The one factor protecting AI from full AGI? Consistency, stated Google DeepMind CEO Demis Hassabis.Hassabis stated on an episode of the “Google for Builders” podcast revealed Tuesday that superior fashions like Google’s Gemini nonetheless stumble over issues most schoolkids may resolve.”It should not be that simple for the common particular person to only discover a trivial flaw within the system,” he stated.He pointed to Gemini fashions enhanced with DeepThink — a reasoning-boosting method — that may win gold medals on the Worldwide Mathematical Olympiad, the world’s most prestigious math competitors.However those self same methods can “nonetheless make easy errors in highschool maths,” he stated, calling them “uneven intelligences” or “jagged intelligences.””Some dimensions, they’re actually good; different dimensions, their weaknesses may be uncovered fairly simply,” he added.Hassabis’s place aligns with Google CEO Sundar Pichai, who has dubbed the present stage of improvement “AJI” — synthetic jagged intelligence. Pichai used this time period on an episode of Lex Fridman’s podcast that aired in June to explain methods that excel in some areas however fail in others.Hassabis stated fixing AI’s points with inconsistency will take greater than scaling up information and computing. “Some lacking capabilities in reasoning and planning in reminiscence” nonetheless have to be cracked, he added.He stated the business additionally wants higher testing and “new, tougher benchmarks” to find out exactly what the fashions excel at, and what they do not.Hassabis and Google didn’t reply to a request for remark from Enterprise Insider.
Please assist BI enhance our Enterprise, Tech, and Innovation protection by sharing a bit about your function — it’s going to assist us tailor content material that issues most to folks such as you.
What’s your job title?
(1 of two)
What services or products are you able to approve for buy in your function?
(2 of two)
Proceed
By offering this info, you agree that Enterprise Insider might use
this information to enhance your website expertise and for focused promoting.
By persevering with you agree that you just settle for the
Phrases of Service
and
Privateness Coverage
.
Thanks for sharing insights about your function.
Massive Tech hasn’t cracked AGIBig Tech gamers like Google and OpenAI are working towards reaching AGI, a theoretical threshold the place AI can motive like people.Hassabis stated in April that AGI will arrive “within the subsequent 5 to 10 years.”AI methods stay vulnerable to hallucinations, misinformation, and fundamental errors.OpenAI CEO Sam Altman had the same take forward of final week’s launch of GPT-5. Whereas calling his agency’s mannequin a big development, he instructed reporters it nonetheless falls in need of true AGI.”That is clearly a mannequin that’s typically clever, though I feel in the way in which that almost all of us outline AGI, we’re nonetheless lacking one thing fairly vital, or many issues fairly vital,” Altman stated throughout a press name on Wednesday earlier than the discharge of GPT-5.Altman added that a type of lacking parts is the mannequin’s capacity to study independently.”One large one is, , this isn’t a mannequin that constantly learns because it’s deployed from the brand new issues it finds, which is one thing that to me appears like AGI. However the stage of intelligence right here, the extent of functionality, it appears like an enormous enchancment,” he stated.
