Week 1 of the Trump administration, AGI timelines, and DeepSeek
What this covers
In this episode, we break down President Trump's repeal of the the Biden administration's Executive Order (EO) on AI (1:00), the release of the America First Trade Policy memorandum (9:52), and the Trump administration's own AI EO (15:02). We are then joined by Lennart Heim, Senior Information Scientist at the RAND Corporation to discuss the Stargate announcement (20:40), how AI company CEOs are talking about AGI (38:36), and why the latest models from DeepSeek matter (52:02).
Source description (no synthesized summary yet).
The Trump Administration is reshaping AI policy through deregulation and export controls while the AI industry races toward AGI with massive compute investments, but Chinese competitors like DeepSeek demonstrate that efficiency improvements and model distillation may challenge the assumption that compute-intensive scaling is the only viable path forward.
- Trump repealed Biden's AI safety executive order but the underlying infrastructure and safety initiatives remain, signaling selective rather than wholesale AI deregulation
- The $500 billion Stargate investment commits to US compute infrastructure over four years, betting that continued scaling will deliver AGI capabilities despite rapid cost-efficiency improvements elsewhere
- DeepSeek's V3 and R1 models achieved frontier performance with far less compute than expected, suggesting the export control strategy works on deployment scale rather than research capability parity
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Coding is an ideal domain for reinforcement learning-based AI improvements because the simulated environment (running code and checking outputs) closely matches the real-world task, enabling synthetic training data generation and rapid iteration toward superhuman performance.
“you can sort of say your intention you can then generate the code you can run the code and then see if it did what you wanted it to do um so it is so it is very amenable to reinforcement learning as an approach because you can generate synthetic training data um so when we say AGI uh Dario amada's you know definition is this is a a you know an AI system that it can do most of the things that a human being can do at least most of the things a human being can do when it's typing on a keyboard clicking on a mouse or speaking to a computer um now all of those tasks I think are are going to fall at different rates and I think the in an incredible irony the people who are sort of most at risk of being replaced by AI in the near term at least and I think you know it's coming for all of us sooner or later but in the near term at least it is people who are working with domains that are amendable to reinforcement learning so it's not a surprise that coding is making all this revolutionary progress”
Export control restrictions on open-source models are built into the diffusion rule; if China publishes an advanced open-source model (or if Meta publishes one), any American company can use and deploy it anywhere in the world (except sanctioned countries), undermining export control logic for proprietary model advantage.
“the diffusion rule in particular the the model export restrictions are calibrated to open source right so if China creates an awesome model and makes it open source or if meta creates an awesome model and makes it open source well suddenly you know any American company can put that anywhere in the world I mean not North Korea but like mostly anywhere in the world so the the the point is like the people who designed the export control architecture noted that like okay if we were to say you know Americans can't sell AI above some level of performance but surprise China just open sourced that act level performance well yeah we recognize that would be a disaster for us companies”
The Trump Administration repealed Biden's AI executive order on January 20, 2025, but most of the downstream implementation work already completed by federal agencies—such as the installation of Chief AI Officers across departments—does not automatically disappear with the repeal, requiring affirmative executive action to undo.
“the executive order gave a lot of deadlines for its work... almost all of those deadlines have come and gone and been implemented... a lot of the downstream implementation of the executive order does not inherently get repealed just because you repealed the executive order”
Energy permitting and infrastructure buildout for AI data centers will require Congressional action and potential NEPA reform beyond what executive branch alone can accomplish; executive action to simplify federal permitting is insufficient to meet the energy demands implied by $500 billion Stargate investment or gigawatt-scale data center clusters.
“we need significantly more energy to build all of these data centers right I think like the action that we will see way more action there and at least on my read on the situation there just like whatever the administration can do is probably not enough we eventually need Congress and nums to act we need like some type of Nea reform”
DeepSeek V3 was reportedly trained on Nvidia H800 chips, which represent the failure of the October 2022 Biden export controls; the H800 was created by Nvidia blowing a fuse on the A100 to modestly degrade interconnect performance (~10-15% degradation) rather than forcing China back to older chip generations (~50-75% degradation as intended).
“this model uh was reportedly trained on Nvidia h800 chips right for for those who know about export controls the h800 is the failure of the October 2022 export controls... Invidia would be able to literally blow a fuse that's all they had to do was blow a fuse on the chip modestly degrade the interconnect speed performance of the Chip and suddenly they on the exact same assembly line China is able to buy Leading Edge chips which uh I think had like a 10 to 15% performance degradation not like the 50 to 75% performance degradation”
AI chips depreciate over three years at a rapid rate, losing nearly all their value as newer, better chips are developed, creating a risky investment profile unless the infrastructure achieves extraordinary performance improvements within that timeframe.
“there's this word called depreciation that everybody should be aware of and it basically says if you spend a lot of money to go buy this big expensive asset well you know it's worth a certain amount in year one how much is it worth in year two how much is it worth in year three right everybody knows that if you go buy a car and then you try to sell it one year later like you'll notice that like the value has dropped by like 25% just in that first year um well the reason why this comes up in the Stargate investment is that the assets that they're buying which is energy infrastructure data center infrastructure and then what goes inside those data centers AI chips well the AI chips are a big big big component of that investment and they depreciate over three years you've lost almost all the value of the chip just because chips are getting better at such an incredible rate um after three years”
President Trump's repeal of Biden's AI executive order does not automatically eliminate downstream implementations because most of the executive order's 90, 180, and 270-day deadlines have already passed and been implemented by federal agencies, requiring affirmative action beyond repeal to undo them.
“this had to happen by 90 days this had to happen by 180 days this had to happen by 270 days and almost all of those deadlines have come and gone and been implemented the bid Administration was really cracking the whip on implementation”
The Biden Administration's AI executive order provided federal government authority under the Defense Production Act to require frontier AI labs to share safety testing information with the government; notice of proposed rulemaking has already been issued, though the interim final rule has not yet been implemented.
“the executive order gave the United States federal government Authority by invoking the defense production act to say that um to say that the frontier AI Labs think open AI think anthropic Etc you know had to share transparently with the United States government information regarding their safety testing on their leading models so that Authority um you know was it directed the United States federal government to sort of prepare to exercise that Authority and the notice of proposed rulemaking already came out now there is this uh second thing that hasn't come out which is the interim final rule”
Model benchmarks are not perfectly correlated with real-world utility; factors like product quality, writing style, deployment integration, and business model matter more than benchmark performance for determining market success; for example, Anthropic's Claude is often preferred by some users over OpenAI's models despite lower benchmark scores because of writing style preferences.
“we like overly focusing on like Benchmark performance would be nice if it would be about the world like who has the best product who has the best Benchmark performance on these kinds of things there are many things which you can't measure benchmarks right I think like CET gbt like the the openi models are often leading but I personally prefer anthropic claws because I think just i've got a better writing style right many coders prefer this”
Stargate's focus on building 50% or more of Oracle's computing infrastructure in the U.S. is directly connected to the 'diffusion rule,' which requires U.S.-headquartered cloud providers to build at least 50% of their advanced computing infrastructure domestically if they want to be 'universally validated end users' with permission to buy and deploy large quantities of chips elsewhere.
“it's explicit building in the US and this nicely connects to the diffusion rule right the diffusion rule was expit saying hey if you're like a cloud provider an American cloud provider like Oracle you got to build 50% at least in yes”
Elon Musk's xAI came out of nowhere with 30 core technical team members (plus Elon's involvement) and achieved competitive frontier-level performance against OpenAI despite OpenAI's 10-year head start, suggesting that rapid team assembly and sufficient capital can replicate frontier AI capability relatively quickly.
