Synthesis

Open models: China, open weights and the frontier

The hosts' claims, mostly Dave's, that Chinese and other open-weight models (GLM, Kimi, DeepSeek) are close to or at the frontier at far lower cost, plus the geopolitics around ch…

Type
Synthesis
Confidence
Medium
Sources
21 videos
First published
Updated

Chinese and open-weight models appear in 17 episodes and the story is Dave’s, with Sam and Jess partly agreeing. Dave’s evidence is mostly back-channel, personal tests and unnamed friends, and he flags some of it. Mar 2026: OpenClaw users run Kimi to save on Anthropic. May: Dave says America has no real open-source take while China prints models and every OpenClaw founder uses them to cut costs. June to July: Dave runs GLM 5.2 on rented B200s (150 tokens/second, about ten times frontier speed) and says he no longer needs frontier models, that open and frontier are ‘neck and neck’ and under two months apart; Sam predicts Chinese open source will do about 90 percent of heavy lifting. Jess pushes back that frontier models have gained share recently. Sep to Oct: DeepSeek said to match a Fable-level model at about 100x less cost; an unnamed host says enterprise spend is moving to open models. Geopolitical threads include smuggled chips, DUV lithography progress and a possible political backlash.

Key points

  • Founders and CFOs swap in cheaper open models: ‘luxury tokens’ undercut, Cursor RL’d on Kimi, OpenRouter growth, token spend moving. — 7 episodes. (Dave · Fire Sam Altman, The End of So (Mar 2026) ▶ 42:53) (Ben Broca · Henrik Werdelin & Ben Broca (Mar 2026) ▶ 42:02) (Josh Wolfe · Peter Deng & Rachel Holt (Aug 2026) ▶ 24:05)
  • Open-weight Chinese models match or approach frontier quality at much lower cost (Dave’s tests and relays). — 4 episodes. (Dave · Every CTO Is Quietly Switching (Jul 2026) ▶ 41:19)
  • The US lacks a strong open model of its own; Thinking Machines’ 975bn open model is described as the first heavyweight. — 4 episodes. (Dave · Why the Met Gala Hates Tech, E (May 2026) ▶ 12:14)
  • Geopolitics: chips, DUV lithography progress, regulatory backlash against Chinese models and open source. — 4 episodes. (Jess · Meta’s AI Campaign, While Riva (Jul 2026) ▶ 28:53)
  • Dave claims US frontier models are hobbled on cyber and bio, so responders used GLM or Kimi (back-channel, uncertain). — 1 episode. (Dave · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 15:10)

Where guests disagree

Are open models taking share from the frontier labs? ⚠️

  • Sam: likely and OpenRouter is growing like crazy, though hard to measure; Dave: neck and neck. (Sam · SpaceX Hit $2.5 Trillion. Now (Jun 2026) ▶ 25:49)
  • Jess: frontier models have gained share over recent months. (Jess · SpaceX Hit $2.5 Trillion. Now (Jun 2026) ▶ 25:45)

Tactics

  • Use a cheap open-weights model for utility agent turns — Dave switched GLM 5.2 in for almost every utilitarian use case in his agents instead of paying for frontier models, because it is described as near frontier and very cheap. (Dave · Open Source AI Wins. Now Labs (Jul 2026) ▶ 06:04)
  • Get a group together to rent GPUs and run an open-weight model. — Dave ran GLM 5.2 on eight rented Nvidia B200s with hacker friends instead of paying frontier prices. (Dave · Chinese AI Model GLM 5.2 Beati (Jun 2026) ▶ 27:24)

Stories & examples

  • GLM 5.2 on eight B200s: Dave and hacker friends from the 2000s rented eight Nvidia B200s and ran the open-weight GLM 5.2; it did 150 tokens per second out of the box and matched or beat frontier models on his coding tasks. Lesson: Open-weight models on shared standard hardware can replace frontier models for some work, which makes big frontier-lab positions look shaky. (Dave · Chinese AI Model GLM 5.2 Beati (Jun 2026) ▶ 26:47)
  • Cursor’s Composer 2: Ben says Cursor dropped Composer 2, which he believes was RL’d on a Chinese open model (Kimi 2.5), as a way to make margins. Lesson: Custom models on open source can get 90% of the value. (Ben Broca · Henrik Werdelin & Ben Broca (Mar 2026) ▶ 42:02)
  • Tech executives say one thing publicly and another privately on China: Jess says execs tell Washington the US must beat China at all costs, but semi-privately say both will lead and dismiss war questions as stupid, then adopt the expedient talking point three weeks later. Lesson: Treat public AI-race rhetoric as positioning, not necessarily belief; she cannot name the people. (Jess · Meta’s AI Campaign, While Riva (Jul 2026) ▶ 30:01)

