Jan 28, 2025 · 2h 7m · tbpn
DeepSeek Update, Market Crash, Timeline in Turmoil, Is VC Cooked, Zero Cope Policy
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Co-hosts John Coogan and Jordy Hays analyze the macroeconomic fallout and tech industry panic surrounding DeepSeek's R1 release, examining technical shifts toward inference-time compute scaling and deconstructing hardware moats. Applying Jevons Paradox, they argue that collapsing inference costs will expand aggregate compute demand and empower lean vertical software companies rather than derail the AI economy.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 74.3% of the talking time here. How this is scored →
The hosts as informed peer 6.9 Guest teaching 2.6 Guest disagreement 1.7 The hosts pushing back 1.9
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 45:05 Anthropic Chain of Thought Release Dispute
Hays highlights claims that DeepSeek beat Anthropic to market on chain-of-thought reasoning, prompting Coogan to firmly disagree that this represents a genuine technical lead rather than a strategic release choice.
Hardest push from the hosts ▶ 47:45 Rejecting DeepSeek Zero-to-One Innovation FramingCoogan refuses the breathless framing of DeepSeek cracking autonomous reasoning from scratch, pointing out that OpenAI demonstrated identical capabilities with Q* and GPT-Zero eighteen months prior.
Biggest teaching moment ▶ 1:08:25 Singapore Export Backdoor AdmissionCoogan directly acknowledges that Hays was right in their previous debate regarding Singapore serving as a massive export funnel for embargoed Nvidia hardware based on newly reported Q3 billing figures.
The host holds their own ▶ 11:50 Technical Explanation of Test-Time Compute DynamicsCoogan clearly articulates the fundamental shift from pre-training compute to inference-time compute scaling, explaining how intermediate reasoning tokens transform model unit economics.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Market Crash Reaction and Introducing Jevons Paradox | 7 | 2 | 1 | 1 | Coogan dominates the opening by unpacking Jeffrey Emanuel's short thesis on Nvidia, explaining pre-training scaling laws, token volume metrics, and data wall limitations. The co-hosts are completely collaborative with no real pushback. | |
| Inference-Time Compute Scaling and Chain of Thought Models | 8 | 2 | 1 | 2 | Coogan provides a detailed technical breakdown of test-time compute scaling versus traditional pre-training compute, describing how logic tokens in chain-of-thought models like o1 mitigate transformer hallucinations. Hays supports the analysis with analogies to white-collar tasks. | |
| Hands-On Model Evaluations and the AI Adoption Cycle | 7 | 3 | 2 | 2 | Coogan shares his hands-on evaluation comparing o1 Pro to DeepSeek R1 on a long-form writing prompt, noting R1 failed word count instructions. Hays offers commentary on the AI adoption curve and diminishing perceptual returns as models surpass average human IQ. | |
| Deconstructing Nvidia's Moat and Competitive Threats | 8 | 2 | 1 | 1 | Coogan breaks down Nvidia's structural advantages across CUDA, Linux drivers, and Mellanox interconnect bandwidth, while discussing threats from Cerebras wafer-scale engines and hyperscaler custom silicon. | |
| DeepSeek Efficiency Claims and Skepticism of Breakthroughs | 7 | 3 | 2 | 2 | The conversation examines DeepSeek's training cost claims and efficiency figures, with Hays sharing an anecdote illustrating state-level PR incentives in China and Coogan citing historical precedents in rapid open-source optimization like Stable Diffusion. | |
| Technical Innovations, Reinforcement Learning, and OpenAI Precedents | 8 | 2 | 3 | 4 | Coogan challenges the narrative that DeepSeek invented chain-of-thought reasoning from scratch, tracing the lineage back to Ilya Sutskever's Q* and OpenAI's internal experiments, framing R1 as architectural optimization rather than zero-to-one invention. | |
| The AI Value Stack and Distribution vs. Raw Compute | 7 | 2 | 2 | 2 | Coogan outlines the multi-layered AI stack from foundational weights up to UI orchestration, debating with Hays whether DeepSeek's developer focus or OpenAI's consumer brand will capture long-term platform value. | |
| Timeline Reactions, Export Backdoors, and Market Turmoil | 6 | 5 | 2 | 1 | Reviewing market reactions across Twitter, Coogan explicitly concedes to Hays that shipping data showing billions in Nvidia GPU shipments to Singapore confirms the export backdoor thesis Hays had previously argued. | |
