I reminds me of very early ChatGPT with mostly correct answers but some nonsense. Given its speed, it might be interesting to run it through a 'thinking' phase where it double checks its answers and/or use search grounding which would make it significantly more useful.