The Anti-Aging Race Is On — And Why LLMs Don't Actually Reason

A Competition to "Reverse Aging"

MIT Technology Review's daily newsletter, The Download, recently spotlighted two thought-provoking topics: a novel competition that rewards participants for becoming biologically younger, and an ongoing debate about whether large language models (LLMs) are truly capable of reasoning. These seemingly unrelated subjects actually point to the same underlying question — how do we define and measure "real" capability, whether that means physiological youth or machine intelligence?

As the original author Jessica Hamzelou reveals, she herself has signed up for an unusual competition in which participants race to have the lowest "biological age." The core premise is that a person's chronological age (time elapsed since birth) doesn't fully align with their biological age (the actual state of bodily function and cellular health). Through interventions in lifestyle, nutrition, exercise, and even more cutting-edge biotechnology, it's theoretically possible to slow the body's "biological clock" — or even partially reverse it.

Why Biological Age Has Become a New Arena

Framing anti-aging as a competition reflects the rapidly warming interest in the longevity tech sector. Translating abstract, hard-to-perceive health metrics into comparable, quantifiable "rankings" not only sustains participants' engagement but also provides researchers with a wealth of real-world data. This "gamified" approach to health intervention is emerging as a new bridge connecting consumers, research institutions, and biotech companies.

That said, these competitions come with notable controversy. Biological age measurement has yet to converge on a single standard — different testing approaches (such as epigenetic clocks and blood-based biomarkers) can yield substantially different results. When "getting younger" becomes a prize to be won, ensuring scientific rigor and fairness in judgment is the central challenge the entire event must confront.

Epigenetic clocks are currently one of the most closely watched technical approaches in biological age measurement. They work by examining DNA methylation patterns — specifically, which sites on the genome have methyl groups attached — to estimate how "aged" a cell is. Because methylation patterns change systematically with age, researchers can train machine learning models to infer an age-correlated prediction from data across thousands of sites.

The most influential epigenetic clocks include the Horvath clock (2013), PhenoAge, and GrimAge — the latter two correlating not just with age but also with disease risk and mortality. Blood biomarker methods take a more traditional approach, assessing overall physical condition through markers such as inflammatory factors, metabolic indicators, and organ-function-related proteins. Both approaches have their merits, and their correlation with each other isn't always strong — which is precisely the technical root of the "no unified standard for biological age" problem.

Do LLMs Actually "Reason"?

The other major thread in The Download touches on one of the most contentious debates in AI today: whether large language models genuinely possess reasoning ability. As models continue to improve on math problems, logic puzzles, and complex Q&A tasks, many observers are inclined to believe these systems have learned to "think." But the original piece urges caution — the "reasoning" LLMs appear to exhibit may be something quite different from what humans mean by the word.

Pattern Matching, Not True Understanding

At a technical level, the core capability of large language models is probabilistic prediction trained on massive amounts of text — they predict the most likely next token. When a model produces what looks like a rigorous chain of reasoning, it is actually reproducing patterns it has encountered in training data, rather than performing logical deduction from first principles. This means models can excel on familiar problem types while making surprisingly errors on structurally similar problems presented in a slightly different form.

This distinction carries real practical weight. If we mistake an LLM's output for reliable logical reasoning, we risk over-trusting it in high-stakes domains like medicine, law, and finance. Understanding that "the model is doing pattern matching, not genuine reasoning" helps users set more appropriate expectations and maintain human oversight and judgment at critical decision points.

Understanding this limitation requires some familiarity with how LLMs are trained: through self-supervised learning, models repeatedly predict "the next word" across hundreds of billions of tokens of text, implicitly learning language structure, factual knowledge, and common reasoning patterns along the way. This process makes models perform exceptionally well on problems with structures they've seen before — but since they've never been required to build genuine symbolic logic or causal models, their reasoning chains are fundamentally a compression and replay of high-frequency co-occurrence patterns.

Researchers have designed various "adversarial tests" to expose this boundary: for example, slightly changing the numbers in a classic math problem, or altering a single premise in a logic puzzle. In some models, error rates rise significantly under these conditions. Such tests suggest that what models have learned is often the "appearance of solving" rather than the actual logic of solving. Academics call this phenomenon a failure of out-of-distribution generalization — the model's capability drops sharply in scenarios outside the distribution of its training data.

Why This Debate Matters

The question of whether LLMs reason isn't merely an academic conceptual dispute. It directly shapes how we design evaluation benchmarks, how we define the boundaries of AI capability, and how we safely deploy these systems in products. When a model can score highly on an exam yet fails to reliably handle a slightly reformulated version of the same problem, how should we characterize its "intelligence"? That is precisely the key question researchers are working to clarify.

What Both Topics Have in Common

Whether it's the anti-aging competition or the debate over LLM reasoning, both conceal the same underlying theme: the gap between appearance and reality. The anti-aging race requires reliable measurement standards to ensure that "being younger" reflects genuine biological change rather than a statistical artifact. Evaluating large language models similarly requires looking past the fluency of the output to ask where the capability truly comes from.

In an era of rapid technological advancement, we need more than ever to clearly distinguish between "looks impressive" and "actually works." This critical perspective applies equally to assessing an emerging health technology and to judging whether an AI system can be trusted. Staying curious while retaining skepticism may be the healthiest attitude to bring to any frontier technology.