By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
a16zWhat DeepSeek Means For The Future Of AI | Tech Veterans Weigh In
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
- 0:00 – 4:08
DeepSeek's Release and Why It Blew Up
Steph Smith asks for a TLDR of the DeepSeek release and why it exploded. Steven Sinofsky and Martin Casado describe a surprise release from a small Chinese team, the cost claims, and the market overreaction that followed.
- •A Chinese hedge fund and research group released a full model that seemed to appear from nowhere
- •The widely cited $5-6 million training cost drew attention, along with the team's paper on training innovations
- •Sinofsky calls the trillion-dollar market-cap move an overreaction
- •Casado ties the hype to excitement around OpenAI's o1 reasoning model and to rumors such as a CCP psyop
- 4:08 – 9:46
Separating Signal from Noise: A Strong Team and Scaling Constraints
Casado argues DeepSeek is a very capable team whose work built on public literature, and that its cost is in line with what others have said they spent. Sinofsky says it exposed the hyperscalers' capital-and-data mindset, and both discuss possible data advantages from the Chinese internet.
- •Casado says V3 was arguably the more impressive feat and that a reasoning model requires a base model like it
- •The $6 million spent on chain of thought is not out of line with what Anthropic and OpenAI have said they spent
- •Sinofsky says US labs defaulted to 'ginormous' scale and saw the problem through capital and data
- •Casado raises two theories: engineering under constraints, and access to Chinese internet data and cheap expert annotation
- 9:46 – 11:21
The MIT License and Released Reasoning Traces
Casado explains two under-reported aspects of R1: its permissive MIT license and the release of reasoning steps. These enable distillation of smaller models and wider proliferation.
- •The license is basically MIT, which Casado says is rare for a state-of-the-art model
- •App-layer companies use many models, so licenses matter
- •OpenAI did not release o1's chain of thought, while DeepSeek did
- •Distillation lets a teacher model train much smaller student models cheaply, so more models can run on smaller devices
- 11:21 – 16:04
Where to Monetize: Models, Licensing, and the Web Server Analogy
Sinofsky compares charging for LLMs to charging for web servers, arguing that real money sits in higher layers and that licensing will matter. Casado takes the other side, noting apps may need to vertically integrate with models, and the two settle on time as the key variable.
- •Sinofsky compares monetizing the LLM layer to the early internet's focus on HTML and HTTP
- •Licensing models for different layers will matter as standardization emerges
- •Casado lays out a fork: models become commoditized, or apps need to own the model
- •Sinofsky says both can be true over time, with new apps emerging and then vertically integrating
- 16:04 – 18:41
New Apps, Not Zero-Sum Replacement
Using internet and mobile history, Sinofsky and Casado argue the important AI apps will be ones that don't exist yet, rather than replacements for old ones. Sinofsky uses PowerPoint to show how AI-native products will differ from what they resemble.
- •Apps that mattered on the internet and mobile did not exist before them
- •Zero-sum thinking is dangerous when new technology divides the spectrum differently
- •Sinofsky says LLMs replacing search will be very hard because models are bad at much of what search does
- •AI-native tools may look like Canva or PowerPoint but will lack the old formatting complexity
- 18:41 – 22:19
Is This Time Different? Capex, Fiber Gluts, and Balance Sheets
Steph Smith asks whether anything differs from past waves. Casado compares today's data center buildout to the fiber buildout but says the funding foundation is much stronger. Sinofsky agrees and says no one will come out unscathed, though not in the expected ways.
- •In the 1990s, investors who wanted internet exposure funded fiber; today big investors want to fund data centers
- •The fiber crash involved a debt-laden company, while AI is funded by big cloud companies with large balance sheets
- •Sinofsky compares current spending to Google's, Facebook's and Microsoft's past investments, such as Bing
- •Both say the damage will not look like past crashes
- 22:19 – 29:22
Scale Up vs. Scale Out and the Lessons of Past Networks
Sinofsky explains scale up versus scale out through mainframes, workstations, and microcomputers, and says model developers have a blind spot about on-device AI. Casado adds the telephone-network versus internet analogy, and Sinofsky shares his QOS story.
- •Big players keep building bigger machines until a cheap, distributed alternative wins
- •Sinofsky says phone MIPS are free while data center MIPS are expensive, and he runs Llama on his Mac
- •Casado says telephone networks' reliability was beaten by best-effort internet delivery that enabled endpoints
- •Casado sees DeepSeek as a step toward AGI in your pocket and rejects shorting Nvidia
- •Sinofsky links specialized on-device models to the browser getting JavaScript
- 29:22 – 32:01
Changing Benchmarks
Steph Smith asks whether benchmarks should shift from bigger and better toward things like cost and device fit. Sinofsky calls the current benchmarks silly, compares them to browser rendering races, and expects application-specific measures like truth for research apps.
- •Early browser benchmarks measured image rendering speed and ended up not mattering
- •Application-specific metrics will matter, such as truth in research tools
- •Research tools lean on information retrieval, footnotes, and vector databases
- •Sinofsky expects ImageNet-style test sets for truthfulness
- 32:01 – 35:24
Building Defensibility: Workflows, Many Models, and Enterprise
Casado says models alone are not defensible because others distill quickly, while stateful apps built around engaged users are. Sinofsky adds that many-model apps and enterprise features like SSO will become moats.
- •Large models are magic on their own but competitors catch up quickly through distillation
- •Defensible companies build retentive, stateful apps around their model once users engage
- •The era is shifting from large models to workflows powered by many models
- •Sinofsky says enterprise needs such as SSO, controls, and alignment features are sticky and a way to price
- •Steph Smith relays Scott Belsky's point that enterprises care about licensed images where consumers don't
- 35:24 – 41:15
Is It AI's Sputnik Moment? Policy and the Internet Lesson
Asked about the Sputnik framing, Casado says the biggest takeaway is how wrong-headed US AI policy has been, including export controls and limits on open source. Sinofsky says the lesson is the internet, not Sputnik, and that regulation should enable diffusion rather than control.
- •Casado says export controls on chips did not stop China and that US policy should fund research and move fast
- •Sinofsky says Al Gore's contribution was regulation that let the internet flourish
- •Past export controls on encryption and PlayStations failed
- •Sinofsky says technology diffusion now takes essentially zero calendar time, citing DeepSeek's usage relative to OpenAI
- •Casado hopes a politician will be the 'Al Gore of this moment'
- 41:15 – 44:29
Impact on Frontier Labs and Participating in the Next Wave
Casado does not see a crisis for OpenAI, Anthropic, or Nvidia, but a wake-up call. Sinofsky says frontier labs should build apps, expects the TAM to be huge, and says the quant-fund origin shows innovation comes from unexpected places.
- •Casado says apps are hard to build and the labs can stay ahead
- •Sinofsky says frontier labs should build apps for the best platform feedback loop
- •Sinofsky expects a TAM of 100X with revenue mainly from apps
- •Sinofsky notes the internet's foundations came from universities, a physics lab, and a failed corporate lab
