On No­vem­ber 30 2022, a sim­ple so­cial me­dia post in­vit­ed peo­ple to in­ter­act with an AI chat­bot. It marked the day that Ope­nAI re­leased Chat­G­PT to the pub­lic. In five days, the chat­bot gained one mil­lion users and crossed 100 mil­lion with­in two months, mak­ing it the fastest adopt­ed con­sumer ap­pli­ca­tion in his­to­ry.

To­day, rough­ly 800 mil­lion peo­ple use the plat­form every week. That means one tenth of the glob­al pop­u­la­tion now reg­u­lar­ly con­sults an ar­ti­fi­cial in­tel­li­gence for help with their work, their ques­tions, or their dai­ly lives.

Sun­dar Pichai, the chief ex­ec­u­tive of Google’s par­ent com­pa­ny Al­pha­bet, calls ar­ti­fi­cial in­tel­li­gence the most pro­found tech­nol­o­gy hu­man­i­ty has ever worked on. When he warns that peo­ple will need to adapt, he is not ex­ag­ger­at­ing, but he is al­so not telling us the whole sto­ry. The pub­lic con­ver­sa­tion about AI fo­cus­es on jobs and au­toma­tion. How­ev­er, some­thing big­ger is hap­pen­ing around AI. That’s what we are go­ing to dis­cuss to­day.

Here is the sim­plest way to un­der­stand what has changed: for the first time in his­to­ry, ma­chines can now process in­for­ma­tion, gen­er­ate ideas and solve prob­lems faster than any hu­man or hu­man in­sti­tu­tion pos­si­bly could.

With prop­er in­puts, an AI sys­tem can read all your doc­u­ments, analyse every is­sue you are deal­ing with and draft re­spons­es to your queries, while you are hav­ing break­fast and prepar­ing for work in the morn­ing. This is not just a faster tool. It is a com­plete­ly dif­fer­ent kind of ca­pa­bil­i­ty, and who­ev­er con­trols it gains an ex­tra­or­di­nary ad­van­tage.

Gen­e­sis

On No­vem­ber 24, the Trump White House, is­sued an ex­ec­u­tive or­der launch­ing some­thing called the Gen­e­sis Mis­sion. The or­der it­self com­pares this ini­tia­tive, in ur­gency and am­bi­tion, to the Man­hat­tan Project. If you re­call this was the se­cret pro­gramme that built the atom­ic bomb dur­ing World War II.

Think about what that com­par­i­son means. The Amer­i­can gov­ern­ment is now treat­ing ar­ti­fi­cial in­tel­li­gence the way it once treat­ed nu­clear weapons: as a tech­nol­o­gy so pow­er­ful that na­tion­al se­cu­ri­ty de­pends on who mas­ters it first. The Gen­e­sis Mis­sion brings to­geth­er na­tion­al lab­o­ra­to­ries, top uni­ver­si­ties, and na­tion­al se­cu­ri­ty fa­cil­i­ties in­to a sin­gle co­or­di­nat­ed ef­fort. Its stat­ed goal is to achieve what the or­der calls glob­al tech­nol­o­gy dom­i­nance.

This is not just an­oth­er pol­i­cy an­nounce­ment or fund­ing pro­gramme. It is an of­fi­cial ad­mis­sion that AI has crossed a line from be­ing a use­ful per­son­al and busi­ness tool to an in­stru­ment of na­tion­al se­cu­ri­ty.

To un­der­stand why this mat­ters, con­sid­er how pow­er has tra­di­tion­al­ly worked. For the past two cen­turies, if you want­ed to build some­thing sig­nif­i­cant you need­ed mon­ey, scale, con­nec­tions and the abil­i­ty to nav­i­gate reg­u­la­tions. Big or­gan­i­sa­tions had nat­ur­al ad­van­tages. They could hire more peo­ple, in­vest more cap­i­tal and lob­by more ef­fec­tive­ly.

The play­ing field was tilt­ed to­ward those who al­ready held pow­er.

