A strict and swift non-permanent pause is certainly one option. Use very aggressive regulation. Perspires slightly.
Pause all development and research everywhere and all at once. Don’t stop totally. Don’t be anti-accelerationism. Just slow the rate acceleration a bit.
Unpause when sufficient safety research has been conducted, safety measures have been identified and agreed upon, and can be practically implemented.
This should take approximately 5-50 years. If the world can mobilize the way it did during the COVID-19 pandemic, 5 years feels reasonable.
The world isn’t so bad that we can’t wait for AGI - or whatever the actual goal of the rat race is.
I would suggest that once paused, the following strategy is used to unpause.
This post is cross-posted from our Substack. Kindly read the description of this sequence to understand the context in which this was written.
The rat race to better and better generative AI systems is a real one. Companies such as OpenAI, Anthropic, and Google Deepmind have a shared goal: creating “benevolent AGI”.
The problem is that not a single one of these companies is currently on track to reach this goal; at least the “benevolent” part. The reason being?
Moloch
They might all want to slow down to focus more on safety concerns, but they can’t. Why?
Moloch
They might all want to spend more of their investor’s money to focus on safety concerns, but they can’t. Why?
Moloch
More than ten years ago, Scott Alexander wrote a phenomenal essay titled “Meditations On Moloch”. Moloch is intro
duced as an answer to one of C.S. Lewis’ questions: What does it? Scott writes, “Earth could be fair, and all men glad and wise. Instead we have prisons, smokestacks, asylums. What sphinx of cement and aluminum breaks open their skulls and eats up their imagination?”
And Allen Ginsberg’s famous poem on Moloch answers it:
The beauty of this poem is that it shows Moloch to be the personification of these terrible forces. Everyone seems to hate the current system, but who perpetuates it? Moloch.
Moloch is the force that coerces competing entities to take actions that might be locally better, but eventually lead to situations where everyone is worse off. Not a single entity wants this, but they have to keep going. Moloch forces them.
Of course, no one literally thinks some kind of ancient demon causes all of this to happen, but to think of the system as an agent actually helps us realize how the system and its perverse incentives is not an agent. Framing Moloch as an agent with a goal and ability to coordinate its actions to achieve it, makes us understand how there is actually nobody at the wheel. Rather, the sad reality is that the system, i.e., Moloch, can produce outcomes that no individual agent within the system desires.
Scott Alexander gives a great many examples, wonderfully illustrating the ways in which Moloch can take form. There is no better substitute than reading the original essay, but for illustrative purposes I will quote one example here:
This scenario is a (rat) race to the bottom.
The same is happening in the current race to “AGI”.
Some people claim it is tough to beat Moloch, others that it’s impossible, but we have only one choice: understand the dynamics within which Moloch operates, and do everything we can to set up frameworks in which humanity thrives, not Moloch. We simply have no other choice than beating Moloch. We do this by (i) actually trying, and (ii) ideally by studying similar dynamics in other contexts and applying relevant knowledge to overcome Moloch in this new situation.
To further illustrate how Moloch plays a role in the context of the current AI rat race, we can take a look at the dynamic I mentioned previously: The frontier labs might all want to slow down to focus more on safety concerns, but they can’t. An example is Dario Amodei, CEO of Anthropic, the company behind the “Claude” models. A TIME magazine article lays out how, in the summer of 2022, Amodei had a difficult decision to make: Release a new chatbot that was far more powerful than anything they had seen before, or work on internal safety testing.
Dario and others at Anthropic decided to wait, focusing more on safety. But then? Moloch! Another actor, OpenAI, launched a similar product called ChatGPT. And then the dynamics of Moloch started: Anthropic had to release their own model in order to stay relevant. Since then, they have had to keep up with the “race to the top”, although they have, according to themselves, tried to make this a “race to the top” with respect to safety.
These dynamics will keep on happening, and while relevant actors such as Anthropic and OpenAI will likely not say that they are being “controlled” by Moloch, they are. When there are no frameworks in place to stop Moloch, Moloch is always present.
How do we beat Moloch? I will try to lay out what I think are the current best options in a future post.