Can AI Write Code Like Humans? Inside AlphaCode & OpenAI’s Theorem Prover

AI Takes on Competitive Programming and Math Proofs

During the Chinese New Year, DeepMind and OpenAI released two high‑profile AI systems: DeepMind’s AlphaCode, a Transformer‑based code‑generation model, and OpenAI’s neural theorem prover that solved two International Math Olympiad problems.

AlphaCode

AlphaCode uses a large‑scale Transformer language model to generate programs and was evaluated on Codeforces challenges. It ranked in the top 54.3 % of participants, beating 46 % of competitors, with an estimated Elo of 1238.

In a typical Codeforces task, AlphaCode samples many candidate solutions, runs them, and filters by output, mimicking the human competition process.

Researchers pre‑trained the model on public GitHub code and fine‑tuned it on a smaller competition‑programming dataset. The system automatically performs compilation, testing, and submission without human intervention.

Although AlphaCode’s performance is far from human‑level on the hardest problems, it demonstrates a substantial step forward for AI in program synthesis.

OpenAI Neural Theorem Prover

OpenAI built a neural theorem prover that uses a language model to search for proofs of formal statements. It achieved state‑of‑the‑art performance on the miniF2F benchmark (41.2 % vs 29.3 %).

The system employs “statement curriculum learning”: a set of auxiliary propositions of increasing difficulty is used as an unsupervised curriculum, allowing the model to iteratively improve by training on newly discovered proofs.

Formal mathematics poses two major challenges for neural methods: an infinite action space of possible proof steps and the lack of a self‑play signal. The authors address the former by sampling actions from a language model and the latter by using the curriculum of auxiliary problems.

Despite impressive results, the prover still lags behind top human contestants on IMO‑level problems, but the work opens new avenues for automated reasoning.

Conclusion

Both AlphaCode and the neural theorem prover illustrate rapid progress in AI’s ability to solve complex, reasoning‑intensive tasks, yet they also highlight current limitations such as error‑prone code generation and the need for better training curricula.