Founder @DeSciLabs. Building AI infrastructure for scientific reasoning @SciWeave. Professor in economics.
Joined March 2013
- Tweets 1,579
- Following 890
- Followers 1,969
- Likes 4,279
56 Photos and videos
Pinned Tweet
Super excited to speak at the @bittensor track at @proofoftalk on Tuesday at 6pm CET!
@bitstarterAI will we streaming this live, so fell free to join!
x.com/bitstarterAI
Jun 1
At this year’s Proof of Talk, DeSci will be part of the center-stage conversation.
Our scientific advisor @PKoellinger, founder of DeSci Labs, will be on stage in Paris discussing how Bittensor can help create infrastructure for verifiable scientific claims.
We believe distributed compute can become a real engine for science by scaling reproducible workloads, aligning incentives, and rewarding useful scientific work.
Philipp has been building toward this future for years, and we’re excited to see him bring that vision to Proof of Talk.
Keep an eye out for his talk!
1
4
395
Jun 13
This is one of many reasons why decentralized alternatives such as @bittensor are so important.
Jun 13
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.
The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Access to all other Claude models is not affected.
We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.
Read our full statement: anthropic.com/news/fable-myt…
1
113
Philipp Koellinger retweeted
Jun 12
Subnet 111. Live on #Bittensor ⚡
Science has a credibility problem.
Papers get published. Results can't be replicated. Nobody gets held accountable.
Claims is changing that - using Bittensor to score, verify and surface research that actually holds up.
Led by Prof. Philipp Koellinger [ @PKoellinger ] and Prof. Christian Roessler: 17,000 citations. Published in Nature and Science. Game theorist specialising in mechanism design. Built @DeSciLabs to improve research incentives.
They won Proof of Pitch. They impressed the judges. They earned the community.
Now they get a subnet.
5
11
58
3,663
Bittensor's co-founder just said something that stopped me cold.
He and his co-founder built the entire protocol from scratch.
And then the subnets started moving faster than both of them combined.
His exact words.
People far smarter than us have started building on top of the system.
The founders do not develop subnets anymore.
The network outgrew them.
Now read why that sentence is the most bullish thing anyone has ever said about $TAO.
Bitcoin's creators did not build every application that runs on Bitcoin.
They built the incentive layer and got out of the way.
The network grew into something far beyond anything two people could have built alone.
That is exactly what just happened to Bittensor.
Decentralised training producing more resilient models than centralised alternatives.
Subnets beating state of the art benchmarks domain after domain.
Corporations like Unity pulling 3D objects directly from Bittensor for commercial video games.
DeepMind researchers publicly acknowledging that decentralised training produces stronger more generalised models.
All of it happening without the founders directing any of it.
TCP/IP does not need its inventors to maintain it anymore.
The internet ran past them decades ago.
Bittensor just crossed the same threshold.
Two founders watching the protocol they built outgrow them in real time is not a red flag.
It is the oldest signal in technology.
The people who recognised that moment in Bitcoin did not need to explain themselves later.
The people who recognised it in Ethereum did not either.
This is still early.
The people who read the docs always buy before the people who read the price.
Screenshot this.
Two people who were early in Bitcoin and early in Ethereum just went on record about $TAO.
One of them wrote a book about Bitcoin in 2013.
The other invested in the Ethereum ICO in 2015.
Both of them started a fund with Jason Calacanis with a single thesis.
Bittensor is the third great open-source substrate after Bitcoin and Ethereum.
Here is the exact framing they used.
In the early 90s Microsoft, AOL, and CompuServe were the well-capitalised incumbents.
Everyone thought they would monopolise and run away with the internet.
Then TCP/IP, Linux, and the World Wide Web came along and everything converged on an open-source substrate.
Bittensor is that open-source substrate for the AI story playing out right now.
OpenAI. Anthropic. Google DeepMind. XAI. Different cast of characters. Same pattern.
And this time you can actually own a piece of the open-source substrate.
Now read the valuation mismatch that should stop you cold.
The four main AI labs combined are worth approximately $1.5 trillion.
Bittensor is worth $1.7 billion.
Ridges subnet competes directly with Claude and Cursor and has beaten them on benchmarks.
Ridges market cap is $30 million.
Cursor is worth $30 billion.
That is not a small dislocation. That is a comical one.
The highest valued subnet in the entire ecosystem is around $80 million.
There has never been a billion dollar subnet yet.
On Ethereum during the ICO mania projects with nowhere near this quality of output were raising hundreds of millions within minutes.
Now think about how many orders of magnitude more capital is chasing AI opportunities today compared to 2017.
When that capital discovers Bittensor the valuation rerating will be violent to the upside.
