A note about intelligence

These posts are my attempt to make sense of the evolution of AI. This one jumps from the nature of intelligence to the latest market updates without smooth transitions — and I'm OK with that.

A note about intelligence

Let's start with the obvious. LLMs are models based on language, and that language must already exist. Language is how we explain our current understanding of things — so LLMs are bounded by the edges of our own knowledge. Even a Super-LLM of all human knowledge wouldn't advance beyond it. AI, in its current incarnation, is limited to our own domain of language.

As we evolve our understanding of the universe, we evolve our language to help us communicate, understand and think. Where this gets funky is that the quest for meaning is motivated by hard-to-define things — ambition, imagination, purpose — that aren't necessarily tied to language. Would imagination exist without language? At what point does AI need something 'else' beyond language to become self-motivating and self-directing?

The thing is, AI's 'intelligence' is not the same as the human mind's. At its core, AI isn't thinking, feeling or reflecting — it's computing, finding patterns in data, making predictions and generating outputs against predefined goals. Natural intelligence gets varied definitions; here's one for size.

Intelligence is one's ability to learn from experience and to adapt to, shape, and select environments.

In AI, intelligence is the capacity to learn from data, adapt to new information and optimise decision-making. These systems have become incredibly adept at narrow tasks — translation, image recognition, even creating art — and increasingly at wider inter-system tasks such as compiling applications.

Google frames AI by development stage: reactive machines (IBM's Deep Blue beating Kasparov), limited memory (most modern AI, improving over time through neural networks), theory of mind (which doesn't yet exist — AI that emulates the human mind and reacts to emotions), and self-aware AI (a mythical machine aware of its own existence). We're still a long way from true general intelligence. All current AI is artificial 'narrow' intelligence: a narrow set of actions based on its prompting, programming and training. Artificial general intelligence — to sense, think and act like a human — does not exist; artificial superintelligence, superior in all ways, less so still.

Looking ahead, AI will evolve in two streams: increasing task specialisation, and a parallel push toward general-purpose systems. We're seeing incredible advances in models like GPT-4o and Gemini, but the leap to AI that understands context and nuance like humans is still several cross-domain breakthroughs away — across language, computing, neuroscience, behavioural psychology, anthropology and philosophy.

The floodgates for AI in decision-making roles are upon us. AI might not be replacing human intelligence any time soon, but it is becoming an essential tool in augmenting it — and that in itself is remarkable. Maybe it will give us more time to think about what is really driving our search for understanding. And there, I need to get back to the cold facts of the market.

Interoperability: deploying AI across systems

If you thought GenAI was merely linearly agentic, look at Anthropic's recent launch — Claude generating not just code but all the tasks required to develop and deploy a site or application. This goes far wider than coding: it's about taking control of a human-operated system to perform inter-system tasks toward a goal. They're calling it Computer Use, demoed compiling a vendor form across multiple systems. The pace into interoperability is super cool.

Mini models: edge AI

On the other end of the spectrum, edge AI: driven by the need for local, privacy-first inference (on-device translation, internet-less smart assistants), Mistral released Les Ministraux, outperforming competitors across benchmarks. This is more fuel for an ecosystem where different scales of model are deployed appropriately for their context — mini-edge models integrated into larger language models, determined partly by compute requirements at the point of execution. As I said back in August: every company will have an AI, but not a single static LLM.

Other market news

• HeyGen raised a $60M Series A.
• ChatGPT rolled out Search in beta.
• Grok / xAI released its API.

A shout-out to the start-ups

Starting up is hard, and it never gets easier; you learn the systems and gain the networks, but the risk is always there alongside the blood, sweat and tears. I give away a bit of time each month mentoring founders at New Native's AI accelerator, and it's humbling to be around people starting up again — particularly first-timers up against investors asking all the hard questions. A big shout-out to the founders out there. And please keep enticing your mentors with swag.