A Google DeepMind preprint maps continuous routes from human-level AGI to artificial superintelligence through four pathways—scaling current systems, paradigm shifts, recursive self-improvement, and multi-agent collectives—while noting associated technical and practical bottlenecks that could alter or delay the process.
How do we go from AGI to Superintelligence? New report discusses four potential pathways: scaling, AI paradigm shifts, recursive improvement, and ASI emerging from large-scale multi- agent collectives. Importantly, it also looks at possible frictions and bottlenecks along these pathways. Instant classic! https://arxiv.org/abs/2606.12683
The analysis flags bottlenecks without spelling out their precise mechanics or timelines, leaving open how severely they might constrain any given pathway.
Researchers focused on governance and AI safety have shared the work, suggesting the framework may feed into strategic planning even before peer review.
Many users praised DeepMind's report mapping pathways from AGI to superintelligence for its insightful framing of bottlenecks like verification and its optimistic outlook on future developments such as artificial ultra intelligence.
Well this goes instantly on the to-read list.
How do we go from AGI to Superintelligence? New report discusses four potential pathways: scaling, AI paradigm shifts, recursive improvement, and ASI emerging from large-scale multi- agent collectives. Importantly, it also looks at possible frictions and bottlenecks along these pathways. Instant classic! https://arxiv.org/abs/2606.12683
AGI is sometimes framed as a single end point; more likely are successive "waves" of technological transformation... 🌊🌊🌊🌊
For more on this topic—and the possibility of Artificial Superintelligence—check out this @GoogleDeepMind report, led by Tim Genewein!
How do we go from AGI to Superintelligence? New report discusses four potential pathways: scaling, AI paradigm shifts, recursive improvement, and ASI emerging from large-scale multi- agent collectives. Importantly, it also looks at possible frictions and bottlenecks along these pathways. Instant classic! https://arxiv.org/abs/2606.12683
Looks like good reading for my train ride tomorrow.
How do we go from AGI to Superintelligence? New report discusses four potential pathways: scaling, AI paradigm shifts, recursive improvement, and ASI emerging from large-scale multi- agent collectives. Importantly, it also looks at possible frictions and bottlenecks along these pathways. Instant classic! https://arxiv.org/abs/2606.12683
@GjMcGowan , my daughter and I will be reading it together, in case she decides she wants to recursively self-improve once she's of age.
Well this goes instantly on the to-read list.
@sebkrier very nice
@mhutter42 @OriolVinyalsML @willmacaskill @Yoshua_Bengio @DavidChalmer @paulfchristiano @dwarkesh_sp @a_korinek @tegmark @karpathy @RichardMCNgo AI paradigm shifts! Surely not!🤭
@sebkrier TO BE CRYSTAL-CLEAR: IN AN ASI WORLD THERE WILL BE NO HUMANS. (If you know anything about ecology, biology, natural selection and the competitive exclusion principle you will accept this FACT)
@sebkrier This is great.
@sebkrier Looking forward to AUI . Artificial ultra intelligence
@S_OhEigeartaigh She's already recursively self-improving!! But she'll cap out at human level, I'm almost certain
I don't think AIXI is the right "superintelligence" limit of practical interest. It's motivated by a philosophical quest to define intelligence but the actual limiting process of interest is more ... economic (?). The most glaring problem is that it treats rewards as a given when reward systems + generalization + the rest are all part of the dynamics we want the limit of.
(the AIXI stuff seems a bit independent of the rest of the discussion in any case)
@mhutter42 @OriolVinyalsML @willmacaskill @Yoshua_Bengio @DavidChalmer @paulfchristiano @dwarkesh_sp @a_korinek @tegmark @karpathy @RichardMCNgo Be nice if we could stop dodging definitions of the first before attempting the second.
Also, legitimate question, where might you be wrong about this acceleration, theoretically speaking?
This posts sounds like a foregone conclusion.
@S_OhEigeartaigh i'm a big fan of its suggestion to feed the paper into a personal ai system for better contextualisation
@anderssandberg Friction points and bottlenecks seem like the most pertinent things to spend thinking time on. AGI/ASI will almost certainly happen, barring the bottom dropping out of NVIDIA. Figuring out what, specifically, a given ASI will be able to do better than us, seems important.
@mhutter42 @OriolVinyalsML @willmacaskill @Yoshua_Bengio @DavidChalmer @paulfchristiano @dwarkesh_sp @a_korinek @tegmark @karpathy @RichardMCNgo Fun!
@sebkrier
@sebkrier Yeah, that framing feels right. Recursive improvement is the one where verification seems like the hard bottleneck: knowing each change actually helps before the loop compounds.