Sep 16, 2026 · 2h 20m
Listen from 1:37:15
Listen at 1:37:15
AI agents meet higher rates and harder questions of trust
Third Party Evaluators, Fed Hikes Rates, Zuck Pushes Back On AI Slowdown | Jeremy Allaire, Tomasz Tunguz, William Layden, Justin Beroz, Eli Wachs, Sean McCarthy, Tom Mueller
The episode connects AI governance and personal-agent economics with the financing conditions shaping startups and infrastructure.
- 1Independent AI evaluation remains difficult as companies debate who can establish trusted oversight.
- 2Personal agents may trade convenience for data, while dedicated hardware still struggles to beat phones and glasses.
- 3Higher rates sharpen the test for AI business models, from supply-chain automation to orbital logistics.
Don't miss
The discussion of Persona’s proactive assistant tests whether an AI device can email contacts, book rides, cancel subscriptions, and order food without overcoming phone-level friction.
The brief
The hosts open with a governance problem: frontier models need credible oversight, but independent evaluators and trusted regulators remain difficult to define.
A 25-basis-point Federal Reserve hike shifts the conversation from AI ambition to startup economics, exposing which business models depended on near-zero rates.
Mark Zuckerberg’s safety argument becomes a test of whether product launches, independent evaluation, and alignment address existential-risk concerns—or talk past them.
Muse, Gemini in Chrome, and Persona frame the personal-agent race around proactive action, distribution, hardware friction, and the value of interaction data.
The episode widens from agents to physical-world execution: BackOps automates supply-chain disputes, while Impulse Space builds orbital logistics for government and lunar infrastructure.
Its central tension is economic as much as technical: AI can automate increasingly consequential work, but financing, trust, and data ownership determine who benefits.
What was said on this episode
62 statements · 39 positive · 9 negative · 1 mixed · 13 neutral
Few groups have long-term expertise understanding frontier model behavior.
“there are not that many groups that have been this invested in, in understanding frontier model behavior for this long. It's just a small group right now.”
Listen at 2:52
Superforecasting and future prediction are important for AI oversight.
“I think taking that seriously is important.”
Listen at 3:11
Frontier forecasting and operational model evaluation are separate disciplines.
“I think that those are actually two separate disciplines, two separate jobs.”
Listen at 3:21
Regulators can be trained to take AI risks seriously without forecasting expertise.
“you can just be like, it is your mandate to take this seriously.”
Listen at 3:47
Meter’s financing and history undermine trust in its independence.
“the financing structure, the history of the various parties there just makes it so that people don't have any trust that they would act like that.”
Listen at 6:20
- Eli Wachson AI model safety evaluationNegative10:42
AI model safety evaluation methods remain unsettled.
“the ways that we detect if a model is safe or not are not, like, set in stone yet.”
Listen at 10:42
- Eli Wachson AI model safety evaluationNeutral10:42
Methods for determining model safety remain unsettled.
“the ways that we detect if a model is safe or not are not, like, set in stone yet.”
Listen at 10:42
Trust and alignment are becoming key differentiators among AI models.
“trust and alignment are quickly becoming the most important capabilities that will differentiate.”
Listen at 20:36
Meta must delay products to ensure billions of users remain safe and secure.
“Meta has been doing forever. You have to do it. You have billions of users. You got to make sure they're safe and secure.”
Listen at 21:58
Some supporters of Meta’s position assign zero probability to AI doom.
“other people that have a PDM of zero”
Listen at 23:29
Tomasz believes Zuckerberg assigns zero probability to AI doom and high abundance probability.
“He has a P. Doom of zero, and his P. Abundance is high.”
Listen at 24:39
Zuckerberg will not stop Meta researchers from pursuing recursive self-improvement.
“there's absolutely zero shot that Zuck walks into MSL and gathers the researchers and says, look, I don't want to make models that make our models better.”
Listen at 25:19
Advancing coding capability is strongly aligned with recursive self-improvement.
“the race to become really, really good at coding is super aligned with RSI.”
Listen at 26:01
Image generation does not appear critical to recursive self-improvement.
“image generation does not seem to be on the RSI critical path.”
Listen at 26:12
Embedding non-employee third-party evaluators is a major governance change for AI labs.
“the prospect of bringing in a non-employee who is there explicitly to whistleblow effectively and like is allowed to talk about anything and like is, is, you know, third party evaluator, like that is a huge step.”
Listen at 28:38
Anthropic is giving Meter Slack access for independent evaluation.
“Meter's going to have Slack access at Anthropic”
Listen at 29:08
Meta’s Muse is receiving highly favorable product reviews.
“Muse seems like it's getting amazing reviews. It seems like an amazing product.”
