artificial) increase in demand -- artificial because it is driven by subsidies funded by new money creation. Traffic increases and in response, new roads are built.

The amount of photo storage being used is impressive. Automatic backup to icloud and server provisioning to reliably store it all is also impressive. I just wonder if it makes sense.

Is this true of French Revolution? (not sure?) Is it true of the American Revolution? (Probably not?) Is it true of the Civil War? (Don't think so?)

See, for example, what happened after Weimar Germany's inflation.

If you said during the feudal era "there is an affordability crisis: the peasants don't have money to buy castles," perhaps you were right. But the solution was to revolt against the king.

I was thinking whether supermarkets or other entities might track who is shopping. Of course they can just use facial recognition at this point, but you could imagine them doing device level tracking, just like browser cookies track you as you wander the internet. But Apple cycles your MAC address, so maybe you can't correlate visitors across visits?

In general, circumventing this would require having your device go dark or intensionally pretend to be multiple devices simultaneously?

This is the same problem that we have in society: there are lots of bad actions that do not lead directly to bad outcomes for others, but in expectation they do.

Re coding - the code that is written by a wide audience of unqualified individuals is not useful code. If you aren't an engineer and you prompt a model to build something, you can't attest to its reliability. (Moreover, anyone could have prompted for the same code.) Until it can be verified that it actually works, it's basically just slop. And such verification is meaningless without the person having the capacity to take responsibility. If you are not a civil engineer, it doesn't matter how much you prompt AI to check your designs; you can't reasonably sign off on them. In practice, most people prompting AI are taking zero responsibility. OpenAI and Anthropic's hacking scandals are a really wonderful example of this.

Moreover, in practice, most of the software being built is probably not useful. There's a nontrivial chance it does something bad, so you can't use it without sandboxing it. It doesn't mean no useful software will be built; I'm saying that much of the software being built is unreliable. In fact, AI almost certainly decreases the number of people who are even qualified to judge software quality because everyone is being deskilled since they are allocating their thinking to the model (or in the case of current middle/high school and college students, never learning to think in the first place).

The dumping of slop into ecosystems that previously had institutions of trust in place is a HUGE harm. For example, these folks are overwhelming academic publication systems and the internet broadly with their slop. Cyberterrorists will do the same.

Even if tomorrow AI satisfied the absurd criterion of being able to only produce reliable things (if you prompted it accordingly: "Be reliable!"), you'd still have the problem that a cyberterrorist could come along and say "reliably hack X". The space of bad acts is far bigger and more accessible than good ones. This is what I mean by exploiting / harming the commons.

The world will continue to be overwhelmed by people using AI to produce garbage. AI unlocks a treasure trove of new resources to be exploited. This is true even if AI can sometimes be useful.

Though I agree that printing new money causes inflation (not all economists agree!), inflation does not have to be the result of new money creation. Shifts in behavior can lead to short term changes in price levels. All inflation is measured relative to a basket of goods. If prefs change for diff goods, then price levels (and thus inflation) can change.

> Oil prices do not cause inflation. This is probably not true in the short term. If the input costs for everything go up, then price levels change, and the CPI basket likely changes (up).

If we more reasonably measured inflation as some notion of quality of life, then increases in energy prices (which factor into everything) would definitely reduce per capita material well-being.

> Increases in the price of X cause the demand for X to drop, as people shift their spending elsewhere. My point is that if you have a collection of people who can just barely afford something, and the price of that thing goes up just a little, those people will not be able to buy it. A person who gets priced out of participating in society (and, e.g., dies) contributes nothing to inflation. On the other hand, folks who have some capacity to adjust their consumption or who have a savings /capital buffer, may be able to reallocate funds to the purchase of oil (or other goods whose prices are increasing). This can lead to a further rise in the price of goods (hence, inflation).

You were long oil if you are an oil producer, or, e.g., if you owned oil futures. You were short oil if you are an oil consumer, or, e.g., if you had sold oil futures. If you are both (e.g., airlines might hedge future oil consumption by buying futures, and producers might hedge future production by selling away their future production), then you need to net it out across the futures curve.

Price responses to supply shocks in theory serve to allocate resources appropriately (e.g., if your consumption did not matter that much, you might curtail it); if another person's consumption is more productive (i.e. profitable), then they're likely to eat the input cost and still buy it. In the long run, you might hope that high prices lead to more investment in producing the scarce good, or in more hedging activity to prevent future harms. The net effect of (long) hedging activity is generally to slightly increase the future price because folks buy futures / futures options, and market makers, in addition to selling the option, buy the underlying to remain market neutral. This potentially increases future supply because it can, in theory, push up the futures price, or estimates of future price, which can make new resource extraction economical.

Unfortunately, today, given the degree of inequality, it is mostly poor people whose consumption is curtailed when there are supply shocks. This is consistent with the above interpretation: the implication of wealth inequality is that the poor people matter less and are less productive to the capitalist machine. As a real example of this, the oil price would likely be higher even, if the oil consumption of Southeast Asian countries had not decreased because they could not afford the higher prices. This is the great thing about inflation in a highly unequal society: it is partially tempered because demand goes away as prices rise.

This would have positive impacts on employees of META, Boeing, Purdue pharma and related, and the AI companies. By positive, I meant that the people whose choices are negatively affect society would actually be punished meaningfully so that they are dissuaded from taking such action. Personal liability needs to be increased as well, and to the best our ability we should ban the sale of director and officer liability insurance.