An AI hardware fantasy, and an IQ erosion attack horror story

10.42, Friday 10 Nov 2023

(This is #2 in what it turns out is an ongoing series of highly speculative, almost entirely unfounded hunches about AI. The first was about alignment and microscopes.)

A fantasy of ‘intelligence too cheap to meter’ in the hardware supply chain

I like to describe AI as a 10 year wormhole into the future (maths here).

We can date it to the release of ChatGPT in November 2022 which is when the technology, UI, and public understanding all came together, but really it was 5 years in the making: the underlying architecture of large language models is the Transformer model, and the original paper came out of Google Research in 2017.

It took OpenAI to do the engineering to scale it though. OpenAI was founded in 2015. So let’s say an overnight 10 year leap with a 7 year run-up.

But now OpenAI is going for it.


What I love about OpenAI is that they hold nothing back.

There’s no clever MBA-authored strategy like holding a feature till next year to maximise profits. Just: bang bang bang. Everything they’ve got, as soon as it’s ready. User-facing in ChatGPT and for developers via the platform APIs.

For instance: here’s Sam Altman’s opening keynote (YouTube) for OpenAI’s developer day last week. It is 45 minutes and tight af.

The full list of announcements (TechCrunch) includes the ability to make custom ChatGPTs that can browse the web and use tools on your behalf, and a app store for them; new APIs for a version of GPT-4 that can see (so fast that it can interpret video), and also for the new image generation model DALLE-3; APIs for great speech synthesis (i.e. everything can talk now) and a bunch more. Like, the Assistants API means it’s easy to build a copilot for any app and you know how I feel about NPCs.

As a developer, this is exactly what you want from your platform company.


So I am not the only person to make a comparison with Apple keynotes.

Which are slick but omg so long and maybe not as action-packed as they used to be. I mean, you think about Apple’s Vision Pro announcement (YouTube) and it’s part of a 2 hour keynote and oh so much explaining.

Which I love for the design nerdery and also is necessary to make sure the media lands right, I know. But you get the impression that Sam Altman would have come on stage wearing the thing, given a brief demo, shared a link to the developer documentation, wrapped up, and the whole slap in the face would have felt like the sonic boom of the future arriving.


Which takes me to a fantasy of combining the Apple and OpenAI approaches.

Apple is a hardware product company. But just suppose it were a hardware _platform__ company, a platform for other people’s hardware, OpenAI-style holding nothing back.

You’d get components that would up-end the supply chain.

Tiny sensors that can do gaze and pointing detection. Microphones with absolutely perfect AI-powered speech recognition built in, and configurable semantic understanding such when someone says “turn on” (or anything similar, while paying attention to it), GPIO pin 1 goes high. Instead of a pseudo-3D lenticular display just for the Vision Pro, one that whichever OEM can build around.

Like any platform company, there would be evaluation boards, but building from the OpenAI playbook, the sensors and components would have plug-and-play versions for individual developers in the form of Raspberry Pi shields and so on. So there would be on-ramps and routes to scale.

This is an old fantasy: in my 2020 post How I would put voice control in everything I set this out…

If I had all the VC money in the world, I would manufacture and sell standardised components – they would connect and act identically to mechanical buttons, switches, and dials, only they would work using embedded ML and have voice, gaze, and pointing detection, for interaction at a distance.

The goal would be to allow manufacturers of every product to upgrade their physical interfaces (add not replace ideally), no matter how trivial or industrial, no matter how cheap or premium.

And this is how we would get to intelligence too cheap to meter and situated, embedded AI. (There are a bunch of examples in that post.)

I want my oven that knows how to cook anything just by looking instead itself and autonomously googling when it recognises the food! I want my telepathic light switches!

But we need AI in the hardware supply chain, not vendors who have to own the whole stack.

Maybe OpenAI will decide to take it on.

Ok. Autumn daydream over.


State-sponsored IQ erosion attacks

Shortly after OpenAI released its new tools, ChatGPT went down together with all the APIs, for several hours.

There was a coding task I was in the middle of that I literally couldn’t complete. Not because I needed API access to GPT-4, but because without ChatGPT I was too dumb to deal with it.

I said on X/Twitter that my IQ had dropped 20 points.

(If you’re a sci-fi fan then it was an experience from Vernor Vinge’s Zones of Thought books – living happily in the Beyond and then being engulfed in a Slow Zone surge.)

And I wonder what the collective intelligence drop was, that day.

Like, if ChatGPT has 180 million monthly active users, could we say something like 1% of the population of the US would have wanted to use it over the down-time?

The US has a population of approx 300 million or, in other units, 30 billion collective IQ points.

So if you ding that by 3 million people at -20 IQ each, that’s 6E7 out of 3E10, or a 0.2% knock on collective intelligence for that day.

By way of comparison, that’s a decent fraction of the effect of leaded fuel. (Everyone born before 1990 has their IQ nerfed by 4.25%.)

And as someone into weird state-sponsored exploits I wonder: would it be worth doing this deliberately?


RELATED TO THIS:

I recently added really smart AI-powered semantic search to my unoffice archive of In Our Time shows. Go to Braggoscope, tap Search in the top nav, and type the biggest planet – the episode about Jupiter comes up. So there’s a kind of knowledge in the large language model, or whatever you want to call it, a sort of relatedness that makes it easy to put ideas together.

Here’s the code on GitHub, open for your interest: you’ll notice in the Cloudflare worker that this “knowledge” comes from a model called baai/bge-base-en-v1.5.

Here’s the model on Hugging Face: BAAI, the creator, is the Beijing Academy of Artificial Intelligence.

Now, I don’t mean to sound paranoid here.

But in Samuel Delaney’s astounding 1966 sci-fi/speculative-linguistics novel Babel-17 (Amazon), [SPOILERS] Babel-17 is a weapon, an artificial language constructed such that intuitive leaps about combat manoeuvres are instantaneous, so it will be adopted virally simply out of utility, but the language itself omits particular connections making certain other ideas topologically impossible.

So: would it be possible to release an AI large language model that is exceptionally good and cheap, maybe, gaining popularity in a target language (English, or Russian, or Korean, or whatever) but makes it really hard to reason about certain concepts?

Not so ridiculous! This has happened once before actually, accidentally!

The argument in Gerovitch’s From Newspeak to Cyberspeak is that, when computer science papers in the 1960s were translated from English to Russian, they were stripped of metaphorical yet inspirational ideas like “memory” and “learning”, constraining the vision of computing to simple calculation. Which is why the Americans figured out the personal computer, our bicycle for the mind, whereas the Soviets did not.

Could you popularise an AI that made conceptual leaps around worker-friendly capitalism much harder? (For example, given that policy makers will be heavy users of future ChatGPTs, and this trend will slowly lead to social unrest.)

Could you wait until a nation were in an intellectual arms race, like a Space Race for the 2030s, say, then knock over the intelligence augmentation infrastructure (i.e. ChatGPT v9) in critical weeks?

I’m not saying that this is what is happening. But any government worth its salt should have a half dozen people figuring out how to perform an IQ erosion attack, precision targeted or otherwise, and another half dozen red-teaming how to respond if one hits.

More posts tagged:
Follow-up posts:

If you enjoyed this post, please consider sharing it by email or on social media. Here’s the link. Thanks, —Matt.