--- title: "AI Co-Scientists, Answered: 20 Questions from a Live Session with a University Research Center" description: "Twenty questions from a live session with a university multiomics center on what K-Dense does, where AI for science helps today, and where it still falls short." updatedAt: "2026-07-22" tags: ["AI", "Research", "Product", "Community"] canonical: "https://www.k-dense.ai/blog/ai-co-scientists-answered-20-questions" --- We recently spent an hour with the researchers and bioinformaticians of a university multiomics center, walking through what K-Dense actually does and taking questions live. No slides, mostly show-and-tell, and a long, honest Q&A about where AI for science is useful today and where it still falls short. The questions were good ones, and most of them come up in every conversation we have with academic labs. If you're evaluating an AI co-scientist for your own group, this is a fair picture of what to expect.

First, a bit of context for the answers below. K-Dense is a Palo Alto startup focused on giving scientists AI tools that accelerate research. We ship two things: K-Dense Web, our hosted, paid platform, and K-Dense BYOK ("bring your own key"), our open-source, MIT-licensed package that you run yourself. Both are powered by the same library of open-source Scientific Agent Skills. Everything below applies to one or both.

## What it is and how to think about it **What is K-Dense, in one sentence?** If you've used ChatGPT or Claude, take that a few steps further: the system can reach roughly 250 public scientific databases (PubChem, HMDB, UniProt, GEO, and most others you'd use for sequencing, proteomics, or metabolomics), write and run its own Python and R code, iterate on that code until it hits your goal, and then write up the results as a paper, a slide deck, a short report, or a cleaned dataset. **How autonomous is it? What's left for the scientist to do?** We deliberately say "AI co-scientist," not "AI scientist." These systems are capable, but they aren't ready to be trusted end-to-end without a researcher in the loop. Think of it as a tool that lets you test ideas quickly and at scale, so you can spend your time on the high-level, creative thinking and less on the execution. It runs autonomously once you approve a plan, because you shouldn't have to babysit a job that takes hours, but it isn't meant to replace your work. Where does full autonomy go from here? Our honest estimate is three to five years before you could treat one of these systems like another member of the lab: point it at your instruments and your problem space and let it find the gaps and propose its own hypotheses. We're not there yet. **Can it handle non-standard analysis, the kind with no established pipeline?** Yes, and that's exactly what it's built for. You don't need a platform like this to run known workflows. The value is in exploring new methodology, having the system test and iterate on ideas, and being able to hand it ten or twenty ideas at once. Think of it as a postdoc-level researcher you can delegate to: you frame the problem, it figures out how to solve it. ## Getting good results **How should I prompt it to get reliable output?** Six things make the biggest difference. We have a full write-up on our blog, but in short: 1. **A clear objective.** The agents iterate until they think they've met your goal. A vague goal means they'll fill in the gaps with assumptions. 2. **Named data sources.** Point it at specific databases, or upload your own data. The more specific, the more accurate. 3. **The deliverable you want.** A 100-page paper, a one-page LaTeX summary, a slide deck, a cleaned dataset, a script. If you don't say, it'll ask a few questions and then make reasonable assumptions. 4. **Your preferred methodology.** If your lab uses specific packages, protocols, or parameters, say so. You can also upload an example script, or even hand it a paper and ask it to replicate that method (it will pull the paper's supplemental code, or reconstruct the method if none exists). 5. **The target audience.** "An interdisciplinary group working in proteomics" produces very different output than "a venture capitalist." 6. **Relevant context.** Prior results, working hypotheses, papers you think matter. Every task starts from scratch, so it knows nothing about you or your prior work unless you tell it. You can absolutely get away with a one-line prompt, but don't expect it to land exactly where you had in mind. **Is there a separation between planning and execution?** Yes. After you submit a prompt, the system asks a round of clarifying questions, then produces a full, phased plan with explicit success criteria. You can approve it as-is or give feedback ("for step 5, use a different technique than SHAP"), and it will revise and ask you to approve again. Only once you say go does it begin executing autonomously, sending progress updates as it works. A task runs anywhere from twenty or thirty minutes to twelve or fourteen hours depending on complexity. **Can it prompt me for input while it's running?** In K-Dense Web, no, by design. After the initial questions and plan approval, it runs set-and-forget so you're not tied to the screen for a multi-hour job. K-Dense BYOK was built with the opposite philosophy: it has an interview-style interface that asks you questions as it executes, some multiple-choice, some open-ended. If you want a closer, more collaborative research assistant, that's BYOK. ## Data, formats, and sources **What data formats does it support?** Anything that can be read by a Python package. If there's a computational way to open the file, the system can work with it. The limit is proprietary, closed formats that can only be read through a vendor's own software. For example, raw mass-spec files from Thermo or Bruker instruments carry proprietary extensions we can't read directly. Convert those to an open standard like MZML or MZXML first, and you're set. **Can it read papers behind a paywall?** Full text for open-access journals, yes. For closed-access papers, it can only see the abstract; we don't have agreements with journal publishers. The workaround: download the papers through your institutional subscription and upload them as PDFs, which the system can then use. **Any examples specific to mass spec and metabolomics?** Two good starting points: our blog post on the [PyOpenMS](/blog/benchmarking-pyopenms-skill-mass-spectrometry) agent skill (PyOpenMS is an open-source package for working with mass-spec data), which shows the skill in action, and the mass-spec use case in our use-case library. But the best test is to run your own data through it. Just remember to convert proprietary instrument formats to an open standard first. ## Outputs and deliverables **Do I get access to the actual code it writes?