It's still in development. It often says the opposite from what is expected. Try Retoor1b chatbot at https://llm.molodetz.nl

This was result after building bot + chat website from scratch including training with embeddings. Design is generated by GPT, I tried my own but all ugly.

It's quite cool huh? Ask it to write some code for you. It's absolutely terrible. If it's down, try again in 5 minutes. I'm still working on it.

What's the result? I finally have a toolkit to make good/serious bots. Code could be bit better, but that's for other day.

Stack: self written webserver (and yes, you can post a gb to it or ddos it. Not sure if it survives the first one. I should limit requests to one mb anyway. Http headers may officially not be more than 4096 in total) since I know http protocol from my head anyway. Python websockets module. Asyncio, chromadb.

It could have xss issues. Don't care.

Let me know what you think

it's down already!

@jestdotty works fine here

@jestdotty, this is what I see:

@jestdotty dns is good. https://check-host.net/check-dns/...

What a great site btw, you can directly use an unix shell for debugging. curl https://llm.molodetz.nl doesn't work. But I've screenshot from @feuerherz that it works from Germany.

I did recently some name server changes. That was yesterday. Maybe smth is still a bit fucked. Will be fixed automatically

@jestdotty I do not think Germans are aware of that. I do not expect from Germany to restrict access to public information. @Germans: is this true?

Fuck, my rant is not clear. This is the result of a day work. All together.

@jestdotty: ask the bot "What says jestdotty about ostream?"

Damn! What model is it running? Is it your own?

@12bitfloat customized version of gwen2.5b0.5. I enriched data with embeddings. The retoor personality is programmed in the model system. It runs on a 4Gb Dual Core VPS. I could probably up to qwen2.5b1.5 but there's more i want to do on that vps. I also run the web version of vscode there for example. What a great product is that. No difference with real vscode. Even codeium runs in the web version! I'm now trying to run qwen2.5-coder:0.5b locally in vscode. Let's see if it's much difference with codeium

@jestdotty an embedding is just a piece of data that you can store in a vector database. I mean that i converted all stuff about dr and stuff to vector thingies and it knows what you're talking about superfast while having quite a big dataset now

@12bitfloat did you try it?

@retoor It unfortunately didn't load for me, it just times out :(

@jestdotty, it's weird that my bot has this in memory: "jestdotty has said on devrant: ostream is just confused and throws his confusion on others. he wants validation / help for his confusion I think." if you ask "Did ostream delete himself?" and it's answer is: "No, ostream did not delete himself. He continues to exist outside of the context provided in the prompt."

What the fuck does that answer have to do with what it had in memory?

Embedding works like this: first you insert a bunch of data. Later you can query them. The prompt used for the model, is the same as you use for the embedding. The embedding data as result is the context for the model.

Yes, the CSS is done by ChatGPT

@12bitfloat so weird, it works superfast here (on wifi and phone 5g) and my friend from Germany can access it as well. Probably we have to sleep a night over it. Still aftermath of the nameserver update i think. I can't run it on other domain either, I updated nameservers for all. I wanted to add cloudflare but than i realized that i don't want a man in the middle. Not scared for ddos anyway. I also have ssh open with password auth. Certificates are just extra for me. fail2ban and firewall active tho. Brute forcing over internet is not doable. Recently checked it. It takes 212321291 combinations to crack a simple word like 'retoor'. No one will do that over web

@retoor Oh that might be it. I use a specific dns provider, maybe the records haven't been flushed yet

UK I'm also getting timeout 😕

Damn retoori why you so mean ;(
Works here in Germany

@Chewbanacas I tried to remove such responses but with no luck. I already happy that it thinks it's a girl and accepted it's name. You can ask devrant related stuff. Who's kiki for example. I know you can't ask it about @Demolishun. One of his data imports went wrong. Ask it to write code for ya, it's humor

Smallest model works best! They're not so preprogrammed. Good for role playing. It's quite sharp. It's 0.5b with code support

Not bad. Btw, someone asked the LLM if i'm hot :P

It knows too much. It must be deleted.

I should post my voice so you can have bots that say malicious things, but with a real voice.

@Wisecrack it's not hard to make a bot speak. I wonder what the bot knows about you.

Currently the bot is terribly slow, it's learning 900 books. I have some asyncio sleeps applied in the book import so chat keeps working but slow. I had to convert all pdf's to text files first. Typically a python job

@retoor I have on the order of a couple hundred books. A respectable collection.

If I had a scanner I could OCR that shit.

If I had the time I could build the scanner.

Alas.

Well, the connection has timed out... Shame. I wanted to ask it if Retoor's hot. She reads hot. But so does Demolishun, and I bet he ugly!

@retoor HTTP timeout, even after putting IP in hosts. Tried yesterday, tried today, same result. :( But no worries, if it's online I'll get there, sooner or later. You tiny hottie, you!

