TBPN
TBPN

Sep 3, 2026 · 2h 26m

Listen from 48:38

Listen at 48:38

AI’s next edge may be specialized data, not bigger models

Model Mayhem, NVIDIA x Hugging Face Deal, GPT-6 Astra | Pablo Torre, Mohit Aron, Akshay Narisetti, Hari Ravichandran, Matt Caldwell & Jordy Leiser, Jeff Thornburg, Charlie O’Neill, Sridhar Ramaswamy

The episode tests whether AI progress will be decided by benchmark scores or by trusted data, infrastructure, oversight, and distribution.

3 key takeaways
  1. 1New model releases intensify competition while exposing the limits of benchmarks, rollout discipline, and enterprise trust.
  2. 2Open-source ecosystems, model routing, and continual learning could make specialized systems more valuable than one universal model.
  3. 3AI’s commercial frontier extends beyond software into recording hardware, consumer security, sports platforms, spacecraft, and enterprise data systems.

Don't miss

The hosts react to GPT-6 Astra’s official launch and its striking ARC AGI 3 result, then question whether the benchmark says anything conclusive about AGI.

The brief

TBPN opens amid a rush of new model releases, an AI-services outage, and debate over whether benchmark gains translate into reliable enterprise value.

The reported NVIDIA-Hugging Face deal puts open-source models, model routing, and ecosystem control at the center of the AI competition.

GPT-6 Astra’s launch and striking ARC AGI 3 result sharpen the episode’s central tension: impressive benchmarks still leave rollout quality and real-world usefulness unsettled.

Guests broaden the frame: Pablo Torre discusses the Clippers investigation, while Mohit Aron, Akshay Narisetti, and Aura’s founder examine context, recording hardware, and consumer security.

The later conversations turn to direct-to-consumer sports, maneuverable spacecraft, continual learning, and Snowflake’s enterprise strategy, all emphasizing specialized data and trusted infrastructure.

The episode’s through line is that AI progress depends on distribution, oversight, and systems that learn from context—not simply on producing a larger model.

What was said on this episode

54 statements · 36 positive · 14 negative · 3 mixed · 1 neutral

  1. Jordi Hayson Anthropic Fable 5.1Positive3:42

    Anthropic’s Fable 5.1 currently leads the three reviewed models on the AI index.

    “Anthropic seems to have the strongest model of the three with Fable 5.1 scoring 66 on the artificial intelligence index.”

    Listen at 3:42

  2. Jordi Hayson Anthropic Fable 5.1Positive3:42

    Anthropic’s Fable 5.1 currently leads the three compared models.

    “Anthropic seems to have the strongest model of the three with Fable 5.1 scoring 66 on the artificial intelligence index.”

    Listen at 3:42

  3. Jordi Hayson Anthropic FableNegative6:23

    Anthropic’s prior data-retention policy may have reduced enterprise Fable adoption.

    “And that may have contributed to lower Fable adoption among enterprises.”

    Listen at 6:23

  4. Jordi Hayson Google Gemini 3.8 FlashNeutral6:38

    Google released its third Gemini Flash model within six weeks.

    “Gemini 3.8 Flash is the company's third Flash release in six weeks.”

    Listen at 6:38

  5. Jordi Hayson Google Gemini 3.8 FlashPositive6:45

    Gemini 3.8 Flash matches costlier models competitively on DeepSwee.

    “It scored 73.7% on DeepSwee, just behind Opus 5, and competitive with models that cost several times more.”

    Listen at 6:45

  6. Jordi Hayson Google Gemini 3.8 FlashPositive7:03

    Gemini 3.8 Flash is fast, inexpensive, and strong at coding on DeepSwee.

    “So very quick, very cheap, and very good at coding, at least on this particular benchmark, DeepSwee.”

    Listen at 7:03

  7. Jordi Hayson NVIDIA acquisition of Hugging FacePositive12:59

    NVIDIA is acquiring Hugging Face for approximately $13 billion.

