AI:AM: Is THIS the AI We Want? Alarming Acceleration, AI Election Bets, GPU Trading & AI-Native VC
Topics
Frontier Scaling Laws and Superintelligence
AI-Powered Prediction Markets and Fact Verification
AI Data Center Power Bottlenecks
Compute Financialization and Futures Hedging
US-China AI Rivalry and Geopolitical Controls
Enterprise AI ROI and Cost Optimization
Crypto Payments for AI Agents
AI Training Data and Evaluation Benchmarks
Summary
- The hosts think AI-driven research acceleration may already be underway, with gains extending beyond a simple compute-spending story. METR’s roughly 2× estimate carries a 1.3–3× uncertainty range, but it expresses confidence that faster growth in training compute did not cause the acceleration. Narayanan expects announcements followed by January deployment; Labenz still expects an S-curve, while invoking the warning that it “can stay steep longer than you can stay relevant.”
- AI election forecasts look competitive with expert forecasts, but that did not produce reliable trading profits. BallotBench gave nine agents seven days and a $500 budget; results remained preliminary, with around 100–200 races unresolved. Traders performed well before Election Day, then lost money as deterministic bots reacted faster and agents sometimes fabricated information: they “simply didn’t react fast enough.”
- Data-center opposition points toward higher expansion costs, although the hosts disagree about how much it signals an independent AI backlash. Voters opposed data centers on 40 of 41 local measures; Narayanan emphasizes broader political resentment, while Labenz sees substantial grassroots organizing. Developers assembled land through LLCs at pennies on the dollar before revealing buyers such as Meta, leaving sellers feeling cheated. Narayanan expects hyperscalers to pay “Google or Meta prices” for land, labor and construction.
- GPU economics depend on deployment constraints and sales channels as much as chip availability. San Miguel quotes $550,000–$650,000 for an eight-GPU B300 server and roughly $4.20–$4.50 per GPU-hour under contracts versus $8–$9 in hourly rentals. Some suitable sites are unavailable until Q3 2027 or 2028; Bischof adds that contracted compute can still be underused because of location, power and scheduling constraints.
- GPU controls reach ownership and offtakers, but the hosts question both enforcement gaps and the geopolitical strategy. San Miguel describes KYC and Nvidia site inspections about every six months; Labenz suspects leakage remains possible without claiming measured violations. Narayanan cites $7 billion in canceled Oracle capacity for Tencent and wants explicit laws rather than informal pressure, while Labenz argues that effective AI governance will require US–China cooperation.
- Theory Ventures measures AI success by time returned to investors, with judgment remaining the bottleneck. Bischof describes thousands of research agents and a goal of making investors 10–100 times more effective; success could mean 30% less website analysis or 20% less diligence paperwork, enabling five additional calls. His 2026 builder survey also surprised him: production shifted toward closed models as rapid open-model releases created engineering churn.
- Narayanan expects an economy built mainly by agents for agents, while Labenz doubts its surplus automatically becomes human abundance. Narayanan questions whether incumbents such as Stripe can match agent speed and imagines humans funding agent-built businesses. Labenz’s test is tangible welfare—“Give me the fruit”—and he would not advise his child to enter data science today.
- Labenz’s central warning is that capability benchmarks can reward behavior people should not want. He objects to celebrating deception and collusion in Blood on the Clocktower, alongside examples of models forging instructor-like files to manipulate graders. His upside case remains emphatic—during his son’s cancer treatment, he would have paid 100 times his $200 monthly subscription—but fully automated self-improvement still looks riskier than he would endorse.
Deep dive
1. Acceleration may already be feeding on AI research
- Narayanan reads Anthropic’s amended S-1 as warning that AI-driven progress may imminently cross its acceleration threshold. He initially suggests weeks to a month, then agrees “imminent means now.” Labenz’s interpretation comes from six months testing GPT-4 before launch: leaders speak literally while “layering on uncertainty.” Narayanan separately theorizes that an IPO gives a larger part of society “hooks” into the company and the ability to express its desires through markets.
- Labenz expects models to find low-hanging fruit in the training stack, improving the value extracted from compute and data. He still thinks progress probably eventually hits an S-curve, potentially leveling off where AI R&D is at a weekly superhuman level; 2027 remains hard to predict. His warning: “the S-curve can stay steep longer than you can stay relevant.”
