True! Though it applies to all the AI sites.
At least this can be run locally.
Yep, it sure is.
Running it locally doesn’t risk anything though; the 7b and 14b models are both quite good; I’ve been running the 14b model on my M3 pro macbook at work, and it’s pretty decent; about 12 tokens/s. Not bad for a laptop without serious hardware.
But on a serious note I’m excited to see what is going to happen with AMD’s new mobile chips (Ryzen AI Max?). They have a similar memory setup to Apple’s with 256bit LPDDR5x-8000 (up to 128GB). They could be a ‘dark horse’ for an affordable AI workstation; i.e. a way to get more than 32GB fast memory under hopefully $ 5000. If a minisforum with this chip and 64/128GB of 256GB/s memory is affordable it could sell like hot cakes to LLM amateurs.
No no, it really isn’t serious hardware.
Remember, Apple is babbys first developer machine.
Only reason I have it is because the alternative (this is a trend apparently) was a $500 Lenovo shitbook running windows 11
lol clearly comparable
Semi Accurate thinks this may have burst the bubble, but in a different way.
https://semiaccurate.com/2025/01/28/did-deepseek-or-microsoft-pop-the-ai-bubble/
They have a good history of performance and with this you should be able to train a model with chain of though RL
https://unsloth.ai/blog/r1-reasoning
While its not DeepSeek, seems like Mistral was inspired by their app and apparently it seems like it can scrapes web content this forum??!
I have barely tried it, but it seems to pull data from web sources. Seems free (as in free beer) at least for now:
Example output
Amusingly Le Chat
refuses to give me the most read stories from theguardian.com, but phind gave me some info and reference links.
I wonder if these sites respect robots.txt? haha…
Oh I guess there is a web version of the DeepSeek Chat - Online & Free phone app, but this free version seems unable to do web scraping.
Is search dead? Are javascript pop-up ads dead? RIP haha…
Is this new? I thought Mistral could pull in web queries for a while now.
Right, no it is not!
I didn’t quite realize it though until last night, as I’ve been under a rock too focused on running local LLMs and hacking on my on toys instead of following the proliferation of ai chat offerings. Most of which now indeed seem to pull in web queries.
My own “agentic workflow” (lol) experiments are basically just some python scripts running:
- duckduckgo search (
duckduckgo-search
) - reranker model (
sentence-transformers w/ mixedbread-ai/mxbai-rerank-*
) - concurrent web scraper (
primp
) - summarization (
litellm
+llama.cpp
)
To steer this back OT, I wonder if the big impact of DeepSeek is not the model itself, but getting many more folks exposed to this kind of “ai chat app”?
My friend showed me her $20/mo OpenAI chats which seem similar, but I bet many new folks tried DeepSeek given no paywall…
I think somethings lined up to set the tone, so to say.
The banning TikTok only seemed to push people to a new Chinese platform… rather than existing players.
Deepseek a Chinese technology, hitting the news as being better and cheaper, then the mobile app hits the top of the store charts.
This is somewhat on-topic and off-topic, but I found this overview by Maarten Grootendorst a good overview, and thought some of you might enjoy it, too. He explains what reasoning LLMs are about, what makes them different from other large language models etc. Maarten works in the AI/LLM field and Open Source initiatives, but he came into this field from his initial work in psychology. Which might have helped him hone his skills in explaining KI and the underlying models. Here the link: A Visual Guide to Reasoning LLMs - by Maarten Grootendorst
they are saying what I said. which undersells what they did a bit, they did do awesome, but nearly every mainstream article got it wrong from the start. we got it right-er
the above article in context of my reply here
The reaction (drop in share prices) to the announcement from DeepSeek wasn’t because their work was so groundbreaking; it was more like adding a crystallization seed to an oversaturated salt solution - one little disturbance, and things start happening.
Sometimes the current situation with AI and the stock market reminds me of the dot com bubble in the late 1990s. Cisco systems was valued at way over 200 times annual earnings, and at least Cisco actually had earnings. But, just like with the Dot Coms then, some of these new players in AI will indeed change how we do things. Except, AI has the potential to be a lot more disruptive.