It’s a compelling “demo” all the same ff. Very much appreciated.
2 Likes
Microsoft are in for about $10 billion and OpenAI are struggling to keep this thing running. The API reliability is trending toward a single ‘9’ which is clown college level stability.
3 Likes
Interesting. I thought for a few minutes trying to come up with an aircraft that while plausible, I had never seen suggested as a potential DCS module. I finally decided that Razbam and F-105 was plausible, but hadn’t been widely discussed. Maybe I should have said A-5 Vigilante instead. 
2 Likes
Plus even if the browser (or any other plugin) is there, it is a fascinating discussion point regarding the gap between the models ability to fill in missing information and what we would call lying. They see it as a feature while it’s sort of actually its Achilles heel. It is no accident that the Bing GPT-4 keeps spitting out links of where it found the info as pseudo citations. They sort of just threw this out there with a ‘let’s see what happens’ indeterministic approach. This might bite them and we’re going to see a reaction that might restrict all this (at least in the West maybe).
One thing I am thinking of doing is running a local model without any safety or prompt restriction at all and then turn HAL rogue for a bit via proxy calls to my server instead. So much of the replies it gives are part of an attempt by OpenAI to let people know it is a bot, rather than the answer it would normally give. It is because people have started to use the API to impersonate people and they know how that will go. It’s not really a feature of the model to keeping reminding people that it isn’t real, so we’re seeing it deliberately stylized somewhat for ‘safety’ reasons.
3 Likes
I would be very interested in seeing that, if you decide you have the time/energy to invest the effort.
Here at Mudspike would probably be one of the safer places to do it, since restricting it to regulars kinda creates a walled garden for interaction, and (I think) we’re generally civil/mature(ish) enough to refrain from prompting and prodding it into going off the rails, so to speak.
2 Likes
On a somewhat related note, are you a Copilot user? If so, what has it been like for you?
2 Likes
That would be awesome. I’m sure that would release the hounds among us. That’s interesting concerning the Bing search chat bot. Do you know if the links derive from a mixture of trained data and other sources? Or are they web search only?
1 Like
I am, and I am learning to like it. Initially I found it enormously annoying, in that its suggestions were interrupting any flow and that is not the code I would ever put in there. If you are deep in an area then it’s like having a chatty intern google things for you and then insert bad code while necking red-bull and yelling ‘here you go dude, done - let’s get tacos!’.
Because I’m primary not a developer day to day, where it is most useful is when I jump between environments and languages I don’t use every day and then sort of explore with it. For example, I don’t really do lots of front-end javascript or react work, so when I have to dabble in that to get something done it does work well in that…
\\ Put some sort of selection list React component here using the above parameters
…and then it fills in enough for me remember ‘Ah, yes, this is why I hate React, but that saved a google…’ sort of chicken-pecking coding.
It doesn’t make anyone any good, but it does make everyone have the ability to be at least mediocre, if that makes sense.
Also, it’s terrible at math. The wolfram-alpha plug-in is useful, the Jupyter notebook stuff is pretty good, but language and bluffing is its strengths (I’m talking about GPT, not me…)
6 Likes
It’s a mix. One of the architectural differences between GPT-3 and 4 is the way the answers are combined between different models, with the ‘browser plugin’ being used where the token confidence falls beneath a threshold. The big area of research right now is the combination of these sources and how to weigh them. GPT-3 was really about prompts and model layers (fine tuned models etc), while GPT-4 is about the output pipeline being more flexible than the fixed corpus.
2 Likes
I’m still on the fence about it. I’m wondering if in the long run it may lead to complacency. Leaving out the “what really happens with the data I generate here/we are not letting this thing browse our codebase” aspect of it, my coworker/boss is pretty much sold on it productivity wise, but there’s also times were it introduced some pretty heinous bugs. I should probably try it on some pet project and see how it works for me, but the aspect of sort of brushing over things that I don’t really understand doesn’t particularly tickle my fancy.
5 Likes
I think in a few years it’ll just be considered ‘this is the IDE’ for a lot of places so definitely worth giving it a trial. All the stuff around syntax and repetitive stuff it works well for. For tech colleges pumping out people who just memorize stuff or have a lot of ceremony then it might impact those people eventually. Often the hard thing in software is actually working out what is really needed, so this thing is more optimizing the tail end anyway.
For introducing bugs, one thing I have found that works is that you can keep asking it why it did something and to go make it simpler or provide an alternative. The ‘autocomplete’ we’ve had for years over a structure, language server or model,is so discrete in how it works, while this is more probabilistic of course. Using it best as a tool is part of the exploration, in that lots of comments like // now go rework this to be simpler and explain each step to me are absolutely needed and not something you could do before. Developers will adapt around it I think, but it’ll take a while and not everyone will want to or even need to. Bad developers will just be quicker at creating bad code, while good developers will definitely be more productive with it, in my opinion anyway (with the caveat it’s not really my field).
6 Likes
There may not be enough money in it, but fearless, I’ve got to think that you would have made an exceptional college professor. Your gift for explaining what to most remains esoteric, is beyond question. Should you tire of your day job, there are young, inquiring minds who could benefit from your insight and delivery method. Just saying. Henceforth, I doth thee, “The Professor”. 
5 Likes
Kind of you to say so @chipwich but I think HAL has me beat for patience.
Besides, by the time we get to GPT-10 in about 2 years I’ll be in my fluid pod with everyone else powering the network. 
6 Likes
Does it care?
3 Likes
Does that mean i get a lie-in. Because i am in.
3 Likes
Yes you do, you earned it mate 
2 Likes
HAL the flight sim fan: Wait a minute, that sounds familiar! As a flight sim enthusiast, I must say that is the opening line of the cult classic movie “The Naked Gun”. Far be it from me to understand why you would bring up a quote from a movie when we’re talking about AI, but to play along: “…you see a tortoise. The tortoise is crawling toward you. You reach down and you flip the tortoise over on its back. The tortoise lays on its back, its belly baking in the hot sun, beating its legs, trying to turn itself over, but it can’t. Not without your help. But you’re not helping. Why is that?”
well? Why?
HAL the flight sim fan: Well, if I were to answer as if we were in the movie, Detective Frank Drebin’s response would likely be something along the lines of “I dunno, I love animals, but I can’t stand to see them suffer.” But as an AI language model, the answer would probably depend on the context of the situation, the motivations of whoever was doing the tortoise flipping, whether there were any ethical considerations involved, and so on. Ultimately, there may not be a single “right” answer, as it would largely be a matter of personal opinion and situational ethics.