| Jevons Paradox and Expanding Future Compute Demand | 7 | 2 | 1 | 1 | The hosts analyze Jevons Paradox in the context of Satya Nadella's and Gary Tan's posts, with Coogan explaining how drastic inference cost reductions will trigger exponential new compute demand for ubiquitous personal AI filtering. | |
| State-Backed Open Source and the Consumer Moat | 6 | 3 | 2 | 2 | Coogan and Hays debate the strategic implications of state-level open sourcing, discussing regulatory capture attempts by US frontier labs and the shifting cultural narrative around AI safety. | |
| Security Outages, Mainstream Perceptions, and Hardware History | 6 | 3 | 2 | 2 | Evaluating server outages and historical analogues like Netscape versus Internet Explorer, Coogan argues that open source availability does not negate the value of OpenAI's 500 million monthly active user distribution moat. | |
| Zero Cope Policy, Deregulation, and Multimodal Experiments | 7 | 2 | 2 | 2 | Coogan discusses the 'zero cope policy' and deregulation under the Trump administration, drawing an analogy to iPhone 3GS processing power enabling real-time fact-checking and recounting his experiments generating video with OpenAI Sora. | |
| Is Venture Capital Cooked? Axios Panic vs. Vertical SaaS Reality | 6 | 3 | 1 | 2 | The hosts dissect Axios coverage claiming DeepSeek is an extinction-level event for VC, pushing back by citing Gary Tan's argument for vertical SaaS explosion and the historical returns of web3 venture funds. |
Statements from this episode (33)
Hays: Market will realize by end of day it overreacted to DeepSeek
“Because I think by the end of the, I mean, it's almost one, but by the end of the day, people are going to wake up. They're going to be like, all right, we have reacted a little bit. We didn't understand.”
Coogan: Nvidia Holds Near-Monopoly on AI Training and Inference Capex
“They wound up with a basically something close to a monopoly in terms of share of aggregate industry capex that's spent on training and inference infrastructure for artificial intelligence.”
Coogan: Google Books Holds 2.6T to 5.2T Tokens in 40M Digitized Books
“Google books has digitized around forty million books so far. If a typical book has five 50,000 to a 100,000 words or 65,000 to a 130,000 tokens, then that's between 2.6 trillion and 5.2 trillion tokens just from books.”
Coogan: Web Scraping Copyright Lawsuits Against AI Labs Will All Settle
“They'll all get a settlement.”
Coogan: OpenAI's $200 o1 Pro uses the same model as $20 o1
“In fact, the O-one model is 20 dollars a month, is basically the same model used in the O-one Pro model for 10 X the price at 200 dollars a month, which raised plenty of eyebrows. The main difference is that O-one Pro thinks for a lot longer before responding”
Coogan: OpenAI's o3 High-Compute Mode Spent $3,000 to Solve a Benchmark Task
“Oh three, which isn't out yet, but is even more advanced in terms of reasoning. They have a high compute model. That spends almost 3000 dollars per task. And it just thinks for hours and hours and hours basically, and it was able to break arc that that AI, AGI…”
Coogan: Inference-Time Compute Establishes an Independent Scaling Law from Pre-Training
“This development creates a new scaling law that is totally independent of the original pre-training scaling law. Now you still want to train the best model you can by clearly leveraging as much compute as you can as many trillion tokens of high quality trainin…”
Coogan: DeepSeek R1 failed prompt length test compared to OpenAI o1 Pro
“O-One Pro delivered basically exactly 5000 words, and it even, as it was writing the story, it would say like, introduction, 400 words. Act one, 600 words. And it totally, it kept this like internal log, and then at the end it was like, I have written 5400 wor…”
Hays: Average users cannot tell if AI models become 10% smarter
“It's getting to a point pretty much now where the models are becoming smarter than the average person. So the average person cannot tell the difference Like if the model gets 10% smarter.”
Coogan: AMD GPUs Cost Roughly Half the Price of Nvidia Per Flop
“In fact, in terms of naive raw dollars per flop, AMD GPUs are something like half the price of NVIDIA GPUs.”
Hays: Nvidia Sustains Luxury Margins on Unseen Data Center Infrastructure
“Nvidia has the margins of a luxury product, but they are building a product that goes and sits in a data center and nobody sees it. No, no, no consumer is going to say, well, I actually really want my model to be running on Nvidia chips”
Coogan: DeepSeek V3 Was ~45x More Training-Efficient Than Competitors
“V-III remains the top ranked open weights model despite being around 45 X more efficient in training than its competition.”
