K-Dense Inc.

This morning I watched Jon Stewart's latest Daily Show segment on AI. He kept asking a simple question: what does the public actually get out of all this? It stuck with me, so I sat down and wrote this piece (with the help of AI, of course). Most Americans are now more worried than excited about AI, and a lot of that worry is earned: layoffs, slop, chatbots failing vulnerable people, higher power bills. Meanwhile, the industry's big showcase this fall was a math proof almost no one can evaluate. The same polls show what people do want: AI forecasting storms, catching cancers early and developing new medicines. There are already real results there, from mammography trials to hurricane forecasts. My argument is simple. If AI wants the public's trust, it should cure something, and let people see it happen. The frontier labs have the compute, talent and money to do far more of it. Curious what you think. https://lnkd.in/g4mHSjjk

Citations play telephone. A paper may cite one thing, but follow the citation two hops back and the original study says something different. Our latest free offering, Citation Telephone, is a free tool that does that tracing for you. Paste a DOI, PMID, PMC ID, or arXiv ID, pick a cited sentence, and it follows the chain back to the paper that first reported the finding, flagging every hop where the claim changed. Try it free: https://lnkd.in/gNcfajzZ

Our paper on K-Dense BYOK is live on arXiv! In our evaluation, K-Dense BYOK outperformed the scientific harnesses we tested on scientific quality and research execution. It also maintains full research provenance through a hash-chained lab notebook, so every result traces back to an auditable record. Paper: https://lnkd.in/gVGK8yg4 Link to the repo in the comments.

We built 503 expert-scientist prompts for AI agents then we tested them against "You are a helpful assistant." The helpful assistant held its own. In our new paper, we evaluate Scientific Agents, the open-source library we built at K-Dense Inc. Each AGENTS.md profile teaches an AI agent how a pharmacologist, structural geologist or bioinformatician works: their workflows, tools, common pitfalls and reporting standards. The setup: 100 matched profiles, 9 science benchmarks, about 4,500 questions, and 5 system prompts per question. The model and the user prompt stayed the same, and answers were graded automatically by fixed rules. With Gemini 3.8 Flash: 📉 Accuracy: no consistent gain. Profiles averaged −0.6 points against the one-line baseline, and no benchmark showed a clear improvement. 💸 Cost: 1.5–2.3× more output tokens and 2.2–4.5× higher cost per call. 🧬 Agentic bioinformatics: 46.7% of problems solved with profiles vs 56.7% without, mostly because the longer prompts hit token and time limits sooner. 🔌 One surprise: long prompts held up much better when the API provider dropped requests (71.6% vs 54.0% correct on the first pass). But generic and mismatched prompts did about as well, so the cause is prompt length or formatting, not expertise. The lesson isn't "expertise doesn't matter." It's that pasting a whole profession into every prompt isn't an effective default. A model doesn't need a pharmacologist's entire career to answer one question. We built this library, and we're publishing the result that says it doesn't improve accuracy by default. 📄 Paper: https://lnkd.in/enwJeEFr 💻 Profiles (open source): https://lnkd.in/eA3z7fbz

Our paper on K-Dense BYOK lead by Aubrey Brueckner is now on arXiv. AI agents can now carry out real analyses. The harder problem is trusting what they tell you. Models overclaim: they report a check as done, or a result as verified, when it wasn't. K-Dense BYOK is our answer. It's a free, open-source research assistant that runs on your own computer, using the model you choose: your own API key, a subscription you already have, or a local model. In that evaluation, K-Dense BYOK outperformed two managed platforms on both scientific quality and research execution. It was also the only one whose deliverables consistently recorded the software environment and included the commands to regenerate the results. We ran this evaluation ourselves, so the paper includes all 20 prompts, the rubric and the per-prompt scores for you to check. Since we submitted it, the same harness has also reached #2 on Genentech's public CompBioBench v1 leaderboard (98% with GPT-6 Astra, as of September 11). Huge thanks to my all the authors including Aubrey Brueckner, Darshil P. and Yuhuan He. 📄 Paper: https://lnkd.in/ezPmxd_p 💻 Code (MIT license): https://lnkd.in/gFFpGHqK

We built 503 expert-scientist prompts for AI agents then we tested them against "You are a helpful assistant." The helpful assistant held its own. In our new paper, we evaluate Scientific Agents, the open-source library we built at K-Dense Inc. Each AGENTS.md profile teaches an AI agent how a pharmacologist, structural geologist or bioinformatician works: their workflows, tools, common pitfalls and reporting standards. The setup: 100 matched profiles, 9 science benchmarks, about 4,500 questions, and 5 system prompts per question. The model and the user prompt stayed the same, and answers were graded automatically by fixed rules. With Gemini 3.8 Flash: 📉 Accuracy: no consistent gain. Profiles averaged −0.6 points against the one-line baseline, and no benchmark showed a clear improvement. 💸 Cost: 1.5–2.3× more output tokens and 2.2–4.5× higher cost per call. 🧬 Agentic bioinformatics: 46.7% of problems solved with profiles vs 56.7% without, mostly because the longer prompts hit token and time limits sooner. 🔌 One surprise: long prompts held up much better when the API provider dropped requests (71.6% vs 54.0% correct on the first pass). But generic and mismatched prompts did about as well, so the cause is prompt length or formatting, not expertise. The lesson isn't "expertise doesn't matter." It's that pasting a whole profession into every prompt isn't an effective default. A model doesn't need a pharmacologist's entire career to answer one question. We built this library, and we're publishing the result that says it doesn't improve accuracy by default. 📄 Paper: https://lnkd.in/enwJeEFr 💻 Profiles (open source): https://lnkd.in/eA3z7fbz

Today we're open-sourcing LabMCP: connectors that let AI assistants run your lab instruments. Balances, stirrers, pumps, spectrometers, potentiostats and more. 32 connectors, 411 tools, with safety limits and a full command log.

Reporting checklists for research papers have existed for years. ARRIVE, CONSORT, STROBE and most journals' own lists all ask the same question: could someone else check this work, or repeat it? Actually checking a paper against them is slow, so it mostly happens late, or not at all. Today we're releasing Rigor Scan, a free reporting check for research papers, built on Jev by TypeSafe AI. Drop in a PDF and within seconds you get: • 74 checks across study design, statistics, materials, data and code, ethics and claims, each linked to the sentence it rests on • Software checks that recompute reported statistics, flag retracted references, match cell lines against Cellosaurus' list of problem lines, and open data and code links • A completeness score, with every question and threshold published • A next step for each gap, like running a power analysis or building an antibody table, that you can hand to K-Dense Web Why Jev? It's a decision model, not a chat model. It answers each check with a probability, and its evidence has to be a sentence that is actually in the paper, so it can't invent a quote. Your PDF is read in your browser and never uploaded, and no account is needed. Try it: https://lnkd.in/eV34STu2 How it works, and why we built it this way: https://lnkd.in/eBw5_Z-z #ResearchIntegrity #Reproducibility #OpenScience