AI changes this equa­tion in ways we are on­ly be­gin­ning to un­der­stand. What mat­ters now is not how many em­ploy­ees you have or how much cap­i­tal you con­trol. What mat­ters is how quick­ly you can train AI sys­tems, how much qual­i­ty da­ta you can feed them, and how ef­fec­tive­ly you can de­ploy them.

A small team with the right AI ca­pa­bil­i­ties can now out­per­form or­gan­i­sa­tions a hun­dred times their size. Con­verse­ly, a large in­sti­tu­tion that fails to in­te­grate AI ef­fec­tive­ly may find its tra­di­tion­al ad­van­tages evap­o­rat­ing.

The re­al dis­rup­tion is not about which jobs will be au­to­mat­ed. It is about what hap­pens when “think­ing” be­comes cheap while the abil­i­ty to act on that think­ing re­mains ex­pen­sive and scarce. Your com­pe­ti­tion is no longer an­oth­er per­son or even an­oth­er com­pa­ny. Your com­pe­ti­tion is a sys­tem that learns and im­proves faster than any hu­man or­gan­i­sa­tion can keep up with.

The prob­lem

Yet with­in this sto­ry of ac­cel­er­at­ing pow­er lies an un­ex­pect­ed vul­ner­a­bil­i­ty. In Oc­to­ber, re­searchers at Texas A&M Uni­ver­si­ty, the Uni­ver­si­ty of Texas at Austin and Pur­due Uni­ver­si­ty pub­lished a study with in­ter­est­ing find­ings. They dis­cov­ered that when AI sys­tems are con­tin­u­ous­ly trained on low qual­i­ty in­ter­net con­tent, the sys­tems ac­tu­al­ly be­come worse.

The re­searchers called this phe­nom­e­non brain rot, bor­row­ing a term usu­al­ly ap­plied to peo­ple who spend too much time con­sum­ing shal­low on­line con­tent. The most alarm­ing find­ing came when the re­searchers tried to fix the prob­lem. Even af­ter re­train­ing the dam­aged sys­tems on high-qual­i­ty, care­ful­ly cu­rat­ed da­ta, the AI mod­els could not ful­ly re­cov­er their orig­i­nal ca­pa­bil­i­ties.

For busi­ness­es, these find­ings cre­ate an im­me­di­ate prob­lem. There is a rush to in­te­grate AI in­to busi­ness op­er­a­tions. It is used for cus­tomer ser­vice, con­tent cre­ation, da­ta analy­sis and de­ci­sion sup­port. Yet if the AI sys­tems busi­ness­es re­ly on are them­selves de­grad­ing, then com­pa­nies face a new kind of risk they have nev­er had to man­age be­fore.

The ques­tion is no longer just whether your AI will oc­ca­sion­al­ly make mis­takes or pro­duce non­sense (what the in­dus­try calls hal­lu­ci­na­tions). The ques­tion is whether your AI runs the risk of los­ing its abil­i­ty to “think” clear­ly over time.

The Gen­e­sis Mis­sion should be un­der­stood in this con­text. The di­rec­tive ex­plic­it­ly man­dates the US gov­ern­ment to move away from train­ing AI on the open in­ter­net. In­stead, it shifts to us­ing fed­er­al sci­en­tif­ic datasets from decades of gov­ern­ment fund­ed re­search in ar­eas like biotech­nol­o­gy, nu­clear en­er­gy, and ma­te­ri­als sci­ence.

We are watch­ing the emer­gence of two dif­fer­ent kinds of ar­ti­fi­cial in­tel­li­gence. On one side, there are the com­mer­cial sys­tems trained on what­ev­er con­tent they can scrape from the in­ter­net. These are sys­tems that may be slow­ly de­te­ri­o­rat­ing with us­age, even as new mod­els im­prove ca­pa­bil­i­ties.

On the oth­er side, there are gov­ern­ment and elite sys­tems trained on care­ful­ly pro­tect­ed, high qual­i­ty da­ta. The gap be­tween these two tiers may widen con­sid­er­ably in the years ahead.

