You don't even need Kimi 2.5 for a decent local LLM setup. - llama.cpp - Unsloth's Qwen 3.5 35B A3B with UD Q4 K XL quants - OpenCode - av/harbor It'll take a while to download/install, but otherwise it's something that mid-range hardware (>32GB RAM, ~8GB VRAM) can run today.

I check HN, GitHub, Reddit, arXiv, and a dozen RSS feeds every morning. That's 8 tabs before coffee. Built a self-hosted dashboard that pulls all of them into one screen. It's called pace. One YAML file. One Docker command. 17 sources. → github.com/av/pace

So, should we use AI? CEOs:

If the LLMs tokenomics comes with some serious pricing issues, why do you think anyone would be able to run AGI at any meaningful scale to make a difference?

I'm convinced the main purpose of this model was to make you crave a larger one.

We’re drowning in information. Every single day, humanity creates more content than any human could consume in a lifetime. You don’t have time to read it. You definitely don't have time to filter it. So I'm building the solution. A smarter way to cut through the noise and only see what you need to see. To help you keep the pace.

I know you have a vibe-coded dashboard for personal stats/tasks/information tracking, show it to me! I'll go first:

I'm releasing Harbor v0.5.0 soon, and it shapes up to be the biggest update in almost a year. Harbor will now spin up native services like MLX/OMLX on macOS and Docker Model Runner on other platforms, in addition to all the backends already present. The app will now be all you need to fully install Harbor and all its dependencies. There are AX improvements for the CLI. Your agents can drive Harbor even better. You can also use Harbor to configure itself, of course. You can now add custom volume mounts on a per-service basis, for example, to mount your custom credentials or system certificates. There are also dozens of smaller changes. Very excited to get it out!

Too much is going on.

Pace will free you up.

I'll just leave it here

Owning a platform comes with some benefits :)

Digital world is hostile now.

You can't keep pace.

Falling stars in mimo code are pretty

Do this. 1. Claude, analyse all the agent sessions you can find on my machines with a haiku subagent: extract all the feature requests, steering, and feedback I gave to an agent. Ask subagent to focus on abstract generalised situations instead of project-specific features. 2. For extracted sessions, use a sonnet subagent to classify kinds of requests. Keep things abstract, non-specific to a given project. 3. For kind of a request, determine when it's applicable based on a project state, do a research about agentic behavior, establish what would lead the agent to an approved/accepted outcome. 4. Save the analysis as a "/be-me" skill, it should guide the agent ot perform the same decision tree as established from the analysis based on the current state of the project. Skill should lead the agent to perform a full clean cycle of software development, including adversarial review, and verification of changes via manual testing (by the agent). The skill should actively rely on the subagents invocation to perform any actual work and keep the main agentic loop purely as an orchestrator. 5. Update "/be-me" skill to include persistence layer via a local log file, one per each specific project. 6. /loop /be-me

Patiently waiting for @ArtificialAnlys to publish the entire cost of the intelligence index eval for Fable 5.

Remember when sonnet was opus and opus was mythos?

Am I using too many agents?

You know what they say? Slow and steady wins the race.

Do less.