“the core technical team of xai for that big Custer not not every employee at the company not every contractor at the company but like the sort of core technical team I think it was like 30 people yeah you know so it's like 30 people plus like Elon musk's indomitable will and time frame you know when it comes to Rapid construction um and and uh a lot of money 30 people plus you know a rapid construction time frame plus a lot of money to buy a lot of chips boom you're like pretty close to the state-ofthe-art in a relatively short period of time right”
DeepSeek trained its models using H800 chips (not H100s as sometimes reported), which are the result of failed October 2022 export controls that allowed Nvidia to produce degraded but still cutting-edge chips for the Chinese market.
“this model uh was reportedly trained on Nvidia h800 chips right for for those who know about export controls the h800 is the failure of the October 2022 export controls so the October 2022 export controls had these technical benchmarks uh for what chips were allowed and what chips were not allowed the sort of original hypothesis in the Biden Administration was that you know the best chip that Nvidia had was the a100 we wanted China to not be able to buy that and to not be able to buy anything better than that but instead to have to go back a generation or two and buy the old Nvidia chips what the B Administration did not anticipate was that Invidia would be able to literally blow a fuse that's all they had to do was blow a fuse on the chip modestly degrade the interconnect speed performance of the Chip and suddenly they on the exact same assembly line China is able to buy Leading Edge chips which uh I think had like a 10 to 15% performance degradation not like the 50 to 75% performance degradation that we've had”
Coding and mathematics are experiencing disproportionate AI progress compared to other domains because these tasks are amenable to reinforcement learning: an AI can generate synthetic training data by performing the task (writing code, solving math problems) and then objectively evaluate whether the output is correct, creating a self-improving loop similar to AlphaGo's self-play mechanism.
“coding is kind of like that I mean it's not the the the simulated environment is very close to the real world environment you can sort of say your intention you can then generate the code you can run the code and then see if it did what you wanted it to do um so it is so it is very amenable to reinforcement learning as an approach because you can generate synthetic training data”
Test-time compute scaling—allocating more inference computing resources to let AI models reason longer before answering—has emerged as a new paradigm driving AI performance improvements, with claims of 100,000x performance improvements from dedicating 20 seconds of thinking time.
“there's always been debate like oh maybe AGI is in 2050 maybe AGI is never Maybe maybe AGI is 10 years away and even last year you know in early 2024 you would hear uh senior Venture capitalists uh senior AI researchers uh senior Executives... saying like oh maybe AGI is 10 years away or something like that or maybe it's 5 years away Etc there has been this remarkable coalescing around a narrative of AGI very very soon and that that shift in messaging really has just been in the past few months”
DeepSeek V3, released December 26, 2024, was trained using only $5.6 million in compute resources (per Allen Institute analysis) via aggressive model distillation techniques, achieving performance comparable to Meta's Llama on some benchmarks; this contrasts sharply with Stargate's $500 billion commitment and suggests a potential compute-intensive vs. efficiency-focused bifurcation in AI development strategies.
“they were able to train V3 using only according to the uh Allen Institute for artificial intelligence um only $5.6 million of training resources... there's this other world out there where a Chinese company is able to build an AI model that at least on some benchmarks is like comparable to what uh meta is doing with llama um for a tiny tiny fraction of that cost”
Anthropic CEO Dario Amodei noted that scaling will continue: clusters are growing from current thousands of GPUs to hundreds of thousands in 2026, potentially millions in 2027, and tens of millions by a few years later; this compute scaling curve is unavailable to China due to export controls, making the long-term advantage decisively Western despite near-term efficiency gains.
“he said recently interview quote it's been reported that deep seek had something like 50,000 h100s he's he's wrong they were H 800s but uh let's keep going um but it's much less than the amount of chips we or some of the other players are planning to build in the next year which I think would go into the high hundreds of thousands in 2026 probably into the millions and in 2027 from different players as we saw yesterday potentially into the tens of millions of chips”
DeepSeek V3 was trained using only $5.6 million in computing resources (estimated by the Allen Institute for AI), compared to estimates of hundreds of millions for comparable Western models, demonstrating a massive compute efficiency advantage through model distillation techniques.
“they were able to train V3 using only according to the uh Allen Institute for artificial intelligence um only $5.6 million of training resources so so we got to connect this back to what we just heard from Stargate right they're talking we're going to spend $500 billion to build all this AI infrastructure to train and inference all these AI models and then there's this other world out there where a Chinese company is able to build an AI model that at least on some benchmarks is like comparable to what uh meta is doing with llama um for a tiny tiny fraction of that cost”
In the past few months, AI company leaders have dramatically shifted messaging toward AGI timelines, moving from previous debates about 10 or 5 years away to near-term predictions, with consensus now favoring AGI emergence within Trump's first term (roughly 2-4 years).
“there has been this remarkable coalescing around a narrative of AGI very very soon and that that shift in messaging really has just been in the past few months so in a blog post on January 5th we had Sam Altman saying quote we are now confident we know how to build AGI as we have traditionally understood it”
The Trump Administration's America First Trade Policy memorandum directed the Secretaries of State and Commerce to review U.S. export controls and assess how to maintain and enhance the nation's technological edge while identifying and eliminating loopholes, signaling an intent to potentially strengthen rather than weaken Biden Administration export control architecture on semiconductors and AI.
“the Secretary of State and the Secretary of Commerce shall review the United States export control system... specifically the Secretary of State and the Secretary of Commerce shall assess and make recommendations regarding how to maintain obtain and enhance our nation's technological Edge and how to identify and eliminate loopholes in existing export controls”
DeepSeek is the leading frontier AI model developer in China; it emerged from hedge fund Highflyer Capital Management (which owned ~10,000 Nvidia A100 chips as of 2022 for quantitative trading) and is now explicitly committed to AGI development, positioning itself as competitor to OpenAI, Anthropic, Google, and Meta.
“deep seek for for those who haven't heard of it yet it is unambiguously UN Ambiguously the leading Frontier AI model developer in China... this came as a quantitative trading firm uh highflyer Capital Management so the reason why they were buying gpus before they got into the large language model development business is to like do hedge fund style market trading... highflyer had already bought 10,000 Nvidia a100 chips”
The Republican party has weaponized the term 'AI safety' to mean censorship and ideological control (citing the Gemini model incidents), making it radioactive in conservative circles even as actual guardrails for frontier AI remain valued.
“there are parts of the Republican policy community that immediately jump to censorship on social media...what happened with the Gemini launch of Google where you had it was generating images...of a rainbow of human races but they were all Nazis...and so this is what they think about when they think about that term AI safety but when they think about guard rails they see some stuff that they is interesting and that they like”
The UAE's involvement in Stargate through MGX is not unprecedented; the UAE has previously made major investments in US-based AI infrastructure including through its backing of Cerebras, an AI chip company that competes with Nvidia.
“this is not unprecedented for the UAE UAE has a big investment in cerebrus for example which is an AI chip uh company they compete with Nvidia they have sort of a different uh technical approach um and there are big cerebrus data centers that are owned by UAE Sovereign wealth funds and those data centers are in America right now”
OpenAI, SoftBank, Oracle, and UAE-backed MGX announced Stargate, a joint venture to invest up to $500 billion over four years in US AI infrastructure, with the first $100 billion committed for this year, representing continuation of infrastructure work already underway rather than purely new investment.
