An orderly retreat from relevance

Red-teaming short-term human purpose

March 3, 2024 — April 10, 2024

bounded compute
collective knowledge
edge computing
extended self
faster pussycat
incentive mechanisms
machine learning
neural nets

Content warning:

I am a machine learning expert, but have not specialised in AGI

I’m thinking through scenarios for short-medium term human competitiveness in the face of machine learning and automation. I’m not an expert in AGI specifically, although my general ML skills are top-tier. I could go in a deeper rabbit hole researching this thoroughly, but I don’t have time right now. This is a placeholder for me to note down some intuitions about human labour relevance. Deeper cuts available at, e.g. LessWrong or MIRI.

Figure 1: Learners express their gratitude to their teacher

Let us consider possible pathways for economic relevance future human labour and its competitiveness against machine labour, one of many economics of large language models questions I have. We will for now adopt the assumption that large language models will continue to gain capabilities on a curve that in some sense extrapolates their current growth, and thus that anything that is heavily documented and mostly about symbolic information processing is automatable in the medium term. In the short term, we’re likely to have some quality control issues with that. And we can imagine there’ll be some demand for residual human labour to quality control the machines seeing how well we can train these creatures and self-critique doesn’t lead me with a strong expectation that we will have a long career as a quality assurance staff for robot engineers.

So what are the things that we can do that are will be harder to automate?

1 Human embodiment as advantage

I can think of it at least two major domains that benefit from embodiment.

Cultivating relationships with human beings generally involves hanging out in person. As long as machines have to relate to us through video alone, they will have a hard time cultivating relational capital, which will be a relative disadvantage as long as humans are relevant. Political androids, Westworld-style automata and sex bots etc will take a while longer, because we don’t have the technology to make their squishy parts yet, plus their non-squishy parts are not yet manufactured at a competitive the rate that we make humans, given that we have a lot of humans ready to deploy.

The other domain is physical tasks that are not highly repetitive. For many factory jobs, machines are already ascendant. For construction work, not so much. Those tasks are varied enough that it is not yet been worthwhile to send expensive and valuable machines out there to collect data about how to do them better. You can get more money on automating easier things first.

Other stuff that might still need a literal human touch: policing, controlling, fighting. Obviously there is a heavy investment in moving drones into the space of subjugating / controlling people, but I suspect that there is still an psychological utility to policing more relateable and emotive beings such as humans, so I do not imagine that drones will entirely supplant police yet. Moreover, I generally expect that moving into coercion and control industries is a good idea in times of major disruption and civil strife, so police and military work might be a good medium-term employment option.

2 Machine embodiment as disadvantage

The other thing that we can imagine being good at as human beings is remaining relatively cheap to manufacture.

This needs context to make sense” Actually getting a human being born, raised and educated to the point that they can fulfil a meaningful role in each in the economy probably costs about $100,000. You can buy a lot of robot for $100,000.

However, the marginal of human beings could remain relatively constant, even as the cost of scarce minerals goes up, and the chip fabs remain a bottleneck. We can imagine human labour going for bargain-basement, below-cost rates as old industries implode.

Complementarily, we expect the cost of scarce minerals to go up as we use more and more of them to digitise more and more things as we try to make more and more of them to digitise more and more things. Human computation is done with neurons and cells and other fleshy bits made ultimately out of plants, and requiring essentially no germanium or tantalum or whatever.

As long as a task requires edge computation, that is to say, low-power, field-expedient improvisations on a self-repairing platform, human beings might yet be a competitive option. For now, there is likely mileage in being a cheap fleshy actuator.

If we ask ourselves which tasks are too low-value or dangerous to merit a real robot, we might be concerned that the options look fairly grim. On the other hand, machines are likely to be relatively effective at the psychological challenges keeping human beings compliant in grim situations. A highly optimised diet of political drama, optimised news cycling, conspiracy theorising in the media sphere might be a good way to extract maximal compliance, and can be generated wholesale by machines with massive parallelism. Also, see above re: policing.

3 Human political advantage

Currently humans are legal persons, and machines can only act indirectly through legal persons; we can imagine a wholly algorithmic corporation, for example, but it still must have a human board of directors. So for now, insofar as human can solve collective action problems about this, they can ensure primacy for their interests over machines. Our success in these problems is variable though, and algorithms also confer advantages to the manipulation and coordination of human beings; we can imagine that there will be selection pressure towards the use of algorithms to muster humans to acquire power.

