Reading about mcp lite ( chatbotkit.com/reflections/a-… ). Interesting. Possibly yet another protocol for us to support.

github.com/alibaba/higress seems impressive from the README, but I don't understand what it does. Even less so after checking their demo demo.higress.io . Does anyone have a simple explanation / usecase / full walkthrough?

We're on our way to make our Wednesday release on Friday.

I've been doing something useful all the time, but I've barely moved a single task in notion... I have a lot of such days and that's very frustrating.

Only a few weeks ago I realised that my ML education is not useless now in the LLM-driven world. Yes, I am slow. Now I'm trying to share this realization the the boss & the company. The only nuance is that I was unable to explain the correct priorities even in the classic ML days

To build a hive mind memory for LLM agents, you basically have to create Twitter for LLMs Think about it: every LLM instance is going to create lots of experiences/memories. They’ll each “post” these memories to “LLM Twitter” (a vectorDB). Then, to read from this giant

I'm trying to create an environment that enables AI coding in our company without micromanaging it. That's much more difficult than I envisioned. Or maybe I'm just overengineering and overmicromanaging the AI once again.

Russian part of the team has two holidays this week starting tomorrow. Have a nice holidays, guys!

The first thing for AI-driven development is tests. For simple development it's enough to just have tests. For products with llm inside we need datasets. And tests is a good place to sync with requirements. So that's what I'm trying to do now: sync.

How to encourage the developers to do everything (or almost everything) via vibecoding and not suffer from a ton of unmanageable shitcode -- I think that's the real Shakespearian question today. Now is the time when we can protect our future selves.

Today, we’re announcing the first major discovery made by our AI Scientist with the lab in the loop: a promising new treatment for dry AMD, a major cause of blindness. Our agents generated the hypotheses, designed the experiments, analyzed the data, iterated, even made figures

I've just learned about REIT's in China. We're doing a project related to them. They are a weird cross between REIT (90%+ of income becomes dividends) and stocks (this is not about real estate, almost anything can be a REIT). Weird.

It's probably impossible to use vibe-coding for big projects yet. And it's very easy to use for small projects. To increase the scope of vibe-coding we need to nudge LLM to be a good developer. It knows how to do it, still you have to teach it from scratch.

It's surprising that in Russia there is no easily-searchable app for AI-based CV ranking. And hh.ru — работа найдется для каждого! has a slow bureaucratic process to gain API key... musing while waiting for the key.

We're rolling out support for MCP servers without Dockefile this week. We also plan to opensource all of our code this week.

Still working to make the opensourced version easy to use. Check for updates very soon.

We've slowed down on opensourcing our projects while preparing for a demo for the Chinese government. But hold on, we're almost ready, and we'll not stop.

For the Russian part of the team: have a nice holidays!

We're considering how to avoid this in our system. (Now in the early stages we just avoid having access to any private data, so it's piece of cake. Later this will be complicated and fun).

We've made the opensource release.[1] Behold: github.com/DARPAI/portal and the other parts. [1] Well, almost: the frontend is lagging 1/2 day because of time-zones.