In many ways, AI-gen­er­at­ed con­tent is poi­son­ing the in­ter­net, mak­ing it less re­li­able to ac­cess clear facts. And while it is in­dis­putable that AI is im­pres­sive, it is al­so pre­dictable. Bren­da Mul­ber­ry am­bi­tious­ly hopes to open the first T-shirt shop on the Moon. Her Min­shal­lesque idea is a per­fect edge case. Her knowl­edge is per­son­al­ly cre­at­ed and his­tor­i­cal­ly pro­duced. It is not em­pir­i­cal­ly gen­er­at­ed. Present AI mod­els can­not gen­er­ate such a bizarre busi­ness propo­si­tion. Her cu­ri­ous idea for a Moon-based T-shirt tourist shop is ex­cep­tion­al­ly hu­man.

In the glow of the flood­lit launch tow­ers of Launch Pad 39B, thou­sands crammed the cause­ways, Co­coa Beach, and mo­tel bal­conies of Flori­da’s Space Coast to wit­ness the launch of Artemis II, sell­ing bar­be­cue and drink­ing Moon­shots. But Bren­da, who has been sell­ing NASA sou­venirs for 40 years, be­lieves that once you get to the moon as a tourist, you will want to buy a T-shirt.

To­day, even the most ground­break­ing mod­els fol­low a fa­mil­iar pat­tern: learn, op­ti­mise, ex­e­cute. Once train­ing is com­plete, the AI sticks to what it knows; when it does not, it hal­lu­ci­nates. With­in the hal­lowed halls of ter­tiary in­sti­tu­tions, stu­dents use AI to solve tu­to­r­i­al sheet prob­lems adroit­ly. This means that the prob­lems that the pro­fes­sors pose aim to mea­sure the gap be­tween what was trans­mit­ted in the lec­ture hall and what the learn­er can re­call.

The test score re­flects the stu­dent’s abil­i­ty to ac­cu­rate­ly re­gur­gi­tate lec­ture notes, con­tent from mul­ti­ple text­book chap­ters, and pol­ished Pow­er­Point sum­maries. This is not ed­u­ca­tion for agency in the in­tel­li­gent age. Learn­ers must be cu­ri­ous, in­quis­i­tive, and will­ing to en­gage with dif­fi­cul­ties, rather than ac­cept the world as it is. A ped­a­gogy built around achiev­ing pre­de­ter­mined out­comes has, un­for­tu­nate­ly, su­per­seded the de­vel­op­ment of trans­for­ma­tive abil­i­ties like Bren­da, which can al­low her to go be­yond the in­for­ma­tion giv­en. The ob­verse of the post-struc­tural­ist dic­tum “to know is to kill” is that “in­quiry frees”. The on­ly pas­sage to in­sight is free­dom in the pres­ence of knowl­edge. Ideas are cre­at­ed piece­meal, ad hoc, from pos­si­bil­i­ties half dis­closed, and from un­ex­plored con­nec­tions and half-con­cealed. In this fer­ment lie the pos­si­bil­i­ties to be made ac­tu­al.

Two big things have hap­pened so far. One is that gen­er­a­tive AI is a “Wine of As­ton­ish­ment” that has awak­ened the pub­lic be­cause it of­fers every­one a suite of con­crete tools. The oth­er is that busi­ness­es have re­alised that there is noth­ing out­side the screen econ­o­my. To­geth­er, these two ad­vances have raised the spec­tre of risks. Ge­of­frey Hin­ton, Yann Le­Cun, and Yoshua Ben­gio have been quite vo­cal about their con­cerns that AI could pose an ex­is­ten­tial risk to hu­man­i­ty. But there are oth­er, more im­me­di­ate risks, in­clud­ing Se­cu­ri­ty for AI and AI-Ac­cel­er­at­ed Cy­ber­at­tacks. Cy­ber­se­cu­ri­ty pro­tects the “con­tain­er” (net­works, servers, ap­pli­ca­tions). AI se­cu­ri­ty ad­dress­es mod­el theft, ma­li­cious da­ta in­jec­tion to al­ter be­hav­iour, and prompt in­jec­tion.

Oth­er trou­bles emerge where the rub­ber meets the road. Prob­lems such as bias, prej­u­dice, mis­in­for­ma­tion, work­force dis­rup­tion, and pri­va­cy in­fringe­ments are among the most ur­gent. To get a glimpse of the fu­ture, we need on­ly to in­ter­ro­gate the present cat­a­logue of fas­ci­nat­ing AI demos and edge cas­es. Most of these demos will fade; oth­ers will drift; and the re­main­der, built with a vi­sion to scale, will sur­vive. Qual­i­ty is not sta­t­ic. The AI fu­ture will be mul­ti-mod­el and mul­ti-in­fra­struc­ture. Agen­tic AI will be­come the de­fault en­ter­prise sys­tem. Swarms of small, spe­cialised mod­els will out­per­form the large gen­er­al­ist mod­els at every task. The or­ches­tra­tion lay­er will be­come the com­pet­i­tive moat as de­vel­op­ers come to un­der­stand that in­tel­li­gence is or­ches­tra­tion.

The eval­u­a­tion pipeline will be­come as crit­i­cal as the train­ing pipeline. In this fu­ture, the eval­u­a­tion in­fra­struc­ture be­comes as crit­i­cal as the train­ing da­ta. Soon, re­spon­si­ble AI will move from pub­lic pol­i­cy doc­u­men­ta­tion and the Law of AI to an en­gi­neer­ing re­quire­ment. Da­ta fly­wheels will feed off user in­ter­ac­tions, be­hav­iour­al sig­nals, da­ta cu­ra­tion and la­belling, and bet­ter re­spons­es. Mod­el builders must in­vest in cu­ra­tion in­fra­struc­ture as much as they in­vest in mod­el in­fra­struc­ture.

This hori­zon is urg­ing na­tion-states to in­vest in pub­lic-sec­tor sci­ence and tech­nol­o­gy. AI is cost­ly, mak­ing it in­ac­ces­si­ble to acad­e­mia. Ter­tiary in­sti­tu­tions have been cre­at­ing AI sand­box­es that al­low a lim­it­ed se­lec­tion of AI Agents and “claws” to op­er­ate. Open­Claw, for in­stance, is an open-source frame­work for de­ploy­ing au­tonomous AI agents that serve as “dig­i­tal cowork­ers” (or “claws”) to man­age, browse, and ex­e­cute tasks across ap­pli­ca­tions, files, and browsers.

What is need­ed in Latin Amer­i­ca and the Caribbean (LAC) is a sig­nif­i­cant in­vest­ment in pub­lic-sec­tor re­search and com­put­ing ca­pa­bil­i­ties, in­clud­ing a Re­gion­al AI Re­search Re­source and Lab sim­i­lar to CERN. It takes a vil­lage to im­prove tech­nol­o­gy. AI can tack­le in­her­it­ed in­equal­i­ty and in­ter­gen­er­a­tional im­mo­bil­i­ty in LAC, but on­ly with a co­or­di­nat­ed ef­fort to en­sure LAC’s lead­er­ship in AI.

Dr Fazal Ali com­plet­ed his Mas­ter's in Phi­los­o­phy at the Uni­ver­si­ty of the West In­dies. He was a Com­mon­wealth Schol­ar who at­tend­ed the Uni­ver­si­ty of Cam­bridge, Hugh­es Hall, the provost of the Uni­ver­si­ty of Trinidad and To­ba­go and the act­ing pres­i­dent, and chair­man of the Teach­ing Ser­vice Com­mis­sion. He is present­ly a con­sul­tant with the IDB.