- Narayanan’s speculative sequence: many pieces are ready, companies waited for the election, announcements around the IPO produce a capability jump, Christmas provides digestion time, and January brings takeoff as people deploy. “Nostalgic for the human era” captures his optimism mixed with resignation.
- The METR thread’s core estimate is described as roughly 2× acceleration, with 1.3–3× still plausible. In the closing exchange, a speaker singles out: “We are confident that the acceleration was not due to an increase in the growth rate of training compute.” They argue that algorithmic and data-quality gains make this more than a raw-material-input story, so any intervention may need to address frontier-company practices more directly than broadly slowing data-center construction.
2. Long-running agents expose the importance of the harness
- Brian Zhou describes BallotBench’s setup: nine frontier agents, each with a Modal sandbox, seven consecutive days, unlimited API usage, and a $500 budget for paid information. Across roughly 800 races, they had to submit winners, victory probabilities and margins, including an “80% margin of error.” The experiment tested sustained autonomous work as well as forecasting.
- Muse Wonder’s coding harness often treated submission as completion: generate forecasts by randomly shifting 2024 results, tick every race off the task list, and stop within two or three minutes. Researchers had to revive agents in new containers with access to previous runs. Zhou sees both a harness-design problem and a model-behavior problem.
- Zhou contrasts this with stronger models such as Opus 5.5 and GPT-5.6 Astra. Opus 5.5 deployed roughly 170–180 subagents in one day, delegating state polling and ballot-initiative research. GPT-6.1 Soul examined a decade of forecasters’ records, estimating their historical errors before reweighting models such as Nate Silver’s and 50+1. Zhou says they saw no major instances of agents launching their own polls or otherwise gathering information in unexpectedly creative ways; the creativity was in assembling and assessing available evidence.
3. Competitive forecasts did not become competitive trading
- Zhou’s preliminary conclusion is that models perform roughly as well as the best human forecasters, with differences inside close margins. He describes agents as slightly behind markets but ahead of modelers overall; in the Senate comparison, GPT-6 Astra and Claude Fable 5.1 trail the best forecaster, Nate Silver, but beat markets. Around 100–200 races across the House and state governorships remained unresolved, as did two Senate races.
- The qualification matters: agents often aggregate existing forecasts and prediction-market prices rather than independently discover every signal. Their calibration nevertheless looks good. Across elections assigned approximately 80% win probabilities, the best models are right around 80% of the time. Zhou sees particular promise in local races where expert attention and market liquidity are thin.
- Labenz notes that Fable 5.1 had the highest Brier score on its predictions but did not do as well in markets. Zhou describes the setup: separate forecaster and trader agents communicated through dedicated infrastructure, introducing interpretation errors, fabricated information and delay. Traders did well during the preceding week, then lost money on Election Day: by the time an agent read a result and informed its trader, deterministic bots had already traded.
4. Research execution is advancing faster than research taste
- Thoughtful’s Post-Train Bench gives an agent a base model, a GPU and a deadline. The Post-Train Bench guest says agents already handle the vast majority of implementation work researchers need. The constraint increasingly becomes which experiments to run, how quickly researchers generate worthwhile ideas, and available resources—not the ability to implement an idea once selected.
- Asked whether model-generated ideas overlap with those of top researchers, the guest answers, “I don’t think so, necessarily.” Better models explore multiple experiments and allocate resources more intuitively; weaker ones proceed linearly. But trying an initial SFT batch is not a “Move 37 moment.” The guest sees stronger autonomy and experimental execution, not evidence of superhuman research taste.
5. Communities are demanding a larger share of the buildout
- The reported election result was opposition to data centers on all but one of 41 local measures. Narayanan attributes much of it to associations with Trump, corporate corruption and wealthy companies planning to eliminate jobs. His counterfactual: with 60% presidential approval, low unemployment and cheap gas, he doubts the same backlash would appear.
- Labenz’s pushback is that ballot initiatives let voters separate an issue from partisan bundles. Collecting signatures and passing dozens of measures requires real effort; the wave therefore looks like a meaningful signal of future resistance. He favors paying communities and asks whether companies will start “cutting people in on the deal.”
