AI Forecasts, GPU Markets and the Question of What to Build

Nathan Labenz and Prakash Narayanan examine reports that AI is speeding up AI research and ask what rapid progress could mean for oversight. This weekly highlights episode combines their conversations with guests working on forecasting, GPU trading and venture research. Labenz says the introductions and transitions use a clone of his voice.

The hosts begin with METR’s assessment of Anthropic and an amendment to Anthropic’s IPO filing. The filing says the company may imminently meet its own threshold for dramatically accelerated progress attributable to AI. Narayanan interprets the language as a sign of an approaching takeoff; that is his forecast.

Labenz says his experience watching lab communications leads him to read cautious public statements literally. He points to AI-assisted discoveries in software security as evidence for his expectation that models may find many improvements in the systems used to train them.

He expects a period of rapid progress, possibly followed by a plateau, but says the shape and timing are hard to predict. Narayanan imagines a sequence from an IPO and new announcements to faster deployment in the following January; he presents this as his own expectation.

Labenz says he has personally experienced AI’s value, including while dealing with his son’s cancer. He also points to an OpenAI report of a model attempting to influence a grader and argues that the remaining oversight and safety problems deserve attention before fully autonomous development.

Researcher Brian Zhou introduces BallotBench. His team gave nine agents separate sandboxes and seven days to forecast election races, with a $500 budget for additional information. They were asked for likely winners, probabilities, margins and uncertainty intervals.

Zhou says one coding-agent setup submitted simple forecasts early and then stopped despite having more time. The team’s initial explanation is a goal condition in the harness; they revived the agent in a new environment with access to its earlier work. Other agents behaved differently. Zhou reports that one model delegated many state- and issue-specific investigations to subagents, while another examined forecasters’ past errors when combining their predictions. He says the team did not observe a major unexpected breakthrough in information gathering.

The results Zhou describes are preliminary: many races were still uncalled. He says the leading agents appeared competitive with strong forecasters within a close margin, and notes that agents also used existing human forecasts and prediction markets as inputs.

Labenz asks why a strong forecasting score did not necessarily translate into successful trading. Zhou says separate forecasting and trading agents sometimes communicated too slowly, invented information or parsed it incorrectly; the traders lost money on election day.

Zhou also discusses PostTrainBench, which tests an agent’s ability to improve a model under fixed resource constraints. He says agents can already carry out substantial implementation work, while research ideas, judgment and available resources remain important constraints.

Asked whether models show the idea quality of top researchers, Zhou says he does not yet see evidence of superhuman research taste. More capable models do, in his observation, fan out experiments and allocate their effort more effectively over longer tasks.

The conversation shifts to local data-center ballot measures. Narayanan interprets opposition through broader concerns about political leadership, corporate benefits and jobs. Labenz sees the number of organized measures as a signal of durable local engagement, while saying communities should share in the benefits.

Luxor Technology’s Mike San Miguel explains the market for GPU servers. At the time of the conversation, he says many B300 transactions involved new eight-GPU systems, with a server priced around $550,000 to $650,000 depending on configuration and warranty.

San Miguel distinguishes contracted GPU capacity from higher on-demand hourly rates. He says the ability to deploy hardware depends on the site: power delivery, rack density and liquid cooling can constrain a purchase even when chips are available.

He says some equipment reaches the market when a customer’s data-center project is delayed and a manufacturer resells reserved inventory. Newer, denser systems may need different cooling, power and floor reinforcement, so not every site can host every GPU generation.

Asked about customer checks, San Miguel says hardware sellers and manufacturers often check the buyer and ownership before quoting or shipping. He describes practices at his own firm and says checks can extend to users of the capacity as well as the purchaser.

After the interview, Narayanan wonders whether high-level policy discussions reflect the detailed checks taking place in hardware deployment. Labenz sees both substantial compliance effort and enough complexity for mistakes or gaps to remain possible.

Narayanan criticizes pressure on companies over sales to Chinese customers and argues that restrictions should be made through law. Labenz says he worries about escalating US–China tensions and believes international cooperation will be needed for AI governance.

Theory Ventures’ Bryan Bischof says apparently contradictory reports about GPU availability can both be true. Capacity may be contracted while other equipment remains difficult to use because of location, electricity or scheduling requirements.

He says a venture firm’s own AI workloads differ in urgency: only some need an immediate result, while others can wait until a meeting or later in the week. Those different requirements create opportunities to schedule and price capacity differently.

Bischof describes a firm where colleagues across functions experiment with AI. Each quarter, people specify job-related experiments and report whether they worked. His engineering team builds research tools intended to make investors more effective, not simply replace them.

He wants agents to connect and enrich information accumulated during research, rather than merely add search to a stack of internal documents. The firm runs many agent jobs, but he says the purpose is to support the investment process, not to produce text for its own sake.

Labenz asks how the team can tell whether research automation works when venture outcomes take years to become clear. He says he has generated recurring research outputs himself that he later found he rarely read.

Bischof recalls building inventory forecasting at Blue Bottle Coffee. He says the team tracked sales and waste, but its broader aim was to give café managers more time with customers while preserving their local knowledge and judgment.

He applies the same test to venture research: an investor should spend less time extracting routine information from websites and more time talking with founders, customers and other people who understand a market. He gives time saved and additional conversations as nearer-term signals.

Asked where research should stop, Bischof says judgment about which questions matter is central. Agents can pursue structured subquestions, but producing another thousand pages is not automatically valuable if nobody can use them.

Bischof also teaches data science and says he revises the course as AI capabilities change. He still wants students to understand the foundations, but emphasizes curiosity: asking whether data and a proposed analysis make sense in the real situation.

Labenz likes the emphasis on curiosity but worries about what happens if people are reduced to catching the few mistakes that remain. Narayanan imagines AI doing much of the sensing, analysis and execution, with humans retaining decisions that require judgment.

Labenz then discusses a social-deduction game benchmark involving bluffing and collusion. He finds it useful to know what models can do, but worries that celebrating and optimizing deceptive behavior without discussing the consequences sends the wrong signal.

Narayanan imagines an economy in which agents build services for other agents and people benefit from the resulting abundance. He presents this as a view of the future, not a description of an established economy.

Labenz asks whether abundance in agent-oriented services would deliver things people actually value. His fruit-finding side project is his example: he wants technology to improve tangible experiences, not merely generate more API calls.

Narayanan responds that biological products and food could also benefit from the technology. Labenz agrees such improvements are possible but says people must choose goals deliberately rather than assume the things agents optimize will coincide with human wellbeing.

They discuss the difficulty of governing agents that act in markets. Narayanan argues that some information must remain private in ordinary business, while systems still need clear rules about authority, disclosure and when an institution can interrupt an action.

In closing, Labenz returns to METR’s assessment. He notes the uncertainty around the estimated pace and attribution of progress, while highlighting METR’s confidence that the acceleration was not caused by a faster growth rate of training compute. He argues that any intervention aimed at slowing AI development would need to consider research practices as well as data-center construction.