Numbers

  • GLM 5.2 speed on that rig: 150 tokens per second (about 10x frontier models) (host’s framing) (Dave · Chinese AI Model GLM 5.2 Beati (Jun 2026) ▶ 26:56)
  • Size of GLM 5.2: 1 trillion parameters (host’s framing) (Dave · Every CTO Is Quietly Switching (Jul 2026) ▶ 41:27)
  • Cost of DeepSeek’s Fable-level model relative to Fable: 100 times cheaper (host’s framing, approx.) (Dave · Anthropic Wants to Slow Down. (Sep 2026) ▶ 06:27)
  • Share of frontier work Chinese open source could do (Sam): 90 percent (host’s framing, approx.) (Sam · SpaceX Hit $2.5 Trillion. Now (Jun 2026) ▶ 24:46)
  • Dave’s Blackwell desktop box price: 5000 dollars (host’s framing, approx.) (Dave · Why the Met Gala Hates Tech, E (May 2026) ▶ 13:29)

Caveats & warnings

  • Dave’s GLM 5.2 evidence rests on back channels and unnamed friends, and the model is only about three weeks old. (Dave · Every CTO Is Quietly Switching (Jul 2026) ▶ 42:10)
  • Dave’s claim that US frontier models are hobbled and that Kimi was needed rests on back-channel engineering talk and an ‘I think’ about K3. (Dave · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 15:00)
  • Jess says she is not an AI researcher and does not know what level of distillation is typical. (Jess · Hype vs Reality: Our 2026 Pred (Dec 2025) ▶ 26:46)

Also covered here: Open versus closed AI

Open versus closed appears in 17 episodes, again led by Dave. The framing: open source is a distribution strategy when you are behind (Mar 2026), with Sam calling the American world closed and Apple-like and open source the Android of AI. By Jul an unattributed host expects three quarters of the world to choose open source ‘because they have to’, and Sam predicts frontier labs lose on cost for most agent work, surviving as an emergency rip cord. Josh Wolfe (a Lux investor in Hugging Face, who says his own stance depends on the cap table) argues enterprises will keep proprietary data silos and run open models on them. Dave and Sam relay that labs are lobbying to make open source illegal or restricted, which Dave quotes Bill Gurley as calling nonsense; Jess says she hears of regulatory buzz and that frontier models have gained share. Practical advice includes routing across models, owning your data and chartering open projects as nonprofits. Evidence is largely anecdotal.

Key points

  • Regulatory fight: reports that labs seek to restrict open source; Sam compares it to Napster and the DMCA. — 5 episodes. (Dave · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 25:40)
  • Founders are swapping open models for frontier ones, and frontier labs must keep raising the bar because open lags. — 4 episodes. (Dave · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 25:00)
  • Governance: open projects die if one company captures them, so charter a nonprofit (OpenClaw Foundation); Dave frames the goal as individuals holding the most powerful AI. — 4 episodes. (Dave · Nvidia’s GTC Was “Taylor Swift (Mar 2026) ▶ 35:06)
  • Open source as a distribution play when behind (Android of AI, China running Meta’s playbook). — 2 episodes. (one of the hosts · Fire Sam Altman, The End of So (Mar 2026) ▶ 45:13)
  • Own your data and point models at it: enterprises keep proprietary silos and swap models; routing preserves optionality. — 2 episodes. (Sam · Peter Deng & Rachel Holt (Aug 2026) ▶ 10:47) (Josh Wolfe · Peter Deng & Rachel Holt (Aug 2026) ▶ 23:01) (Matt Mazzeo · Matt Mazzeo (Sep 2026) ▶ 21:51)
  • Position follows stake: ‘where you stand depends on where you sit on the cap table’. — 1 episode. (Josh Wolfe · Peter Deng & Rachel Holt (Aug 2026) ▶ 22:29)