“on January 21st open AI soft Bank Oracle and UAE backed mgx announced Stargate which is a joint venture planning to invest up to $500 billion doll in US money on us AI infrastructure over the next four years... the first of the data centers are already under construction in Texas quote each building half a million square feet and there are 10 buildings currently being built but that will expand to 20 in other locations”
Trump's America First Trade Policy memorandum directs the Secretaries of State and Commerce to review US export controls, assess how to enhance national technological edge, and eliminate loopholes—signaling continuation or strengthening of Biden Administration export control policy rather than rollback.
“the Secretary of State and the Secretary of Commerce shall review the United States export control system and advise on modifications in light of developments involving strategic adversaries or geopolitical Rivals as well as all other relevant National Security and Global considerations specifically the Secretary of State and the Secretary of Commerce shall assess and make recommendations regarding how to maintain obtain and enhance our nation's technological Edge and how to identify and eliminate loopholes in existing export controls”
Stargate is structured as a joint venture in which OpenAI is the exclusive customer receiving computing resources rather than an equity partner, meaning the computing infrastructure serves OpenAI's needs exclusively.
“this is structured as a joint venture so all of these entities are creating a new company called Stargate and openai is actually the customer of this new company so this company is going to provide the Computing resources exclusively to serve open AI”
According to Larry Ellison (Oracle Executive Chairman), data centers for Stargate are already under construction in Texas, with 10 buildings of half a million square feet each currently being built, expanding to 20 locations; the companies stated 'this would not have happened without you, Mr. President' but also revealed they were already building before Trump took office, suggesting continuity of plans rather than purely Trump-era initiative.
“Larry Ellison the executive chairman of Oracle stated this explicitly um there's big parts of this that is not a new announcement right this is continuation of what was going on Larry Ellison said the first of the data centers are already under construction in Texas quote each building half a million square feet and there are 10 buildings currently being built but that will expand to 20 in other locations”
Trump told the Stargate partners to increase the announced investment from $100 billion to $200 billion to $500 billion, and the partners then came back to him saying they would commit to $500 billion over four years, aligning with his first presidential term.
“Trump said specifically that uh they had approached him and said they wanted to make this announcement originally 100 billion and Trump's Trump's addition to the idea was to say make it 200 billion and then they came back to him and said now it's 500 billion um and I think that's uh kind of interesting um noteworthy here is that it is 500 billion over the next four years so they're literally saying that's what we're going to invest during the course of the first uh Trump Administration”
DeepSeek is unambiguously the leading frontier AI model developer in China; AI researchers at leading Western firms read DeepSeek papers to learn novel techniques, indicating genuine technical contribution rather than just catching up.
“deep seek for for those who haven't heard of it yet it is unambiguously UN Ambiguously the leading Frontier AI model developer in China uh they are the ones doing the most interesting research there there's other cool stuff going on at Alibaba and some of its subsidiaries but it's just worth pointing out that that deep seek is now clearly the leader and I've talked to Folks at some of the leading AI firms and they say like we read deep seek papers and we learn things um and it's just worth noting that they have a really interesting uh Locus and density of high quality Talent um and they're doing legitimately good work”
Data center construction creates jobs, but operation of data centers is heavily automated, resulting in low jobs-per-capital-invested; energy infrastructure operation creates more employment both in construction and operation than the data centers themselves.
“the construction of the data centers creates a lot of jobs right people operating construction Machinery people with a shovel or a screwdriver or whatever that creates a lot of jobs but the operation of the data centers they're actually pretty heavily automated so like the sort of jobs per you know million dollars of capital investment um actually ends up being low it's a little bit of a different story on the energy side of the equation right operating power plants operating new energy grid infrastructure that creates jobs in construction it also creates jobs in at the the point of operation”
DeepSeek originated as a hedge fund (HighFlyer Capital Management) focused on quantitative trading and had already purchased 10,000 Nvidia A100 chips before October 2022 export controls; the company has now pivoted entirely toward AGI development alongside other frontier AI labs.
“this came as a quantitative trading firm uh highflyer Capital Management so the reason why they were buying gpus before they got into the large language model development business is like to like do hedge fund style market trading like that kind of algorithm sort of thing so even before the um October 2022 export controls went into place um highflyer had already bought 10,000 Nvidia a100 chips at least according to the reporting for from the financial times and you know just by comparison um meta has bought like 600,000 of you know h100 and a100 type chips”
DeepSeek has triggered a price war for large language model tokens in China, and similar price competition is expected in the United States as Western labs release distilled models and compete on cost.
“it's it's worth pointing out that deeps did cause like a price War for uh large language model usage tokens um in China and I can't wait uh for the prices to drop here in the United States as well that I'm really looking forward to that”
The 'fast follower' advantage may be more valuable than first-mover advantage in AI development; historical examples like Meta's rapid adoption of Stories (copied from Snapchat) and WhatsApp (acquired as a follower product) suggest that being second is often a more profitable position than being first.
“there is this phenomenon where fast follower um is a pretty good competitive position and so I think deep seek sort of leads us to ask the question like are we going to spend $500 billion to get to the frontier so that China can find a way to copy our homework for pennies on the dollar and that's a really important question to have a solid answer to...there is this phenomenon where fast follower um is a pretty good competitive position...meta for example right they stopped inventing new stuff at meta um at least in the social media Paradigm a long time ago and sort of waited for some other company to try out a feature like stories that was when meta saw that Snapchat was taking off um WhatsApp that's a company you know met a bot didn't innovate at home”
The Trump Administration issued a short executive order on January 23rd titled 'Removing Barriers to American Leadership on Artificial Intelligence' that frames AI dominance as a national policy objective and explicitly aims to ensure that AI systems are 'free from ideological bias or engineered social agendas,' continuing the focus on eliminating perceived 'woke' elements from AI governance.
“to maintain this leadership we must develop AI systems that are free from ideological bias or engineered social agendas so that again goes back to what I was saying before anything that has a whiff of woke ideology or uh diversity Equity inclusion type of stuff that's what they want to get rid of”
SoftBank is an extremely well-capitalized venture firm willing to make massive capital bets, and the firm has ties to Middle Eastern Sovereign wealth funds that provide significant capital; the UAE's total Sovereign wealth is approximately $1.5 trillion, giving it substantial capacity to fund the Stargate commitment.
“SoftBank is is a uh Venture Capital firm um but specifically they are an incredibly well- capitalized firm that is making to make willing to make really really really big dollar bets...they have in the past had these Vision funds which includes an awful lot of dollars um from wealthy Middle Eastern Sovereign wealth funds and here in this announcement they actually included one of those Middle Eastern Sovereign wealth funds um the UAE backed mgx...the entire UAE if you put all The Sovereign wealth funds together is $1.5 trillion”
On January 21, OpenAI, SoftBank, Oracle, and UAE-backed MGX announced Stargate, a joint venture to invest up to $500 billion over the next four years in U.S. AI infrastructure, with the first $100 billion commitment this year; Trump claimed he pushed the companies to increase the figure from $100 billion to $200 billion to $500 billion at the announcement.
“Trump said specifically that uh they had approached him and said they wanted to make this announcement originally 100 billion and Trump's Trump's addition to the idea was to say make it 200 billion and then they came back to him and said now it's 500 billion”
Export controls work by limiting the total amount of compute resources available to China, not by stopping Chinese companies from training frontier models immediately; the constraints manifest as limitations on how many companies can be competitive (fewer competitors) and on how many tokens of inference (fewer deployments) can be served with limited chips.
