AI Speeds Discovery, but Kevin Kelly’s ‘Thinkism’ Still Stands in the Lab

Kevin Kelly’s 2008 critique of “thinkism” still holds against newer evidence from AI-assisted science. Systems can now generate hypotheses, design protocols and automate parts of computational research, but the strongest biomedical results still pass through experiments, expert review and calendar time.

What has changed is the location of the bottleneck. AI can compress literature work, reasoning, coding and some forms of evaluation far more aggressively than Kelly’s original argument anticipated. It has not demonstrated that intelligence alone can produce validated treatments, reverse human ageing or deliver the instantaneous material progress associated with the strongest versions of the Singularity.

Kelly’s target was instant progress, not intelligent machines

Kelly did not argue that machines could never become smarter than people. His narrower objection was that Singularity narratives often confuse greater intelligence with completed scientific work: a powerful mind might form better explanations, but explanations do not automatically become reliable knowledge or functioning technology.

In Kelly’s original “Thinkism” essay, published on September 29, 2008, he applied that distinction directly to cancer, cellular ageing, fusion and immortality. These problems require observations that do not yet exist, physical apparatus, failed prototypes and repeated tests on living systems. Faster reasoning can choose better experiments, but it cannot obtain an experimental result merely by imagining the missing data.

This is why “thinkism” is more precise than a general complaint about AI hype. It identifies a particular unsupported leap: moving from a system that can reason beyond human ability to the conclusion that it can immediately control biology and matter. Kelly’s claim leaves ample room for acceleration while rejecting the idea that acceleration must become instantaneous.

AI co-scientists strengthen both sides of the argument

Recent systems show that AI can do more than summarize published work. Google’s AI co-scientist was designed to generate and rank research hypotheses and propose experimental protocols. That is a meaningful expansion of what machines contribute to discovery, especially when a field contains more literature and candidate relationships than one research team can examine.

Yet the project’s most persuasive evidence also illustrates Kelly’s boundary. Google Research’s February 2025 report says predictions for acute myeloid leukaemia drug repurposing were assessed through computational biology, clinician feedback and in-vitro experiments; the broader evaluations used expert guidance and laboratory work across three biomedical applications. Google presented the system as an assistive collaborator and announced access through a Trusted Tester program, not as an autonomous engine that had completed drug development.

The distinction matters. Generating a plausible candidate is not equivalent to establishing safety, effectiveness or a treatment for patients. A result in cell lines may justify further research, but it does not erase the subsequent stages in which organisms, doses, side effects and long-term outcomes must be observed.

Some experiments really can move into software

The strongest update to Kelly’s framework is that “the real world” is not a single, fixed bottleneck. When a scientific problem can be represented by suitable data, executable code and an objective score, AI may perform thousands of iterations without waiting for cells to divide or hardware to be constructed. In that domain, computation is not merely thinking about an experiment; it can execute and evaluate the relevant empirical loop.

This boundary became clearer in 2026. Google Research’s updated account of Empirical Research Assistance describes a system that generates evaluation code, explores thousands of code variants and achieved expert-level performance on six benchmark problems spanning fields including genomics, public health, neuroscience and numerical analysis. The tasks were deliberately “scorable”: each supplied a problem, an evaluation metric and appropriate data.

That result does not refute Kelly so much as refine the categories. AI can dramatically accelerate experiments when their environment, evidence and success criterion are already computable. The acceleration becomes harder to transfer when the decisive evidence must come from a new physical measurement, an organism’s response or an intervention observed over time.

Immortality remains outside the scorable-task shortcut

Human longevity is not one well-defined benchmark with a complete dataset and an agreed score. It joins many nested questions: which biological changes are causes rather than correlations, whether an intervention works across tissues, what harmful effects emerge, and whether benefits persist in people. Even a system that proposes an excellent mechanism has not thereby established the outcome.

Simulation can narrow the search and help researchers reject weak candidates before costly testing. Laboratory automation can run more experiments in parallel, while better models can adapt the next experiment to earlier results. These are genuine speed gains, but they increase the throughput of empirical work rather than abolish the need for it.

The key test is where new evidence enters the chain. If the answer can be scored against existing data or generated inside a faithful computational environment, AI may compress the cycle sharply. If the answer depends on an unknown response from a living body, a new material or a physical device, reality still has to return the result.

What Kelly’s critique does—and does not—settle

“Thinkism” does not prove that a technological Singularity is impossible, nor does it establish a permanent ceiling on scientific progress. It challenges the proposed mechanism behind an abrupt version of the event: intelligence recursively improves, then rapidly derives every important invention and converts those ideas into working outcomes.

The evidence now supports a less theatrical transition. AI systems can enlarge the space of hypotheses, write experimental software and help scientists prioritize physical tests. Their impact may be profound even if no single day cleanly separates a pre-Singularity world from a post-Singularity one.

For claims about immortality, the practical question is therefore not how impressive an AI appears in conversation or on a reasoning benchmark. It is whether the system produces a testable intervention, whether independent evidence survives successive stages of validation, and whether the measured benefit applies to humans. Until those steps are completed, intelligence has accelerated the search—not delivered the promised destination.

Also read:

• Ozempic Microdosing Promises Less Drug, but the Evidence Still Falls Short
• Venting to AI Can Ease Distress—but the Evidence Is Still Short-Term
• China Is Winning the AI Race Not Just with Open-Source Models — But with Unmatched Speed of Real-World Implementation