Dario vs Jensen on Open Weights, OpenAI & Anthropic in DC, Xi Exports AI to Global South | EP #275
The crew dissects the open-vs-closed AI firestorm sparked by Jensen Huang's Open Secure AI Alliance letter and Dario Amodei's biosecurity-focused rebuttal, then turns to reports that OpenAI and Anthropic are jointly lobbying Washington for a federal review regime, which the panel largely reads as regulatory capture. They cover the global release of the open-weight Kimi K3 model (notable for dropping position embeddings entirely) and Anthropic's Claude Opus 5 launch, including suspicions of benchmark gaming on Arc-AGI-3. Geopolitical themes include China's AI-diplomacy push into the Global South and a debate over Elon Musk's prediction that money will stop mattering by 2036. The episode closes with Starship 13's successful splashdown, two brain-computer-interface breakthroughs (Science Corp's Prima vision implant and Neuralink's thought-controlled wheelchair), and an audience AMA covering longevity escape velocity, synthetic data, and startup investing.
Jensen Huang's Open Secure AI Alliance and the open-vs-closed AI fightAI▶ 0:00
Jensen Huang's first-ever tweet, signed by 77 companies, argues open models strengthen safety, cybersecurity, and sovereignty, citing the Hugging Face breach where closed models failed to help forensics and an open-weight model (GLM-2.5) succeeded. He launches the Open Secure AI Alliance; the panel compares it to the 1998 Open Source Initiative and Nvidia's incentive to commoditize the model layer.
Dario Amodei's rebuttal: biosecurity over open-vs-closedAI▶ 5:38
After three days of silence, Dario reframes the debate as being about whether authoritarian states (China) reach the AI frontier, not open vs. closed per se, citing bioweapon risk from his biophysics background. He proposes blocking chips to China, cracking down on distillation, and mandatory safety testing for all powerful models. The panel is split on whether the biosecurity argument holds up.
Nvidia vs. frontier labs: commoditizing the model layerCompute▶ 8:52
Alex frames the conflict as the infrastructure layer (Nvidia/GPUs) moving from detente to open war with the model layer (OpenAI/Anthropic), applying the aggregator principle of 'commoditize your complement.' The group debates why Nvidia isn't also aggressively financing chip-fabrication competitors to TSMC and Samsung, and whether Elon/Tesla is secretly doing so.
OpenAI and Anthropic's Washington lobbying and regulatory captureGeopolitics▶ 21:07
The Information reports OpenAI and Anthropic are cooperating on lobbying ahead of an August 1st Trump administration deadline, pushing a voluntary 30-day federal review for powerful models that would also bind Meta, xAI, and startups. Alex calls it 'too clever by half' since open- and closed-weight deployment risk differs; Salim frames it as classic regulatory capture akin to railroads, banks, and telcos, but doubts it will succeed given the pace of open models.
Four-layer AI business model and where profit accruesEconomy▶ 24:39
Peter proposes a four-layer stack for how frontier labs monetize despite free open-weight competition: unreleased top-tier models used internally for breakthroughs, Pareto-frontier paid models, fully commoditized open-source models, and the application layer (WhatsApp, Gemini, ChatGPT) where usage scale generates revenue. Alex counters that power laws mean one business model will likely capture the vast majority of profit rather than an even split.
Kimi K3 open-weight release and its NoPE architectureAI▶ 34:33
Kimi K3 launched globally on Hugging Face with no gatekeeper, racking up 2,500 downloads in 2 hours and roughly 100,000 in 24 hours. Alex highlights that it eliminates global position embeddings entirely ('NoPE'), instead smuggling positional/fading-memory information through its Kimi Delta Attention (KDA) mechanism, and later layers can drop attention altogether while preserving output quality.
