philomath_mn

I am both skeptical and interested:

- Looks like you turn the knob to set your page

- This, and your book title, appear to be the only input data to the system (no other sensors)

- Does this mean they aim to have a database of the text of every edition of every book, indexed by page?

If they can pull it off I think it could be useful, but I think they will need to compromise on quality if they hit copyright issues.

Would ChatGPT use tools to analyze this? Or are we talking raw data in the context window?

Have you done this? Any interesting results?

Are your positions really that sensitive? A lot of hedge funds don't even care enough to have client-side decryption.

Self advocacy is part of the job.

And just because you’re an introvert doesn’t mean you are incapable of building soft skills. Talking to people is absolutely exhausting for me, but I force myself to do it and practice at it because I know it is important for my career.

Right -- I mean the misery at work is what I cannot relate to.

Furthermore, if the best way to support my family and to live my non-professional life is to take a punch in the stomach each morning, I think I'd find a way to deal with it?

Working a soul-sucking but cushy corporate job is not trauma.

I suppose I was reacting mostly to the "my job is so bad it is akin to being physically assaulted"

It is interesting to hear how your perspective resonates with the essay, I was mostly wondering how universal that resonance is.

Anyone else feel that they cannot relate to the premise of this essay in any way?

I mean I'd rather be on a road trip with my kids, but my work is generally pleasant outside infrequent periods of high stress.

But this all comes back to the self-interest and game theory discussion.

Let's suppose that, like you, both Moscow and Beijing do not want AGI to exist. What could they do about it? Why should they trust that the rest of the world will also pause their AI development?

This whole discussion is basically a variation on the prisoner's dilemma. Either you cooperate and AI risks are mitigated, or you do not cooperate and try to take the best outcome for yourself.

I think we can expect the latter. Not because it is the right thing or because it is the optimal decision for humanity, but because each individual will deem it their best choice, even after accounting for P(doom).

> Let's start there

Ok, I can imagine a domestic policy like you describe. Through the might and force of the US government, I can see this happening in the US (after considerable effort).

But how do you enforce something like that globally? When I say "not really possible" I am leaving out "except by excessive force, up to and including outright war".

For the reasons I've mentioned above, lots of people around the world will want this technology. I haven't seen an argument for how we can guarantee that everyone will agree with your level of "acceptable" P(doom). So all we are left with is "bombing the datacenters", which, if your P(doom) is high enough, is internally consistent.

I guess what it comes down to is: my P(doom) for AI developed by the US is less than my P(doom) from the war we'd need to stop AI development globally.

Not a great analogy. If

- there were a range of expert opinions that P(destroy-the-world) < 100 AND

- the chemical could turn lead into gold AND

- the chemical would give you a militaristic advantage over your adversaries AND

- the US were in the race and could use the chemical to keep other people from making / using the the chemical

Then I think we'd be in the same situation as we are with AI: stopping it isn't really a choice, we need to do the best we can with the hand we've been dealt.

I think it is extremely unlikely we are going to be able to convince every interested party that they should give up the golden goose for the sake of possible calamity. I think there are risks here, not trying to minimize that, but the coordination problem becomes untenable when the risks/benefits are so large.

It is essentially the same problem as the atom bomb: it would have been better if we all agreed not to do it, but thats just not possible. Why should China trust the US or vice versa? Who wants to live in a world where your competitors have world-changing technology but you don't? But here we have a technology with immense militaristic and economic value, so the everyone-wants-it problem is even more pronounced.

I don't _like_ this, I just don't see how we can achieve an AI moratorium outside of bombing the data centers (which I also don't think is a good idea).

We need to choose the policy with the best distribution of possible outcomes:

- The US leads an effort to stop AI development: too much risk that other parties do it anyway

- The US continues to lead AI development: hope that P(takeoff) is low and that the good intentions of some US labs are able to achieve safe development

I prefer the latter -- this is far from the best hypothetical outcome, but I think it is the best we can do when constrained by reality.

Ok, let's suppose that is true.

What bearing does that have on China's interest in developing AGI? Does the risk posed by OpenAI et al. mean that China would not use AI as a tool to advance their self interest?

