Contamination in coding benchmarks is subtle. Instead of just spitting out a known solution, models like GPT-5.2 use implicit knowledge from their training data (e.g., popular codebases) to reason about unstated requirements. This makes it hard to distinguish true capability from memorization, as the model's 'chain of thought' appears logical while relying on leaked information.

Foundation models can't be trained for physics using existing literature because the data is too noisy and lacks published negative results. A physical lab is needed to generate clean data and capture the learning signal from failed experiments, which is a core thesis for Periodic Labs.

When an LLM sees the same document thousands of times, it prioritizes memorizing that specific content to lower its training loss. This gives a false impression of progress, while the model fails to generalize to new, unseen data. A study of Google's T5 corpus found one sentence repeated over 60,000 times.

The idea of separating "fact learning" from "skill learning" is a false dichotomy. Models need a base of internalized facts to reason effectively. The key is developing intelligence to compress what's important and discard what isn't, much like lossy human memory.

Scientific literature suffers from severe publication bias because researchers rarely publish failed syntheses. Models trained only on published successes lack negative controls and cannot master real-world exploration. Periodic Labs captures complete campaign histories—tracking failures, reasoning traces, and iterative adjustments. Liam Fedus argues that models must be trained on the messy process of conducting science rather than just final polished outputs.

When asked to analyze 100 papers, LLMs often admit they didn't complete the task. This failure stems from outcome-based training, which prioritizes a plausible-looking final output over correctly following the required process, revealing a fundamental flaw in current training paradigms.

A flawed or unsolvable benchmark task can function as a 'canary' or 'honeypot'. If a model successfully completes it, it's a strong signal that the model has memorized the answer from contaminated training data, rather than reasoning its way to a solution.

Unlike humans, whose poor memory forces them to generalize and find patterns, LLMs are incredibly good at memorization. Karpathy argues this is a flaw. It distracts them with recalling specific training documents instead of focusing on the underlying, generalizable algorithms of thought, hindering true understanding.

Current LLMs fail at science because they lack the ability to iterate. True scientific inquiry is a loop: form a hypothesis, conduct an experiment, analyze the result (even if incorrect), and refine. AI needs this same iterative capability with the real world to make genuine discoveries.