Timothy Kassis

Since `CLAUDE.md`, `AGENTS.md`, and Agent Skills emerged, we’ve been thinking about a deeper question: how do you represent an expert’s experience in a form LLM agents can actually use? Take a world-class cancer biologist. What makes them an expert, and a pioneer in their field, is not just the knowledge they’ve accumulated. It’s their way of thinking: their mindset, how they frame problems, the principles they rely on, the shortcuts they trust, and the traps they avoid. How do you capture that? `mimeo` is our attempt to do exactly that. Give `mimeo` the name of an expert, and it searches the internet for interviews, essays, opinion pieces, talks, YouTube videos, and other public materials. It then codifies what it finds into an `AGENTS.md` or `SKILL.md` file, ready for your agent to use. It extracts principles, mental models, heuristics, frameworks, and anti-patterns. We call these human representations `mimeographs`. The tool: github.com/K-Dense-AI/mim… Examples: github.com/K-Dense-AI/mim… Live demo: pantheon.k-dense.ai

Heading to Google HQ in June for the 2nd @GoogleStartups Gemini Startup Forum. Repping @k_dense_ai, where we're building a multi-agent AI co-scientist to accelerate scientific discovery. Excited to swap notes with fellow AI founders and the @googlecloud + @GoogleDeepMind crews.

We're thrilled to be part of the 2nd @GoogleStartups Gemini Startup Forum. K-Dense is building a multi-agent AI co-scientist to accelerate scientific discovery. Excited to go deep with fellow AI startups and experts from @googlecloud and @GoogleDeepMind at Google HQ in June.

Scientific agents are only as good as the literature they can read. Today, @ExaAILabs joins K-Dense's open-source Scientific Agent Skills library — bringing neural search and URL extraction to every agent built on the skills. Why this is a big deal for AI-driven research: ➡️ Embedding-based retrieval finds the right paper even when the query and the paper use completely different words ➡️ Scholarly filtering is first-class, not bolted on ➡️ A two-pass pattern (allowlist → open web) is documented out of the box, so agents don't have to reinvent retrieval strategy Bring an API key and a question to try it out in K-Dense BYOK today. k-dense.ai/blog/towards-s…

@SawyerMerritt @Starlink Yes, it's the difference between having a productive flight and getting nothing done.

Would having @Starlink high-speed Wi-Fi rank high on your list when deciding which airline to fly with?

Learn how to sandbox your AI co-scientist with @NVIDIAAI OpenShell: k-dense.ai/blog/sandboxed…

How mimeo works. Try it yourself at github.com/K-Dense-AI/mim…

Excited to be part of the K-Dense team leveraging @Google’s new Multimodal File Search to deliver more powerful, verifiable AI capabilities for scientific research.

Big thanks to the scientists and researchers using K-Dense Web. Your feedback shaped the UI improvements we just shipped. Keep it coming.

@jrkelly @Ginkgo Well said

It is telling that the 2nd largest private biotech financing in history is going to a biotech that outsources all their lab work. We need to have our “TSMC moment” for scientific lab work. Companies like @ginkgo can focus on reducing the cost of lab work via automation, while others can leverage that platform for drug discovery, ag products, and more. We will be a more competitive industry that way, let’s get started! Article below 👇

Mapping the RAS Mutation Landscape in Pancreatic Cancer with K-Dense Web k-dense.ai/blog/ras-mutat…

I was discussing @sentra_app with engineers at one of the biggest financial services firms in the world. They asked for our repo so they could "inspect the inner workings." to build trust. Then, almost in passing, a different engineer mentioned they had already taken an MIT-licensed graph memory project and then vibe-coded it into a full internal deployment. Two of their engineers built it in under a week, deployed inside their network, with no vendor in the loop. Since then I have heard one version or another of the same story in no less than 7 other firms. Those conversations have stuck deep within me because it threatens one of the major open-core playbook has worked very well over the last several decades. Release a strong core, let the community kick the tires, build trust around it, then sell the polished enterprise version with compliance reviews and a vendor on the hook when things break. That model built Red Hat, MongoDB, Elastic, and many others. AI coding agents are quietly breaking it, and I do not think founders (or investors) have priced this in yet. Here is the thing I keep coming back to. The expensive part of the open-core loop was never the code itself. It was the senior engineering effort it took to harden, integrate, and operate that code inside a real enterprise. That labor is what justified the enterprise license. AI coding tools just made it dramatically cheaper (and easier). Two engineers with good prompts are not as good as a dedicated platform team, and they do not need to be. They only need to be good enough for a risk committee to sign off, and "good enough" has dropped a lot in the last 18 months. The most risk-averse buyers in the world are starting to do this math out loud in pitch meetings, which is the part that should worry every founder selling open-core infra right now. Open source itself is not in trouble. The collaboration and verification benefits are real, and they are not going away. What is breaking is the specific upsell flywheel that turned open code into enterprise revenue. The fastest-moving AI-native categories are where it cracks first, especially memory systems and retrieval infrastructure. The hardest part of those products is no longer the part the community can stress-test for you. It is the part competitors can now rebuild over a long weekend. After that meeting, I was glad we had not open-sourced the core of Sentra's memory system. Though its still something I am unsure about and might be convinced to rethink. In an AI-first world, the winning move is to open what builds trust and protect what actually compounds. Those used to be the same thing, but I don't think they are anymore. Curious if others are watching some version of this play out.

@JustonBrazda I prefer the speed profiles in 14.2.2.5

On a 100 mile drive 14.3.2 in standard mode keeps going 62 in a 70 on a wide open 4 Lane highway I bumped it up to hurry and it doesn’t go any faster. I’ve hit the accelerator at least 30 times in 30 miles.

We asked K-Dense Web, with a single prompt, to analyze 116 newly declassified UAP files from the PURSUE release. It produced a full report across ~833K words of source material. Headline finding: lots of unexplained aerial activity, but no concrete evidence of extraterrestrial biological life. PDF: k-dense.ai/examples/sessi…

At @k_dense_ai we agree!

I just spent $29 on a singular @k_dense_ai prompt: build a 2-5 year equities strat assuming @leopoldasch's AGI timelines (from situational awareness) are right 1.25 hours later, it gave me full 32-page investment report. and now giving it to you for free

Hey @Stanford 👋 To thank our neighbors across the street, we're giving every stanford.edu signup $60 in free K-Dense Web credits. Run real research with an AI co-scientist - literature review, hypothesis generation, analysis, the works. Through Sunday only. k-dense.ai

We did some analysis. Take home question: why aren't you using K-Dense?

@OpenRouter Nice! Would also love 'search' and 'deep research' endpoints from Parallel Web Systems, Exa and Valya

1/ Audio is now first-class on OpenRouter. Two new endpoints live today: 📢 /api/v1/audio/speech — text-to-speech (TTS) 🎤 /api/v1/audio/transcriptions — speech-to-text (SST) Same routing, billing, and keys you already use for text, image, and video.