Artificial general intelligence is a theory, a hunch and a belief. Nobody has a map from here to there, because we don't know what makes intelligence general.
There are many definitions, and one of them has money attached. OpenAI defines AGI as highly autonomous systems that outperform humans at most economically valuable work, and its contract with Microsoft depends on that line. A definition in a contract can't be vague, so you can be sure this is what they work toward: replacing labour with software.
The philosopher Joscha Bach describes intelligence as a set of vectors, among them knowledge, language, learning, reasoning, perception, creativity and autonomy. Today's models score high on some and low on others. For AGI to be real, all of them have to come up together.
My own bet is that AGI will not be a single bigger language model. It will be several kinds of AI working together: reactive, predictive, generative and autonomous systems, plus kinds we have not seen yet, with one able to call on another at machine speed. That will look general, and it will be.
Google DeepMind proposed levels: emerging, competent, expert, virtuoso and superhuman, where a superhuman system outperforms every person. Past human level is superintelligence, and what lies beyond that is anyone's guess. It is a singularity in the literal sense: we can't see over the edge.
Forecasts keep moving closer. In 2024 the Metaculus community put AGI around 2032, down from the 2040s a few years before. The paper that made me start a newsletter was Sparks of AGI (2023), where Microsoft researchers argued that GPT-4 already reasoned beyond its instructions.
Some companies already run AGI-like. Ant Group serves around a billion users with roughly 10,000 to 15,000 employees. Its lending follows the 3-1-0 rule: three minutes to apply, one second to approve, zero humans involved.
The computer scientist Edsger Dijkstra said the question of whether machines can think is about as relevant as the question of whether submarines can swim. Maybe that is the right attitude to AGI as well.
Since 2024
Reasoning models pushed scores on hard maths and science tests far past where they were when I recorded this talk. The argument has moved from whether AGI is coming to when, and to who decides that it has arrived.
Sources
- OpenAI, Charter
- Morris et al., Levels of AGI (Google DeepMind, 2023)
- Bubeck et al., Sparks of Artificial General Intelligence (2023)
- Metaculus, Date of artificial general intelligence
- Sam Altman, The Intelligence Age (2024), the essay behind the recording's title
- Edsger W. Dijkstra on Wikiquote