    “NVIDIA is going even bigger on open source with the $13 billion acquisition for hugging face.”

    Listen at 12:59

  8. Jordi Hayson NVIDIA open-source AI ecosystem strategyPositive15:05

    Jensen Huang wants NVIDIA to anchor open-source model selection and GPU deployment.

    “he wants to be the place where developers and companies go to pick their models and then hopefully rack them on NVIDIA GPUs.”

    Listen at 15:05

  9. Jordi Hayson Hugging FacePositive19:24

    Hugging Face became the default home for AI models.

    “Hugging Face very quickly became the default. home for AI models.”

    Listen at 19:24

  10. Jordi Hayson Hugging FacePositive19:24

    Hugging Face became the default home for AI models.

    “Hugging Face very quickly became the default. home for AI models.”

    Listen at 19:24

  11. Jordi Hayson Hugging FacePositive21:21

    Hugging Face raised under $400 million before a roughly $12 billion outcome.

    “Huggy Face, I'm going to say only raised under $400 million, which is a lot of money, but not at a $12 billion outcome.”

    Listen at 21:21

  12. Jordi Hayson Hugging FacePositive21:43

    Hugging Face became profitable in 2025 while retaining half its raised capital.

    “They became profitable in 2025, still had half the money that they raised.”

    Listen at 21:43

  13. John Cooganon OpenAI AstraPositive24:09

    Leaked Astra benchmarks reportedly show an ARC AGI score of 398.6.

    “Arc AGI 398.6.”

    Listen at 24:09

  14. John Cooganon OpenAI AstraPositive24:28

    Astra’s leaked benchmark results appear strong despite no official launch.

    “So seems pretty good, but still no official announcement.”

    Listen at 24:28

  15. Jordi Hayson GrokNegative28:01

    Grok’s AI companions may disappear within six months rather than become dominant.

    “But this is the same thing as the Sora discourse, where everyone was caught up in Sora is either going to be the most powerful thing ever, and we're not going to be able to stop watching it, or it's going to be... good and so fun and it's going to be amazing and dominant, but no one was counting just like, oh, it might just go away in six months.”

    Listen at 28:01

  16. John Cooganon AI romantic companions marketNegative28:37

    The romantic-companion market does not generate trillion-dollar revenue.

    “there's just not a trillion dollars of revenue.”

    Listen at 28:37

  17. Pablo Torreon Kawhi LeonardNegative35:51

    Kawhi Leonard received $48 million under an unannounced agreement without performing work.

    “Kawhi Leonard was paid, according to this agreement, a total of $48 million, 20 in stock, 28 in cash, to do nothing for a deal that never got announced.”

    Listen at 35:51

  18. Pablo Torreon Steve BallmerNegative36:07

    Steve Ballmer used prohibited wealth transfers to circumvent the NBA salary cap.

    “there's a massive salary cap circumvention scheme in which the richest owner in American sports, Steve Ballmer, is trying to use all of his wealth in ways he's not allowed to to get money to a guy”

    Listen at 36:07

  19. Pablo Torreon Kawhi LeonardNegative42:04

    The Clippers-related arrangements allegedly used fake jobs and consulting agreements for Kawhi Leonard.

    “What they really did was they created fake jobs for Kawhi Leonard and fake consulting agreements, consulting fees for the companies.”

    Listen at 42:04

  20. Pablo Torreon Steve BallmerNegative43:31

    The NBA punishment was reportedly the largest ever imposed on an American sports owner.

    “this was the biggest punishment for an owner in the history of, I think, American professional sports.”

    Listen at 43:31

  21. The Clippers will not select another first-rounder for many years.

    “The next Clippers first rounder that they will be able to select hasn't even sniffed puberty yet.”

    Listen at 46:51

  22. Pablo Torreon Los Angeles Clippers valuationNegative48:38

    The Clippers would sell for less than ten billion dollars.

    “they're not getting double digit. They're not getting age figures.”

