4 hours ago, MattNischan said:

They don't actually know the correctness or soundness of any given statement; they can't do math, they can't understand what things are true or false, they don't have any concept or idea of logic, they're just reiterating what seems like a conversationally plausible response to a text prompt.

Weirdly, that sounds like a step up from the default MSFS ATC

My two cents is that air traffic control is mostly pretty easy. It's like "fly this route, unless another aircraft gets in the way, then move the aircraft so their safety bubbles don't touch, send one into a hold if you can't maintain separation without moving it more than 15% off its flight planned route" or something like that. That, and assistance with energy management on descents. Home computers have actually been doing this part pretty well since Radar Contact back in the early 2000's. After all, ATC instructions pretty much are scripted in any given real-world situation unless other traffic is around, and even then the controllers usually handle busy times in similar ways. Crowded skies see vectors for extended downwinds, with the length of the downwind depending on how many aircraft need to be sequenced, that kind of thing. Arrivals in this sector, departures in that sector. 

The AI part is using natural language to parse pilot requests. You can imagine how much computational horsepower you would need if the pilot says "Albuquerque Center, N123TL request, we need to divert because of an erroneous indication here in the cockpit, and we'd prefer an airport in VMC with at least a 10000 ft runway, but if there isn't one within 30 minutes of our present position then we'd like vectors to the nearest ILS at an airport with rescue equipment on the field."

Huge investment for the language parsing, but the control part is pretty easy. In the above example, break it down to: VMC within 30 minutes? Yes, filter runways to >=10000 feet, re-clear to the nearest using standard routes. No, then vectors to the nearest ILS. Easy from a control perspective, but hard when natural language is used to phrase the request instead of drop-down menus.

Edited May 15, 20242 yr by prolixindec

19 hours ago, prolixindec said:

My two cents is that air traffic control is mostly pretty easy. It's like "fly this route, unless another aircraft gets in the way, then move the aircraft so their safety bubbles don't touch, send one into a hold if you can't maintain separation without moving it more than 15% off its flight planned route" or something like that.

If this were true, then we would have seen the success of computerized ATC integrated into the modern workflow. However, many companies have tried for the past 30 years, and all have failed. The problem is that the solutions to such a problem set are unbounded.

Thankfully, us sim users generally forgive an occasional safe space intrusion or collision not involving our plane. So, the home version can take shortcuts, like being in direct control of the non-user planes, having infinite knowledge of their state, and in dire situations, forcibly moving those aircraft to an optimal location. 🙂 Even with all those shortcuts, it's a difficult problem space.

I like the part that mentions time to response as that was one thing that was holding me back...I have watched videos and noticed there was a longer response time from ATC

12 hours ago, MattNischan said:

If this were true, then we would have seen the success of computerized ATC integrated into the modern workflow. However, many companies have tried for the past 30 years, and all have failed. The problem is that the solutions to such a problem set are unbounded.

Thankfully, us sim users generally forgive an occasional safe space intrusion or collision not involving our plane. So, the home version can take shortcuts, like being in direct control of the non-user planes, having infinite knowledge of their state, and in dire situations, forcibly moving those aircraft to an optimal location. 🙂 Even with all those shortcuts, it's a difficult problem space.

 

I was really hoping that MSFS 2024 would embrace AI for Missions, ATC, Flight Instruction or even onBoard communication. So many different Careers where teased last year. All require dynamic generation for briefing, communication during the missions, react on weather, etc. Will this all be scripted like in MSFS 2020? I am confused. What else then AI could be the unique technical selling point for MSFS 2024. If there is a company that can and should make it happen, then it's Microsoft. I mean, look at Bing. The chatbot is called Co-Pilot !!!!

Upcoming livestream (8 hrs from now) by SI's lead dev:

 

 

On 5/14/2024 at 10:02 PM, MattNischan said:

Because they're treating it as an investment and a research tool. The more you use it, the more data they have to tune it.

LLMs are presently one of the most expensive compute workloads that exists at this time. Even before the recent explosion, OpenAI was spending somewhere between $1-2M per day in compute costs. And the estimates just for training GPT-4o are in the hundreds of millions range. These models are massive and need a ton of really expensive hardware power to make them go.

A super basic GPT-3 setup useful for a small handful of users is quoted around this specification:

• 4x compute nodes (probably with something like A800s) - $40-50K/ea
• 8x InfiniBand (200Gbps) per node - $16K/ea for cards, $20K for the switch
• Storage Server - $35K, plus $20K for connectivity infrastructure

So, just for a small setup sized for a basic research lab, you're already in for $300+K of static hardware costs before enclosures, racking, cooling, and electricity. Will the costs come down? Incrementally, sure, but there are no magic wands to wave here nor any freebies to be had. A single VM of this configuration in Azure (before storage and bandwidth costs) is $20K/mo.


If you want to explore powerful AI chat models yourself without such heavy costs or setup, platforms like https://overchat.ai/chat/best-free-ai-chat provide one of the best free AI chat experiences available today. It’s a great way to experiment with advanced conversational AI without investing in expensive hardware or infrastructure.

 

I fully agree that working with LLM is a huge expense and a serious investment. Such models require incredible resources not only for training, but also for maintenance - and even relatively small configurations cost hundreds of thousands of dollars in hardware alone.

Even with the gradual reduction in hardware costs and optimization of work, the scale of calculations remains impressive. This explains why access to such models is often limited and why companies try to get as much data and experience as possible to improve the models to justify such investments.

Of course, LLM is not just a "product", but a complex research project with a very high cost of support.

I never thought I'd see someone necroing a post about an MSFS addon😂

4 hours ago, Crepare said:

 

I fully agree that working with LLM is a huge expense and a serious investment. Such models require incredible resources not only for training, but also for maintenance - and even relatively small configurations cost hundreds of thousands of dollars in hardware alone.

Even with the gradual reduction in hardware costs and optimization of work, the scale of calculations remains impressive. This explains why access to such models is often limited and why companies try to get as much data and experience as possible to improve the models to justify such investments.

Of course, LLM is not just a "product", but a complex research project with a very high cost of support.

why did you reply to a thread that no-one has posted in for over a year?

2 hours ago, Luis Hernandez said:

I never thought I'd see someone necroing a post about an MSFS addon😂

Happens all the time, but the em dash makes me suspicious that even the post itself was generated using an LLM lol