White House AI and Crypto Czar David Sacks: China Is ‘Hot on Our Heels’ in AI Race

MACCALLUM So here is ░░░░░░░░ President Trump stands on ░░░░░░░░ possibly pulling ahead in ░░░░░░░░ AI race Watch this ░░░░░░░░ starts Reporter Karoline you ░░░░░░░░ restore U S dominance ░░░░░░░░ there a fear that ░░░░░░░░ U S either is ░░░░░░░░ or has fallen behind ░░░░░░░░ how would the president ░░░░░░░░ sure U S stays ░░░░░░░░ LEAVITT Now the president ░░░░░░░░ confident that we will ░░░░░░░░ American dominance in AI ░░░░░░░░ ends MACCALLUM OK who ░░░░░░░░ to talk to about ░░░░░░░░ than White House AI ░░░░░░░░ crypto czar David Sacks ░░░░░░░░ has been at the ░░░░░░░░ of this issue in ░░░░░░░░ Valley co founded a ░░░░░░░░ the company that became ░░░░░░░░ and many other very ░░░░░░░░ crucial tech investments David ░░░░░░░░ to have you with ░░░░░░░░ today Thank you very ░░░░░░░░ for joining me SACKS ░░░░░░░░ to be here Martha ░░░░░░░░ So you hear that ░░░░░░░░ in the White House ░░░░░░░░ office thank you and ░░░░░░░░ Levitt saying that the ░░░░░░░░ wants to restore U ░░░░░░░░ dominance in AI How ░░░░░░░░ ground has been lost ░░░░░░░░ this development SACKS Well ░░░░░░░░ think the Chinese companies ░░░░░░░░ catching up very fast ░░░░░░░░ think that s the ░░░░░░░░ I would say that ░░░░░░░░ haven t lost our ░░░░░░░░ here The DeepSeek R1 ░░░░░░░░ is basically comparable in ░░░░░░░░ to the OpenAI o1 ░░░░░░░░ which came out about ░░░░░░░░ months ago So they ░░░░░░░░ kind of hot on ░░░░░░░░ heels here and I ░░░░░░░░ we basically have somewhere ░░░░░░░░ three and six month ░░░░░░░░ on them But they ░░░░░░░░ catching up very very ░░░░░░░░ And I think President ░░░░░░░░ is entirely right to ░░░░░░░░ concerned that they are ░░░░░░░░ up and we want ░░░░░░░░ do everything we can ░░░░░░░░ maintain our AI leadership ░░░░░░░░ dominance MACCALLUM You know ░░░░░░░░ heard about this 100 ░░░░░░░░ investment in Stargate by ░░░░░░░░ and SoftBank and Sam ░░░░░░░░ You know one of ░░░░░░░░ questions is are we ░░░░░░░░ too much This version ░░░░░░░░ cheaper And apparently it ░░░░░░░░ much faster to train ░░░░░░░░ build David SACKS Yeah ░░░░░░░░ mean this is where ░░░░░░░░ s innovation is is ░░░░░░░░ have managed to get ░░░░░░░░ efficiency out of chips ░░░░░░░░ re squeezing more juice ░░░░░░░░ of the chips and ░░░░░░░░ kind of have to ░░░░░░░░ of the export controls ░░░░░░░░ advanced chips going to ░░░░░░░░ they ve had to ░░░░░░░░ how to be more ░░░░░░░░ So necessity has been ░░░░░░░░ mother of invention for ░░░░░░░░ That being said there ░░░░░░░░ still great advantages to ░░░░░░░░ a lot of chips ░░░░░░░░ number of chips And ░░░░░░░░ think that this is ░░░░░░░░ area where America can ░░░░░░░░ to lead is in ░░░░░░░░ build out of this ░░░░░░░░ and having the most ░░░░░░░░ chips And so I ░░░░░░░░ it s a little ░░░░░░░░ of an overreaction to ░░░░░░░░ that America doesn t ░░░░░░░░ big AI data centers ░░░░░░░░ I think we need ░░░░░░░░ build out of these ░░░░░░░░ AI data centers I ░░░░░░░░ President Trump has expressed ░░░░░░░░ for them and I ░░░░░░░░ we need to make ░░░░░░░░ easier to build those ░░░░░░░░ centers to get them ░░░░░░░░ and licensed and to ░░░░░░░░ power generated electricity for ░░░░░░░░ data centers If we ░░░░░░░░ the biggest data centers ░░░░░░░░ is still an advantage ░░░░░░░░ is true that DeepSeek ░░░░░░░░ shown new ways for ░░░░░░░░ models to be efficient ░░░░░░░░ I think our AI ░░░░░░░░ are going to learn ░░░░░░░░ adopt those efficiency techniques ░░░░░░░░ well But you still ░░░░░░░░ to be able to ░░░░░░░░ compute and the data ░░░░░░░░ are essential for that ░░░░░░░░ Is there IP theft ░░░░░░░░ Did they steal some ░░░░░░░░ what they were able ░░░░░░░░ scale up here from ░░░░░░░░ And do we need ░░░░░░░░ close the doors on ░░░░░░░░ we re doing in ░░░░░░░░ of invention and technology ░░░░░░░░ Well it s possible ░░░░░░░░ s a technique in ░░░░░░░░ called distillation which you ░░░░░░░░ going to hear a ░░░░░░░░ about And it s ░░░░░░░░ one model learns from ░░░░░░░░ model Effectively what happens ░░░░░░░░ that the student model ░░░░░░░░ the parent model a ░░░░░░░░ of questions just like ░░░░░░░░ human would learn But ░░░░░░░░ can do this asking ░░░░░░░░ of questions and they ░░░░░░░░ essentially mimic the reasoning ░░░░░░░░ that they learn from ░░░░░░░░ parent model and they ░░░░░░░░ kind of suck the ░░░░░░░░ of the parent model ░░░░░░░░ there s substantial evidence ░░░░░░░░ what DeepSeek did here ░░░░░░░░ they distilled the knowledge ░░░░░░░░ of OpenAI s models ░░░░░░░░ I don t think ░░░░░░░░ is very happy about ░░░░░░░░ And I think one ░░░░░░░░ the things you re ░░░░░░░░ to see over the ░░░░░░░░ few months is our ░░░░░░░░ AI companies taking steps ░░░░░░░░ try and prevent distillation ░░░░░░░░ so we ll see ░░░░░░░░ the leading AI companies ░░░░░░░░ prevent distillation by third ░░░░░░░░ companies that would definitely ░░░░░░░░ down some of these ░░░░░░░░ models.”