Profile avatar
campbell.fi
Data enthusiast, Father, consultant
41 posts 98 followers 1,079 following
Prolific Poster
Conversation Starter

I posted my guide to running LLMs on your local computer starting from scratch. This covers uv, llm, and llm-mlx. Many thanks to @simonwillison.net for the brilliant LLM tools and shared wisdom.

My LLM command-line tool and Python library now has support for tool calling! You can define tools as Python functions or bundle them in plugins, and LLM can then make them available to models. OpenAI, Anthropic, Gemini and Ollama are supported so far. simonwillison.net/2025/May/27/...

Three things made the US a rich and powerful nation: the rule of law, its science & innovation system, and openness to foreign talent. Remarkable how Trump has taken a sledgehammer to all three. No enemy of this country could do more.

I realized something recently, which is that we're building AI tools completely backwards. While I was aware that there's some de-skilling happening already, I finally clicked a bunch of pieces together and decided to write this to talk about how we can do *better* hazelweakly.me/blog/stop-bu...

Apologies in advance for what might look like a rambling diatribe going nowhere. But the one thing I know really well - better than almost anyone else - is boxing. And there’s a really valuable lesson from boxing that the Democratic Party needs to learn in a hurry. Let’s start with the boxing bit.

Continuous Thought Machines | Discussion

llama.cpp has vision language model support now! ❤️‍🔥 get started with sota VLMs (gemma 3, Qwen2.5VL, InternVL3 & more) and serve them wherever you want 🤩 learn more github.com/ggml-org/lla... 📖

llama.cpp shipped new support for vision models this morning, including macOS binaries (albeit quarantined so you have to take extra steps to run them) that let you run vision models in a terminal or as a localhost web UI My notes on how to get it running on a Mac […]

🚨New Preprint! Did you know that steering vectors from one LM can be transferred and re-used in another LM? We argue this is because token embeddings across LMs share many “global” and “local” geometric similarities!

Have we talked about this paper yet on bsky? It's really good, and ought to be cited widely by people working at the humanities-AI interface. TLDR: If AI is going to be applied to fuzzy human problems, social science measurement theory becomes essential. #MLSky

Jimin Kang shares an overview of the MapReduce programming model and how it can be used to optimize large-scale data processing.

ColBERT (a.k.a. multi-vector, late-interaction) models are extremely strong search models, often outperforming dense embedding models. And @lightonai.bsky.social just released a new state-of-the-art one: GTE-ModernColBERT-v1! Details in 🧵

Yesterday I shared how to run fast ModernBERT inference with vLLM. Today I’m running the same pipeline serverlessly on @hf.co Jobs — using a single script powered by uv. Inline metadata handles all dependencies. Just hfjobs run + uv run. That’s it!

Lisbon restaurant/bar/activity recommendations, go!

Run Gemma 3 27B on your desktop GPU 🔥 Our new QAT-optimized int4 models slash VRAM needs (54GB -> 14.1GB) while maintaining quality. Now accessible on consumer cards like the NVIDIA RTX 3090 via ollama, hugging face, lmstudio, kaggle and llama.cpp developers.googleblog.com/en/gemma-3-q...

🚀 Introducing **Bauplan** A serverless, code-native platform for building data and AI pipelines — directly on your object store. No clusters. No notebooks. No GUI based workflows. Just Python + SQL + S3. 👉 www.bauplanlabs.com/blog/hello-b...

Harvard study shows we can measure leadership skills by seeing how folks manage GPT-4o simulated people AI assessments strongly correlate (r=0.81) with human assessments. Effective leaders ask questions & do conversational turn-taking and have fluid & social intelligence www.nber.org/papers/w33662

Show HN: HNSW index for vector embeddings in approx 500 LOC Discussion

Can't wait for when I can vibe code a production recommender system. Until then, here's some system designs: • Retrieval vs. Ranking: eugeneyan.com/writing/syst... • Real-time retrieval: eugeneyan.com/writing/real... • Personalization: eugeneyan.com/writing/patt...

I created a small DuckDB extension to choose file interactively using native file open dialogs. Now you can install this via Community Extensions repository! github.com/yutannihilat...

Kids, I don't care how you feel about the Palestine protests. Targeting pro-Palestine activism is just the sharp end of the wedge. They aren't going to stop. Not with Palestine. Not with students. Not with immigrants. www.washingtonpost.com/immigration/...

I feel like I'm having a stroke docs.mcp.run/blog/2025/03...

Datasets like this could have a lot of value for people who want to train their own models.

Oh, it's been a while since I posted about this, but there's an honest to goodness WEBRING for anyone who has a site that contains data-related content, whether a blog, project site, list of bookmarks, whatever. just submit a PR. details here: #dataBS github.com/randyau/data...

i tore this apart this morning, the gist: - yes, it separates knowledge from reasoning 🎉 - it substitutes MHA computational complexity for knowledge graph schema design 🤔 i’m partly ecstatic, this is huge, but also disappointed bc KG design is largely unsolved

🚀 Querying #MicrosoftFabric DWH Data from Your Laptop with #DuckDB UI 🦆💻 I am not saying it is a good idea to do so, but you can if you want. www.youtube.com/watch?v=q8Dp...

Thread: I've been generally bearish on Europe for a long time - aging population, high taxes, high regulation, failed energy policy, resource dependence on Russia, US & China does not make for a dynamic economy. But for the first time in a long time, I am seeing reasons to be optimistic:

I’ve found this reporting to be indispensable recently (*glares at CR*), so I just signed up as a new member to support their work. If you can do that too, please consider it.

Life improved immensely once I learned to stop telling Cursor "it still doesn't work" and instead tell it to pause, don't write any code, investigate the issue properly, come up with 3-5 theories for what's going wrong, and validate theories with logs. I also give it web + docs access, if relevant.

Probly: Spreadsheets and Python and AI, right in the browser Discussion

dlt+ Cache is an example of offline processing that I think is going to become a data engineering + analytics standard. @mmullins.coginiti.co has shown similar ideas for Coginiti on processing Iceberg data and caching it locally.