Where guests disagree

Is open source eating frontier-lab value? ⚠️

  • Sam, Dave, an unidentified host: yes for most agent work on cost. (Sam · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 34:26)
  • Jess: frontier models have been gaining share; Dave also says open lags so frontier labs must keep making the next model. (Jess · SpaceX Hit $2.5 Trillion. Now (Jun 2026) ▶ 25:45)

Frameworks

  • Labs as stealth bombers: zero-day models held in reserve — Sam’s thought experiment, not a pitched business model: open-source models serve almost all everyday agent work cheaply, while a few labs keep unseen, very expensive models and capacity that governments and companies call in when a bot rebellion breaks out. (Sam · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 32:22)
    • Everyone uses cheap open-source models for nearly all work
    • Frontier labs lose that commodity work on cost
    • Labs keep secret, extremely powerful models and on-demand compute
    • When a rebellion or incident hits, the labs are the only ones to call and can charge enormous sums (he floats a hundred billion dollars)

Stories & examples

  • Founder event: everyone swapping in open source models: Dave watched 10 to 20 founder lightning talks from sizeable businesses and says everyone was trying to replace as much frontier model use with open source as possible. Lesson: Real implementers treat open source as the margin lever. (Dave · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 25:00)
  • Napster and the DMCA as a model for regulating AI: Sam compares the open-weights fight to Napster: the record labels got control over moving music, leaving digital music hobbled; he adds that the labs ripped the whole internet to build version one and now complain about distillation. Lesson: Attempts to control copy-able digital goods tend to produce hobbled systems. (Sam · Open Source AI Wins. Now Labs (Jul 2026) ▶ 17:52)

Numbers

  • Android share of all smartphones, as claimed in the iOS/Android comparison: 73 percent (self-reported, approx.) (one of the hosts · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 30:29)
  • Hypothetical fee for a lab to put down a bot rebellion in Sam’s thought experiment: 100 billion dollars (host’s framing, approx.) (Sam · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 33:32)
  • Parameters of the Thinking Machines open weights model (Dave): 975 billion parameters (host’s framing, approx.) (Dave · Open Source AI Wins. Now Labs (Jul 2026) ▶ 16:21)

Caveats & warnings

  • Guest has a stake: Josh says where you stand depends on the cap table and that Lux is a large investor in Hugging Face. (Josh Wolfe · Peter Deng & Rachel Holt (Aug 2026) ▶ 22:29)
  • Dave says open models under about 70 billion parameters are weaker: hard to keep talking to and tool calling is only okay. (Dave · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 29:41)
  • Dave’s GLM 5.2 evidence rests on back channels and unnamed friends, and the model is only about three weeks old. (Dave · Every CTO Is Quietly Switching (Jul 2026) ▶ 42:10)
  • Sam says not all tokens are the same so baselining is hard, and the field lacks language for measuring productivity; he thinks the economics of tokens is a major research opportunity. (Sam · SpaceX Hit $2.5 Trillion. Now (Jun 2026) ▶ 27:39)

How much to trust this

  • Evidence: 25 episodes, 5 independent speakers (the four hosts and unidentified host speech count as one voice, the show; guests counted individually).

Blueprints that use this

Sources

  1. (01)

    Why the Met Gala Hates Tech, Elon vs OpenAI Drama, and the Rise of Chinese AI Models

    YouTube

    12:14 into “Why the Met Gala Hates Tech, Elon vs OpenAI Drama, and the Rise of Chinese AI Models”

  2. (02)

    OpenClaw vs Meta vs OpenAI: The Personal Agent Wars Heat Up

    YouTube

    6:07 into “OpenClaw vs Meta vs OpenAI: The Personal Agent Wars Heat Up”

  3. (03)

    Stripe Is Trying To Buy PayPal For $150b | OpenAI Cyber Incident, Frontier Model Economics, Peptides

    YouTube

    15:00 into “Stripe Is Trying To Buy PayPal For $150b | OpenAI Cyber Incident, Frontier Model Economics, Peptides”

  4. (04)

    iPhone Duo, Meta Muse & Why AI Finally Scared Dave Morin

    YouTube

    32:54 into “iPhone Duo, Meta Muse & Why AI Finally Scared Dave Morin”

  5. (05)

    Open Source AI Wins. Now Labs Are Running to Washington |Thinking Machines, Johnny Ive, Realtime API

