AI revolution.

 

There’s a colossal shift going on in arti­fi­cial intel­li­gence — but it’s not the one some might think. While advanced lan­guage-gen­er­at­ing sys­tems and chat­bots have dom­in­ated news head­lines, private AI com­pan­ies have quietly entrenched their power. Recent devel­op­ments mean that a hand­ful of indi­vidu­als and cor­por­a­tions now con­trol much of the resources and know­ledge in the sec­tor — and will ulti­mately shape its impact on our col­lect­ive future.

The phe­nomenon, which AI experts refer to as «indus­trial cap­ture», was quan­ti­fied in a paper pub­lished by research­ers from the Mas­sachu­setts Insti­tute of Tech­no­logy in the journal Sci­ence this month, call­ing on poli­cy­makers to pay closer atten­tion. Its data is increas­ingly cru­cial. Gen­er­at­ive AI — the tech­no­logy under­ly­ing the likes of Chat­GPT — is being embed­ded into soft­ware used by bil­lions of people, such as Microsoft Office, Google Docs and Gmail. And busi­nesses from law firms to the media and edu­ca­tional insti­tu­tions are being upen­ded by its intro­duc­tion.

The MIT researchers found that almost 70 per cent of AI PhDs went to work for com­pan­ies in 2020, com­pared with 21 per cent in 2004. In par­tic­u­lar, he said aca­dem­ics were unable to build large lan­guage mod­els like GPT-4, a type of AI soft­ware that gen­er­ates plaus­ible and detailed text by pre­dict­ing the next word in a sen­tence with high accur­acy. The tech­nique requires enorm­ous amounts of data and com­put­ing power that primar­ily only large tech­no­logy com­pan­ies like Google, Microsoft and Amazon have access to. Ahmed found com­pan­ies’ share of the biggest AI mod­els had gone from 11 per cent in 2010 to 96 per cent in 2021.

A lack of access means research­ers can­not rep­lic­ate the mod­els built-in cor­por­ate labs and can, there­fore, neither probe nor audit them for poten­tial harms and biases very eas­ily.

The paper’s data also showed a big dis­par­ity between pub­lic and private invest­ment in AI tech­no­logy. In 2021, non-defence US gov­ern­ment agen­cies alloc­ated $1.5bn to AI. The European Com­mis­sion planned to spend €1bn. Mean­while, the private sec­tor inves­ted more than $340bn on AI in 2021.

In 2019, OpenAI pivoted from a non­profit into a profit­mak­ing enter­prise with a $1bn invest­ment from Microsoft, cit­ing a need «to increase our invest­ments in com­pute and tal­ent rap­idly».

The con­sequences of this shift are man­i­fold. First, it means pub­lic altern­at­ives to cor­por­ate AI tech, such as mod­els and data sets, are becom­ing increas­ingly scarce. And new applic­a­tions are likely to be com­mer­cially driven rather than in the broader pub­lic interest, sev­eral research­ers poin­ted out. Hanna, whose work is fun­ded by non-profits, agrees. «If you want to work on spe­cial­ised AI tasks like ensur­ing biod­iversity, or cli­mate sci­ence or agri­cul­ture, there is not a lot of appet­ite for that,» she said.

Meredith Whit­taker, pres­id­ent of encryp­ted app Sig­nal, has com­pared the situ­ation with the US mil­it­ary’s dom­in­ance over sci­entific research dur­ing the cold war in a sem­inal paper in 2021. «It is here, in these darker his­tor­ies, that we con­front the steep cost of cap­ture — whether mil­it­ary or indus­trial,» she wrote. «And its per­il­ous implic­a­tions for aca­demic free­dom . . . cap­able of hold­ing power to account».

Resoomed from original article of Mrs Madhumita Murgia Financial Times