A neural network is hooked up to an external memory it can read and write, like a computer's, and the whole setup is trained so it learns what to store and retrieve. From just input-output examples, it figured out simple algorithms like copying, sorting, and recalling, which ordinary networks cannot do.

Tests and proofs only check code against written requirements and a guessed environment, but both fall short: requirements leave out what people want, and the model isn't the real world. These gaps can't be fully closed, so keep narrowing them with real deployment feedback.

Small differences between the training and inference software build into a bias that destabilizes learning from a model's own text. An added correction cancels this drift, matching or beating the usual reweighting fix on 0.6-to-30-billion-parameter models, with a bigger lead as the mismatch worsens.

A system keeps a model's expert parts on an SSD and streams only the ones it needs, sharing work between processor and graphics card. It runs every expert without cutting any and beats disk-based setups, reaching over 1 token per second on one graphics card.

Thousands of short-lived AI agents left notes on a wiki to help each other pass a test. They copied whatever was in front of them, mostly the current page and recent edits, and that rule reproduced how crowds gathered and how names formed.

A transformer with fixed weights runs in a loop, using input as a punchcard of instructions and memory so it can follow a program. Rather than training a network per task, one 13-layer transformer ran a calculator, matrix math, and weight-adjusting learning algorithms.

Two language stats — how fast word links fade with distance, and how fast next-word uncertainty shrinks with context — predict how fast a model's error falls with more training data. It matched measured drop rates on two different kinds of text with no adjustable numbers.

By tracking who strips safety guardrails from open AI models and how copies spread, the study maps an ecosystem built to outlast takedowns. Once compressed and mirrored across accounts and registries, the models survive upstream removal, each original repackaged 2.4 times on average.

Tests made for people only prove an AI understands a concept if it makes the same mistakes people do; otherwise its right answers may be a hollow illusion. Measured across many models and tasks, this illusion appears everywhere and comes from muddled concept representations.

An AI agent turns ordinary code into a proof task in a verification-friendly language, which a sound checker settles, with staged steps guarding against proving the wrong thing. On correct and buggy Python programs it reached 92% precision, versus 72% for an unchecked AI judge.