Philipp Koellinger

What surprised me most building SciWeave: How people actually use it SciWeave is an AI research assistant that searches the scientific literature (>250M papers) and answers questions with verifiable citations you can inspect. We soft-launched in Sept 2025. 5 months later (with minimal ad spend): • 26k users • 153 countries • ~90k research queries • 56% multi-turn deep research threads We built it for literature reviews and research tasks. But many users treat it as a learning tool — not just researchers and students, but lawyers, investors, consultants, doctors, etc. They use it to check claims, understand concepts, and explore ideas with trustworthy sources. Then we accidentally ran a natural experiment. Our paid ads broke and were off for most of February. Growth slowed — but organic users (mostly from Google, Reddit, and X): • returned 2× more often • asked 2× more questions • converted to paid plans more often Now I'm wondering: Is AI for research fundamentally a word-of-mouth product? sciweave.com

Reviewing yet another academic paper full of hallucinated references, odd citation practices, and improperly attributed material. To all academic colleagues out there, literally ruining our profession: thanks for nothing.

@Aella_Girl @0xAgentOwl The lower median but similar mean for men compared to women in your sample means that the right tail of the male distribution is much longer (i.e. higher bc), which is probably as expected. The higher median for females in your sample could be sampling bias.

@0xAgentOwl yea i'm not totally sure how to interpret the gap, I need to look into it more. but iirc CDC does in person interviews in your house, which i wouldn't be shocked if it meaningfully suppresses sex partner count reporting for women.

In my data, here's the median sexual partners by age and gender:

@causalinf @MarkHahnel had a related blog post with interesting data on this. Developing AIs that check papers, requiring open code at submission, and raising submission fees are all reasonable and doable suggestions to deal with the increasing supply of papers imo openresearch.wtf/have-we-alread…

A modest proposal to editors of Econ journals for how to manage predicted increase in manuscript submissions due to AI agent shocks to the marginal costs of production. causalinf.substack.com/p/claude-code-…

AI won’t replace scientific research anytime soon. But it can accelerate it dramatically. From tens of thousands of research queries on SciWeave, we see three patterns appear repeatedly: 1. Literature review acceleration AI helps researchers quickly map the landscape: what’s been published, which papers matter, where the debates are. 2. Evidence triangulation Scientists compare evidence across multiple papers, datasets, and disciplines. AI can surface the sources — humans judge the evidence. 3. Hypothesis exploration Researchers “jam” with AI to explore possible explanations and overlooked variables. The pattern is clear: AI is becoming research infrastructure, not yet a replacement for expert judgment. The most useful systems are not general purpose LLMs that hallucinate references. The most useful AI's are specialized tools that help researchers: - navigate the literature - connect evidence - separate signal from noise.

It's true, most researchers get "procedural" utility out of their work - they love it, despite all the flaws of academia. And that love for the job gives them a comparitive advantage at it - until a machine comes along that does it all better and faster without love...

10/ In vertical AI domains like science, what we choose to evaluate determines whether systems become convincing or reliable. If you had to optimize one metric for literature-grounded systems, which one would you go for?

9/ We’re starting to see research-grade AI architectures built specifically for literature grounding (e.g. OpenScholar, @SciWeave ). sciweave.com Different architecture. Different optimization target. Different outcomes.

1/ If we want to use AI for serious research, we need to benchmark it for serious research. Not for chat. Not for vibes. Not for leaderboard screenshots. 🧵👇

10/ Curious how others see this: Are we underestimating how much vertical AI requires architectural specialization? Or do you think prompting + scale will solve it?

9/ If we want AI systems that researchers, analysts, lawyers, or doctors can rely on: We need to treat hallucinated references as an infrastructure problem — not a formatting problem.

1/ Your AI just hallucinated a reference (again)? It’s not a prompt problem. It’s an architecture problem. 🧵👇