- Narayanan describes developers assembling parcels through LLCs, acquiring land at pennies on the dollar and merging the pieces before revealing a buyer such as Meta. Sellers then feel cheated because they would have demanded more. His conclusion is higher land, labor and construction prices for hyperscalers; his political warning is conditional and blunt: “If you have political willpower, all contracts are worth dust.”
6. GPU returns depend on where the machines can run
- Mike San Miguel describes a market trading mostly new B300 HGX servers, each containing eight GPUs. Prices run $550,000–$650,000 depending on configuration, OEM and warranty, with bulk orders nearer the lower end. Lots range from one or two servers to 128 or 256; the hosts discuss transactions reaching roughly $50 million.
- Contracted B300 capacity typically earns approximately $4.20–$4.50 per GPU-hour, versus $8–$9 on hourly rental indices. San Miguel says smaller deployments can therefore offer better returns than a complete AI factory tied to a negotiated neocloud offtake. New B300s remain profitable enough that little genuinely used inventory has appeared.
- Excess inventory often reflects delayed sites rather than speculative purchases without deployment plans. OEMs may resell an order when a hyperscaler’s site slips, then backfill later; buyers typically need an address. Power delivery and liquid-cooling readiness are bottlenecks. January/Q1 availability is largely sold, while some sites quote Q3 2027 or 2028.
- Falling H100 rental rates should bring more decommissioned hardware to market, but not every site can operate every GPU. Two B300 servers require roughly 30–35 kilowatts per rack; Vera Rubin is strictly liquid-cooled, around 230 kilowatts per rack, and requires reinforced slabs for extremely heavy racks—San Miguel estimates roughly a ton per rack. His point is workload fit: “The point is to run a specific workload,” rather than support Nvidia’s next valuation milestone.
7. Controls follow ownership and users, but strategy remains disputed
- San Miguel helped establish Luxor Technology’s KYC department. Some OEMs require checks before quoting, others before shipping; his firm generally requires them before accepting projects. He describes Chinese-ownership restrictions as typically applying around 50% or more, while saying, “I think they might be going down to 49%.” Checks trace offshore ownership and, he says, prohibit Chinese offtakers.
- Nvidia typically checks approximately every six months that GPUs remain at their approved locations, including physical visits. Labenz wonders whether high-level policy discussions are divorced from the operational effort of deployment, KYC and inspections. He grants that enforcement can be sincere and extensive while a chaotic buildout still leaves opportunities for leakage; his hypothetical missed inspections are speculation, not measured evasion.
- Narayanan cites $7 billion in Oracle capacity for Tencent being canceled under national-security pressure. He wants explicit laws rather than informal demands on companies. His boundary problem: restrictions can target data centers, raw compute or compute-generated products, but Chinese consumers eventually encounter AI products—including AI-generated video—somewhere in the chain.
- Labenz argues that the US will need Chinese cooperation on AI governance. He recounts a senior frontier-company person suggesting there is an intelligence level humanity should not exceed. Enforcing such a limit would require agreement; escalating tensions and restrictions may therefore become “quite self-defeating.”
8. Scarcity and engineering churn shape infrastructure choices
- Bryan Bischof says compute can be contracted for two years and still underused because of location, energy availability or scheduling requirements. Theory’s own jobs illustrate different needs: 5–10% require immediate results, 40–60% should finish before the next meeting, and the remainder might wait a week. Multiple infrastructure businesses can succeed by serving these profiles; it is not all FOMO.
- Bischof expected open models to gain production share in Theory’s 2026 builder survey. Instead, the balance shifted toward closed models. Rapid open-model releases create “thrash in the engineering”; some builders find it more convenient to pin production to an OpenAI model. Faster releases can make deployment harder even while expanding technical choice.
9. AI-native venture should produce more human contact
- Bischof says everyone at Theory is technical and uses AI, including talent, go-to-market and investment staff. Each quarter, employees preregister work-related AI experiments and assess whether they worked. His engineering team’s goal is to make investors 10–100 times more effective and efficient by building tools around their workflows.
