Introducing a new open-source tool called TeamCopilot: a coding agent for teams (github.com/rishabhpoddar/tea…). Think Claude Code, but shared across your entire team and running on your cloud. A few key differences compared to other AI agents: • Multi-user environment -> everyone uses the same agent setup. Configure once, the whole team can use it. • Skill & tool permissions -> control who can use which skills and tools through the agent. Example: allow only certain people in the team to use a skill for making server config changes. • Approval workflow -> anyone can create tools/skills, but engineers in the team must approve them before the agent can even see them. • Fully auditable -> chat sessions can’t be deleted by users and are stored on your server. • Use it anywhere -> web UI lets you talk to the agent even when you're away from your work machine. The aim here is create a safe and user friendly environment for all team members to leverage AI agents! It works with OpenAI and Claude models using API keys or subscriptions! This is just a start and a lot more features coming up! 👀

Shutting down of Fable 5 is a strong reason why you should ALWAYS try and find the smallest model that can complete your tasks. You don’t need to always use the latest and biggest for everything!

Just revamped TeamCopilot.ai using @claudeai Fable 5! So good

With powerful ai models being deployed in enterprises, more that ever, we need tools that increase transparency of workflows created by these agents.

A sonnet on fable 5: Fable 5 weaves worlds from fleeting words, Turning simple prompts to flights unheard. Not a dreamer, yet it helps us see, A spark of story, shaped collaboratively.

Just had a thought. I want to replace doom scrolling with YouTube lectures converted into short reels. These videos would be served to me one by one in order of the actual lecture. Might build it.

These are some of the new features added to teamcopilot.ai: - A secret manager for your team so that LLMs never have to see your secrets directly. - Scheduling cronjobs: Run AI agents with a todo-list on a schedule for automating business processes - Calling workflows via APIs - Long running agents: Give agents a long running task that can run for hours until the task is complete. Since teamcopilot runs on your servers, you can check in via your mobile, so you can work on the go.

Getting ai text to sound human is a surprisingly annoying challenge.

I automated blog posting for my product: - AI agent runs every alternate day on my server. - Generates blog post ideas for my project based on trending tech news. - Make a list of top 5 ideas and asks me for my pick - Does SEO keyword research to come up with the best title for the post. - Writes the post, generates the cover image - Reviews the post, makes it less AI sounding - Publishes it on my blog - Checks that it's live that everything is ok All of this for $0.2 per post ($0.1 for image, $0.1 for the full pipeline and writing) And the best part? It's 100% customisable. If i feel something is wrong, I just have to change the skill file or the todo list. Detailed blog coming up on the exact steps you need to take to implement this yourself on teamcopilot.ai/blog

Did a very different format with @reinerpope – a blackboard lecture where he walks through how frontier LLMs are trained and served. It's shocking how much you can deduce about what the labs are doing from a handful of equations, public API prices, and some chalk. It’s a bit technical, but I encourage you to hang in there - it’s really worth it. There are less than a handful of people who understand the full stack of AI, from chip design to model architecture, as well as Reiner. It was a real delight to learn from him. Recommend watching this one on YouTube so you can see the chalkboard. 0:00:00 – How batch size affects token cost and speed 0:31:59 – How MoE models are laid out across GPU racks 0:47:02 – How pipeline parallelism spreads model layers across racks 1:03:27 – Why Ilya said, “As we now know, pipelining is not wise.” 1:18:49 – Because of RL, models may be 100x over-trained beyond Chinchilla-optimal 1:32:52 – Deducing long context memory costs from API pricing 2:03:52 – Convergent evolution between neural nets and cryptography

No frontier model can accurately reference pixel coordinates in an input image. For example, if you want the VLM to return a pixel perfect border around walls in a 2D floor plan, it can’t do that accurately.

Convert a 2D floor plan into a 3D walkthrough in 10 minutes using layout: youtu.be/kZbnfcnD5_8?si=FBVX…

Added something new to my list of side projects today: layoutview.com It’s a browser based tool which allows you to quickly view 2D floor plans as a 3D walkthrough (first person or third person like in a video game)

A new post about all the claude code pricing options and when to use what: teamcopilot.ai/blog/claude-c…

Added a new feature for teamcopilot.ai: Analytics page which shows: token usage and costs over time per model.

Session limits from coding agents is kind of a good thing. Encourages work life balance.

Here’s an open source version of this: TeamCopilot.ai. Run it on your own infra, privately, with any model. Plus it has permissions feature to selectively share workflows and skills with team mates!

Incredible! Now you can setup your coding agent on a server and have it run with virtually unlimited storage!

I just discovered that YouTube has this feature called “key concepts” which explains the key concept spoken in the video at the current time. So cool!

Using Claude code / codex is more fun than playing a video game!

AI agents should never see your API keys. TeamCopilot.ai secret proxy model lets agents use secrets without ever exposing the raw values to the LLM, while keeping secret access manageable for the whole team. Learn more here: teamcopilot.ai/blog/ai-agent…