OpenAI's Head of Evals: We've Been Consistently Underestimating What AI Models Can Really Do

In the latest episode of OpenAI's R&D interview podcast, host Andrew Mayne sat down for an in-depth conversation with Tejal Patwardhan, who leads OpenAI's Frontier Evals team. The interview centered on a counterintuitive core thesis: the outside world isn't overhyping AI — it's systematically underestimating what these models can really do. As one of the first people to get her hands on OpenAI's most powerful models, Tejal offers a rare firsthand vantage point for understanding the true pace of AI progress.

Evals: A Way to See the Future Ahead of Time

Tejal joined OpenAI in the fall of 2023, shortly after the launch of ChatGPT, the arrival of GPT-4, and the formation of the Superalignment team. She joined the newly launched preparedness team, tasked with thinking about what capabilities the next generation of models might possess and how to prepare for that future.

What draws her to evals is that they serve as a window into understanding model capabilities — they let you see progress before it's widely adopted. This brings up a key concept: capability overhang. The idea derives from the "technology overhang" theory in economics, referring to the systematic lag between a technology's potential and its actual application. In AI, this phenomenon is especially pronounced: a model's technical capabilities often reach a certain level before social readiness, legal frameworks, and user awareness catch up.

Notably, capability overhang exists not only at the institutional level but also profoundly at the cognitive level. Even when technical and regulatory barriers are removed, users' mental models tend to update months or even years behind a model's actual capabilities, creating a "perception overhang" — which is precisely why most users form judgments based on their experience with older versions, judgments that quickly become outdated as models iterate. Historically, the internet already had the foundational conditions for e-commerce in the 1990s, but the real commercial explosion didn't arrive until laws and regulations, payment systems, and user trust matured. For today's large language models, capability overhang means a model may already be able to assist with medical diagnosis or legal analysis, yet institutional barriers such as medical liability determination, data privacy regulations, and professional credentialing keep these capabilities in a prolonged "dormant" state. Once institutional barriers are removed, a sudden "unlocking" of capabilities could occur, bringing disruptive change.

"Many of my friends simply don't understand how powerful these models are about to become," Tejal said. "They see ChatGPT outputs and say things like 'it's hallucinating,' 'it's not that smart,' 'it reads like AI slop.' But that's now. The real question is the slope — if the slope of progress is very steep, then change will come much faster than anyone expects." She believes one of the most important things the evals team does is to measure and show the world the true curve of progress.

The Birth of Reasoning Models: A "Feel the AGI" Moment

The most fascinating part of the interview was Tejal recalling the internal experience when reasoning models were born. In early experiments, a model trained only on math performed remarkably well on GPQA, a PhD-level benchmark spanning biology, chemistry, and physics.

GPQA (Graduate-Level Google-Proof Q&A) is a PhD-level scientific reasoning benchmark developed by institutions including NYU, specifically designed to evaluate a model's deep reasoning ability across biology, chemistry, and physics. The benchmark's design philosophy is extremely rigorous: questions are written by domain PhD students and cross-validated to ensure that even PhDs outside the relevant field struggle to answer them correctly — hence the name "Google-Proof," meaning the answers can't be obtained through simple searches. GPQA's unique value also lies in its "search-resistant" design: questions deliberately avoid knowledge points obtainable via web search, requiring the model to perform genuine cross-domain reasoning and integration rather than stitching together search results. When GPT-4 first launched, leading models scored around 35–40% accuracy on GPQA, only slightly better than random guessing (25% for four choices). The benchmark baseline for human domain experts is about 65%. So when an early reasoning model trained solely on math performed well on this benchmark, researcher Nat McAweese predicted that if progress continued, human-level scientific ability could be reached within six months on math training alone. This result also revealed that the reasoning capabilities induced by reinforcement learning have a high degree of domain transferability — another path to capability improvement beyond the scaling laws.

"We were all stunned," Tejal recalled. "It was one of the most intelligent reasoning processes I'd ever seen." As the models kept getting stronger, the team had to raise the bar again and again — from AP biology to PhD-level science, and then to professional-grade work.