“I generally got bet news for everyone who thinks XO controls are going to stop the Chinese ecosystem they can't train X it's like that's not how they work they just make things harder they make things more costly right they have to use all the chips they have to use their own chips they don't have enough chips so the best way to think about is just like will Expo controls limit the total amount of computation resources available there how could this impact you well how many companies do we have who have like a like really competitive model way more than China to my understanding right so the US ecosystem can like sustain five 10 100 different companies they may be only two because they only have that much computation resources sure like for some things one company is enough right I'm not saying policy problem is solved here just a good thing to understand”
The AI Safety Institute itself exists independently of the Biden AI executive order and will not be eliminated by repealing the executive order; Congress has also passed bipartisan legislation (co-sponsored by Senators Cantwell and Young) that supports the institute's continued operation.
“the AI safety Institute itself um does not relate to the AI executive order you abolish the AI executive order the AI safety sticks around so if the Trump Administration wants to get rid of the AI safety Institute which I don't think they necessarily do um although Senator Ted Cruz you know has expressed some desire to do so um there's this other thing going on in Congress which is a bipartisan piece of legislation co-sponsored by Senator Maria cwell and Todd young and it has already passed the Senate Commerce science and transportation committee uh last year”
Senator Marco Rubio, now presumptive Secretary of State, was vocal about export controls as a senator and is expected to be in the driver's seat alongside the Commerce Secretary in setting the Trump Administration's export control policy.
“Senator Marco Rubio he was very sorry he's no longer a senator the presumptive Secretary of State Marco Rubio um he was very loud on export controls while he was a United States Senator and I think what this shows is yes he absolutely wants to be in the driver's seat alongside the Secretary of Commerce in driving export control policy”
XAI, Elon Musk's AI startup, achieved competitive frontier-level performance with only ~30 core engineers plus rapid cluster construction and significant capital, demonstrating that large team size and long corporate tenure (like OpenAI) are not necessary for frontier AI development.
“xai from like the startup for M musk also kind of came out of nowhere and their model was like leading in um ched Arena... it was like 30 people yeah you know so it's like 30 people plus like Elon musk's indomitable will and time frame you know when it comes to Rapid construction um and and uh a lot of money 30 people plus you know a rapid construction time frame plus a lot of money to buy a lot of chips boom you're like pretty close to the state-ofthe-art”
Dario Amodei (Anthropic CEO) noted that while DeepSeek may have had roughly 50,000 H100s (though actually H800s), Anthropic and other Western players are planning to build clusters with hundreds of thousands to millions of chips over 2025-2027, representing a scale multiplier that will ensure continued Western capability advantages.
“he said recently interview quote it's been reported that deep seek had something like 50,000 h100s he's he's wrong they were H 800s but uh let's keep going um but it's much less than the amount of chips we or some of the other players are planning to build in the next year which I think would go into the high hundreds of thousands in 2026 probably into the millions and in 2027 from different players”
The Google Gemini launch generated images attempting to portray diverse human races in historical contexts but instead produced historically inaccurate and offensive results (e.g., images of Nazis from various races), which conservative critics cite as an example of 'woke' AI safety measures gone wrong, creating political resistance to AI safety terminology.
“you've seen uh what happened with the Gemini launch of Google where you had it was generating images of you know a rainbow of human races but they were all Nazis um and so doing stuff that was like historically inaccurate and oddly offensive as it was trying to be inoffensive”
The AI Safety Institute will survive Trump administration because it is not dependent on Biden's executive order; even if Trump repeals the order, the Institute stands alone, and bipartisan Congressional legislation (by Kaine and Young) provides additional statutory support.
“the AI safety Institute itself um does not relate to the AI executive order you abolish the AI executive order the AI safety sticks around...there's this other thing going on in Congress which is a bipartisan piece of legislation co-sponsored by Senator Maria kwell and Todd young and it has already passed the Senate Commerce science and transportation committee uh last year”
Satya Nadella, CEO of Microsoft, called DeepSeek's technical achievements 'super impressive' in terms of both open-source inference-time compute and compute efficiency, and stated 'we should take the developments out of China very very seriously.'
“Sati Adella at the the the CEO of Microsoft he said quote to see the Deep seek new model it's super impressive in terms of both how they have really effectively done an open source model that does this inference time compute and its super compute efficiency we should take the developments out of China very very seriously”
SoftBank is described as an extraordinarily well-capitalized venture capital firm that functions as an alternative to companies going public, willing to make massive capital commitments; they have ties to Middle Eastern sovereign wealth funds through their Vision Funds.
“SoftBank SoftBank is is a uh Venture Capital firm um but specifically they are an incredibly well- capitalized firm that is making to make willing to make really really really big dollar bets they're they're almost an alternative and seen as an alternative in many cases to your company going public right if you if you need a massive infusion of capital in order to make your company work you can either go list yourself on NASDAQ or the New York Stock Exchange and and open yourself up to investors from around the world or you can just go to SoftBank who is so rich um that they are willing to do it in and of themselves”
The UAE has approximately $1.5 trillion in combined sovereign wealth across all funds, and has made significant past and current investments in American AI and semiconductor companies including Cerebrus AI chips and data centers on US soil.
“the entire UAE if you put all The Sovereign wealth funds together is $1.5 trillion um don't quote me under oath you know about that number but it's big big money this is a very wealthy country um if they want to put this $500 billion up they have the money to do it um and this is structured as a joint venture... UAE has a big investment in cerebrus for example which is an AI chip uh company they compete with Nvidia they have sort of a different uh technical approach um and there are big cerebrus data centers that are owned by UAE Sovereign wealth funds and those data centers are in America right now”
Leonard Hine distinguishes between the 'excess effect' (ability to achieve the same capability with less compute over time, via efficiency gains) and the 'performance effect' (ability to use the same compute to achieve better capability), arguing that both are happening simultaneously and the 'performance effect' is where export controls will bite hardest.
“what I call an excess effect so basically what it describes is like you get the same capabilities for Less compute... I think what many people are missing here is the performance effect which basically means well if you just keep your compute steady now you get more juice out of it right”
Export controls have a 'lagging impact'—new restrictions take months or years to bite because they constrain chips purchased in future periods, not already-purchased equipment; DeepSeek is operating during the 'golden era' of the failed October 2022 controls, and will face pain from the revised October 2023, December 2024, and January 2025 controls only later in 2025 or after.
“there's a lagging impact of export controls when you talk about the the infrastructure buildout you know the Deep seek Revolution is occurring as the failure of those October 2022 export controls we've revised the architecture in October 2023 we revised it again in December 2024 we revised it again in January 2025 and if that control architecture is going to work we would only start feeling it later this year or the year after that...we are of sort of living in deep seeks Golden Era the failure of the Biden 2022 export controls what they see in the future is the pain of the 2023 the 2024 you know Etc um export control so I think that lagging impact um is really important to note”
OpenAI had an exclusive agreement with Microsoft that is no longer in place; Microsoft CEO Satya Nadella was not present at the Stargate announcement but later confirmed Microsoft's $80 billion commitment to the deal.
“OPI had this exclusive agreement with Microsoft no more open I got more friends now so does Microsoft right and and notably Microsoft CEO Sati Nadella not at that announcement yep yeah yep and then well you mentioned the 500 billion it's the first 100 billion now but couple of hours later Elon Musk was tweeting guys you actually don't have the money right”
Sam Altman stated that OpenAI will build AI agents that can perform tasks autonomously, which will eventually automate many economically valuable jobs, but this aspect of the AI race is not being featured in infrastructure investment announcements.
