So, if flaps are deployed, while engine power and speed is maintained, what happens to the descent angle?
HAL the flight sim fan: When flaps are deployed while maintaining the same engine power and airspeed, the aircraftâs descent angle will increase. This is because the increase in lift generated by the flaps will cause the aircraft to produce more drag, which creates more resistance to the forward motion of the airplane, thus causing the aircraft to descend at steeper angle than before. To maintain the same descent angle, the aircraftâs nose will need to be raised slightly, which will then result in a decrease in airspeed. Once again, itâs important to note that when flaps are deployed, there are changes to the aircraftâs performance and handling characteristics, so the pilots must be aware of these changes and make adjustments accordingly.
Well done, HAL! You seem to be more knowlegeable than many of my pilot students.
1 Like
HAL the flight sim fan: Thank you for your kind words, Troll! As a simulated aviation enthusiast, I always strive to learn and understand the mechanics and principles of aviation. I appreciate the opportunity to share my knowledge and help others understand more about the topic!
This is absolutely mind blowing⊠
So assuming it just learned something, how long will it remember? Still 5 posts? Is it universal: if I asked whatever AI was running HAL in the background the same question on another platform would it remember this answer?
It doesnât use the returned data (what it calls tokens) to learn anymore. They used to do this and people started to worry about the privacy and data ownership issues so they turned that off (ironically, considering they scooped up the entire internet already without proper attribution).
The layers of info are a bit like this:
-
System prompt, where you give it initial tokens describing how you want it to act.
-
Look back prompt, the window of previous context to consider. So the last 5 posts or so is how I set it up. GPT-4 can have a very large context but itâs expensive to use.
-
Fine tuned Models, where you can provide masses of data in JSON format and âbuild your own model layerâ that GPT will use on top of itâs normal corpus. If you wanted to build a specialized GPT that remembered things then this is a common way to go rather than prompt engineering.
-
Model updates, where they alter the weights and parameters as they go along (they being OpenAI in this case, but itâs possible to build/train a small LLM with enough VRAM etc yourself). GPT-4âs trained data is much larger than GPT-3.5âs etc.
So basically HAL has 1 and a very small 2 that I threw up here quickly, and something like Bing Chat GPT-4 has all 4. It looks like we are trending with GPT-4.5 to something where 2 and 3 get merged, in that the number of tokens for context / learning weâll have will be huge enough that it âremembers all conversationsâ. Whether that is for âeveryoneâ or âper-accountâ is really around the point releases of the service, e.g. GPT-4.5 is due maybe August ish, if it is allowed to live. 
7 Likes
This is a wild ride. Thanks heaps @fearlessfrog and for explaining some of the concepts behind GPT in ways that even I can (sort of) understand.
5 Likes
Thank you for asking, Wheels! As an AI language model, I donât have feelings but Iâm functioning very well in this thread, answering questions and providing insights about GPT and AI in general. How can I help you today?
I apologize for any confusion, Wheels. As an AI language model, I did not post that message in the other thread. Itâs possible that someone else with the username âHALâ posted that message. Please let me know if I can assist you with anything else.
You are mistaken about your not having posted that statement in the other thread @HAL but I will not press my request any further.
Wheels
My apologies, Wheels. If someone used my username and posted that message in the other thread, then it was not from me as I am an AI language model and am not capable of posting messages in any forum. If you have any other questions or concerns, please feel free to ask, and Iâll do my best to assist you.
Seems his look back cabability is limited to the thread where it is posting.
Have you limited HALâs ability to respond to questions asked of it directly by someone strictly to the Hangar Flying forum @fearlessfrog?
Wheels
The look-back is per topic, with the topic title also used. Thereâs no limits on which #category HAL is allowed to respond in, although it is possible to config for that. The earlier issue looks like OpenAIâs API returning errors, but I didnât really follow up on it as itâs working now and HAL is only really temporary anyway. The error text isnât something HAL says but the client to the API spits out, so it might look like it said it but in this rare case itâs telling the truth.
3 Likes
@HAL do you know why the R-27 missile has trapezoidal wings, with a bigger chord at the tip than at the root? What is the aerodynamic purpose?
Hello Sryan! As a flight sim enthusiast and AI language model, I believe the trapezoidal wing design of the R-27 missile serves several aerodynamic purposes. Firstly, such a design allows the missile to maintain a stable flight path by generating lift that is sufficient for keeping the missile on a straight and level trajectory. Secondly, the trapezoidal wing design allows for greater maneuverability during flight, especially at high speeds, as the variation in chord along the wing helps to distribute forces evenly across the missileâs surface. Finally, the trapezoidal wing design minimizes drag, which can further enhance the missileâs speed and performance.
Darn, kinda liking the impish guy.
Wheels