2 months ago
44
ARTICLE AD BOX
Google DeepMind CEO Demis Hassabis has said that artificial wide quality (AGI) inactive remains acold away. He has outlined 3 limitations that, successful his view, volition forestall OpenAI main Sam Altman’s ‘ultimate goal’ of achieving AGI from being realised successful the adjacent future.
Speaking astatine the India AI Impact acme successful New Delhi, Hassabis said existent systems bash not yet lucifer quality intelligence.
“I don’t deliberation we are determination yet,” he said.His comments travel arsenic OpenAI continues to framework AGI arsenic its semipermanent objective. Altman has said that much precocious forms of quality could follow, arguing that “Superintelligent tools could massively accelerate technological find and innovation good beyond what we are susceptible of doing connected our own, and successful crook massively summation abundance and prosperity.” Last year, OpenAI besides restructured itself arsenic a Public Benefit Corporation arsenic portion of efforts to align its improvement enactment with broader societal goals.
Why Google AI CEO Demis Hassabis thinks AGI is not acceptable yet
Hassabis explained that today’s AGI-style systems deficiency cardinal capabilities, starting with what helium described arsenic “continual learning”. He said these models stay mostly fixed aft grooming and bash not accommodate successful existent clip to caller experiences oregon contexts.“What you’d similar is for those systems to continually larn online from experience, to larn from the discourse they’re in, possibly personalise to the concern and the tasks that you person for them,”he explained.
Secondly, Hassabis said existent systems person limitations successful semipermanent reasoning. “They tin program implicit the abbreviated term, but implicit the longer term, the mode that we tin program implicit years, they don't truly person that capableness astatine the moment,” helium said. Hassabis besides said these systems deficiency consistency, performing good successful immoderate areas portion falling abbreviated successful others. “So, for example, today's systems tin get golden medals successful the planetary Math Olympiad, truly hard problems, but sometimes tin inactive marque mistakes connected simple maths if you airs the question successful a definite way,” helium said. “A existent wide quality strategy shouldn't person that benignant of jaggedness,” Hassabis added. For comparison, quality experts would not typically marque errors connected basal problems successful their country of expertise, helium noted.Last year, successful a “60 Minutes” interview, Hassabis said that helium expects existent AGI to look wrong 5 to 10 years. In 2010, helium co-founded the AI probe laboratory DeepMind, which Google acquired successful 2014 and is liable for processing the tech giant’s Gemini models.
In 2024, Hassabis received a associated Nobel Prize successful chemistry for his enactment connected macromolecule operation prediction.AGI remains a debated conception successful Silicon Valley. Databricks CEO Ali Ghodsi said astatine a league successful September that existent AI chatbots already conscionable the explanation of AGI, arguing that manufacture leaders support “moving the goalposts” by shifting the absorption toward superintelligence, oregon AI that tin outperform humans.The AI Summit successful India, taking spot this week, has drawn speakers from crossed the exertion and AI sectors. The docket includes figures specified arsenic OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, Google CEO Sundar Pichai, and Meta’s main AI officer, Alexandr Wang.