    Listen at 48:38

  23. Pablo Torreon Los Angeles Clippers valuationPositive48:50

    The Clippers would sell for more than observers expect despite franchise problems.

    “they're going to make more than people, I think, are prepared to intuit based on the fact that, intuit is a pun not even intended, based on how horrific this franchise has been.”

    Listen at 48:50

  24. Pablo Torreon Steve Ballmer purchase of another NBA teamNegative54:33

    Ballmer’s conflict with NBA owners could hinder approval to buy another team.

    “one way to be approved to buy a new team is to not go to war with the 29 other ownership groups that will need to approve that.”

    Listen at 54:33

  25. Hari Ravichandranon Cyfin go-to-market strategyPositive57:18

    Cyfin will target both large enterprises and smaller companies.

    “My sweet point is obviously going to be the big enterprise, but I don't want to leave the smaller companies behind. So it's going to be all of them.”

    Listen at 57:18

  26. Hari Ravichandranon Go-to-market operationsNegative58:17

    Go-to-market teams commonly operate with insufficient context.

    “We execute our go-to-market operations without enough context.”

    Listen at 58:17

  27. Hari Ravichandranon CyfinPositive59:47

    Cyfin will gather data automatically, retain human oversight, and provide real-time coaching.

    “The solution includes all of that.”

    Listen at 59:47

  28. Akshay Narisettion Pocket recording devicePositive1:02:58

    Pocket users used its recording device ten times more than its app.

    “people were using the device 10 times more than the app.”

    Listen at 1:02:58

  29. Akshay Narisettion Pocket recording devicePositive1:07:31

    Recording convenience determines whether users record conversations.

    “convenience beats intentions.”

    Listen at 1:07:31

  30. Akshay Narisettion Pocket recording conveniencePositive1:07:40

    Making recording convenient substantially increases users’ likelihood of recording.

    “if you actually make it convenient enough to record and they want to record even a little bit, they'll a hundred percent record.”

    Listen at 1:07:40

  31. Akshay Narisettion Rabbit and Humane AI devicesMixed1:10:28

    Rabbit and Humane failed, creating lessons that benefited Pocket.

    “Rabbit and Humane died so Pocket could live.”

    Listen at 1:10:28

  32. Akshay Narisettion AI note-taking devicesPositive1:11:22

    AI note-taking achieved product-market fit while broader device features did not.

    “note-taking was essentially a product market fit.”

    Listen at 1:11:22

  33. Akshay Narisettion Pocket AI transcriptionPositive1:15:44

    Pocket claims approximately 75% margins on its AI transcription business.

    “Probably 75% margin in topical description, essentially.”

    Listen at 1:15:44

  34. Hari Ravichandranon Consumer scamsNegative1:19:57

    Aura sees average scam losses rise to $25,000 per scam.

    “Now we see that's gone up to $25,000 per scam on average.”

    Listen at 1:19:57

  35. Hari Ravichandranon AuraPositive1:25:34

    Aura centers its security product on a family relationship graph.

    “we're probably the only product that puts the actual family at the center.”

    Listen at 1:25:34

  36. Hari Ravichandranon AuraPositive1:26:16

    Aura has approximately $340 million in annual recurring revenue.

    “We're about $340-some-odd million of ARR.”

    Listen at 1:26:16

  37. Jordi Hayson OpenAI AstraPositive1:27:03

    GPT-6 Astra scored 99.9% on ARC AGI 3.

    “The score is 99.9%.”

    Listen at 1:27:03

  38. Jordi Hayson OpenAI AstraPositive1:28:06

    GPT-6 Astra scored 64.6% on the stated reasoning benchmark.

    “GPT-6 Astra scored 64.6% on reasoning effort max.”

    Listen at 1:28:06

  39. Akshay Narisettion Jump sports ticketing platformPositive1:34:39

    Jump’s platform gives sports teams direct access to ticketing and fan data.

