Unlimited, almost free intelligence: What does it mean?

Capability improvement in generative AI is an important story. But another, perhaps more overlooked one is cost-reduction:

The chart above shows cost reduction in GPT-3 and GPT-3.5 class models. A year ago, these were state of the art. Today, using Mixtral, a GPT-3.5 quality open-source model, price has dropped 33X. A long blog post now costs fractions of a penny–$0.0006.

The cost of model output should drop to the marginal cost of electricity. Once you train a model and build a datacenter, generating tokens just takes power. What might that mean for the future?

A good analogy for what's happening is digital photography. When digital cameras first came out, quality didn't match film. But they won out for cost and convenience. Then they improved over time.

Today, photos are so cheap we never think about cost. This destroyed traditional film companies like Kodak. And it enabled things we never previously considered, like Instagram, cloud photo storage, and image models like Midjourney.

The same will be true for widespread, nearly free intelligence on demand. We can't be sure what will happen, but I'd guess it will include:

• Accelerated job disruption. Even second-tier, GPT-3.5 class models can perform quality work that displaces humans. Recent research, for example, shows that translator earnings on online job platforms dropped 30.2% after the launch of ChatGPT. This will accelerate as models get better and cheaper while education gets ever more expensive.
• AI-first solutions. Why build complex applications or interfaces when you can just throw AI at a problem? It's cheaper, and you benefit from constantly improving models. This could undermine the value of SaaS companies and reduce the complexity of software builds.
• "Infinite monkeys" opportunities. When intelligence becomes cheap, you can solve problems by generating a lot of output and seeing what sticks. This is partly how Google's AlphaCode 2 solves complex code challenges. It generates up to a million code samples per problem, then filters down to the best options.
• Frivolous (seemingly) use. Just as we take more frivolous photographs now that it's too cheap to meter, so too will we frivolously use cheap intelligence. ChatGPToaster that knows the right settings for your food, and can write a poem about it? Why not.

I could go on (and probably will in other posts)—more spam, smarter spam blockers, personalized everything, and so on—but hopefully that's enough to get others thinking about this as well.

What might you do with unlimited, nearly free, on-demand intelligence? Because it's already here.