Artificial general intelligence is far from being a reality because AI systems of today do not use intrinsic knowledge about the world to predict and plan
Kyle Wiggers / VentureBeat : Tweets: @bradfordcross and @bgurley Tweets: Bradford Cross / @bradfordcross : An important starting point for lay people looking into AI is understanding that we do not have a consensus definition of what ‘intelligence’ means or how to test it. Nonsense about AGI can take a back seat until we define the ‘I’ or the ‘GI’...then we can talk about the ‘A’ ;-) http://twitter.com/... Bill Gurley / @bgurley : From all of my reading this is the consensus among scientists in the field. So surprising that people chose to unnecessarily spread dystopian fear on this subject. Geoffrey Hinton and Demis Hassabis: AGI is nowhere close to being a reality http://venturebeat.com/...
Context & Ripple Effects
In late 2018 the skeptical case against AGI was close to unanimous among the voices here: Bradford Cross argued there is no consensus definition of 'intelligence' to test against, Bill Gurley called AGI skepticism the scientific consensus and attacked dystopian fearmongering, and Geoffrey Hinton and Demis Hassabis agreed AGI is nowhere close. VentureBeat's technical framing was that current systems lack the intrinsic world knowledge needed to predict and plan.
Five years of subsequent coverage complicate that consensus rather than refute it outright: Noema now argues advanced LLMs from OpenAI and others will be recognized decades hence as the first true AGI, much as ENIAC was for computing, while OpenAI itself operates as if successfully creating AGI carries existential risk. The sharpest counterpoint comes from The Road to AI We Can Trust, which warns that discounting present-day harms because AGI hasn't arrived leaves society exposed to phishing, fraud, and propaganda now.
First-order effects
- For investors and builders taking their cue from Gurley and Cross, the actionable takeaway in 2018 is to ignore AGI timelines entirely and fund narrow applications — a stance that made sense when deep learning systems were, per MIT Technology Review, neither understandable to creators nor accountable to users.
- Hinton and Hassabis lending their names to the 'nowhere close' position gives labs and policymakers cover to treat AGI as a non-issue in near-term planning.
Second-order effects
- Once capable LLMs arrive, the same 'AGI isn't here' logic gets weaponized in both directions: skeptics use it to dismiss regulation, while labs like OpenAI justify extreme caution by preparing for AGI as an open possibility.
- Cross's definitional objection ages into the central battleground — without an agreed test for 'intelligence', claims like Noema's ENIAC analogy can only be settled retrospectively, keeping the debate unfalsifiable.
Third-order effects
- If the pattern holds, AGI becomes a moving definitional goalpost: each generation of systems is dismissed as 'not real intelligence' until hindsight reclassifies it, meaning governance built around an AGI arrival date will always be mistimed.
- The durable lesson for policy is the one The Road to AI We Can Trust draws — regulate the harms systems actually cause today rather than gating action on a consensus definition of general intelligence that may never come.
The trend: The AGI debate is shifting from a scientific consensus that general intelligence is distant toward a contest over definitions, as each new system generation forces retrospective reclassification of what counts as AGI.