Built an Energy-Based Model without training – just GloVe and GPT embeddings (github.com)

Researchers (or a developer) built a training-free, toy “JEPA+EBM” style evaluator by combining pre-trained GloVe (300d) and GPT embeddings (1536d) to score semantic compatibility via cosine similarity. Given the context “A child is playing with a red ball in the park.” and the candidate “The kid happily throws the bright red ball across the playground.” the system reports per-embedding similarities (GloVe ≈ 0.846, GPT ≈ 0.710), a joint cosine ≈ 0.778, and an energy ≈ 0.222 using energy ≈ 1 − joint_similarity. The pipeline loaded 400k GloVe vectors, computes token-level vectors, aggregates context/candidate embeddings, and exposes the scoring utility as a tool intended for a GPT-5.1 responses API; the authors emphasize this is a simulation with no EBM training. This is significant because it demonstrates a fast, interpretable, and training-free way to approximate semantic compatibility for ranking, filtering, or re-ranking candidate continuations using off-the-shelf embeddings. Key technical implications: combining heterogeneous embeddings (dense 1536-d GPT vs. 300-d GloVe) can boost robustness to lexical and contextual matches, but the setup is fundamentally limited — it’s not a learned energy landscape, ignores conditional dynamics, and depends on embedding norms and aggregation choices. Useful for lightweight tooling and debugging, but it should be treated as a heuristic rather than a substitute for trained JEPA/EBM models when modeling complex conditional distributions.