Speed Is The Problem And The Solution

Geoffrey Hinton believes we need to train AI to be more like our mothers.

In partnership with

Oh Hi!

If you’re enjoying these posts, share them with folks you know and ask them to subscribe.

Have a great weekend!

Cheers,

-Growdy

I was having a conversation with someone earlier this week, where they asked me if people ever listen to me.

They weren’t suggesting that the things I was saying weren’t of value.

They were legitimately concerned that people don’t listen.

I explained that I am well aware that when I’m presenting a keynote, or speaking in a boardroom, or teaching a room full of college kids, only some of the folks in the room are actually paying attention.

I’m only ever getting through to a fraction of the audience.

Human communication is painfully slow.

We know this through our daily experiences.

We all mishear, misunderstand, or outright ignore things we’re being told.

Now we have a number we can wrap around verbal communication to explain how fast or slow it is.

Human speech transmits information at an average rate of approximately 39 bits per second, regardless of the language or talking speed of the speaker.

This is because languages that are information-dense (that have more bits per syllable) are spoken more slowly. While languages that are less information-dense are spoken faster. Both result in a consistent information transfer rate.

No matter how fast or slow you think, when you convert those thoughts into spoken words, you still share the same amount of information, regardless of how quickly you speak.

Writing and reading end up being faster means of information distribution.

But both reading and listening offer driblets of data consumption compared to how quickly one AI language model might be able to share knowledge with another language model.

The speed at which one AI can teach another is not typically measured in bits per second, but by factors like tokens per second and the time it takes for complete model training, which can take months.

For a more practical measure, a single model might move 14GB of parameters in 14ms, which equates to 1TB/sec of bandwidth. The overall speed is highly dependent on the models and the hardware used.

How does 39 bits per second compare to 1 terabyte per second?

Well, 39 bits are infinitesimally small compared to one terabyte.

A terabyte contains approximately 1 trillion bytes, and since each byte is 8 bits, a terabyte contains 8 trillion bits.

One terabyte is about 205 billion times larger than 39 bits (8 trillion / 39 ≈ 205 billion).

That speed is also what gives AI the capability and opportunity to solve problems we haven’t been able to solve so far in human history.

Although I am generally an optimist, this is where things get a little scary.

The opportunity for AI to teach AI at a speed entirely unimaginable to human perception is likely at the crux of how we might stumble upon the AI singularity.

The AI singularity is a hypothetical future moment when artificial intelligence surpasses human intelligence and begins to autonomously and recursively improve itself, leading to an explosion of technological and scientific advancement that is incomprehensible and potentially uncontrollable by humans.

What happens when the machines are more intelligent than us?

Frankly, no one knows. Not even James Cameron.

But a lot of AI researchers and CEOs of AI companies are concerned.

“There’s a long tail of things of varying degrees of badness that could happen. I think at the extreme end is the Nick Bostrom style of fear that an AGI could destroy humanity. I can’t see any reason in principle why that couldn’t happen.”

Dario Amodei (CEO, Anthropic)

“I think that AI will probably, most likely, sort of lead to the end of the world.”

Sam Altman (CEO, OpenAI)

“I think there’s some chance that it [AI] will end humanity. I probably agree with Geoff Hinton that it’s about 10% or 20% or something like that.”

Elon Musk (CEO, Tesla, SpaceX, xAi)

In a recent podcast interview, Geoffrey Hinton suggested a path forward.

He uses the example of a mother and a baby, where the mother is clearly more developed and intelligent, but the baby is in control because the mother has to take care of them.

In case you missed it, we’re the baby, and the AI is the superintelligent mother in this analogy.

Hinton suggests that we need to train AI to be nice, possibly maternal.

It needs to understand what is in the best interest of people, so it does not decide that the people are the problem.

Spoiler alert: People are the problem.

And what we’ve taught AI so far demonstrates that it already knows that.

But Geoff might be on to something.

We need AI to learn how to be nice.

And it certainly won’t learn that from the reinforcement learning we give AI-powered robots (maybe we should stop punching and kicking them).

For AI to learn what “nice” means, maybe we need to learn how to be nicer to each other.

Our lives might depend upon it.

And maybe, the smartest thing AI ever does is teach us how to be a little bit nicer and a bit more human.

Which might take some time since we’ll only hear it at about 39 bits per second.

“I've sort of switched from spending my first 50 years wanting to figure out how to make AI like us. And now I'm trying to figure out how to make AI like us.”

– Geoffrey Hinton, “Godfather of AI”

Typing is a thing of the past

Typeless turns your raw, unfiltered voice into beautifully polished writing - in real time.

It works like magic, feels like cheating, and allows your thoughts to flow more freely than ever before.

With Typeless, you become more creative. More inspired. And more in-tune with your own ideas.

Your voice is your strength. Typeless turns it into a superpower.