Is there a limit to what LLMs can do? You have to wonder

Even though I work with this technology every day and constantly discuss it with researchers and developers, I am still amazed by the power of LLMs.

Here's an example. A friend of mine recently completed a solo flight from Ohio to Massachusetts. I stalked him on FlightAware, and when his flight was completed, I uploaded a screenshot - just a screenshot from my phone - of the flight summary to ChatGPT using the 4o model.

Here's the conversation that followed ...

---

🧑🏻Me: This is the log of a flight by a new pilot. Do you have any insights?

🤖 ChatGPT: The flight log shows the altitude (green line) and speed (brown line) over time. Here's a breakdown of the key insights:

1. Takeoff and Climb: The initial climb appears smooth and consistent, reaching a cruising altitude of around 31,000 feet. The pilot maintained a good climb rate with a steady increase in speed, suggesting a controlled takeoff and ascent.

2. Cruising Phase: The flight maintained a steady cruising altitude and speed for the majority of the flight, indicating stable conditions and possibly good handling of the aircraft.

3. Descent: The descent begins sharply around 11:03, with a simultaneous decrease in speed. The descent rate seems controlled and follows a steady path, which is crucial for maintaining passenger comfort and aircraft safety.

4. Final Approach and Landing: The speed decreases significantly as the altitude drops, which is typical during the final approach to landing. The landing appears to be smooth with a gradual reduction in both speed and altitude.

Overall, this flight log suggests that the new pilot managed a consistent climb, steady cruise, and controlled descent, which are all positive indicators of a well-executed flight. If this was one of their first flights, it shows good handling and understanding of key flight phases.

🧑🏻Me: Do you have any criticisms of the flight?

🤖 ChatGPT: From the flight log, there are a few areas that could be critiqued or considered for improvement:

1. Climb Rate Fluctuations: During the climb phase, the rate seems to fluctuate a bit before stabilizing at cruising altitude. A more gradual and steady climb rate can be more fuel-efficient and comfortable for passengers. The climb rate starts off quite aggressively, peaking quickly before leveling off.

2. Cruising Speed: While the cruising phase shows a mostly steady altitude, there is a slight fluctuation in speed. Ideally, maintaining a more consistent cruising speed is preferred for optimal fuel efficiency and passenger comfort.

3. Descent Profile: The descent starts off relatively sharp, which is normal, but maintaining a consistent descent rate is important. If the descent rate fluctuates too much, it can lead to passenger discomfort or operational inefficiencies. The descent here is steep initially, followed by a leveling off before a steeper final descent. It might benefit from a more gradual and consistent descent to avoid rapid altitude changes near the end.

4. Final Approach Speed Management: As the plane approaches the final stages of descent, the reduction in speed is quite steep. Managing a more gradual decrease in speed during the approach phase can help ensure a smoother landing and better handling of the aircraft.

These observations are not necessarily indicative of major issues, especially for a new pilot, but they are areas where practice and experience can lead to smoother, more efficient flights. Overall, it seems like the flight was managed well, but focusing on smoother transitions between phases and maintaining more consistent rates could be beneficial for future flights.

---

My friend took the criticisms well and felt they were entirely appropriate. For my part, I think this is astonishing. Never mind if this is stochastic parroting without "real" understanding or creativity: this is dazzling, thrilling technology, and I am delighted to see this come to fruition in my lifetime.