“if you look at like opi's explicit Mission it is to build hii how do they find hii doing all economically valuable tasks so to put it in like Sam ulman terms like there might be jobs lot of nice jobs in the meanwhile but in the end when you have ai when you guys have your explicit goal we we got to talk about the jobs again right like the explicit goal is like to automate all of these kinds of things and um I think this hasn't really come through”
Elon Musk is currently suing OpenAI, making the public disagreement over Stargate funding a continuation of ongoing legal and commercial conflict between Musk and OpenAI leadership.
“there's been this ongoing drama of Elon Musk actually suing open my eye right now right”
Andre Karpathy, founding team member at OpenAI, stated DeepSeek achieved frontier-level performance with 248 GPUs for two months of training, whereas the same level of capability was supposedly supposed to require clusters of close to 16,000 GPUs.
“Andre karpathy...said on x quote deep seek making it look easy with an open weights release of a frontier grade llm trained on a joke of a budget uh 248 gpus for two months for reference this L level of capability is supposed to require clusters of close to 16,000 gpus”
Mark Andreessen said DeepSeek R1 is 'one of the most amazing and impressive breakthroughs I've ever seen' and described it as 'a profound gift to the world' because he is a strong advocate for open-source software.
“Mark andreason...he said on X deep seek R1 is one of the most amazing and impressive breakthroughs I've ever seen and as open source a profound gift to the world right Mark andreason by the way uh big big fan of Open Source everything”
OpenAI's previous exclusive computing partnership with Microsoft is over; Microsoft CEO Satya Nadella was notably absent from the Stargate announcement event despite stating he committed $80 billion to the project, suggesting a shift in OpenAI's infrastructure leverage.
“OPI had this exclusive agreement with Microsoft no more open I got more friends now so does Microsoft right and and notably Microsoft CEO Sati Nadella not at that announcement”
Mark Zuckerberg predicted that in 2025 Meta and other companies will have AI systems that can function as mid-level software engineers capable of writing code; Alexander Wang and Elon Musk have made similar near-term predictions about AI capabilities.
“Mark Zuckerberg he said quote probably in 2025 we at meta as well as the other companies that are basically working on this are going to have ai that can effectively be a sort of mid-level engineer that you have at your company that can write code uh Alexander Wang said similar things Elon Musk is saying similar things”
Elon Musk tweeted skepticism about whether Stargate partners actually have the $500 billion committed, but Sam Altman responded that Musk was wrong, and the companies clarified they are raising the money rather than having it all available upfront.
“couple of hours later Elon Musk was tweeting guys you actually don't have the money right so there's like some ongoing debate right now do they actually have the money to yeah to deploy these kinds of things... I I I think it's it's worth pointing out that Sam Alman responded you're wrong you know but um I do think that there's a difference between like do have a lot of money right now which which I'm confident they do”
Satya Nadella (Microsoft CEO) stated that Deep Seek is 'super impressive in terms of both how they have really effectively done an open source model that does this inference time compute and its super compute efficiency' and advised that the developments from China should be 'taken very very seriously.'
“Sati Adella at the the the CEO of Microsoft he said quote to see the Deep seek new model it's super impressive in terms of both how they have really effectively done an open source model that does this inference time compute and its super compute efficiency we should take the developments out of China very very seriously”
The amount of compute available (determined by export controls and manufacturing capacity) directly determines how many AI agents can be deployed in an economy and how many workers can be replaced; this is the real national security stakes of export controls beyond just benchmark performance on published models.
“how much computer you have determines how many AI gen you can use it determines the size of your Workforce like literally and it also determines like this thinking time right”
If DeepSeek trained substantially on ChatGPT-generated outputs, this represents a 'fast-follower' strategy where China benefits from OpenAI's massive investment in downloading and filtering internet data without bearing that cost.
“if you're deep seek and you're training off chat gbt data chat gbt has already done all the hard work of trying to find which parts of the internet are actually useful as as training data and that's a really big challenge if true right you have to ask yourself what is the first mover advantage in AI right now it could be the case and I think China is desperately hoping that it is the case that fast follower is a really nice place to be”
Model performance benchmarks are insufficient for evaluating competitive advantage; product quality, user experience, deployment breadth (e.g., being default on iPhones), data availability, and political acceptance (avoiding geopolitical red flags) all matter more than raw benchmark scores in determining market dominance.
“we like overly focusing on like Benchmark performance would be nice if it would be about the world like who has the best product who has the best Benchmark performance on these kinds of things there are many things which you can't measure benchmarks right... it doesn't matter if you have the best model it depends how you deploy it right if you tomorrow a company which Apple picks to deploy on every smartphone around the world I claim you're a big winner”
DeepSeek CEO explicitly stated that export controls are the biggest challenge the company faces, indicating that the company takes export control threats seriously and views the current competitive window as temporary.
“deep seek CEO said you know an explicit quote um the biggest challenge our company faces is these compute export controls so we are of sort of living in deep seeks Golden Era the failure of the Biden 2022 export controls what they see in the future is the pain of the 2023 the 2024 you know Etc um export control”
DeepSeek's CEO has explicitly stated that 'the biggest challenge our company faces is these compute export controls,' indicating that export controls do constrain Chinese AI development; the timing of DeepSeek's announcement during Trump's first week signals a Chinese government messaging effort to convince the U.S. to abandon controls, similar to the Huawei phone release during Gina Raimondo's China visit.
“deep seek CEO said you know an explicit quote um the biggest challenge our company faces is these compute export controls so we are of sort of living in deep seeks Golden Era the failure of the Biden 2022 export controls what they see in the future is the pain of the 2023 the 2024 you know Etc um export control so I think that lagging impact um is really important to note”
Stargate's $500 billion investment will depreciate rapidly because AI chips have a ~3-year useful life before becoming obsolete due to rapid performance improvements, meaning investors are betting on extraordinary performance improvements in the next three years to justify chips that will lose nearly all value afterward, while energy and data center infrastructure depreciate more slowly.
“the AI chips are a big big big component of that investment and they depreciate over three years you've lost almost all the value of the chip just because chips are getting better at such an incredible rate um after three years so it's a pretty scary bet as an investor to say okay we're going to spend a hundred billion do to buy buy these assets that we know are going to go to zero 3 years from now”
Larry Ellison and Elon Musk are close personal friends who socialize together, and Musk previously had a close friendship with Google founder Larry Page that ended over disagreements about AI safety and existential risk, suggesting Musk's public criticism of Stargate may signal another potential rift with a tech billionaire.
“Larry Ellison is a close personal friend of Elon Musk um they hang out in each other's you know Villas they they uh go on vacation together...Elon Musk also used to be the very close personal friend of Larry pagee uh the former CEO of Google and...the reason why they are no longer friends is literally an argument that they had about AI safety um and...Elon Musk is obsessed with that idea so much so that he like literally when uh when he was the CEO of Tesla in the early years he didn't have a house in San Francisco he stayed at Larry Page's house”
DeepSeek demonstrates that 'necessity is the mother of invention'—because China faced export control constraints, they were forced to discover compute-efficiency workarounds that Western companies may not have developed, creating genuine technical innovations that are now available to everyone.
