⟍ı⋂ᑕο−ހ⅂∪ᒧ+х⊃Γ·L−⟍ހᒧހ·ހ−ހı⅂⟍⅂⟍ހ⊃ހ−⋂⟍⅂οހı⅂ᑕ⅂οހ−⅂ᑕ⋂⟍ހ⅂ހLހLހο⋂⟍ހLހ·⋂⟍⅂⟍ހᒧ⅂+⅂+ހı
In the past, you had to hand draw promotional pictures to convey the marketing message. Nowadays, you just generate the same pictures with AI. Either way, the pictures and stories are equally fake.
I guess the biggest difference is the amount of time a human spent with the picture. Nowadays, some marketing pictures clearly went through exactly zero seconds of quality control.
Dangers like that can be identified quite easily. It’s a qualitative thing, and qualitative chemistry is pretty robust. For instance, we can say that there’s a risk that a particular reaction will produce hydrogen under specific circumstances. We’ll just build the plant accordingly instead of trusting that we can always operate the plant correctly. Sooner or later, you’ll end up running the plant in the wrong way, and you’ll produce some hydrogen, so it’s good to have a plant that can detect and deal with it safely.
However, usually the idea is to produce something entirely different, and do so efficiently. Those sorts of questions are quantitative, and that’s where things can and will go wrong all the time. Like, how do you ensure that your expensive catalyst isn’t covered in goo, or corrosion doesn’t eat your fancy impeller? How do you ensure that the amount of impurities in the product will remain reasonably low? It’s all about the quantities and reaction rates, and that’s the hard part with inorganic chemistry.
Inorganic chemistry is fairly simple and fun as long as you keep it in the lab. Industrial-scale inorganic chemistry gets ridiculously complicated because all the reactants and products are complicated and messy. Also, a large reactor will have all sorts of gradients, which means that the reactions take place in unfavourable conditions all the time. None of it is ideal, and none of it follows simplified laws or rules very well.
Sure, we have all sorts of fancy calculations, but none of them predict very accurately what’s going to happen and when. Even the best models and theories give approximate and crude answers when you’re dealing with messy industrial-scale chemistry.
Models give you a rough idea, lab experiments give you a decent idea, but running the process at full scale is the only way to find out exactly how those reactions really work in real life.
Turns out, our theories are too simple to handle complicated solutions. They can predict the behaviour of simple solutions very well, but that’s not good enough. In real life, you rarely have well behaved clean reagents.
LOL, that would put a lot of linux bloggers out of business.
When experimenting, you shouldn’t use your finest beans. When I’m getting to know a new method, I usually start with the cheapest light roast I can find in a supermarket. It’s nowhere near a specialty coffee, but it’s still a lot nicer than any medium roast I’ve ever found. As it costs about 18 €/kg, making mistakes doesn’t get too expensive. This method is ok if you’re testing a new method and you’re brewing coffee for a large crowd at the same time. Alternatively, you could brew for yourself, but you would need to drink more than a few cups.
Another option would be to use decaf, which comes with different compromises. It’s about twice the price, but you can drink a lot of it. When I was getting to know my cheap espresso machine, I went with decaf. I ended up making six double shots in a day, and that amount of regular espresso would have been way too much for me. With decaf though, I was fine. If I had used regular coffee, I would have needed to spread these experiments over several days, but with decaf I was able to do all of it during a single day.
If you have a list of commands to type in the terminal, but you need them only once, you aren’t going to make a bash script for that. If you think you’ll need to type that thing every week, making a script suddenly begins to make sense.
Same thing with agents. Do you think you’ll be asking these kinds of questions again in the future? If so, making an agent for it could make sense.
Yeah, it sounds like it was a disaster waiting to happen.
Oh and the man pages of terminal commands. Usually it feels like those were written for the people who already know all about the command, but simply forgot which flag does what. Like, was that -n or -t again? Oh, let’s check the man pages. Got it, it’s -n, so let’s go with that.
What about those who have never used that command before? I feel like the man pages are the wrong place for people like that. You could spend 20 minutes reading and get very little value out of it. Instead, you could spend 2 minutes reading a tutorial blog, adapt the command to your use and get on with your life.
