Hypesters sometimes pretend that AGI has already been achieved, but that’s not certainly true by traditional understandings of what that term means.
A new paper from Google DeepMind, Waterloo, ANU and UCL makes clear that what the new paper calls “competent AGI” has NOT yet been achieved, let alone what they are calling “expert” or “superhuman AGI”.
I fully concur; all claims to the contrary are just marketing.
Users dismissed claims about current models like ChatGPT reaching AGI because they lack world models, common sense, and rely on workarounds rather than genuine capabilities.
No Digg Deeper questions have been answered for this story yet.
https://arxiv.org/pdf/2606.12683
Hypesters sometimes pretend that AGI has already been achieved, but that’s not certainly true by traditional understandings of what that term means.
A new paper from Google DeepMind, Waterloo, ANU and UCL makes clear that what the new paper calls “competent AGI” has NOT yet been achieved, let alone what they are calling “expert” or “superhuman AGI”.
I fully concur; all claims to the contrary are just marketing.
@GaryMarcus Is it a certainty that we will ever achieve competent, expert or superhuman AGI?
@GaryMarcus so competent agi just means the definition moved up a level again
next year well hear we need "execution capable AGI"
@GaryMarcus And even calling ChatGPT or Claude *any kind of AGI* is a leap.
No world model, no common sense, propped up by overrides, no understanding of its own outputs, mathematically certain to hallucinate, and can't reliably follow even given "core" instructions like guardrails.
You have to watch the RoboCop movies to contemplate that AGI may never ever exist, but that an AI augmented human brain may be where peak intelligence is actually achievable.
Nature plus tools gave us today's world.
I'm not seeing anything today that is leaning towards a tools only world where natural intelligence is superceded by AI tools on their own.
@GaryMarcus I don't understand.
First, Big Sam told me to feel the AGI.
Then they told me scaling laws didn't require continued, exponential expense for diminishing returns.
Then they told me that GPT 5 was AGI.
Then they told me Mythos was too scary to let out, but let it out anyway.
@GaryMarcus None of these things are possible with the current 2022 architecture they are still recycling.
@suzynuze11 @GaryMarcus Under the pure LLM scaling model, it's looking likely that we will NEVER get as far as "expert", except in select disciplines.
LLMs show diminishing returns for ruinous costs, and are mathematically unable to be 100% accurate.
@GaryMarcus Correct. And worth noting: the DeepMind taxonomy has no axis for intent verifiability.
Not what the system can do. Whether you can determine -- not just detect -- what it actually intends.
Capability levels 0-5 don't close that gap. They compound it.
@gonlenidefi @GaryMarcus 2028: "Fixable AGI"
"AGI you can give a task to in the trust that when it fucks it up, it does so in a way that you can at least un-fuck -it-up."
@suzynuze11 @GaryMarcus no, but nonetheless it's likely because the incentives are huge
@GaryMarcus Hmmm there should probably be something between emerging and competent, because a multi-agent coding agent swarm that also solves unsolved math problems is on another level than the very stupid 2023 chatgpt.
@GaryMarcus 🤣 AGI = AI + man