“Arvin sovos uh apologies if I got that name wrong but he's the CEO of perplexity AI he said quote necessity is the mother of invention because they had to figure out workaround they actually ended up building something that has a lot more efficient the workarounds I should point out that he's referring to are the export controls”
The Trump Administration faces a political choice regarding AI safety measures: while the term 'AI safety' is radioactive in conservative circles due to associations with perceived social bias and DEI initiatives, some Republican officials like Cara Frederick recognize that regulatory guardrails matter.
“the word AI safety has been poisoned it is radioactive in some very conservative circles so I think that's a very reasonable take so I think there's the the question of like when they hear AI safety there are parts of the Republican policy community that immediately jump to censorship on social media”
The Stargate infrastructure requires staggering amounts of energy to operate, connecting to the Biden Administration's infrastructure executive order which simplified permitting for data center and energy infrastructure development; Trump Administration is endorsing the underlying permitting reform ideas despite opposing the executive order itself.
“we saw the diffusion Rule and the diffusion rule explicitly says well you please build an us but to build an us we need energy and I think all when we talk about 500 billion you need staggering amounts of energy right just like doing this we saw Microsoft wanting to activate fre m Island again and what this last exit over order basically tries to do is just like well simplify permitting making it easier to get access to the energy”
The difference between publishing a research paper and deploying a production model at scale is critical: Deep Seek demonstrated capability research, but the Western labs are planning deployment at orders of magnitude larger scale, which determines economic value and real-world impact.
“the amount of compute that it takes to publish a research paper and the amount of compute that it takes to be the dominant AI model provider to the customers of planet Earth those are different amounts of compute absolutely and I think when we talk about the 500 billion for Stargate well I hope they want to make money at some point right so they got to deploy it right people are actually got to use it”
DeepSeek's emergence raises the question of whether export controls have failed if China can discover new compute-light approaches to frontier AI that become the dominant paradigm, challenging the assumption underlying the export control strategy that compute scarcity constrains Chinese AI capabilities.
“secondary this is something Leonard and I and and Briel were going to have to get into is like what does this mean for the export controls um because if our theory of the universe was AI is computation intensive we can stop China from getting enough computation to compete um the plenty of folks in Washington DC are asking me this question of does this mean the export controls have failed like has China discovered a new more compute light approach to AI um that is actually going to be the dominant Paradigm as opposed to everything that we were betting on um in ideas like Stargate and in ideas like uh the export control architecture”
The semiconductor industry took 80+ years to mature; it is much easier to catch up in newer industries (like AI, which has only been on the frontier for ~4 years) compared to mature industries; therefore, catching up in AI is expected to remain easier for longer than in semiconductors.
“the best example to think about is like when we think about the semiconduct industry well it's e way easier to catch up if you're in the first 10 years of a new industry and it's a semiconduct industry but if we 80 years down the pipe I don't expect any startup Tomorrow come out of nowhere and build um asml euv machine so like challenge tsmc right whereas like well we like nness AI I companies are thing like what four years are they even profitable right now like is the question so like it's like way easier to catch up on these kinds of things”
Xi Jinping explicitly spent 25% of his time in meetings with President Biden requesting that export controls be lifted, indicating he views the controls as actually working and threatening to China's AI competitiveness.
“I have heard that you know uh from you know Biden Administration officials that in meetings between uh Joe Biden and XI jingping like XI jingping spent 25% of the time saying that these export controls need to stop they need to stop right now right he's not using 25% of that meeting to ask them to stop because he doesn't think they're working he's asking them to stop because he thinks they are working”
The dominant trend in AI technology right now is 'smaller, faster, cheaper' capabilities, creating a tension between companies needing to justify massive compute investments (Stargate at $500B) and the reality that models are becoming more efficient.
“for folks who don't listen to the Strater podcast there was a really good interview that Ben Thompson did with Nat fredman and Daniel gross where they said one of the dominant trends that appears to be going on in AI technology right now is smaller faster cheaper um and the question sort of becomes well who are the customers who can really justify laying out a lot of money for the absolute best performance and who are the customers for which you know this cheaper model which is still pretty good is going to be good enough and all of that then raises the question is Stargate crazy right are they getting ready to make a crazy crazy investment at a time in which the future might be cheaper and smaller”
The inference time compute paradigm is itself compute-intensive, which explains why Stargate's massive investment in compute infrastructure is still necessary despite the efficiency gains, because models will need massive compute clusters to provide inference time compute at deployment scale.
“and this is why everybody is sort of losing their minds on AI is we've got this new approach to driving performance improvements and we have barely even begun squeezing the first grapes right to make uh you know to make wine out of that new uh model and this is related to what we were just talking about at Stargate this new paradigm shocker it's going to be compute intensive and that is why they're sort of saying this is the right moment to spend hundred billion doar on an assets that that's going to depreciate over three years and why we're willing to make that hundred billion doll investment you know every year for the next four years it's because they see we're we're close we're close to AGI now is the sort of time to to strike and get a really strong first mover advantage”
If everyone in the world starts using Deep Seek models, the company will face a GPU crunch and cannot deploy enough instances to serve all users, facing deployment constraints that limit revenue and reduce incentive to invest in next-generation training runs.
“I will I will promise you if everybody now starts using deep seek they will have a GPU crunch they simply cannot deploy the model that many instances like serve all of these users does this matter well maybe it matters it matters you make less money Therefore your next training run is smaller and you buy less gpus right”
Occupations and tasks most amenable to reinforcement learning (coding, mathematics) are ironically the ones most at risk of near-term AI replacement, and this threatens the people currently doing AI research and development; AI companies are strategically automating coding first both because it's technically easible and because it benefits their own R&D operations.
“it's not a surprise that coding is making all this revolutionary progress um it's not a surprise that math is seeing all of this evolutionary progress because those are domains in which the simulator is a near equivalent representation of the real world um and so reinforcement learning is so good and so that's another reason why folks you know who are out there who are saying like hey I played around with chat GPT in October and I played around with chat gbt in January and I'm not seeing all that much progress meanwhile you have Silicon Valley you know screaming their heads off about how much progress they've seen it's like well are you a coder are you using chat gbd to generate software because there we have seen an awful lot of progress”
Greg Allen distinguishes between having capital on hand right now and having the capacity to raise capital; the Stargate announcement was partly a prospectus to other potential investors, and the fact that they're already building in Texas suggests they have raised substantial sums already, even if not the full $500 billion upfront.
“I do think like this this was kind of the advertising perspectus for you know other people who want to contribute the money look look we've got the backing of the American president look we've said loudly and proudly we're going to do this so uh if you're you know the head of SoftBank and you want to raise all this money that event was very helpful to you um they clearly got a lot of money because as Larry Ellison said they're already building in Texas um I would be surprised if they already had you know $500 billion like in a bank somewhere ready for this project but it's clear that they're going to raise jaw-dropping amounts of money”
Frontier AI companies value their partnership with the U.S. government's AI Safety Institute because it provides an independent third-party validation that they can use as a counterargument against European regulators who propose stricter AI controls, framing the AI Safety Institute as evidence that America has credible guardrails without Draconian overreach.
“the companies have sort of said they like the partnership that they have with the AI safety Institute they like that the government is this sort of independent third party who they have and they really really like the fact that it gives the United States a talking point as it's negotiating with Europe and saying hey Europe you're proposing to regulate American AI companies in all these ways that we view as Draconian and Regulatory overreach and our rebuttal to that is can either be nothing or it can be the AI safety Institute this other sort of transparency measures etc etc”
The semiconductor industry required 80+ years of evolution to reach current maturity (ASML EUV machines represent decades of embedded expertise); by contrast, AI companies are only ~4 years old and competition is still in early stages, making catch-up easier in AI than in semiconductors where lock-in is extreme.