Alternatively, you could ask an LLM to generate a command that does what you want and… hope for the best. Hallucinations do exist. However, I’ve also discovered some awesome commands this way. In simple cases, it’s surprisingly fast. In complex cases, you’ll find yourself on a wild goose chase again.
Well, that was just a quick summary of an idea. You can’t just tell someone (or a an LLM) to think critically and expect it to work magically. However, you can follow certain guidelines to take a few steps in the right direction. There is a method to it you know. It starts with simple things like investigating whether or not a claim is supported by the evidence, are there unstated assumptions and so on. If you follow rules like this, you can avoid some of the simplest reasoning mistakes. Putting all of that into a prompt takes some time and effort. Doing it properly will probably result in a 500 line long manual of critical thinking.
I’ve done something similar with programming, and the LLM usually follows most of my style guide instructions reasonably well. It’s not perfect though, and it will deviate from some rules no matter how hard you try. Also, it requires constant hand holding, because it’s an LLM. Anyway, I don’t expect a critical thinking agent to be a fool-proof solution. As long as it can avoid some of the stupidest reasoning mistakes, it could be a bit more useful than the default version, and might actually be helpful in troubleshooting.
Oh and the high-impact language they use feels really jarring and out of place. Everything is just so intense with these LLMs.
I have made a few agents in Mistral, and I think I also added a line about asking clarifying questions. Should probably make another one just for troubleshooting. There should be many lines about critical thinking, ruling out the obvious, avoiding assumptions, and asking questions.
LOL. Just throw more money at the problem until it goes away. Sure… that worked great last time.
It’s been a mixed bag for me. The case has to be a relatively simple one for it to work. LLMs need a lot of hand holding to be useful.
If it’s connected, it can be hacked. All you need is an Internet connection and some skill.
The chances are very high. People always have technical problems, and they are also impatient enough to use an LLM. Why spend hours reading stackoverflow, and try solutions that aren’t exactly what you’re currently facing? That takes time, effort and improvisation skills.
Meanwhile, you could type a question to an LLM, and get an answer in a few seconds. In the best case scenario, you’ll get something useful out of it, but you could also start another wild goose chase. Humans are lazy, so they’ll fall into this trap very easily. It’s a gamble.
It’s usually some random thought that is in the right neighborhood, but not quite spot on. A human troubleshooter would straight up say that it’s impossible to tell you what the problem is, so you would need to narrow it down by testing a few things.
An LLM just says that you need to update drivers or whatever. If the problem is caused by something obscure (like this one), an LLM will never be able to figure it out. This kind of stuff apparently just doesn’t exist in the training data, so there’s no way for the model to extrapolate and reach the right conclusion. Instead, it will continuously interpolate with the data it has, and you’ll end up with an infinite list of wrong answers.
Sycophancy really doesn’t help either. If you have any ideas what might cause the problem, the LLM will cling to those, no matter how wrong you might be. Troubleshooting requires critical thinking and LLMs don’t seem to be very good at that.
Yeah, that’s a pretty safe approach. You never know which parts are crude averages of the training data and which parts got interpolated.
If you happen have subject matter expertise, you can tell where it all went wrong. On the other hand, at that point you already know what the right answer is so why bother asking an LLM in the first place. It’s a very paradoxical situation.
You need to brew very gently. Moka pots are finicky and they don’t forgive your mistakes.
If you’re ok with any random coffee, it’s really easy, and that’s why they are so popular. If you want to make good coffee, you really need to know what you’re doing.
To operate the hardware. We’ve had firmware as long as we’ve had electronics.
The real question is: Why does a dishwasher need to be connected to the internet? That just opens it up to all sorts of online threats and creates more problems than it solves.
It’s 150-500 °C. If your factory happens to have that kind of waste heat, this could be very interesting. On the other hand, anything that hot can also be used for a variety of different things. Your factory is probably already dumping all that heat into the district heating network, so why would you bother with any of this?
It’s a breakthrough when you consider what the older methods required. When you think about industrial waste heat in general, this development probably won’t be of much interest, because normal waste heat is nowhere near hot enough.