I just tested here and surprisingly, it's not a complete mess.

For example:

Write a Python proram calculating cosine similarity between two sentences.

import numpy as np
def cosine_similarity(sentence1, sentence2):
    # Convert sentences to vectors
    vectors = np.array([sentence1.lower().split() for sentence in sentence1.split()], dtype=np.float32)
    vectors = np.array([sentence.lower().split() for sentence in sentence2.split()], dtype=np.float32)

    # Compute dot product
    dot_product = np.dot(vectors, vectors.T)

    # Compute magnitudes of vectors
    magnitude1 = np.linalg.norm(vectors)
    magnitude2 = np.linalg.norm(vectors.T)

    # Calculate cosine similarity
    similarity = dot_product / (magnitude1 * magnitude2)

    return similarity

# Example usage
sentence1 = "Python is a powerful programming language."
sentence2 = "JavaScript is a high-level language used for client-side web development."

similarity = cosine_similarity(sentence1, sentence2)
print(f"The cosine similarity between the two sentences is: {similarity}")

Let's be real... any other 1b model would be falling apart.