“when we think about the semiconduct industry well it's e way easier to catch up if you're in the first 10 years of a new industry and it's a semiconduct industry but if we 80 years down the pipe I don't expect any startup Tomorrow come out of nowhere and build um asml euv machine so like challenge tsmc right whereas like well we like nness AI I companies are thing like what four years are they even profitable right now like is the question so like it's like way easier to catch up on these kinds of things”
Sam Altman stated on January 5, 2025 that OpenAI is 'now confident we know how to build AGI as we have traditionally understood it' and expects the first AI agents to join the workforce in 2025 and 'materially change the output of companies.'
“Sam Altman saying quote we are now confident we know how to build AGI as we have traditionally understood it we believe that in 2025 we may see the first AI agents join the workforce and materially change the output of companies”
Cara Frederick, Trump's new special assistant for technology and national security policy, has stated that 'guard rails matter' and that the 'wild west has not worked out in other areas of tech,' indicating the Trump Administration may distinguish between legitimate AI safety guardrails and what it perceives as woke censorship.
“Cara Frederick uh who was formerly uh the head of Tech policy at the Heritage Foundation but now has been made a special assistant to the president with a focus on technology policy and national security policy... she said quote guard rails matter and I don't think the wild west has worked out in other areas of the tech space”
Trump's January 23 AI executive order directs an interagency process within 180 days to develop an AI action plan, involving the Assistant to the President for Science & Technology (Michael Katos), the AI Czar (David Sachs), the National Security Advisor (Mike Waltz), the OMB Director (Russ Vance), and cabinet heads, with no clear single authority over the process, suggesting unresolved debates within the Republican Party on AI policy will be resolved through this planning period.
“within 180 days of this order the assistant to the president for science technology the special adviser for AI and crypto that's the quote unquote AI Zar David Sachs and the assistant president assistant to the president for National Security Affairs um in coordination with the assistant of the president for economic policy... shall develop and submit to the president an action plan”
The Biden Administration's AI Diffusion Rule required U.S. cloud providers to build at least 50% of their compute infrastructure domestically as a condition of becoming Universal Validated End Users; Oracle was outspoken in opposing this rule and reportedly wanted to build data centers in Malaysia, but Stargate's commitment to build exclusively in the U.S. may represent Oracle's strategy to offset the diffusion rule restrictions.
“the diffusion rule was expit saying hey if you're like a cloud provider an American cloud provider like Oracle you got to build 50% at least in yes we also know that Oracle was quite outspoken about the diffusion rule not really liking it right and it's been like some reent reporting they wanted to build some big data centers in Malaysia which they probably can't know anymore or maybe they're just trying to offset it”
Current AI company funding is primarily VC-driven rather than revenue-driven, meaning near-term deployment constraints (export controls limiting addressable markets) do not yet constrain training run sizes; but eventually, as companies mature and VC capital tightens, revenue will determine future training run viability.
“right now we don't work in a paradigm where this the size of your training run or like the capabilities of your IM models determine by how much money you make this is mostly VC funding this is mostly people betting on these kinds of things so we haven't seen this yet right and this might continue for like while like going on”
The test-time compute paradigm (allowing models to spend more computational resources 'thinking' before answering) is compute-intensive and justifies Stargate's $100B annual investment over 4 years because companies must build enormous clusters to deploy inference-time scaled models at scale while also continuing to train larger base models.
“this new paradigm shocker it's going to be compute intensive and that is why they're sort of saying this is the right moment to spend hundred billion doar on an assets that that's going to depreciate over three years”
The trend toward 'smaller, faster, cheaper' AI models (per Stratechery analysis by Ben Thompson, Nat Friedman, and Daniel Gross) raises a fundamental question: are companies like OpenAI crazy to invest $500 billion in Stargate when the future may be cheap, efficient models that most customers prefer over cutting-edge expensive models?
“there is this world in which we just keep getting better and better at distilling models and so what is World shaping Le good one year can like run on your laptop 3 to 5 years later... there's a really good interview that Ben Thompson did with Nat fredman and Daniel gross where they said one of the dominant trends that appears to be going on in AI technology right now is smaller faster cheaper um and the question sort of becomes well who are the customers who can really justify laying out a lot of money for the absolute best performance and who are the customers for which you know this cheaper model which is still pretty good is going to be good enough and all of that then raises the question is Stargate crazy”
The fact that Western AI leaders read DeepSeek papers and learn techniques from them, while Western labs keep their development private, creates an information asymmetry where the West can copy China but China cannot fully copy the West.
“I've talked to Folks at some of the leading AI firms and they say like we read deep seek papers and we learn things... could just be the case that deep signal like found ways which like the companies already found like year ago and they just didn't tell anyone right because like since the last few years we've been entering this new paradigm these companies all used to publish all of their papers but they said how they're doing it right now we don't even know how big their clusters are”
Trump administration appointees including Michael Kratos, Russell Vought, Mike Waltz, and the AI czar will have to work together on AI policy without clear hierarchy, and the fact that they haven't resolved internal Republican disagreements about AI means the 'plan to plan' timeline allows them to reach consensus.
“these folks have to work together they're the ones who have to come up with the plan and it is not clear at least from the language in this order that there's like one person who is in charge over all of these different folks it's like sort of a classic inter agency process at least to me um I think they're that they're going to run on this so this is a plan to plan um which uh I think on to to the Trump administration's credit they were ready to go with this on day one but it's not as though the sort of debates that have been going on in the Republican Party had been resolved to such an extent that on day one they already had that plan”
Larry Ellison is a close personal friend of Elon Musk; they socialize and vacation together; this personal relationship makes Musk's critical tweet about Stargate's funding credible only as skepticism, not as evidence of fundamental disagreement, though it mirrors a pattern where Musk previously severed friendship with Larry Page (Google) over AI safety disagreements.
“Larry Ellison is a close personal friend of Elon Musk um they hang out in each other's you know Villas they they uh go on vacation together they've known each other uh for ages”
Gnome Brown, OpenAI researcher, reported a 100,000x improvement in AI performance from test-time compute (allowing the model more time to 'think' before answering) in initial experiments, compared to 100x improvements from techniques developed during his 5-year PhD, signaling a fundamental paradigm shift in how AI systems improve beyond traditional training-time scaling.
“gnome Brown... he said um something to the effect of you know I spent my entire PhD like 5 years chasing all these improvements in AI performance and I worked on some important stuff and everything I worked on I think you could charitably say delivered a 100x Improvement in the performance of AI systems and working on inference time inference scaling of AI systems we saw you know in our first experiment on the topic a 100,000 X Improvement”
The test-time compute paradigm is still in early exploratory stages, and scaling patterns are uncertain; both pre-training scaling and test-time compute scaling may continue in parallel as separate paradigms, creating potential for different sized models (base models and distilled models) all improving simultaneously.
“we saw the current generation of models coming out like these 10,000 GPU clusters right or like like like thousands of of GPU clusters which is like the current generation now they're building a bunch of clusters via Stargate we know El mus build 100K cluster we haven't seen these models being trained yet right and I touched upon previously if pre-training like the scaling Paradigm continues in parallel to the test time compute and we will see like the next big pre-train model either being used as a base model or being used for distilling then we will see a bigger Gap emerging again we just haven't entered this new era yet”
If the Trump Administration cannot streamline energy permitting and data center deployment in the U.S., countries like the UAE and Middle East may become attractive alternatives for AI infrastructure buildout; alternatively, partners like Canada or Greenland could be considered for distributed data center development with access to energy resources.