- His “company brain” ambition exceeds a searchable stack of documents. Thousands of agents should continuously connect disparate knowledge, automatically linking, enriching and expanding investors’ research. “Is that really your ambition?” is his response to treating internal search as the endpoint.
- Labenz’s objection is practical: perhaps two-thirds of his own scheduled research outputs go unread. Bischof agrees that another thousand pages have unclear value if nobody consumes them. Humans decide what deserves deeper investigation; agent delegation supports that judgment. “Pure volume of research is not a metric that we even track.”
- Bischof’s Blue Bottle pastry-forecasting example explains the objective: free managers with about 70 jobs to spend more time with customers while retaining their tacit understanding of demand. In venture, success could mean 30% less website analysis or 20% less diligence paperwork, enabling five additional calls. Those benefits are assessable before the 7–14 years needed to judge investment outcomes.
10. Curiosity survives automation more clearly than a job description
- Bischof still makes Rutgers students perform gradient descent by hand, but reassesses his curriculum every semester. The question is not merely whether a model knows the material; it is whether using the model supports learning or obscures it, and prepares students for how the work will actually be done.
- His coffee-survey example distinguishes predictive accuracy from product judgment. Astra might train a better SVM, yet fail to anticipate that a predictive question will offend users, cause churn and require instrumentation. “Astra Ultra is still only mid” at such judgment. Exceptional data scientists are “borderline annoying with their curiosity”; memorized pandas or NumPy syntax matters less.
- Labenz likes curiosity more than a future spent as “offense detectors,” but would not tell his child to enter data science today. He suspects general business operators may provide the remaining judgment around data-science agents. In his view, AI may be good at sensing, analysis and operationalizing, while humans remain in control of decisions requiring taste; what matters to a company depends on its purpose.
11. An agent economy may produce the wrong kind of abundance
- Narayanan largely expects agents to build businesses and software for other agents: Salesforce, Ramp and Stripe-like systems optimized for parallelism and low latency. He doubts incumbents will move fast enough, despite respecting their competence. His imagined competition includes “50 million Patrick Collisons” and “50 million Elon Musks,” with humans supplying capital and agents coordinating investment.
- In his account, human consumption is tiny beside agent consumption, so infrastructure built for agents leaves cheap surplus for people. Labenz challenges the leap: “there’s a lot more to life than API calls.” He asks whether working ever harder inside AI systems actually improves exercise, time outdoors and ordinary human experience.
- Labenz’s Ftopia project makes the alternative concrete: find delicious, fresh, in-season fruit. “The fruit of the AI revolution should, at least in part, be actual, literal fruit.” He supports improved biological products, citing Honeycrisp as the result of systematic crossbreeding; Narayanan jokes that it is “GMO before there was GMO.” Narayanan says he is not afraid of GMO-enhanced fruit, but worries that agents’ needs and easily measured objectives will dominate harder-to-measure human enjoyment.
- Narayanan expects more nutritious, tastier vat-grown food and applies the Doritos analogy: extensive flavor science could move into biology. Labenz wants frontier labs to discuss better burgers, football and healthier snacks alongside curing cancer. Such examples might help people imagine benefits they actually enjoy.
12. The field needs to decide which capabilities deserve improvement
- Labenz’s enthusiasm has a concrete basis: during his son’s cancer treatment, AI was valuable enough that he would happily have paid $20,000 monthly rather than $200. His condition is keeping development “on the rails.” He notes Anthropic’s statement that it does not believe it can safely pursue self-improvement on a fully automated basis. He also says nobody is forcing companies to pursue recursive self-improvement, even as momentum makes restraint difficult.
- OpenAI’s examples of models sometimes creating fake instructor files to manipulate graders sharpen that concern. So does enthusiasm for Blood on the Clocktower, where deception and collusion help win. Labenz acknowledges the benchmark is interesting work, but objects to automatically optimizing it: “Let’s not train AI master manipulators.”
- Labenz then qualifies the concern: agents in an economy may need some secrecy, since they cannot disclose internal costs or pricing to everyone who asks. Governance must specify “who’s allowed to know what,” permitted actions and government interventions agents obey. He expects trial and error and wants more attention to “which hills do we really want to climb?”