Grant Williams

tl;dr - don't just use Dense Vector Retrieval! While this has been common advice since ~2023, it's great to have some solid benchmarking and proof points that dense vector retrieval alone ain't it! Whether you pair it with something like BM25, reranking, alternate search methods (grep/find), or add some spice to the queries, you're going to find success more generally by using a hybrid approach. So, once again, RAG dodges the bullet - and lives on another day. https://lnkd.in/gUQpW3ae

38

4 Comments

Building with AI agents is great, but let's be honest: sometimes they get "hallucination-happy" and ship code that doesn't actually work. If you’re a senior engineer or a small biz owner, you don’t just want AI to write code—you want it to verify that it works before it ever touches your repo. I just dug into 𝗭𝗲𝗿𝗼𝘀𝗵𝗼𝘁, and it’s essentially like hiring an autonomous engineering team that lives in your CLI. Instead of one agent trying to do everything (and failing), Zeroshot runs a multi-agent workflow: 𝗧𝗵𝗲 𝗖𝗼𝗻𝗱𝘂𝗰𝘁𝗼𝗿: Plans the task. 𝗧𝗵𝗲 𝗪𝗼𝗿𝗸𝗲𝗿: Implements the code in an isolated environment (Docker/Worktree). 𝗧𝗵𝗲 𝗩𝗮𝗹𝗶𝗱𝗮𝘁𝗼𝗿𝘀: Independent agents that test and verify the work. 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗶𝘀 𝗮 𝗴𝗮𝗺𝗲-𝗰𝗵𝗮𝗻𝗴𝗲𝗿 𝗳𝗼𝗿 𝘀𝗺𝗮𝗹𝗹 𝘁𝗲𝗮𝗺𝘀: 𝗕𝗹𝗶𝗻𝗱 𝗩𝗮𝗹𝗶𝗱𝗮𝘁𝗶𝗼𝗻: The agents checking the code aren't the ones who wrote it. They don't share context, so they can't "lie" or be biased toward the solution. 𝗔𝘂𝘁𝗼-𝗜𝘁𝗲𝗿𝗮𝘁𝗶𝗼𝗻: If a test fails, it doesn't just stop. It loops back, fixes the error, and tries again until it passes. 𝗦𝗵𝗶𝗽 𝗠𝗼𝗱𝗲: You can literally tell it to open a PR and auto-merge once approved. It’s built for the "correctness over speed" mindset. If you have a clear Jira or GitHub issue, you point Zeroshot at it, walk away, and come back to production-grade code. https://lnkd.in/dxWqgfkM #SoftwareEngineering #AI #SmallBusiness #Agents #BuildInPublic #DevOps

12

We built alloy, an open-source Rust CLI that merges LLM weights 45% faster than existing tools. As more teams deploy fine-tuned 1-3B parameter models, weight merging has become a common way to combine capabilities from multiple fine-tunes into one model without retraining. alloy handles the most common merge methods about 45% faster than existing Python-based tooling by skipping the runtime overhead entirely. 10 merge algorithms, ~4 MB binary, reads existing mergekit YAML configs natively. Had fun building this with Adam Abdalla and Reyab Saluja! Writeup and repo attached below. alloy.how https://lnkd.in/eJcUz8qX

69

20 Comments

This is an insightful observation. I hand inspect all code LLMs produce for me and guide the architecture, testing approach, etc.

3

Backstage interview: Yann Lechelle Ian explained how :probabl. turned scikit-learn (a widely used open-source tool) into a sustainable business while keeping it open. His message: Open source isn’t just code; it’s Europe’s best shot at tech independence, from open data to open hardware. https://lnkd.in/e_i287uq

13

"The performance gap between open source and closed source models is closing" Neils Rogge, ML Engineer at Hugging Face. #AINativeDevCon

3

“I rewrote my code in 30 minutes — and it ran 200× faster.” Bryan Catanzaro, VP of Applied Deep Learning Research at NVIDIA on the moment GPUs changed everything: “NVIDIA showed up in our lab and said, ‘You should try CUDA.’ I plugged in a GPU, rewrote my SVM training code, and it ran 200× faster than my CPU version.” “I thought, that’s it. This is dramatically easier and clearly the future of machine learning.” “The vision was simple: accelerate the world’s most important computations by 10× or 100×, and use that to power AI.” “The compute required for intelligence is essentially unbounded.”

3