** Yes, all of it. When a run finishes you get a session directory containing every Python and R script in execution order, the logs, all generated figures, the written output, a manifest of which files were used, and a README summarizing the run. It's a one-click download as a ZIP, and it's entirely yours to run elsewhere, share, or archive. **Can it interpret the figures it generates, not just produce them?** Yes. When it writes a paper, it uses vision capability to actually look at each figure, pull out what matters, and write the caption and description from that. In K-Dense Web you don't have much room to interrogate a figure interactively, though you can always open a separate session, upload a figure, and ask about it. BYOK is far more interactive here: you can upload a figure, ask what it shows, and even have it help you redraw the figure to highlight a specific point. **Can I get Word documents instead of LaTeX?** Not yet, but it's a common request and it's on our roadmap. Papers currently come out as PDF and LaTeX, so you can continue editing in BYOK's built-in LaTeX editor, Overleaf, or wherever you prefer. The bibliography is included. **Can I get editable PowerPoint slides?** Also on the roadmap. Right now slide requests come back as a PDF deck. We've heard the request for editable `.pptx` with real, movable objects rather than flat images on each page, and we agree that editability is the whole point. Nothing the system produces is meant to be final; it's a strong first draft you shape in your own style. ## Models **Which foundation model does it use?** In **BYOK, you choose.** The default is Opus 4.8 (Anthropic), which we recommend alongside GPT 5.6 Sol (OpenAI) and Gemini 3.1 Pro. For simpler tasks that are more about writing and running code than creative reasoning, cheaper models like Grok 4.5 or open-weight models like GLM 5.2 work well. Because you're paying the model providers directly through your own key, it's a genuine cost-versus-quality choice. In **K-Dense Web, we pick the state-of-the-art model per task.** **Can I switch models mid-session?** In BYOK, yes, at any point in a session or project. There's a small dropdown at the bottom to select any supported model, and another for your compute. ## Privacy, security, and memory **Does the system learn from me over time?** No, and that's intentional. We've experimented with memory, but for scientific work an imperfect memory does more harm than good. A common failure: a workflow fails once because of a bad data point, memory records "this workflow is bad," and the system then avoids a perfectly good method in future without understanding why it failed. There's a possible future feature we'd welcome feedback on — project-level preferences — where you'd store a methodology and parameters once and have them apply to every session in that project. **Does it use what I do to answer other users? Do you train on my data?** No on both counts. We don't carry learning from one user to another, which matters especially for industry users protective of their methods, and we don't train on your data. We do keep encrypted access to prompts and outputs so we can see what's working and ship improvements (we release updates roughly twice a week), but none of it is tied to specific context, molecules, or methodology. **How much does a new session remember from a previous one?** Nothing; there's no cross-session memory, and every session starts fresh. This is actually a feature. Because these models have limited context, a very long session bloated with files, figures, and logs degrades performance. Our recommended best practice: when a session finishes, download it, pick out only the files you actually need, and upload those into a new session. You keep the context that matters without the bloat. **We work with patient data. Is uploading clinical data acceptable?** We don't hold clinical or HIPAA certification, so we don't officially endorse uploading patient data. We follow strong security practices throughout, and publicly available or non-human clinical research data is fine. If you need to work with clinical data, de-identify it first: scrub names, birthdays, and other identifiers before uploading. When in doubt, always defer to your organization's data handling policies. ## Running it and getting access **Do analyses run on your servers or on my machine?** K-Dense Web runs on our servers, on a compute instance in Google Cloud (US) with access to two GPUs. It's not a GPU cluster, so heavy deep-learning training won't be blazing fast, but it works. K-Dense BYOK runs on your own resources, or can outsource GPU- and CPU-heavy jobs to Modal, a third-party instance provider, then pull the results back to your machine. BYOK gives you more flexibility over compute; Web gives you the convenience of just logging in. **What's the trade-off between BYOK and Web?** Web is built for scientists who want an assistant that's ready to use: log in and go, with high-quality end-to-end output including full paper drafts. BYOK asks more of you up front (download, install, add an OpenRouter API key) and its writing output isn't a polished full paper, but it gives you more control, more interactivity, and a lot of power for day-to-day analysis. Our recommendation is to use both; they suit different moments. For reference, we've benchmarked BYOK against the leading comparable platform and it performs better on our benchmarks. **How is BYOK different from just loading the skills into Claude or Gemini?** **Scientific Agent Skills** work anywhere, including plain Claude Code or Codex. But BYOK was built from scratch for scientists, so the skills perform noticeably better there, thanks to prompts and sub-agents we tuned specifically for them. On top of that you get a file viewer and editor supporting most scientific formats (MZML, sequencing, protein files), visualization, the interactive interview mode, optional Modal compute, and built-in LaTeX and PDF editors with compilation. It's meant to be a one-stop shop. **Can BYOK be installed on an HPC cluster?** In principle, yes, but that depends entirely on how your institution's IT has the cluster set up, so it's really a question for your IT team. **Are both products free? How do I get access?** BYOK is completely open source, MIT-licensed, and free; do whatever you want with it. Web is paid, but we run a research grant program that gives selected labs, centers, and nonprofit resource centers a substantial pool of research credits and the ability to add team members. If your group is interested, apply through the program on our pricing page. ## Resources A few links worth bookmarking: - **[Scientific Agent Skills](https://github.com/K-Dense-AI/scientific-agent-skills)** — our open-source skills repository (~148 capabilities, 31,000+ GitHub stars, ~160,000 users), usable in any agentic harness. - **[Research Grant Program](/research-program)** — apply for up to 90% off our K-Dense Web annual Team plan. Have a question we didn't cover? Reach out; we'd genuinely like to hear what you're working on.