@retoor @cprn oh, now it's really down for a sec. Is it possible that I'm using ipv6 and you ipv4? I can't ping with ipv4 to llm.molodetz.nl either. With ipv6 I can. I have server running again now, but I don't know for how long. My bot got personality issues at the moment. It takes the name mentioned in books. Like "My name is Bing Xu.". I have to see to prioritize data. It's very slow now during embedding of 900 books

@retoor No worries, I'll get to it when it's done reading, lol. And TBH with you, the only thing I know about IPv6 so far is how to disable it. So yeah, it's absolutely possible you're using IPv6 and I use IPv4. Shouldn't matter for the domain, though, should it?

@cprn I think ipv4 doesn't work because it doesn't ping. I'll look at config now. It does resolve the right ip tho. Maybe my hoster disabled the pinging or smth, or my setup. You do not really get a fresh vps install. It's some image from the hoster. At the cloud company I worked, people could just select the iso and they have guarantee nobody fucked with it

@retoor Just as an update: It still times out for me here in germany

times out still in canada

Now it's really down, will continue on it this weekend. I have new hardware. So I can train a whole new model again, the aquasition for the right model will take some time. Often you spend hours in it to find out it has an unfixable flaw. I have other ideas as a bot as well, something innovative. Top secret, so we'll probably know it next week already 😂 Today I made such beautiful stuff. It was python + vim day. I made a kinda proxy and weirdly enough the webservice is faster WITH my proxy in the middle than without. There's must be smth right about my buffer size / concurrency that the webservice likes. That service is written in ruby 3.3 IIRC. Slow shit imho. Soon, I'll publish my repositories. Projects do need some cleaning. 50% was written while high as a kite. Somethimes it benefits and smth it doesn't

@retoor I wanna ask what drugs and talk about effects of them on code but maybe I should not reel you into that

@jestdotty amphetamine is the best, it's like almost designed for coding. It doesn't make you crazy and gives you much focus but don't do it too much, you'll be busy the whole day programming issues or take shortcuts even. You should not do it more than three days a week for VERY benefiting effects on coding. For people having issues with such stuff, MANY (especially high educated) students take ritalin to study and try to let them diagnose for ADHD to get it. It's the same, but it shows how effective it is. Warning: it's good for learning things fast, not very good, your brain becomes very selective in what it will store for long term. But without it, you wouldn't have studied it at all and now you know the features of everything you want to learn. You don't have to need WHAT a good buffer size for certain stuff is, you only have to be aware that there is such thing. So, you learn at least a lot of subjects. Many things are learned by doing, doing is what you do on amphetamine

@jestdotty comparing green tea with amphetamine is like comparing a Brabus with a Kia. Only one is to die for

@retoor Word!

I started getting sick of monkey patching my python side projects, so I rewrote the whole thing into an ASP.NET equivalent. Before I always thought of Clean code being easy to read, nor have I touched .NET in like 5 years.

Problem is the chronic underestimation of my competencies is so stressful.

„That you don’t know how to debug the container logs that I know, but stop projecting your complexes into me fuckah „

@12bitfloat I figured out why nobody could connect. ALL ip addresses were wrong in DNS regarding V4. For some people (and for me) it connects over IPV6. That worked. You could try again. It has now a random model loaded. a small one

@Chewbanacas can understand that .NET feels a bit decent than Python because it's compiled. You have some sort of guarantee. I didn't do python for a long while and I notice that I make completely different type of bugs as I would in C. In C i go quite fast, but in Python too fast. So many mistakes because Python makes me lazy as f.

I love python, but if I would want to make smth really decent, I would go for smth like C# or Java too personally. Sadly, jobs are mostly Python.

Python is patching patching patching indeed. To easy to edit or smth

@retoor srsly? I started learning C# because apparently this is what everyone’s looking for here. Are these job boards pulling the same shit like dating apps as in inducing scarcity? Fuck them man

Yea it also felt so uncomfortable how you can’t properly encapsulate properties in Python. I hated my IDE suggesting me methods where it shouldn’t have access to at all.

Still I love this language, since it was my first. 🐍

@retoor It does work now!

@Chewbanacas python was for me one of the last. I was once member of a dutch developer community and there was only ONE doing python and it was a bit weird guy so I decided for the rest of my life not to do Python. But yeah, i ended up in outsource doing every mainstream language almost. Python is in my opinion best language ever written and I do even love the ecosystem. But problem is, it's kinda boring. A while ago i wrote a load balancer in python in ~10 minutes and never changed again. Just worked. I hosted a year long my hard drive on port 443 using such script btw. It redirected SSH to openssh and what it didn't recognize to https server. Two services on one port! Never had an issue with it. All made with asyncio. But python uses so freaking much processor always. Sick

@12bitfloat does it say smth usefull? I think you're talking to the batshit fast 136m model now

@retoor I'm not a big AI guy, even chatgpt rarely says something useful for me :D