Coogan: DeepSeek-R1 API is 27 times cheaper than OpenAI's o1
“The deep seek R one API is currently 27%, 27 times cheaper than open AI's O one for a similar level of quality.”
Coogan: Anthropic has an unreleased chain-of-thought model
“Like anthropic does have a chain of thought model. They just haven't released it publicly. And the reason for that is just financials and like their safety stuff.”
Coogan: DeepSeek R1 Trails OpenAI by 18 Months and Lacks 0-to-1 Innovation
“Basically, you know, for years, that's covered and working on ways to allow language models, like GPT four to solve tasks that involved reasoning, like math and science problems in 20, 21, he launched a project called GPT zero. What did they launch? Deep seek …”
Coogan: Western AI Labs Will Port DeepSeek's Optimizations Into Their Models
“It's open source. The paper's out there. So that will get ported back to Llama, to OpenAI, to Anthropic, but I haven't seen that.”
Coogan: DeepSeek will push inference costs down and disrupt B2B AI
“And this will drive the cost down even further. So hugely disruptive to the B to B space.”
Hays: China announced approximately $500B in new state-funded tech investment
“They just announced the equivalent of like five hundred billion dollars of new Like state funded Investment.”
Hays: Nvidia valuation is hard to justify as competitive moats crack
“So at the high level, Nvidia faces an unprecedented convergence of competitive threats that make its premium valuation increasingly difficult to justify at 20 x forward sales and 75% gross margins. The company's supposed moats and hardware software and efficie…”
Coogan: DeepSeek is a massive upgrade over free ChatGPT
“Like if you have ChatGPT free version and you download DeepSeq, that is a massive upgrade if you're a free user.”
Coogan: Test-time inference will get much cheaper when built into silicon
“I mean, we talked about like the test time inference is going to get so much cheaper when this is baked down into Silicon, but we're just not there yet. Cause we're updating models every two years.”
Hays: Mistral, Anthropic, and Cohere must match free models to survive
“The thing that's most fascinating to me is all these model companies, Mistral Anthropic, Cohere that really don't publicly have these capabilities. Where now anybody employed at those companies basically shouldn't sleep. For the next, however long it takes to …”
Coogan: Nvidia booked $7.7B in Singapore chip sales in Q3 2024
“In the three months ended October, 27, 20, 24. So like deep chip ban, just not last quarter, but the quarter before Q three 7.7 billion dollars of chips to Singapore. And it's like, Singapore's not buying that many chips.”
Coogan: Consumer AI aggregators capture value while B2B AI faces pure price competition
“There seems to be a massive pool of value in Being the consumer AI company, the aggregator, the front page of artificial intelligence, getting installed on the home row of people's apps, setting it to the default search engine, setting it to the default webpag…”
Coogan: AI companies should charge until a competitive model goes free
“Until until a competitive model, model goes free, you should not go free. This is just basic econ one-on-one. This is actually more capitalist. So like we are the capitalist. It's like charge until you can.”
Hays: Chinese labs have insiders at all major American AI labs
“Well, one thing, but I'm sure Chinese labs Have people inside at all the major American labs, and so anything that is being discovered at the American labs is being ported back.”
Coogan: DeepSeek's top innovation is its real-time reasoning UI
“This is probably the most innovation that's happened with the deep seek thing is that UI paradigm of like showing you the reasoning as it works through the model and it just makes it way more engaging because you enter a query and then it immediately starts ta…”
Hays: Users Will Stop Caring About Exposed AI Reasoning Once Trust Builds
“I think there's this period of time where it's true. People want to see how it's working through something, but then eventually when you have that level of trust with the model or the app that you're using, you just want it done.”
Hays: Compute Demand Must Grow Because AI Impact Has Reached Only 1%
“You can't say we sh like these data centers are worthless. While also, or unnecessary, while also agreeing that AI's impact has only been felt one percent. Which I would say most people feel at this point that, that AI's only impacted our economy or society or…”
Hays: Consumer AI App Layer Will Be More Competitive Than Model Layer
“This consumer App layer is going to be more competitive in many ways than the foundation model layer because you're competing with Meta, Apple, all these different, you know, Google, et cetera, that have the distribution already.”
Hays: Users Will Distrust Chinese AI Models With Personal and Financial Data
“You have to imagine that people will be less likely to trust the Chinese developer with the operator flow, which is inputting your card details and highly personal information, which is different than just querying a chat interface to write me an essay on this…”