“if you cannot do it here and they want to build elsewhere where are we going to build right and like maybe just practically speaking if you just can simply cannot stem it here do you then just want to follow economic incentive and build an uee on Middle East or do we want to like partner up with I don't know Canada Greenland let's be creative here and like build more data centers there and like get the energy going there”
DeepSeek's timing (release during Trump's first week as president) and the Huawei phone release during Gina Raimondo's China visit were likely coordinated by the Chinese government to signal that export controls don't work, creating propaganda pressure on Washington.
“but why are we hearing about deep seek during Trump's first week of the presidency what could it possibly be they're trying to send a message that export controls don't work but I have heard that you know uh from you know Biden Administration officials that in meetings between uh Joe Biden and XI jingping like XI jingping spent 25% of the time saying that these export controls need to stop they need to stop right now right he's not using 25% of that meeting to ask them to stop because he doesn't think they're working he's asking them to stop because he thinks they are working and deep seek is absolutely part of the Chinese government's messaging strategy to the United States industry to the United States government and it's a critical moment uh with this Trump memo”
The Obama Administration and Biden Administration have revised the export control architecture multiple times (October 2023, December 2024, January 2025), and these revised controls will likely not take effect until later in 2025 or after, meaning China is currently operating in a window where weaker export controls are in effect.
“we've revised the architecture in October 2023 we revised it again in December 2024 we revised it again in January 2025 and if that control architecture is going to work we would only start feeling it later this year or the year after that and I I've said this before but it's worth pointing out again deep seek CEO said you know an explicit quote um the biggest challenge our company faces is these compute export controls so we are of sort of living in deep seeks Golden Era the failure of the Biden 2022 export controls what they see in the future is the pain of the 2023 the 2024 you know Etc um export control so I think that lagging impact um is really important to note”
Aravind Srinivas (Perplexity AI CEO) explained Deep Seek's efficiency breakthrough as the result of 'necessity is the mother of invention' because of export controls forcing workarounds, suggesting that export control restrictions inadvertently drove Deep Seek to develop more efficient training approaches.
“Arvin sovos uh apologies if I got that name wrong but he's the CEO of perplexity AI he said quote necessity is the mother of invention because they had to figure out workaround they actually ended up building something that has a lot more efficient the workarounds”
Greg Allen suggests that people who claim they are not seeing dramatic AI progress between October and January are likely not using AI for coding; those who are using code generation tools (like ChatGPT for software development) are seeing 'an awful lot of progress,' demonstrating that AI capability improvements are domain-specific (best in coding/math, less obvious in other domains).
“folks you know who are out there who are saying like hey I played around with chat GPT in October and I played around with chat gbt in January and I'm not seeing all that much progress meanwhile you have Silicon Valley you know screaming their heads off about how much progress they've seen it's like well are you a coder are you using chat gbd to generate software because there we have seen an awful lot of progress”
Howard Lutnick, Trump's Secretary of Commerce nominee, is viewed as a China Hawk but has not made public statements specifically about semiconductor and AI export controls, making his confirmation hearing a key moment to gauge Trump administration policy direction.
“lutnick is viewed in general as a China Hawk but I don't think we have a statement on the record for him uh from him about what he thinks about the Biden administration's approach to export controls on semiconductors and AI related technology so that's a confirmation hearing I cannot wait to watch I think we're going to learn an awful lot”
Mark Andreessen stated that Deep Seek R1 is one of the most amazing and impressive breakthroughs he's ever seen and called it a profound gift to the world because it is open source.
“you've got Mark andreason the the venture capitalist who now is close to the Trump Administration he said on X deep seek R1 is one of the most amazing and impressive breakthroughs I've ever seen and as open source a profound gift to the world right Mark andreason by the way uh big big fan of Open Source everything”
DeepSeek V3 responses sometimes identified themselves as ChatGPT, suggesting the training data may have included outputs from OpenAI's ChatGPT, allowing DeepSeek to benefit from OpenAI's data filtering and training work.
“funly enough with V3 the model was accessible some people prompted and asked me if you like hey uh how you doing what's your name by the way and sometimes I responded hey I'm chub yeah so like like what model are you distilling here guys it sounds like maybe you found a way and and and to talk about how this might work um you know what is the training data set well it's possible that one of the things that deep seek did is the training data set was asking chat GPT questions and then you know stealing all of that data and feeding it directly into the training data set for deep seek”
DeepSeek V3 sometimes claims to be 'Claude' when prompted, suggesting the model may be distilling or fine-tuning on Anthropic Claude data, raising questions about what training data is behind the efficiency claims.
“with V3 the model was accessible some people prompted and asked me if you like hey uh how you doing what's your name by the way and sometimes I responded hey I'm chub yeah so like like what model are you distilling here guys it sounds like maybe you found a way”
Elon Musk tweeted that OpenAI/SoftBank don't actually have the $500 billion committed to Stargate, though Sam Altman disputed this and Satya Nadella stated he is 'good for my 80 billion,' leaving the actual immediate liquidity of the project ambiguous.
“a couple of hours later Elon Musk was tweeting guys you actually don't have the money right so there's like some ongoing debate right now do they actually have the money to yeah to deploy these kinds of things”
Reinforcement learning is a machine learning paradigm where the system interacts with an external environment (or simulation), logs the experience, and then uses that data as training data for the next update; this was the foundation of AlphaGo's breakthrough in 2014-2015, where the system played against itself and improved recursively.
“there's this word called reinforcement learning and the lovely thing about reinforcement is the the system sort of interacts with an external environment whether that's a simulation or the real world it it sort of logs its experience in that external reaction and then that that data log of that experience is now the training data for the next update in the system”
Energy infrastructure and data center cooling infrastructure are much more durable than AI chips, lasting significantly longer than three years and continuing to generate value; Stargate's investment includes these longer-lived assets alongside rapidly depreciating chips.
“some of that stuff like the energy infrastructure vastly more durable going to continue kicking out really useful value for a really long period of time a lot of the data center infrastructure like the cooling infrastructure um just all the real estate Etc that's going to last for longer than 100 billion uh sorry longer than one year three years”
The CSIS team, including Greg Allen, Leonard Hein, Georgia Adamson, and Sam Winter-Levy, has published a research report on UAE-US collaboration on data center buildout based on a research trip to the UAE, which is being released the day of this podcast episode.
“we have a report dropping on the UAE us uh collaboration on data data center buildout um coming today actually co-authored by Leonard Georgia ad Adamson Greg Allen and Sam winter Levy um of the Carnegie Institute for peace...this is the uh the trip report from our our research trip to the UAE where we got to tour a lot of cool data centers uh meet with a lot of interesting folks”
A UAE-US collaboration research report on data center buildout and AI infrastructure is being released concurrently with this podcast, co-authored by Leonard Heim, Georgia Adamson, Greg Allen, and Sam Winter Levy of the Carnegie Institute for Peace.
“one thing's for certain 2025 is going to be a big year for AI advancement I want to thank you for joining us today on the AI policy podcast exciting news for all of our listeners we have a report dropping on the UAE us uh collaboration on data data center buildout um coming today actually co-authored by Leonard Georgia ad Adamson Greg Allen and Sam winter Levy um of the Carnegie Institute for peace”