Listen at 31:15
Users genuinely like Muse, which appears to be a strong product.
“People genuinely love Muse. I think it seems to be. a great product”
Listen at 34:08
Meta’s platforms provide a significant distribution advantage for personal agents.
“the distribution advantage from meta platforms is significant”
Listen at 35:29
Google’s Gemini agent remains poorly understood publicly.
“It's the world's most mysterious agent yet.”
Listen at 36:38
Wristbands may achieve more continuous wear than smart glasses.
“people don't wear the glasses all the time. Whereas a wristband, like a whoop, people wear that all the time.”
Listen at 46:40
Cryptographic trust and verification technology is necessary for agentic economies.
“We can't move forward in the agentic economy without technology like this.”
Listen at 52:44
Stablecoins enable near-instant global transfers at sub-cent cost.
“you can move them programmatically instantly anywhere in the world in a fraction of a second for a fraction of a cent.”
Listen at 53:43
Stablecoins will commoditize payments and settlement.
“payments and settlement becomes commoditized.”
Listen at 53:51
Stablecoin rails will reduce longstanding payment intermediary rent extraction.
“that compresses out this rent extraction that's existed for a very, very long time”
Listen at 54:19
Tokenized stocks can give emerging-market users instant trading and options access.
“I can now do that and I can purchase it and trade it instantly. I can take out options on it instantly.”
Listen at 55:57
USDC and similar stablecoins become legally recognized digital dollars in January 2027.
“as of January 2027, stablecoins like USDC become legal digital dollars in the U.S. financial system and in the global financial system.”
Listen at 57:27
U.S. financial regulators will advance crypto regulation without waiting for Congress.
“the agencies that regulate these things like the CFTC, the SEC, even the bank regulators, too, they're all just going to move forward and regulate it themselves.”
Listen at 58:11
- William Laydenon Company namesNeutral1:01:58
Company names often predict whether companies will succeed or fail.
“the names of companies are often very predictive of future failure or success.”
Listen at 1:01:58
- William Laydenon Company namesNeutral1:01:58
Company names often predict whether companies will succeed or fail.
“the names of companies are often very predictive of future failure or success.”
Listen at 1:01:58
- William Laydenon Facebook versus MySpacePositive1:03:46
Facebook ultimately dominated MySpace because of product-name implications.
“Facebook would ultimately dominate MySpace.”
Listen at 1:03:46
- Justin Berozon AI control systemsPositive1:08:18
Zuckerberg’s position supports control systems around AI models.
“I think he's arguing for control systems in a way.”
Listen at 1:08:18
- Justin Berozon AI agent liabilityNeutral1:09:49
Users will ultimately bear responsibility for agents they initiate.
“you will ultimately be responsible for the agent that you start.”
Listen at 1:09:49
- Justin Berozon Safety-critical transportation systemsNeutral1:10:17
Society accepts nonzero failure probabilities in cars and airplanes.
“The reality is like doom exists in a car. Doom exists in an airplane. And so we accept some probability of failure in those.”
Listen at 1:10:17
- Justin Berozon U.S. data centersPositive1:11:49
Data centers contribute two-thirds of U.S. GDP growth.
“data centers contribute two thirds to US GDP growth.”
Listen at 1:11:49
- Justin Berozon AI inference softwarePositive1:12:58
Inference is currently software’s largest market.
“the single biggest market in software today is inference.”
Listen at 1:12:58
- Justin Berozon Open-source AIPositive1:16:16
Open-source AI will be essential to the future AI ecosystem.
“We think open source is an absolutely essential component of the ecosystem going forward.”
Listen at 1:16:16
- Justin Berozon AI harness companiesPositive1:18:17
Fast-growing AI harness companies trade at 100–150 times current ARR.
“the AI harnesses are now the fastest growing ones are trading between 100 to 150 times current ARR”
Listen at 1:18:17
- Justin Berozon Enterprise AI agentsPositive1:18:43
Employers began paying AI agents at human market rates or higher in 2026.
“2026 would be the first year where employers would pay agents at market level for a person or more. And that happened”
Listen at 1:18:43
- Justin Berozon Legacy software companiesPositive1:20:57
Legacy software acquisitions offer a multiple-arbitrage opportunity.
“there's a multiple arbitrage opportunity.”
Listen at 1:20:57
- Justin Berozon Venture capital liquidityNeutral1:22:32
Managing liquidity ranks among LPs’ top two or three concerns.
“In every ELP conversation that we have, managing liquidity is among the top two or three topics.”
Listen at 1:22:32
- Justin Berozon Personal-agent companiesNeutral1:24:03
Personal-agent companies initially prioritize acquiring interaction data.
“The first phase is really about data acquisition.”