    YouTube

    5:58 into “Open Source AI Wins. Now Labs Are Running to Washington |Thinking Machines, Johnny Ive, Realtime API”

  6. (06)

    Elon Musk vs OpenAI Trial, Google Cloud Surge, Meta’s Blocked Acquisition, Anthropic Winning

    YouTube

    25:00 into “Elon Musk vs OpenAI Trial, Google Cloud Surge, Meta’s Blocked Acquisition, Anthropic Winning”

  7. (07)

    OpenAI's $852 Billion Valuation Makes Zero Sense (ClawCon Tokyo, Anthropic's Leak, Goodbye Internet)

    YouTube

    37:38 into “OpenAI's $852 Billion Valuation Makes Zero Sense (ClawCon Tokyo, Anthropic's Leak, Goodbye Internet)”

  8. (08)

    AI Can Change Everything and Still Be a Terrible Business + Peptide 101 from Matt Mazzeo

    YouTube

    21:45 into “AI Can Change Everything and Still Be a Terrible Business + Peptide 101 from Matt Mazzeo”

  9. (09)

    Meta's AI Campaign, While Rivals Ask Washington To Slow it Down | Lilian Weng Saga, China Cracks DUV

    YouTube

    25:31 into “Meta's AI Campaign, While Rivals Ask Washington To Slow it Down | Lilian Weng Saga, China Cracks DUV”

  10. (10)

    Nvidia’s GTC Was “Taylor Swift for Straight Men” (Apple, OpenAI, and OpenClaw News)

    YouTube

    1:46 into “Nvidia’s GTC Was “Taylor Swift for Straight Men” (Apple, OpenAI, and OpenClaw News)”

  11. (11)

    Paramount vs Netflix: Who Really Wins Warner Bros And What's Next?

    YouTube

    40:46 into “Paramount vs Netflix: Who Really Wins Warner Bros And What's Next?”

  12. (12)

    Fire Sam Altman, The End of Software Engineers, and Why AI Is All Narrative

    YouTube

    42:05 into “Fire Sam Altman, The End of Software Engineers, and Why AI Is All Narrative”

  13. (13)

    The Internet Is Fighting Back Against Your AI Agent

    YouTube

    10:26 into “The Internet Is Fighting Back Against Your AI Agent”

  14. (14)

    The Risk Of Securitizing GPUs Like Long-Lived Infrastructure & More - A MOL Special Team Edition

    YouTube

    10:47 into “The Risk Of Securitizing GPUs Like Long-Lived Infrastructure & More - A MOL Special Team Edition”

  15. (15)

    SpaceX Hit $2.5 Trillion. Now Watch What Elon Buys Next.

    YouTube

    24:44 into “SpaceX Hit $2.5 Trillion. Now Watch What Elon Buys Next.”

  16. (16)

    Chinese AI Model GLM 5.2 Beating Frontier Models | Meta Glasses, Polymarket Scandal, AI Talent War

    YouTube

    26:47 into “Chinese AI Model GLM 5.2 Beating Frontier Models | Meta Glasses, Polymarket Scandal, AI Talent War”

  17. (17)

    Hype vs Reality: Our 2026 Predictions for Tech, Media, and Culture

    YouTube

    26:46 into “Hype vs Reality: Our 2026 Predictions for Tech, Media, and Culture”

  18. (18)

    Meta Muse vs Instinct vs OpenAI Dots: Do AI Agents Have Any Moat?

    YouTube

    20:07 into “Meta Muse vs Instinct vs OpenAI Dots: Do AI Agents Have Any Moat?”

  19. (19)

    Anthropic Wants to Slow Down. Is It Because They're Losing?

    YouTube

    6:19 into “Anthropic Wants to Slow Down. Is It Because They're Losing?”

  20. (20)

    The New Startup Stack: One Founder + Agents | Henrik Werdelin (Audos) and Ben Broca (Polsia)

    YouTube

    42:02 into “The New Startup Stack: One Founder + Agents | Henrik Werdelin (Audos) and Ben Broca (Polsia)”

  21. (21)

    Every CTO Is Quietly Switching Off Anthropic and OpenAI | GLM 5.2, Vox Americanus, Claude Fable 5

    YouTube

    41:19 into “Every CTO Is Quietly Switching Off Anthropic and OpenAI | GLM 5.2, Vox Americanus, Claude Fable 5”

(By email)

Compounding Insights