An interview with DeepMind's Shane Legg and Meredith Ringel Morris, who propose five levels of AGI taxonomy and say the first, “emerging”, has been achieved
Will Douglas Heaven / MIT Technology Review : Mastodon: @michellemanafy@journa.host . X: @mlamons1 , @the_magrathean , and @elieraad Mastodon: Michelle Manafy / @michellemanafy@journa.host : Google DeepMind researchers put out a paper that cuts through the cross talk with not just one new definition for artificial general intelligence, but a whole taxonomy of them. “AGI must not only be able to do a range of tasks, it must also be able to learn how to do those tasks, assess its performance, and ask for assistance when needed. … X: Matthew Lamons / @mlamons1 : What is Artificial General Intelligence? At the moment, it depends on who you ask, but Google DeepMind researchers published a paper giving AGI a whole taxonomy. #AI #ML #futurism #IntelligenceFactory #digitaltransformation #DX https://www.technologyreview.com/ ... @the_magrathean : ‘AGI typically means artificial intelligence that matches (or outmatches) humans on a range of tasks’ No it doesn't. Shifting the goal posts yet again. First they redefined AI to mean ‘chatbot’, now AGI is being hastily redefined. https://www.technologyreview.com/ ... Elie R𐤀𐤀d / @elieraad : As AI capabilities advance, we no longer feel the need to precisely define or categorize systems as “AI” (past 2023). The focus will likely shift from AI to AGI. https://www.technologyreview.com/ ...
Context & Ripple Effects
The interview follows Microsoft’s “early signs” AGI argument for GPT-4, reflecting a growing effort by frontier labs to attach milestones to a term that had been used loosely across very different capability claims.
DeepMind’s five-level framework makes the debate less about a single finish line and more about which capabilities—learning, self-assessment, and seeking help—should count at each stage. That framing anticipates OpenAI’s own five-level progress framework, even though the firms define the path differently.
First-order effects
- DeepMind gains a public vocabulary for presenting its systems as having reached an “emerging” AGI threshold without claiming full AGI, while researchers and commentators get a more granular basis for challenging that assessment.
- The taxonomy raises the importance of evidence for meta-capabilities such as learning, evaluating performance, and requesting assistance—not simply broad task performance.
Second-order effects
- Rival labs face pressure to state their own thresholds and evaluations, rather than relying on broad AGI rhetoric; OpenAI’s later staged model illustrates how capability ladders can become competitive positioning tools.
- Buyers, policymakers, and evaluators may increasingly distinguish between general-purpose models that perform many tasks and systems that can reliably adapt, monitor themselves, and escalate when needed.
Third-order effects
- If multiple labs and external evaluators converge on staged definitions, AGI may become governed as a sequence of capability thresholds rather than a binary label—though competing taxonomies could also preserve ambiguity.
- The durable shift is toward institutionalized measurement of frontier-model capabilities, where claims of progress invite demands for standardized tests and clearer assurance practices.
The trend: AGI discourse is moving from headline-level declarations toward staged capability frameworks that can organize competition, evaluation, and governance.