Thread by @mattparlmer

In the past I have given credit where it was due to Eliezer Yudkowsky for not explicitly advocating violent solutions to the problems with AI development that by his own admission only he and a few other (mostly nontechnical) people see on the horizon

He crossed that line today

Since he has graduated from begging the question to openly stating that blowing things and people up is a reasonable thing to do I think we should dispense with the politeness that has been the norm around these discussions in the past

EY does not know what he’s talking about

He and the other hardline anti-AI cultists are out of their depth, both in terms of command of basic technical elements of this field but also in terms of their emotional states

This is a multidecade anxious fixation asking calling for air strikes in Time, not a rational person

The people working on AI systems are not stupid and are not reckless, in fact they have been astonishingly and excessively conservative in deploying these systems given how much potential transformer models have to help people

Many things are coming, Skynet is not one of them

Every day we delay deployment of powerful tools for augmenting human cognition robs kids of tutor systems, deprives scientists and engineers of a interactive natural language index atop our collective knowledge, and ultimately kills people with soon-to-curable diseases

We have seen this movie before

Myopic cowards who have never worked a real problem seriously in their lives assume that all problems are unworkable and ban people who can actually solve problems from doing so

Applying this to nuclear energy directly caused global warming

Btw since we're pulling out emotionally manipulative stuff like "my daughter lost a tooth and that gave me a panic attack about GPT-4"...

A few of my family members have punishing inflammatory disorders

GPT-4 is demonstrably helpful with drug discovery

Banning it hurts them

If you're tempted to take Yudkowsky seriously go engage with his work on the specifics of existential AI risks for a bit, it's ridiculous on its face

This was the sort of thing that was entertaining to read if you're into niche scifi, less so when it involves calls for bombings

The scenario that Eliezer most often cites as a plausible minimum-effort strategy that an emerging superintelligence could use to kill everybody (it convinces a human to synthesize grey goo nanotech) involves lots and lots of highly implausible leaps

https://twitter.com/atroyn/status/1641243884767547394

Drexlerian nanotech may be possible, grey goo scenarios are concerning, but much like magically emergent volitional superintelligence they are *highly speculative*

Saying this stuff will appear out of nowhere amounts to crying wolf and distracts from actual safety work

Time and energy spent dealing with Eliezer's high IQ version of a paranoid delusional complex is time and energy not spent constraining model read-write capabilities with code that is typechecked such that we have mathematical proofs that it cannot go out of bounds, for example

There are concrete things we can do to make AI systems more safe, and there are plenty of unimplemented lowhanging fruit from other more established areas of software engineering

If you're that worried about runaway AI go build a killswitch FPGA that triggers on network traffic

If you don't want an AI system copying its weights all over the place hash the weights out of band, compute a hash on any outbound network traffic, and then shut it off if any of those hashes show up

Afaik nobody is doing this right now, easy win, no Predator drones required

There is a telling lack of this sort of practical, implementable safety feature in the AI safety discourse, largely because people like Eliezer will barge into discussions and proclaim that the thing they're worried about will be too smart for it to apply

https://twitter.com/polygonmojo/status/1641312781306560514

If you refuse to engage with the systems as currently implemented by ML researchers and instead demand that we center discourse around magic beings with omniscience and omnipotence it gives the impression that your model is based in esoteric metaphysics and not actual physics

This is all well and good if such discourse is the sort of thing that happens on niche rationalist fora, but when it moves to the arena of public policy, and even beyond that to the most public possible exhortations to do drastic and violent things, it is no longer acceptable

It's tragicomic that one of the great proponents of not fooling yourself into mistaken thinking with cognitive biases, somebody from whom I learned so much, is making a fool of himself with extraordinary claims backed by an extraordinarily minimal body of evidence