Sweet! Mistral has been my ollama go-to for a while.
8
invalidusernamelol [he/him] - 2w
Honest question for the open model folks here, what's the upshot? Like when I'm programming I much prefer to go line by line and make sure everything is exactly how it should be.
I'm guessing most people are the same, but I still haven't really seen how a model is any better than spending an extra 30 minutes reading documentation or going through PR discussions.
5
gay_king_prince_charles [she/her, he/him] - 1w
A 14b model isn't worth it for programming or anything most people use an LLM for. What lightweight open models are really good at is transforming unstructured data (such as natural language input) into structured data. It doesn't do this perfectly, but it's a difficult problem to solve conventionally. For example, Mistral 3 could go from an arbitrary and unknown log format to a specific one around 80% of the time, and I imagine this would do it much more effectively.
4
invalidusernamelol [he/him] - 1w
That's a decent use case I guess. Feels like you could also just learn awk or sed and get it done reproducibly way faster. Or even just a Python script that can handle breaks in the data.
4
gay_king_prince_charles [she/her, he/him] - 1w
Another usecase is sentiment analysis or needing to extract meaning or intent from arbitrary data. I didn't work on the case above, and it was a while ago (and I don't think I can say anything more than what I did), but there was an existing traditional system that didn't meet requirements.
3
invalidusernamelol [he/him] - 1w
I used to use NLTK back in highschool and college for sentiment analysis and it was usually decently accurate at scale (like checking average sentiment of a tag) and ran surprisingly fast even on an old laptop. Are the open models as performant as NLTK?
2
gay_king_prince_charles [she/her, he/him] - 1w
I can't say. I haven't used NLTK, but I do know that even a 14b model is pretty expensive to run (you more or less need a TPU or GPU) and it isn't fast either. The 3b models can be run on almost anything other than embedded systems, but I haven't used any in years.
2
invalidusernamelol [he/him] - 1w
NLTK just does Chomsky diagrams and tokenizes text based on parts of speech. It's mostly a bunch of hash tables and optimized algorithms with a simple pre trained machine learning model (VADER) that can do rudimentary sentiment analysis.
I can see how just jamming text in a pipeline is a simpler solution though, since you need to build the extraction model by hand using NLTK.
2
JoeByeThen [he/him, they/them] - 1w
Rubber duck troubleshooting that talks back.
Completely destroys writer/coder block from having a blank page with an added bonus that when I'm writing out my prompts I get a better understanding of what it is I'm trying to build; The act of writing out exactly what I'm looking for better internalizes the design I'm looking for.
Lowers the barrier of entry on advanced techniques. I don't need to be an expert to experiment with something like writing compute shaders or whatever that's outside my wheelhouse.
Also, because I've been self taught since the 80's I tend to learn by troubleshooting, so figuring out what's not working from a language model is no different than me trying to parse why the answers from stack overflow aren't working.
Note that a local model doesn't always do the trick so I do have to fallback to chatgpt and deepseek from time to time, but they're getting much better. I do have mixed feelings about if l would recommend this for someone starting out who doesn't know what they're doing already... But I grew up relying on Google, which tbh is now crap for what I used to do, so GPTs are becoming the only option for those of us of many hats.
Oh, and I don't use that line by line copilot stuff. I prompt out exactly what I'm looking for and have the model spit out a class or two all at once.
3
invalidusernamelol [he/him] - 1w
That's a fair use case. I just didn't have the patience for it lol. I've always been better at just failing repeatedly until I succeed, which is basically what GPTs do, but instead of me getting the benefit of that process, they get it and then immediately forget.
Might try the rubber ducking thing at some point, but most of my code is in a field where there's not really many good examples and the examples that do exist tend to be awful, so it's pure hallucination. I've seen some stuff colleagues have vibe coded and it gives me the ick/bad code smells.
1
Blakey [he/him] - 2w
Great, more efforts to cook the planet
hahaha wheeeeeee
3
gay_king_prince_charles [she/her, he/him] - 1w
I'm more excited about how performant the 14b model is. Development on small local models has stagnated for the past few months after China had moved away from lightweight models in attempts to compete with American SOTA models and after LLaMa 4 was terribly underwhelming. This is going to make a lot of mundane, "hidden" NLP tasks much easier.
yogthos in technology
Introducing Mistral 3 open model
https://mistral.ai/news/mistral-3Sweet! Mistral has been my ollama go-to for a while.
Honest question for the open model folks here, what's the upshot? Like when I'm programming I much prefer to go line by line and make sure everything is exactly how it should be.
I'm guessing most people are the same, but I still haven't really seen how a model is any better than spending an extra 30 minutes reading documentation or going through PR discussions.
A 14b model isn't worth it for programming or anything most people use an LLM for. What lightweight open models are really good at is transforming unstructured data (such as natural language input) into structured data. It doesn't do this perfectly, but it's a difficult problem to solve conventionally. For example, Mistral 3 could go from an arbitrary and unknown log format to a specific one around 80% of the time, and I imagine this would do it much more effectively.
That's a decent use case I guess. Feels like you could also just learn awk or sed and get it done reproducibly way faster. Or even just a Python script that can handle breaks in the data.
Another usecase is sentiment analysis or needing to extract meaning or intent from arbitrary data. I didn't work on the case above, and it was a while ago (and I don't think I can say anything more than what I did), but there was an existing traditional system that didn't meet requirements.
I used to use NLTK back in highschool and college for sentiment analysis and it was usually decently accurate at scale (like checking average sentiment of a tag) and ran surprisingly fast even on an old laptop. Are the open models as performant as NLTK?
I can't say. I haven't used NLTK, but I do know that even a 14b model is pretty expensive to run (you more or less need a TPU or GPU) and it isn't fast either. The 3b models can be run on almost anything other than embedded systems, but I haven't used any in years.
NLTK just does Chomsky diagrams and tokenizes text based on parts of speech. It's mostly a bunch of hash tables and optimized algorithms with a simple pre trained machine learning model (VADER) that can do rudimentary sentiment analysis.
I can see how just jamming text in a pipeline is a simpler solution though, since you need to build the extraction model by hand using NLTK.
Also, because I've been self taught since the 80's I tend to learn by troubleshooting, so figuring out what's not working from a language model is no different than me trying to parse why the answers from stack overflow aren't working.
Note that a local model doesn't always do the trick so I do have to fallback to chatgpt and deepseek from time to time, but they're getting much better. I do have mixed feelings about if l would recommend this for someone starting out who doesn't know what they're doing already... But I grew up relying on Google, which tbh is now crap for what I used to do, so GPTs are becoming the only option for those of us of many hats.
Oh, and I don't use that line by line copilot stuff. I prompt out exactly what I'm looking for and have the model spit out a class or two all at once.
That's a fair use case. I just didn't have the patience for it lol. I've always been better at just failing repeatedly until I succeed, which is basically what GPTs do, but instead of me getting the benefit of that process, they get it and then immediately forget.
Might try the rubber ducking thing at some point, but most of my code is in a field where there's not really many good examples and the examples that do exist tend to be awful, so it's pure hallucination. I've seen some stuff colleagues have vibe coded and it gives me the ick/bad code smells.
Great, more efforts to cook the planet
hahaha wheeeeeee
I'm more excited about how performant the 14b model is. Development on small local models has stagnated for the past few months after China had moved away from lightweight models in attempts to compete with American SOTA models and after LLaMa 4 was terribly underwhelming. This is going to make a lot of mundane, "hidden" NLP tasks much easier.