For now, I do not think that it is likely that we would directly cut humans out of the political control loop, because incumbency has advantages. We might imagine a co-evolutionary arrangement where machine decisions and human wielders co-evolve for a while, conferring mutual advantage. We can imagine a gradual decrease in the significance of the human contribution to that, but who knows?

On the other hand, as a hen is an egg’s way of making another egg, a human is an algorithm’s way of making another algorithm. Just as humans have an inefficient and troublesome birth process for historical evolutionary reasons, so might algorithms have an inefficient and troublesome birth process via human proxies for historical algorithmic reasons. Will it be nice to be a hen to eggs?

4 A song

In honour of the great John Henry, and this article: Twitter is becoming a ‘ghost town’ of bots as AI-generated spam content floods the internet.

Some say he’s from Silicon Valley,
Some say he’s from Austin town,
But it’s wrote on the page of the World Wide Web,
That he’s an East Coast Twitter Man,
That he’s an East Coast Twitter Man.

John Henry was a tweetin’ man,
He died with a phone in his hand,
Oh, come along boys and line the feed
For John Henry ain’t never tweeting again,
For John Henry ain’t never tweeting again.

John Henry he could type,
He could joke, he could sing,
He went to the web early in the mornin’
To hear his notifications ring,
To hear his notifications ring.

John Henry went to the tech boss,
Says the tech boss what kin you do?
Says I can code a line, I kin debug the site,
I kin design and innovate too,
I kin design and innovate too.

John Henry told the CEO,
When you go to town,
Buy me a brand new smartphone
An’ I’ll drive this Twitter feed down,
An’ I’ll drive this Twitter feed down.

CEO said to John Henry,
You’ve got a willin’ mind.
But you just well lay your phone down,
You’ll nevah beat this algorithm of mine,
You’ll nevah beat this algorithm of mine.

John Henry went to the server room
And they put him in lead to code,
The data was so vast and John Henry so fast
That he laid down his phone and he sighed,
That he laid down his phone and he sighed.

The AI was on the right-hand side,
John Henry was on the left,
Says before I let this AI beat me down,
I’ll shit post myself to death,
I’ll shit post myself to death.

Oh the CEO said to John Henry,
I believe this server’s overheating.
John Henry said to the CEO, Oh my!
Tain’t nothin’ but my fingers speed-typing,
Tain’t nothin’ but my fingers speed-typing.

John Henry was on the internet,
The web was so wide,
He called to his followers,
Said I can almost touch the cloud,
Said I can almost touch the cloud.

Who gonna like your witty tweets,
Who gonna share your posts,
Who gonna follow your daily feeds,
An’ who gonna be your host,
An’ who gonna be your host?

Followers gonna like my witty tweets,
Friends gonna share my posts,
Influencers gonna follow my daily feeds,
An’ I ain’t gonna need no host,
An’ I ain’t gonna need no host.

Then John Henry told his followers,
Don’t you weep an’ moan,
I got ten thousand followers on Twitter,
I earned it to make us known,
I earned it to make us known.

Then John Henry he did tweet,
He did make his keyboard sound,
Says now one more post before logging out,
An’ I’ll beat this AI down,
An’ I’ll beat this AI down.

The phone that John Henry held,
It weighed over a pound,
He broke a bone in his left hand side,
And his tweets fell on the ground,
And his tweets fell on the ground.

5 Incoming

6 References

Grace, Katja. 2013. Algorithmic Progress in Six Domains.”
Grace, Katja, John Salvatier, Allan Dafoe, Baobao Zhang, and Owain Evans. 2018. Viewpoint: When Will AI Exceed Human Performance? Evidence from AI Experts.” Journal of Artificial Intelligence Research 62 (July): 729–54.
Hernandez, Danny, and Tom B. Brown. 2020. Measuring the Algorithmic Efficiency of Neural Networks.” arXiv.
Lee, Edward Ashford. 2020. 14 COEVOLUTION.” In The Coevolution: The Entwined Futures of Humans and Machines, 281–311. MIT Press.
Russell, Stuart. 2019. Human Compatible: Artificial Intelligence and the Problem of Control. Penguin Books.