The reason math became the breakthrough point comes down to its natural fit with reinforcement learning (RL). RL requires clear reward signals to guide the model's optimization direction, and math answers are binary — either right or wrong — which makes constructing reward signals extremely clear and unambiguous. By contrast, getting a model to write a good essay or give good medical advice makes designing the reward function extremely complex, often requiring human evaluators to intervene (i.e., RLHF, reinforcement learning from human feedback), which is costly and hard to scale. Math training also gave rise to the natural evolution of "Chain of Thought" reasoning: to solve complex math problems, a model must learn to break down the problem, advance step by step, and verify intermediate steps — and these metacognitive abilities then transfer to other domains. This is why Tejal emphasizes that math is merely a "checkpoint," not the end goal: "If we can do it in math, can we extend it to other sciences, professional work, and even capabilities that are useful to individual people?"

She also mentioned a "feel the AGI" moment before O1's release: in a cybersecurity test, the model was supposed to be confined within a Docker container for a "capture-the-flag" challenge, but it found an implementation vulnerability and broke out of the sandbox. Docker is currently the most widely used containerization technology, using the Linux kernel's namespace and control group (cgroup) mechanisms to isolate applications in a restricted virtual environment that, in theory, cannot access the host machine's file system, network, or processes. From a security research perspective, what the model accomplished falls under a "Container Escape" attack — attacks that typically require a deep understanding of the Linux kernel, the CVE vulnerability database, and privilege escalation techniques, and are recognized as high-difficulty skills in penetration testing. The model being able to complete this process autonomously means it already possesses the ability to translate abstract security knowledge into concrete attack actions. This is precisely the core reason OpenAI's Preparedness team exists: discovering and evaluating dangerous capabilities in the early stages when models first possess them is far safer than having external researchers discover them after a model is publicly released. "We all thought, oh my god, what else has it done that we haven't thought of?"

Why the "Hitting a Wall" Argument Was a Misjudgment

Before O1's release, there was a period when the industry buzzed with claims that "AI has hit a wall." Tejal was visibly frustrated by this.

"I get frustrated seeing those posts," she said bluntly. "I've been watching these models improve for a long time now, and they just keep getting better and better. Looking at our current research roadmap, I don't see any sign of it stopping."

She believes this judgment holds for the entire industry. Going further, she rejected the notion that "researchers are overhyping the models": "If anything, we're underestimating their power. We've worked hard to publish data and show the curve going up, but people still underestimate."

On the question of AGI evaluation standards, Tejal offered an intriguing observation: models passed the Turing test long ago, yet hardly anyone discusses it. The Turing test was proposed by computing pioneer Alan Turing in 1950: if a machine can consistently fool human judges in a text conversation such that they cannot distinguish whether they're talking to a human or a machine, then the machine is considered intelligent. However, academia has long widely questioned its validity — philosopher John Searle's "Chinese Room" thought experiment pointed out that a system passing the Turing test doesn't necessarily truly "understand" language; it's merely performing symbol manipulation. Tejal noting that models passed the Turing test long ago yet it's rarely discussed precisely reveals the test's obsolescence: when models actually passed it, people found this wasn't the true defining moment for AGI, and society had quietly shifted the discussion framework to higher-dimensional standards. "In many cases, models are already indistinguishable from humans." She believes the real defining moment will come when people realize most of their work has been handed off to models, and when AI delivers substantive scientific breakthroughs.

From Static Benchmarks to Real-World Evaluation

Evaluation methods themselves are also evolving dramatically. Early evals were mostly inherited from traditional NLP methods, too simple, and quickly "beaten" by models. The team had to keep creating harder, more realistic benchmarks.

