Finetuning for a specific goal is a very iterative process. Instead of managing 15 different notebooks for each model and babysitting the script. You can just use the git like workflow we have at oxen. Want to test out different models on a 1TB dataset? just create a new branch and trigger a finetune. We cowrote a case study about this with the llama team. https://lnkd.in/e_rpw56b

Do you know what a mini-loci system is and how to use it? Before a pitch or keynote, sketch the whole story as a few simple shapes boxes, arrows, symbols or even the solution. Then learn the visuals and the narrative to explain the visuals, not the words. On stage, you don’t recall sentences - just “see” the sketch in your mind and the narrative unlocks itself. No robotic delivery. No blank moments. None of that irritating script reading I see too often. Just a clear structure that feels natural. If scripts don’t work, try drawing your story instead. It’s the simplest memory hack not many people use.

How do you optimize the model-facing surface of a tool for a given agent or task? You use tool masking. A simple concept, but as always, the devil is in the details. Frank Wittkampf and Lucas Vieira define the novel concept of tool masking.

⚙️ Treating AI like just another software rollout? That’s the problem. It’s time to build an operating model that aligns people, processes & purpose. Explore more by diving into Raft's latest blog here 🔗 https://lnkd.in/eNnQ9kGy

Thrilled to launch our Engineering blog, where Finch engineers will account first hand what they're building. The first one is written by none other than Moyez Mansoor, our Founding Engineer and Naya loyalist. Our first post dives into how we replaced our Celery/Django task queue with Temporal, beefing up observability 👀 and reliability 🗿 of the workflows that orchestrate our AI agents. Link in comments.

What labeling a million hours of video really teaches you Scaling video annotation isn’t about how many frames you can label. It’s about how consistently thousands of annotators interpret edge cases. Every pixel tagged differently changes how a model learns movement or intent. Quality control at scale isn’t an afterthought - it’s the core product. The most advanced models are born from workflows that don’t crack under volume. When quality scales, models stop overfitting and start understanding.

Recently watched a live stream from a known AI voice in the industry... In the chat I asked about PRDs, requirements and documentation before diving into the coding... The response... No PRDs... just take your idea and start building. For some organizations, that's how things used to be... Interesting swing back in the direction of simplicity and speed. Obviously, if you are beholden to the documentation rigor of a big corporation, then you do what you have to do... but it seems the attraction to speed is regaining momentum.

How do you optimize the model-facing surface of a tool for a given agent or task? This article by Frank Wittkampf and Lucas Vieira defines the novel concept of tool masking.

Knowledge doesn’t scale if it’s buried. AI-driven knowledge layers are turning static documentation into living, searchable memory. Imagine if your company could actually use everything it already knows.

From costly human error to AI-powered inspections | A clip from my Substack post, Stop Wasting Time Googling!. Watch it here: https://lnkd.in/g3-xdxcm