Or are you saying that the risks from OpenAI et al. will come to fruition before we need to worry about China's AI use? That still wouldn't prevent China from pursuing AI up until that happens.

I am still not convinced that there is a policy which can prevent AI from developing outside of the US with high probability.

Given the compute and energy requirements to train & run current SOTA models, I think the current political rulers are more likely to have control of the first AGI.

AGI would then be a very effective tool for maintaining the current authoritative regime.

Also, to be clear: I reject your pushback based on my understanding of the incentives/goals/interests of nation states like China.

This is completely separate from my personal preferences or hopes about the future of AI.

No, I do not expect things to just work out. I just think our best chance is for the US to be a leader in AI development and hope that we're able to develop it safely.

I don't have a lot of confidence that this will be the case, but I think the US continuing to develop AI is the decision with the best distribution of possible outcomes.

That is a massive bet based on the supposed psychology of a world super power.

There are many other less-superficial reasons why Beijing may be interested in AI, plus China may not trust that we actually banned our own AI development.

I wouldn't take that bet in a million years.

> It's possible to stop developing things

If the US were willing to compromise some of it's core values, then we could probably stop AI development domestically.

But what about the rest of the world? If China or India want to reap the benefits of enhanced AI capability, how could we stop them? We can hit them with sanctions and other severe measures, but that hasn't stopped Russia in Ukraine -- plus the prospect of world-leading AI capability has a lot more economic value than what Ukraine can offer.

So if we can't stop the world from developing these things, why hamstring ourselves and let our competitors have all of the benefits?

> it's worth trying to stop it

OP's point has nothing to do with this, OP's point is that you can't stop it.

The methods and materials are too diffuse and the biggest players (nation states) have a strong incentive to be first. Do you really expect China to coordinate with the West on this?

There are some ebbs and flows, but usually everything is happening all at once. There is usually never a single product, and even then different parts of a product will be in different stages at the same time.

So you could be doing heavy prototyping or feature work on a mature product if you are enhancing it in some way, or you could be doing some maintenance / refactoring / bug-fixes on an early stage product.

The distribution of time between those tasks will change over the life of the product, but there is rarely a "feature complete" stage in my experience.

Long term, I think it is a mix of expectations that:

- AGI/ASI would be a breakthrough worth trillions in market value

- whoever gets AGI/ASI first may have a permanent advantage (goes for other labs and other nations)

Looks like this project was recently restarted [0]

[0] https://www.zawya.com/en/projects/construction/kingdom-holdi...

> I think violence is going to become more common

What kind and why?

I thought the _Learning Spark_ book was a pretty good introduction. Databricks offers it for free here [0]

[0] https://pages.databricks.com/rs/094-YMS-629/images/LearningS...

> but all electronic devices in the kid's room exist in a baffling state of perpetual reorganization, everything with a plug being unplugged, moved around, and maybe or maybe not plugged back in, seemingly every day, for no apparent reason

This is so accurate. We have at least a dozen USB power bricks floating around our house, even my charger isn't safe from being "borrowed".

a) What are the benefits? Most of these devices are savings mere seconds of labor

b) What are the costs? Just about every person I know who got into IoT devices has regretted it within a few years (things break, manufacturers go out business, just keeping up with updates can be a pain)

Why even bother?

Idk man, survival seems like a pretty blunt selection force for such an intricate mechanism

I have the same general sense: the days feel slightly more productive but I am not sure I am producing more on the scale of months.

One area where I am absolutely seeing gains is prototyping. AI is so good at getting the 80% you need for a demo but it has a hard time getting beyond that (for now, anyway).

I am curious if others have tips for scaling up AI tools to larger scopes (working with full codebases, architecture, moving the needle on deploying features, etc)

Can you easily combine it into one file? You could use PEP 723 and `uv`'s script running capability [0].

If your project is not flat, you could have a "build" step to combine everything into one file (assuming everything is simple).

`uv tool` would be great but you'd need to manage a private repo and credentials -- the "it just works" magic kinda goes away at that point.

[0] https://docs.astral.sh/uv/guides/scripts/#declaring-script-d...

How big is your distribution file? I tried this but my project had numpy and pandas and the distributable was like 350 MB!