The smaller models have been creeping upward. They don't make headlines because they aren't leapfrogging the mainline models from the big companies, but they are all very capable.
I loaded up a random 12B model on ollama the other day and couldn't believe how good it competent it seemed and how fast it was given the machine I was on. A year or so ago, that would have not been the case.
yeah especially that this simplifies e.g. doing mobile app for 3rd party developers - not extra cost, no need to setup proxy server, monitoring usage to detect abuse, don't need to make complicated subscription plan per usage.
We just need Google or Apple to provide their own equivalent of both: Ollama and OpenRouter so user either use inference for free with local models or BringYourOwnKey and pay themself for tokens/electricity bill. We then just charge smaller fee for renting or buying our cars.
Not just local models but bespoke apps. The number of bespoke apps I've created shot up dramatically in the last 6 months. I use one to do my recipes/meal plan every week. I have one that goes through all my email addresses and summarizes everything daily. I just finished an intelligent planner / scheduler for my irrigation system that takes into account weather forecast and soil moisture levels. If something is annoying and there is no commercial solution or open-source solution that has the features I want I just make it now and it's fantastic.
I've had friends/family ask to use some of them; I declined. I don't want to do support / feature requests.
As someone who hasn't used AI for "real" app development (mainly just getting ChatGPT to generate small functions & scripts), do you have any recommendations on what tools or resources I should use to get started with this?
Cursor/Cline/Windsurf are my recommendations for clients. For models stay away from Sonnet 3.7. I find it just lies to you. I'd rather you a slightly less capable model like Sonnet 3.5 where I know it will just make mistakes that won't compile.
I do my planning with a combination of Grok3, and higher power OpenAI models. Once I have plan of what I want to build, I create an implemenation_plan.md with all the steps to build my solution. (Generated by the higher power models) I carefully review this plan and if it looks good, I throw it into agent mode and get to work.
Including figuring out which more expensive models to use when needed instead of doing that by default. Early LLMs were not great at reasoning and not great at using tools. And also not great at reproducing knowledge. Small models are too small to reliably reproduce knowledge but when trained properly they are decent enough for simple reasoning tasks. Like deciding whether to use a smarter/slower/more expensive model.
I spend a lot of time working with smaller models, I often had to split the problem into smaller subtasks to make it give acceptable accuracy. With the big models in the cloud you can often get things working much faster, it seems like a tradeoff in engineering time. What was your experience?
I figure this will take the same shape as package distribution. If you have ever used a linux distribution you’ll always see a couple .edu domains serving you packages. Big tech might be able to have specialized models, but following the linux paradigm, it will likely have more cutting edge but temperamental models from university research
Most open source projects don't need the kinds of resources that ML development does. Access to huge GPU clusters is the obvious one, but it's easy to forget that the big players are also using huge amounts of soulcrushing human labor for data acquisition, cleaning, labeling and fine tuning, and begrudgingly paying for data they can't scrape. People coding in their free time won't get very far without that supporting infrastructure.
I think ML is more akin to open source hardware, in the sense that even when there are people with the relevent skills willing to donate their time for free, the cost of actually realizing their ideas is still so high that it's rarely feasible to keep up with commercial projects.
Easier said than done, training is usually done on "big iron" GPUs which are a cut above any hardware that consumers have lying around, and the clusters run on multi-hundred-gigabit networks. Even if you scaled it down to run on gaming cards, and gathered enough volunteers, the low bandwidth and high latency of the internet would still be a problem.
For the bigger open source projects, companies who use that code for making money. Such as Microsoft and Google and IBM (and many others) supporting Linux because they use it extensively. The same answer may end up applying to these models though - if they really become something that gets integrated into products and internal workflows, there will be a market for companies to collaborate on maintaining a good implementation rather than competing needlessly.
Last time I did that I was also impressed, for a start.
Problem was that of a top ten book recommendations only the first 3 existed and the rest was a casually blended hallucination delivered in perfect English without skipping a beat.
"You like magic? Try reading the Harlew Porthouse series by JRR Marrow, following the orphan magicians adventures in Hogwesteros"
And the further towards the context limit it goes the deeper this descent into creative derivative madness it goes.
An interesting development to look forward to will be hooking them up to search engines. The proprietary models already do this, and the open equivalents are not far behind; the recent Qwen models are not as great at knowledge, but are some of the best at agentic functionality. Exciting times ahead!
Many tasks that one might want to give a model end up implicitly including search as a subtask. For example, "plan me a trip to Santiago" obviously requires the model to understand details about the real city of Santiago. Less obviously, "write me a Python script to do ..." requires they understand APIs, libraries, etc., the same things you might ask a search engine to pull up. The tasks which do not require a coherent + mostly-correct exterior-world-model are relatively few -- text processing (e.g. "proofread this") is a big one; calculation tasks fit, but LLMs are also bad at those.
Exactly, I think all those base models should be weeded out from this nonsense, kardashian-like labyrinths of knowledge complexities that just makes them dumber by taking space and compute time. If you can google out some nonsense news, it should stay there in search engines for retrieval. Models should be good at using search tools, not at trying to replicate their results. They should start from logic, math, programming, physics and so on, similar to how education system is suppose to equip you with. IMHO small models can give this speed advantage (faster to experiment ie. with parallel diverging results, ability to munch through more data etc). Stripped to this bare minimum they can likely be much smaller with impressive results, tunable, allow for huge context etc.
I really like Gemma 3. Some quantized version of the 27B will be good enough for a lot of things. You can also take some abliterated version[0] with zero (like zero zero) guardrails and make it write you a very interesting crime story without having to deal with the infamous "sorry but I'm a friendly and safe model and cannot do that and also think about the children" response.
Qwen3 and some of the smaller gemma's are pretty good and fast. I have a gist with my benchmark #'s here on my m4 pro max (with a whole ton of ram, but most small models will fit on a well spec'ed dev mac.)
>> These benchmark numbers cannot be real for a 7b model
> LLM benchmarks are mostly bullshit right now. Wait a few years until the hype cycle returns to sanity.
This could mean a lot of things. Can you be a bit more specific? It's one thing to say benchmarks are gamed. Another to say models end up being trained on the benchmark indirectly. Another to say they the particular experimental setup during the benchmark is unclear. Another to say mapping a benchmark to a real use case is hard. Are you saying some/all of these claims?
Have you plotted MiMo versus others? Another comment suggests smaller models are performing better than expected. Any comment on that?