Claude Opus 5 launch and benchmark scrutinyAI▶ 45:16
Anthropic ships Claude Opus 5 at $5/$25 per million input/output tokens (unchanged from Opus 4.8), claiming it approaches Fable 5's intelligence at half the price. Alex is only 'somewhat excited,' noting the biggest benchmark jumps cluster in code-gen and vision-adjacent tasks (Arc-AGI-3 jumped from 1.5% to 30.2%), while Frontier Math actually declined relative to Fable 5, raising suspicion of targeted distillation or mild bench-maxing.
Claude shared-chat privacy leak via Google indexingAI▶ 59:37
A Reddit user discovered that searching 'site:claude.ai/share' on Google surfaced publicly indexed Claude conversations containing personal health records, private documents, and phone numbers, originating from Claude's link-sharing feature. Alex sees both a privacy failure and an inherent tension in a feature designed to be social; the group ties it to the broader argument for on-premise AI to avoid competitors seeing proprietary data.
Xi Jinping's AI diplomacy in the Global South (Pax Silicon)Geopolitics▶ 1:02:45
The Financial Times reports China is exporting AI models and infrastructure to the developing world as a tool of statecraft ('Pax Silicon'), while Washington debates open vs. closed policy internally. Salim warns that restricting US open models would be a strategic 'self-own,' ceding the Global South's AI ecosystem to China; Alex compares it to the F-16 export dilemma and warns foreign-controlled AI is more invasive than telecom infrastructure since it 'thinks for you.'
Starship 13 launch, splashdown, and falling launch costsSpace▶ 1:08:37
SpaceX's Starship 13 deployed 20 operational Starlink V3 satellites, performed an in-orbit Raptor engine relight for Artemis prep, and achieved a soft splashdown in the Indian Ocean with the vehicle intact. The panel reviews the collapsing cost-per-kilogram of launch (Space Shuttle ~$54,000/kg, Falcon 9 ~$2,000/kg, Starship targeting $10-100/kg) and Elon's tweet that flight 14 may attempt a Mechazilla tower catch.
Brain-computer interfaces: Science Corp's Prima and Neuralink's thought-controlled wheelchairBiotech▶ 1:16:24
Science Corporation's Prima implant, which restores central vision in age-related macular degeneration via a retinal chip and camera glasses, received a CE mark for launch in Europe, with patients gaining five lines of improvement on eye charts after 12 months. Separately, Neuralink demoed paralyzed patients steering a powered wheelchair purely by thought. Alex extrapolates toward full-body exoskeleton control and eventual motor-cortex-based partial brain uploading within five to ten years.
Elon Musk's post-capitalist prediction: money won't matter by 2036Economy▶ 1:33:41
In an Economist interview clip, Elon argues that once robots produce more goods and services than humans can consume, money becomes irrelevant, and predicts deflation rather than inflation. Salim counters money still functions as a medium of exchange for scarce goods; Alex frames it as 'Star Trek economics' where daily necessities demonetize but scarce/positional goods resist it; Dave ties it to Jeff Booth's observation that GDP growth has required four times as much debt growth, an unsustainable structural mismatch with fiat currency.
“we might hit we're going to hit AGI in 2 to 3 years”
Your call:
openElon Musk (clip): Money will stop mattering as robots produce more goods and services than any human could consume, and deflation (not inflation) will be the dominant economic issue
openAlex Wissner-Gross: Most aspects of daily living (food, shelter, healthcare, utilities, education, entertainment) will be effectively demonetized within about 10 years, but full economy-wide demonetization, including scarce/positional goods, will take roughly 30 years
“If I had to put my finger to the wind and guess when demonetization hits the total economy... I would guess approximately 30 years out.”
Your call:
openAlex Wissner-Gross: Full exoskeleton/full-body BCI control and some form of partial brain uploading (via decoded motor cortex) will be demonstrated within the next five years, certainly by the end of this decade
openAlex Wissner-Gross: Given how precise Starship 13's splashdown was, SpaceX's flight 14 is likely to attempt catching the booster with the Mechazilla tower arms
“My guess is Grok jumps to the top of this leaderboard in the next next release.”