    “it's a way to have the data on your side of the table, on your app”

    Listen at 1:34:39

  40. John Cooganon personalized AI modelsPositive1:54:46

    AI will evolve toward hundreds of millions of personalized models, potentially one per person.

    “we're going to end up in a world with some great models from the Frontier Close Source Labs. But we're going to end up with not 5 million, but probably hundreds of millions of models. Possibly even one model for each person.”

    Listen at 1:54:46

  41. John Cooganon Personalized AI modelsPositive1:54:46

    The future may contain hundreds of millions of personalized AI models.

    “we're going to end up in a world with some great models from the Frontier Close Source Labs. But we're going to end up with not 5 million, but probably hundreds of millions of models. Possibly even one model for each person.”

    Listen at 1:54:46

  42. Hari Ravichandranon Open-source artificial intelligencePositive1:58:27

    After reaching task capability thresholds, organizations should generally switch to open-source models.

    “once you have the base level of intelligence required to do it, you probably do want to swap to open source.”

    Listen at 1:58:27

  43. Hari Ravichandranon continual model learningPositive1:59:28

    Model training will eventually become more organic and customer-driven.

    “I think eventually we'll get there where it's much more organic.”

    Listen at 1:59:28

  44. Hari Ravichandranon AI model evaluationPositive2:03:33

    Real-world economic usage is the best basis for evaluating AI models.

    “The best way to do this is aggregate the economy.”

    Listen at 2:03:33

  45. Sridhar Ramaswamyon AI model evaluationPositive2:04:17

    Real economic usage is a better model-evaluation signal than niche benchmarks.

    “The best way to do this is to look at how everyone is using them in the economy for real things that people are paying for”

    Listen at 2:04:17

  46. Sridhar Ramaswamyon Continual learningNegative2:05:40

    Continual learning research remains pre-paradigmatic and lacks consensus.

    “I think we are pre-paradigm when it comes to things like continual learning.”

    Listen at 2:05:40

  47. John Cooganon continual learningMixed2:07:34

    Continual learning is either unnecessary or requires brute-force scaling.

    “I actually think that continual learning is either unnecessary or it is brute forceful.”

    Listen at 2:07:34

  48. John Cooganon Continual learningMixed2:07:34

    Continual learning is either unnecessary or requires brute-force retraining.

    “I actually think that continual learning is either unnecessary or it is brute forceful.”

    Listen at 2:07:34

  49. Sridhar Ramaswamyon continual learning for enterprise agentsNegative2:08:22

    Continual learning currently breaks down for specialized enterprise agents.

    “Continual learning breaks down. We don't have an answer to it.”

    Listen at 2:08:22

  50. Sridhar Ramaswamyon AI automationPositive2:12:29

    AI can largely automate work previously requiring individual employees.

    “What AI makes possible is stuff that needed individuals, people to go and do can largely be automated.”

    Listen at 2:12:29

  51. Sridhar Ramaswamyon AI automationPositive2:12:29

    AI can largely automate work previously requiring individual employees.

    “What AI makes possible is stuff that needed individuals, people to go and do can largely be automated.”

    Listen at 2:12:29

  52. Sridhar Ramaswamyon SnowflakePositive2:17:26

    Snowflake considers its AI token spending worthwhile.

    “The money that we are spending on AI tokens is well worth the cost.”

    Listen at 2:17:26

  53. Sridhar Ramaswamyon Snowflake CoworkPositive2:20:10

    Snowflake’s Cowork is being deployed to thousands of enterprise users.

    “We are seeing deployments. of co-work go to thousands of users within companies.”

    Listen at 2:20:10

  54. Jordi Hayson GPT-6Positive2:24:30

    GPT-6 has been released.

    “GPT-6 is out.”

    Listen at 2:24:30

Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.

Weekly email

State of the Pod, every Monday

The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.

Listen to the full episode and explore every guest, topic, and moment on PodLume.

AI’s next edge may be specialized data, not bigger models · PodLume