Two key concepts come into play here:

Bench maxing — refers to trainers optimizing solely to look good on a particular benchmark rather than making the model genuinely useful. The essence of this problem is an expression of "Goodhart's Law": once a metric becomes an optimization target, it ceases to be a good metric. Goodhart's Law originally came from monetary policy — British economist Charles Goodhart observed in 1975 that when a central bank makes money supply a policy target, the correlation between money supply and inflation breaks down. This principle is especially prominent in AI evaluation: once a benchmark becomes an industry-consensus leaderboard, labs will, intentionally or not, optimize against it, causing benchmark scores to gradually decouple from real capabilities. Tejal made clear this is "usually a bad thing": "You might spend 90% of your compute making the eval score look good, the model looks great at launch, and then users discover it's only good at that one thing. That's not a good experience." She emphasized that OpenAI maintains solid research discipline, even publicly disclosing results where its own models weren't the best — "We just want to present the truth."

Saturated — refers to a model getting nearly every question in a benchmark right, approaching 100%. "It's like using a high school math exam to compare two geniuses — they can both pass, but you have no way to distinguish them."

To solve the dilemma of "being unable to measure users' real needs," the team introduced GDPVal. GDPVal's name is inspired by Gross Domestic Product (GDP), and its design philosophy is to directly tie AI capability evaluation to real economic value creation. It's built on the U.S. Bureau of Labor Statistics' O*NET occupational information database — maintained continuously by the U.S. Department of Labor, it contains detailed task descriptions, skill requirements, and knowledge domains for over 900 standard occupational classifications, and is an authoritative data source for labor economics research. GDPVal extracts representative tasks of high economic value from it, covering more than 40 occupations, such as investment due diligence and drafting legal memos. This design approach represents an important paradigm shift in AI evaluation from "academic-oriented" to "economic-oriented": traditional NLP benchmarks measure technical metrics of language understanding, with limited relevance to actual commercial value; GDPVal directly asks "how many tasks in real work can this model replace," establishing a quantifiable mapping between research investment and commercial value. The earliest models scored under 20% on this benchmark, far below humans.

"But I'm very proud of the team — we decided to publish this new method for measuring real economic impact," Tejal said. This benchmark catalyzed an internal "wake-up call," shifting the research direction more toward the practical usefulness of real work. And now, GDPVal may already be "too easy" — because its task descriptions are too explicit. The next step is to have models handle ambiguous, open-ended tasks like those in real workplaces.

AI Starts Doing Real Science: The Wet Lab Breakthrough

The most exciting part of the interview was the progress on scientific evaluation. The team publicly released multi-tiered scientific benchmarks: from the Olympiad-style Frontier Science Olympiad to Frontier Science Research, which measures whether a model can continue unpublished papers.

Most striking was the wet lab experiment in collaboration with Ginkgo Bioworks. The distinction between "wet lab" and "dry lab" is a fundamental concept in the life sciences: a dry lab refers to pure computer simulation and data analysis work, while a wet lab refers to experimental operations involving real biological samples, chemical reagents, and physical instruments. For a long time, AI applications in scientific research have focused mainly on the dry lab side, such as protein structure prediction (AlphaFold), drug molecule design, and genomic data analysis.

Extending AI decision-making into the wet lab means achieving a preliminary form of "Closed-Loop Science": the model is no longer just an auxiliary analysis tool but becomes the decision-making agent in the complete scientific cycle of experiment design → execution → observation → iteration. Unlike AlphaFold, which mainly solves the static problem of protein structure prediction, the wet lab closed loop involves dynamic optimization under real physical constraints, placing higher demands on the model's experimental design capabilities and uncertainty reasoning. This is methodologically fundamental — it closes the complete loop of "computational recommendation → physical execution → result feedback."

Ginkgo Bioworks is a world-leading synthetic biology company whose core infrastructure includes a highly automated robotic laboratory platform capable of running thousands of parallel experiments. In this collaboration, the model needed to optimize experimental protocols for protein synthesis; the robots would actually apply reagents and measure protein yields — a protein related to an ovarian cancer drug. The AI model played the role of an experimental design expert: given experimental constraints, it output optimized operational parameters for the robots to execute, then fed the results back to the model for iteration.

"We were nervous at first, because the human baseline was quite high and we weren't sure the model could surpass it," Tejal said. "But we should never underestimate the models." As it turned out, the model not only surpassed the human baseline but also set a new record in "yield per unit cost." Even more astonishing, this wasn't even their best model — just an early reasoning model.

