RT @Albertorizzoli: 🏛️ Knowledge Hubs is RAG on steroids, but even more so, a full CRM replacement. If AI can understand your files and wh….

Summer is starting a training job (or two) in the cloud (@modal_labs), then heading out to the forest for some raspberry picking while it runs. Later, coming home to check the loss curves and make raspberry pie

Join millions who have switched to Grok.

RT @PartechPartners: AI models are taking the world by storm, and while they work for basic tasks, would you trust it to do your back-offic….

new blog post here:

Is Gemini 2.5 Pro good at bounding boxes? Sort of.

New Spegel version is out, now powered by @LiteLLM! . Try any model you like! Feedback appreciated

There is also some pretty insightful comments on HN

Spegel - A Terminal Browser That Uses LLMs to Rewrite Webpages

It's been a year since we launched Go, and it's genuinely exciting seeing customers build and deploy agents solving very real and very complex problems. Now releasing Concierge, making agents accessible across your org, even if you've never written a prompt.

This was a fun one to build. At ~3s, it felt broken. At ~1s (2.0 flash), it just clicked. Funny how much of a difference 2 seconds can make.

My small Ecto-esque DSL for @qdrant_engine in #Elixir is starting to take shape.

Worth noting that o1 doesn't handle system messages yet. Let's see if I can just put everything in user. .

I had a lot of fun working on this app. The 'intention' model was a ResNet-ish distilled model weighing in at less than 20 MB (the only way to get real-time image processing back in the day without burning up your phone). It was mainly a combination of metal shaders and neon.

Without a comprehensive benchmark, changing model versions, provider or prompts can be scary and error-prone. A good benchmark tailored to your specific task makes all the difference. Public benchmarks are useful but not always applicable.

Also be aware that gemini assumes that the image is 1000x1000 px, so you will have to normalise on that too.

Changing the bounding box coordinate order from .{ y_min, x_min, y_max, x_max} to { x_min, y_min, x_max, y_max} (just the order).drops the performance to 0.71.

Tried Gemini 1.5 (Pro) for object localization (salient objects) using bounding boxes on the first 100 images in the DUTS dataset. The results are pretty good (avg IOU 0.84), not real difference between Pro and Flash. But you need to be very careful on how you prompt!

Spent 30 mins today puzzled as to why my Phoenix project was slowing down as the number of parallel requests increased (locally). Turns out, with code_reloader: true, each request triggers a sequential code reload via Phoenix.CodeReloader. Disabling it brought response times.

RT @ArtificialLawya: Pinsent Masons Deploys ‘V7 Go’ GenAI Platform . > Interesting BYO move by @Pinsent_Masons - AL takes a look. HT @V7Lab….