As global tech giants and capital flood into the AI agent space, former OpenAI co-founder and leading AI expert Andrej Karpathy has issued a deeply counterintuitive warning: the cutting edge of agent capabilities currently lies not with cash-rich corporations, but with independent developers and entrepreneurs. At the same time, major AI labs are attempting to compensate for the structural deficiencies of foundation models in end-to-end reasoning through ecosystem positioning, igniting a melee that has shifted from “competing on model parameters” to “locking down workflow ecosystems.”
In a recent internal sharing session, Karpathy pinpointed the industry’s core mistake: people are forcing agents to work while completely ignoring the need to first master the underlying foundation models. Drawing on his firsthand experience at OpenAI, he revealed that as early as 2016, OpenAI attempted to have agents perform everyday tasks like booking flights and ordering food via reinforcement learning. The project failed completely, costing the company five full years.
“At that time, the technology simply wasn’t ready. The only hammer the team had was reinforcement learning. The correct move at that point should have been to completely forget about AI agents and focus all energy on building language models,” Karpathy emphasized. His core logic follows three steps: first, immediately stop fantasizing about agents doing everything and first understand the underlying model; second, recognize the industry reality—demos are extremely easy, but building products takes a decade; third, understand that agents themselves are not the product at all—the foundation model is the true core. Once the foundation is solid, agents will naturally emerge.
This thesis finds subtle confirmation in the latest moves by tech giants in the AI-for-science (AI4S) arena. On June 30, Anthropic and OpenAI coincidentally placed key bets in the AI4S space, and their choices happen to expose the foundation model shortcomings Karpathy warned about.
Anthropic released Claude Science, a scientific research agent workbench, explicitly stating it “does not rely on new models.” Instead, it integrates existing capabilities through workflows to handle scientists’ daily research processes. The workbench connects to over 60 scientific databases, comes pre-built with toolkits for genomics, protein structures, and chemistry, and features a main AI assistant that breaks down tasks like a project manager, with a fact-checker performing cross-validation. The technical essence involves calling external vertical models via the MCP protocol to execute specific computations, with Claude itself only handling natural language understanding, task decomposition, and result interpretation.
OpenAI, meanwhile, launched GeneBench-Pro, an evaluation benchmark covering 10 fields including genomics and quantitative biology. Its test data shows that across 129 real-world scientific research workflow problems, even the strongest model, GPT-5.6 Sol, achieved only a 28.7% end-to-end pass rate under Max inference settings. Among non-GPT models, the strongest performer, Claude Opus 4.8, managed only a 16.0% pass rate.
This data reveals a critical flaw OpenAI has dubbed the “notice-act gap”: models can notice data anomalies and identify localized diagnostic signals, but cannot translate that awareness into downstream methodological adjustments or make the corresponding correct analytical decisions. As Wu Hao, founder and CEO of Lumitech, analyzed, general-purpose large language models face three structural shortcomings in the life sciences: difficulty directly understanding the unique structure of raw biological data; the inability to simply apply text tokenization rules to biological phenomena; and the prevalence of large amounts of unknown missing values in biological data.
The three major AI labs have adopted starkly different strategies in AI4S, reflecting their respective judgments on where the ceiling lies. Anthropic’s approach is the most straightforward—essentially “owning” an entire lane by using engineering to compensate for model unreliability. Claude Science is available to Pro, Max, Team, and Enterprise subscribers. The company also recently launched a $30,000 grant program targeting 50 postdoctoral and graduate student projects, aiming to lock in the academic habits of young scientists before they become independent PIs.
OpenAI’s logic is to use GeneBench-Pro as the referee, defining “what good AI4S looks like,” and then use GPT-Rosalind—a specialized biological reasoning fine-tuned model launched four months ago—as the athlete to chase high scores. The model is available as a research preview to qualified U.S. enterprise customers, subject to security review.
Google DeepMind holds a unique trump card. It possesses foundational scientific models like AlphaFold and AlphaGenome as proprietary assets, deeply bundled with Gemini for Science and integrated with over 30 life science databases. Models that other players can only access as external tools are, at Google, part of their own underlying infrastructure.
Karpathy offered a highly disruptive assessment of this ecosystem competition: “When a paper on a new type of agent is published, teams at major labs also find it eye-opening, because they haven’t been secretly developing in that specific branch for five years. This means the giants must compete on equal footing with all grassroots entrepreneurs and hackers in this space.” He advised developers to draw fresh inspiration from neuroscience, referencing brain structures like the hippocampus and thalamus to design memory, planning, and conflict-resolution mechanisms for digital entities.
Notably, the top-tier clients in the AI4S space have yet to be locked in by any single giant. Pharmaceutical behemoth Novo Nordisk appears simultaneously on Anthropic’s Claude Science case study client list and OpenAI’s list of early Rosalind partners. The same client is trialing multiple vendors’ solutions in parallel, indicating that the market remains in an open competitive phase—no single company’s toolchain is yet strong enough to convince scientists to migrate their complete workflows.
Karpathy’s warning and the giants’ actions both point to a shared reality: model capabilities have hit the ceiling of the “notice-act gap.” The old path of stacking compute power does not work in complex scenarios like scientific research. Engineering integration, ecosystem positioning, and data sovereignty have become more pragmatic breakthrough points. But as he noted, autonomous driving has already proven the decade-long chasm between demo and product, and agents fit the same pattern—imagining and demoing is easy, but to truly make them work, developers must be prepared to grind for a decade.
For ordinary developers currently building agents, Karpathy’s conclusion may serve as both a sobering reality check and a shot of confidence: “You are at the very forefront of this transformative technology.”