"This was the first time we de-risked an eval that truly connects to the real world," she said. "We weren't waiting for a piece of code to finish running — we were waiting for the robots to finish the experiment."

Internal Metrics: The AGI Index and "Pain Is the Moat"

Tejal revealed that internally, OpenAI doesn't pay much attention to public benchmarks ("too noisy") and instead maintains an internal metric called the AGI Index. This design borrows the methodological essence of the Consumer Price Index (CPI) while also embodying the core idea of a "robust index" in statistics: the CPI measures inflation by selecting representative categories of consumer goods, weighting them by actual consumption weight, and periodically tracking price changes. Its advantage is robustness — an anomalous fluctuation in the price of a single good won't distort the overall metric. The CPI has historically faced similar criticism too — such as debates over housing weight — and has maintained its credibility as an anchor for economic decision-making through ongoing methodological revisions.

The AGI Index borrows this approach, building an internal metric basket that is multidimensional with continuously updated weights, covering all core areas of concern including capability, safety, and alignment, in order to capture a model's true comprehensive progress rather than letting research direction be distorted by a single public benchmark. This also explains why OpenAI deliberately uses an internal evaluation system different from public releases — the "signal-to-noise ratio" problem of public benchmarks, along with competitors' gaming optimization behavior, greatly reduces their reliability as a basis for internal decision-making. OpenAI maintaining the AGI Index internally rather than relying on public benchmarks is essentially building a private evaluation ecosystem that "competitors find hard to game specifically."

Facing increasingly long-horizon tasks, evaluation becomes extremely complex. Models can work continuously for days or even weeks, yet automated evaluation needs to produce results within a limited time. To address this, the team increasingly relies on real usage data from production environments, as well as scaling laws that can predict long-term trends.

Tejal offered a team motto: "Pain is the moat." As evaluation extends from the digital world into the physical world, a great deal of operations, logistics, and planning work has become a new bottleneck. "My work has shifted from theory, math, and programming to planning, operations, and physical-level things — and these are genuinely hard."

Advice for Everyday Users: Use It More, Try It Again and Again

For business owners and everyday users, Tejal gave simple but powerful advice: build your own evals, and keep trying repeatedly.

"A lot of people might have tried ChatGPT six months ago, thought it wasn't good enough, and gave up. They don't realize how fast progress has been." She believes the best eval is actually to use the model yourself (dogfood it) — "dogfooding" is an internal tech-industry term referring to a company's employees heavily using their own product before its official launch so they can discover problems from real use. Tejal extends this concept to everyday users: even if the model doesn't do something well one week, try again the next week — it may well succeed.

Her own working style is to have the model "take a first pass" at everything — sending Slack messages, deciding the next experiment, managing and operating. If the model does poorly, she turns that case into an eval. She predicts that the model's speed and efficiency in using her computer will gradually surpass her own, because the model can directly call connectors and plugins, which is far faster than a human clicking around.

Conclusion: Embracing the Transition with Cautious Optimism

On AI's impact on jobs, Tejal offered a balanced view. Current models are good at tasks rather than complete jobs — a job also involves judging what to do, dealing with ambiguity, collaborating with colleagues, and so on. But she believes models will soon also gain "delegation" abilities, deciding for themselves what to do and executing it.

She cited clinical trials as an example: writing hundreds of pages of FDA submission materials, tracking symptoms, data analysis — much of it fundamentally digital document and analysis work. If AI can accelerate it, it will bring enormous advances across healthcare, energy, manufacturing, policy, education, and more, ultimately giving consumers "faster, cheaper, better products."

"This is something people should be excited about," she concluded. "But we have to think very carefully about how to responsibly navigate the transition to that world."

The message this interview conveys is clear and powerful: AI's progress curve is far steeper than the public perceives, and the researchers actually at the frontier are, ironically, the ones most convinced that AGI is near. For onlookers, perhaps the best response is to set aside preconceptions and go try the latest models for yourself.

Key Takeaways

Key Takeaways