Your call:
openDave Blundin: Companies that fail to optimize their AI usage will see token costs exceed payroll within about a year, and reach roughly 10 times payroll within two years
“my token costs are actually going to be bigger than my payroll by the end of the year. And then if I forecast out 2 years from now, my token costs are 10 times my payroll.”
Your call:
openPeter Diamandis: World population will reach approximately 9.5 to 10 billion and then decline rapidly, so overpopulation and resource-scarcity fears are misplaced
“I think the numbers are going to reach 9 and a half, 10 billion, and then very rapidly decrease.”
Your call:
Numbers that matter
77 companiesSigned Jensen Huang's open letter on the value of open AI models.
17,000 actionsNumber of actions logged by the intrusive agent that breached Hugging Face's clusters, escalating privileges and harvesting credentials.
13% drop (~$230 billion)Reported secondary-market decline in Anthropic's valuation after Kimi K3's announcement.
99.95% cheaperApproximate drop in the cost of intelligence/AI inference over the last 3.5 years, per Peter's numbers.
2,500 downloads in first 2 hoursKimi K3's initial download count on Hugging Face after release.
~100,000 downloads in 24 hoursKimi K3's download count on Hugging Face one day after release.
~$55/hourApproximate cost to run Kimi K3 at full throttle on dedicated Modal GPUs.
$5 / $25 per million tokensClaude Opus 5 pricing (input/output per million tokens), unchanged from Opus 4.8.
1.5% to 30.2%Arc-AGI-3 score jump from Claude Opus 4.8 to Opus 5, the largest benchmark gain highlighted, later questioned as possible bench-maxing.
63.9% to 64.7%Humanity's Last Exam (with tools) score from Fable 5 to Opus 5, a comparatively modest gain versus the Arc-AGI-3 jump.
10-week to 8-week doubling timeCompression of the AI price-performance doubling period this year.
every 6 daysAverage rate of frontier model releases across labs over the last few years.
$54,000/kg vs ~$2,000/kg vs $10-100/kgCost per kilogram to orbit: Space Shuttle vs. Falcon 9 vs. Starship's target.
17,500 miles per hourOrbital reentry speed Starship must dissipate energy from during descent.
five lines of improvementVision improvement on a standard eye chart for macular degeneration patients 12 months after Science Corp's Prima implant.
Worth digging into
🕳️ Why isn't Nvidia vertically integrating into chip fabrication?
Alex raises an unanswered question: if Nvidia wants to commoditize the layers above it, why isn't it aggressively financing TSMC/Samsung competitors, leaving that move to Elon's rumored 'Terafab' instead.
🕳️ Kimi K3's NoPE + Kimi Delta Attention architecture
Described as possibly the most frontier publicly-inspectable transformer architecture available, dropping global position embeddings entirely in favor of a fading-memory attention mechanism.
🕳️ Did Anthropic bench-max Claude Opus 5 on Arc-AGI-3?
The headline 1.5%-to-30.2% jump on Arc-AGI-3 wasn't replicated by a third-party evaluation on a similar benchmark, and Frontier Math scores actually declined versus Fable 5, suggesting selective optimization.
🕳️ Max Hodak's undisclosed neuron-growth BCI approach at Science Corporation
Described as an alternative to Neuralink-style rigid electrode arrays, growing neurons directly into brain tissue to interface with electrical circuits, currently in animal models only.
🕳️ China's 'Pax Silicon' AI-diplomacy strategy in the Global South
The Financial Times reports China is exporting AI models and infrastructure as leverage in international diplomacy while the US debates open-model policy internally, with the panel warning of long-term strategic consequences.
🕳️ Jeff Booth's debt/deflation thesis as a structural obstacle to abundance
The claim that every $1 of GDP growth has required $4 of debt growth over 50 years suggests fiat currency systems may be fundamentally incompatible with an AI-driven abundance economy.