Posts tagged: llm

I’ve been using gitingest web ui [[ thoughts-516 ]] for quite awhile to serialize git repo into llm friendly text files. This gives tools context about repos that are not in the training data so that it knows about it and how to use the code in the repo. gitingest also has a python library [[ thoughts-517 ]]

I had a use case for a project not yet on git, and found yek.

Their instructions tell you to curl to bash.

curl -fsSL https://bodo.run/yek.sh | bash

I don’t like curl to bash from random sites, so I have my own self...

Do you remember regression models from college: given some data, you find a best fit line that allows you to predict Y given X. At the end of the day, ChatGPT, and LLMs in general, are the same thing as the regression model – it’s just that ChatGPT is the largest and fanciest model we currently have to model language and information.

I really am coming to the idea of calling it a “word calculator”, this seems to be the most succinct description of llms that the lay person can comprehend and relate to.

ChatGPT does not hallucinate or become unhinged

I think Steve goes much deeper on this in his intervew on fafo.fm. They describe it more as a pleaser or “yes man” essentially all the companies that are building these models want to give the “best” answer, better than their competitors. With this comes the risk of it being completely wrong, they are designed to always give an answer.

...

Under 2000 everything is happy, green field. Any decision you have made is relatively easy to back out of (barring you making a library with downstream users), but as you go, regret kicks in. Regret we didn’t make that pydantic 2 upgrade earlier, as new features become more apealing. Regret that we chose sqlite for simplicity, speed, agility, and now we might need robust and distributed. Regret that you chose a front end framework, or to have a front end at all to a backend problem. Regret that you put 6 layers of abstraction on your db early on and now that you understand the problem you want different abstractions, but all of your endpoints deeply depend on the current one.

Vibe coding will not save you, it will only make these wrong decisions for you without the context that you have. You will hate it’s decisions more because you had no input into some of them.

Steve is such a great listen, the neurospicy 🌶️ rambles this episode goes on is so relatable. I feel like I really missed out on some great takes on intellij vs neovim, but got some really great knowledge about vector db’s, embedding, text compression, similarities to vector algegra like infinite craft.

Just popped open infinitecraft and I’ve definitely played this with my kids before, super fun, just could not remember the name of this one. I do remember an android one as well that is alchemist or something like that, which we have also played a lot.

Interesting how confidently he says we can easily go to the top. really makes you wonder what we the normies are leaving on the table by using these general purpose models and what could be achieved with really tuned in models. Could I make an automatic blog tagger more accurately, maybe smaller, maybe tuned so well it runs fine on cpu?

jina reader is a pretty sweet tool to convert a site to ai compatible text. There are other web to markdown types of tools, but the convenience of just adding r.jina.ai to the front of any page makes it so easy to grab for one page of docs.

the racked up 4 framework mainboards sound wild. connected with usb4 and 5gig ethernet. they said they can run big models quantized down from 600Gb to within the 512GB limit they have. This seems wild to bring this level of capability to such a low price point. It will be really cool to start to see demos come out.

OpenTools is an index for mcp servers to work with new agentic workflows like roo code and windsurf.

Damn prime makes an interesting point near then end of this video. He’s seen a bunch of people able to just throw down charts and shit at their company and end up being “the coding guy” cause they proompted something once. In a way I can relate, I got into software in a similar way, but at a time that it took a lot more hard work, understanding , and copy past from the right stack overflow. Based on some of the people around me at the time I can only imagine how some people must feel like they got pushed into it without wanting it, and now are building something they don’t know anything about with no care about it or care to build any expertise. Is the future proompted charts from enterprise chatgpt or do we only continue growing more need for software from here.

There is a glimmer of hope out there that normal people can scrap together enough gpu to really run the latest models themselves. The ui really appears to be having huge leaps forward such that doing things like rag is no longer such a research project that it was just a few years ago. So excited to see Prime go through this homelab exercise.

Damn these deepseek memes go hard. Wild to see openai get played by their own game.

It’s crazy that the normie news that I have seen on deepseek shows that the Chinese made what the Americans did at a fraction of the price, without taking notice that they are building on the shoulders of openai.

👏👏👏 This one is really good. I’m right there with him on most of this. I am very hesitant on subscription models, and all the ai tools feel like they are getting ready to be the next round of death by a thousand cuts, this time with pretty limited free tier and relatively high prices to run. I’m sure we will see companies get taken by huge bills soon by building off of someone else’s service.

On the flip side I’m definitely the guy that gets in a rut of just copy paste to the ai, wait for codeium to to inject. I feel like I have issues of momentum more than anything. When I’m on one side or the other I tend to stick it out for too long, but less so on going without because that llm drug is calling you when you hit a hard problem.

I’m excited to see him build out a homelab for llm stuff that he mentioned at the top. I’m interested, but probably not building one out for myself until we start to see some cheaper maybe used hardware to do it.

Don’t stop learning! Stop trying because you have a doomer outlook on ai, llms, industry and think they are taking over. If you have no hope for the future, if you stop now you are cementing in that you will be no good and the ai will be better. Many, maybe most of us in this industry go here by hard work, long nights of learning, trying to solve problems that our job had. If llms take over then the world is going to be a whole lot different, it will be a world you cannot predict or plan for. For now put your head down and succeed in the world we have today.

TEEJ has some great thoughts on this whole sentiment, put this on for you morning walk or whatever you do.

interesting UI for RAG based workflows, i.e. chatting with your documents. It looks like it can run a number of models, feels like ollama with RAG and a nice web ui.

Looks like a great start to a rules file for fastapi.

I have never heard anyone say this. It feels weird to me. The other early return, find and handle errors early all make sense to me, but happy path last is new to me.

Really interesting way to generate a rules file for agentic workflows based on your current repo. John uses gitingest here, looks like a fantastic tool, but probably not useful for most private repos. I’m sure you can replicate the same thing in a private repo wtih a small amount of effort the few times you need to do it. gitingest looks like a great way to pull in some extra context for some open source dependencies that you have though.

Gitingest has a python package on pypi that you can run with uvx, and it accepts the same arguments as the web version, right in your terminal

Replace hub with ingest in any github and get a prompt friendly codebase ready to feed into any llm. It combines the entire codebase, based on a gitignore style glob that you pass in, into a single TXT file.

This is so easy compared to self hosting stable diffusion yourself. It even has a nice api that you can hit with curl or python. The pricing seems competitive as well. Bookmarking this to try next time I need something like it.

I recently updated ollama, and it now installs a systemd service that I was not expecting. Seems like a great option, but I hadn’t expeted this and I was able to kill it previously. It was using up gpu, and I do other things on my machine with a gpu. I tried pkill, kill, and everything, it was still coming back.

No matter what it comes back

# stop it systemctl stop ollama.service # disable it if you want systemctl disable ollama.service # confirm its status systemctl status ollama.service

You can confirm this with the following command.

# checking running processes ps aux | grep ollama pgrep ollama # checking gpu processes gpustat --show-cmd --show-pid

Next time you want to start you can do it as before with ollama serve.