Listen at 1:24:03
- Justin Berozon Personal-agent companiesPositive1:24:22
Personal-agent companies may subsidize products to acquire training data.
“they need the data acquisition. It makes sense to subsidize it for a while.”
Listen at 1:24:22
- Justin Berozon Personal AI agentsNeutral1:24:32
Personal agents will prioritize distribution, cost scaling, then monetization.
“Priority one is distribution. Priority two is scaling the cost. And then third is monetization.”
Listen at 1:24:32
- Justin Berozon AI agent targeting dataPositive1:25:07
New advertising targeting criteria can support companies worth hundreds of billions.
“Anytime you have a new targeting criterion, you can have a multi-hundred billion dollar company.”
Listen at 1:25:07
- Justin Berozon Google agentic systemsPositive1:25:16
Agentic systems could double or triple Google’s U.S. average revenue per user.
“The average user on Google in the U.S. generates about $120 in ARPU. it's very easy to see a doubling or tripling of that with these agentic systems.”
Listen at 1:25:16
- Justin Berozon Agentic advertising systemsPositive1:25:20
Agentic systems could double or triple Google’s U.S. user ARPU.
“it's very easy to see a doubling or tripling of that with these agentic systems.”
Listen at 1:25:20
Excess generation is an inherent feature of renewable energy systems.
“that abundance or excess is actually a feature of renewable energy and not a bug.”
Listen at 1:27:23
- Justin Berozon Rune solar data centersPositive1:28:54
Rune expects to deploy 100–200 MW of data-center capacity at a 400-MW solar facility.
“we're working with like a 400 megawatt solar facility. We think we can probably put, you know, 100, 200 megawatts of solar there or of data center capacity there.”
Listen at 1:28:54
- Justin Berozon Rune RelicPositive1:29:37
Rune can deploy a modular data-center unit in about 60 minutes.
“We can deploy, you know, a unit in about 60 minutes.”
Listen at 1:29:37
- Justin Berozon Rune solar conversion technologyPositive1:30:45
Rune can convert every solar power plant into a data center.
“every solar power plant is a latent data center and we can convert it with our technology.”
Listen at 1:30:45
- Justin Berozon Solar energy deploymentPositive1:33:57
The solar market is far from saturation in deployed capacity.
“We're nowhere near saturating the solar market in terms of how much we've deployed.”
Listen at 1:33:57
Navier–Stokes theory lacks proof that infinite velocity cannot occur.
“we don't have any firm mathematical proof whether that can or cannot happen.”
Listen at 1:37:15
- Justin Berozon OpenAI Navier–Stokes proofNeutral1:37:32
OpenAI submitted a Navier–Stokes proof that may take years to validate.
“they've submitted a proof, but it may take years for the actual experts to review this thing and figure it out.”
Listen at 1:37:32
- Justin Berozon Turbulence modelingNegative1:39:20
A complete mathematical framework for turbulence remains unresolved.
“Having a complete mathematical framework to describe that, that is also an open question.”
Listen at 1:39:20
- Justin Berozon Fluid-flow simulationsNegative1:40:20
Current fluid-flow simulations are not reliably predictive.
“the simulations are really not predictive.”
Listen at 1:40:20
- Justin Berozon Turbulence software companyPositive1:43:06
The turbulence company plans software first and a commercial product in several years.
“the company's going to develop in three stages. So we're starting first with better software. We're a couple years away from releasing our first commercial product.”
Listen at 1:43:06
- Justin Berozon Footprint financial-crime platformPositive1:50:12
Footprint enables financial-crime systems to retain every investigated case.
“we give them the ability to remember every case that they've seen.”
Listen at 1:50:12
- Justin Berozon Footprint financial investigation agentPositive1:51:13
Footprint’s agent traced a flagged Russian individual through Ukrainian and U.S. records.
“our agent went back into Ukrainian newspapers. read the initial Cyrillic because Cyrillic can be translated improperly, then tracked down this person's wife to a home in New Jersey and found a real estate listing in their name.”
Listen at 1:51:13
- Justin Berozon Footprint fraud-model evaluationsPositive1:54:08
Footprint evaluates models using synthetic cases derived from real cases.
“we have a huge library of synthetic cases that are based off real cases”
Listen at 1:54:08
- Justin Berozon AML transaction alertsPositive1:55:56
Footprint aims for 99.999% false positives among flagged AML transactions.
“we actually want that to be 99.999% because we're only looking at what's flagged as the highest risk transactions.”
Listen at 1:55:56
- Justin Berozon FootprintPositive1:56:20
Footprint’s mission is to undermine cartels and corrupt regimes.
“our mission is to topple cartels and corrupt regimes.”
Listen at 1:56:20
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.