“From initial testing, g1 accurately solves simple logic problems 60-80% of the time that usually stump LLMs.”
Accuracy “has yet to be formally evaluated.” The strawberry test scored 70% over 10 trials, against 0% for plain Llama-3.1-70b and 30% for GPT-4o. Ten trials is an anecdote. Groq sells speed, and a loop that burns 5 to 10 calls per answer is a demo for buying more tokens.