Im­pli­ca­tions

For com­pa­nies, the im­pli­ca­tions are un­com­fort­able. Treat­ing AI as a sim­ple sub­scrip­tion ser­vice, some­thing you buy from a ven­dor and plug in­to your op­er­a­tions, may be a risk. Firms that de­pend on gen­er­al pur­pose AI sys­tems trained on the open in­ter­net may see the un­der­ly­ing qual­i­ty de­grades over time.

Com­pa­nies that want to re­main com­pet­i­tive may need to in­vest in their own da­ta in­fra­struc­ture, build their own care­ful­ly cu­rat­ed train­ing ma­te­ri­als, and con­tin­u­ous­ly test their AI sys­tems for signs of de­cline.

The deep­er ques­tion, which nei­ther gov­ern­ment pro­grammes nor cor­po­rate strate­gies ful­ly ad­dress, con­cerns or­di­nary peo­ple caught in the mid­dle of this tran­si­tion. Pichai is right that peo­ple will need to adapt. But adapt­ing is dif­fi­cult when the ground keeps shift­ing be­neath your feet.

The re­al­i­ty is that the sys­tems that help you think are them­selves chang­ing faster than you can learn to use them. As a re­sult adap­ta­tion be­comes less a skill you can ac­quire and more a pos­ture you must con­stant­ly main­tain.

The re­al­i­ty is you can­not win this race by sim­ply work­ing hard­er or learn­ing more. The sys­tems you are com­pet­ing against can work in­fi­nite­ly hard­er and learn in­fi­nite­ly more than you ever could. I would sug­gest that the on­ly vi­able strat­e­gy is to learn how to work with these sys­tems.

This means learn­ing to in­te­grate them in­to how you think and op­er­ate. The goal is to mul­ti­ply your own ca­pa­bil­i­ties rather than try­ing to com­pete di­rect­ly against ca­pa­bil­i­ties you can­not match.

The pub­lic con­ver­sa­tion frames AI pri­mar­i­ly as a threat to jobs. That fram­ing miss­es the big­ger pic­ture. AI is more like a grav­i­ta­tion­al force, pulling val­ue and op­por­tu­ni­ty to­ward those who learn to use it ef­fec­tive­ly and away from those who do not.

The peo­ple who will thrive in any pro­fes­sion are those who fig­ure out how to make AI an ex­ten­sion of their own abil­i­ties. Those who re­sist or ig­nore it will find their skills worth pro­gres­sive­ly less, re­gard­less of how hard they work.

This will show up on­ly in un­em­ploy­ment sta­tis­tics. It will al­so show up as a widen­ing gap in out­comes. That is in earn­ings, in op­por­tu­ni­ties, in qual­i­ty of life, that our ex­ist­ing sys­tems of ed­u­ca­tion, em­ploy­ment, and so­cial sup­port were nev­er de­signed to han­dle.

Three years af­ter Chat­G­PT, we find our­selves in a strange in-be­tween pe­ri­od. The old ways of or­gan­is­ing work and pow­er have not col­lapsed, but they are vis­i­bly strain­ing. The new arrange­ments are not yet clear, but their out­lines are start­ing to emerge.

Chat­G­PT was re­leased as what Ope­nAI called a re­search pre­view. The ask was to try it out, es­sen­tial­ly an ex­per­i­ment to see how peo­ple would use it.

Three years lat­er, it has be­come the in­ter­face through which hun­dreds of mil­lions of peo­ple en­counter ar­ti­fi­cial think­ing for the first time. What be­gan as an ex­per­i­ment has be­come in­fra­struc­ture.

In the fu­ture it is like­ly to be­come a na­tion­al as­set for coun­tries that prop­er­ly har­ness­es it. This is a con­ver­sa­tion that we have yet to have here in T&T.

Ian Nar­ine is a fi­nan­cial con­sul­tant who be­lieves that in­tel­li­gence isn’t ar­ti­fi­cial. Send your com­ments to ian@ian­nar­ine.com