Their exact words. Not mine.
The man who called $TAO at $3,000 by end of 2026 said it directly.
By 2030 it will be a trillion dollar ecosystem.
Every molecule in my body is screaming this is another one.
The people who read the docs always buy before the people who read the price.
This is still early.
8
27
155
9,512
You can still get into our new Claims subnet at discounted early investor conditions.
@DeSciClaims
#Bittensor
Jun 5
Paris saw @DeSciClaims win Proof of Pitch. 🏆
But the crowdfund isn't over. Both teams still need your support to power their launches on Bittensor.
If @provenonce_ai doesn't secure a subnet, all pledged TAO will be returned.
Back a team ↓
1
3
116
Philipp Koellinger retweeted
Jun 4
The judges voted. You voted.
After 45 minutes live at the Louvre, @DeSciClaims had raised 184 TAO.
Claims wins Proof of Pitch at @proofoftalk 🏆
And now they're launching a subnet on Bittensor - auditing the science that the world runs on.
2
17
78
4,053
Pitching our new baby @DeSciClaims at @proofoftalk in the Louvre and WINNING this adrenaline-raising competition was such an honor. Joining the amazing @bittensor community as a subnet owner is literally a dream come true for me. Pitch deck:
bit.ly/4odS2TS
Jun 3
Proof Of Pitch: Bitstarter crowdfund x.com/i/broadcasts/1dGYllLqM…
3
144
May 14
SciWeave: 0% hallucinated citations on ScholarQABench (Asai et al., Nature 2026).
GPT-5.2: 59%.
Claude Opus 4.6: 62%.
100 queries across biomedicine and neuroscience.
The 0% isn't a model trick. SciWeave doesn't ask the LLM to recall papers. It retrieves passages from OpenAlex (~300M scientific works) and synthesizes the answer from what comes back. The architecture rules out fabrication.
The benchmark, the scoring script, and the NLI judge are all public. We ran SciWeave against them.
Hallucinated citations are already showing up in peer-reviewed bibliographies. The failure rate deserved a number.
desci.com/blog/zero-hallucin…
2
2
305
Apr 22
Encyclopedias on CD-ROM looked like the right idea. Then the browser showed up.
Many researchers are still looking for an "AI research tool": a destination site they open, paste questions into, wait.
That category is dissolving.
The AI you already use (Claude, ChatGPT, Cursor, Codex, Windsurf) can reach into 300M peer-reviewed papers mid-conversation without leaving the chat.
The destination is the browser tab you already have open.
SciWeave is plumbing. You install it once, into whatever model you work in.
Now spinning up 4 parallel sub-agents for platform-native variants.
sciweave.com/web/mcp
220
Apr 21
We've just launched an MCP server integration for SciWeave - I'm excited!
You've tried doing research tasks with Claude, ChatGPT, CODEX etc. They're amazing tools - but they still hallucinate >50% of academic references. Now you can combine them with the SciWeave MCP and your favorite AI turns into a hyper-productive research companion.
For example, you can use this to:
- Draft a grant proposal with real citations
- Write a syllabus for a course you're teaching
- Do a systematic literature review
- Brainstorm plausible hypotheses that haven't been tested yet
- Build automated research workflows that are grounded in the published literature
Try it out and let me know what you think! We'd love to get feedback.
Set-up guide: sciweave.com/web/mcp
2
2
230
Apr 21
Free to start. 30-second install.
🔗 mcp.sciweave.com
Blog post:
🔗 sciweave.com/blog/sciweave-m…
98
Philipp Koellinger retweeted
Apr 6
Instead of watching an hour of Netflix, watch this 2-hour Stanford lecture on AI careers. It will teach you more about winning in the AI race than all the AI content you’ve scrolled past this year.
161
3,124
14,107
2,005,449
SciWeave is now available for Apple mobile devices! I like using it on the go to quickly check facts based on scientific evidence and for health-related questions. #Science
@SciWeave
apps.apple.com/eg/app/sciwea…
1
84
Mar 31
Empirical evidence from a series of behavioral experiments with students at @Wharton suggest that AI is beginning to influence human thinking and decision making.
Many people "surrender" their judgement to AI instead of critically questioning and treating it as a sparing partner.
"Surrendering" improves confidence while decreasing cognitive performance.
Critical engagement with AI yields much better outcomes.
@gidin
papers.ssrn.com/sol3/papers.…
4
154
Mar 29
RT @JonHaidt: If you come across someone asserting there is "no scientific evidence" that social media is causing harm, please send them th…
467
Philipp Koellinger retweeted
Mar 24
Aysu Okbay and I are hiring a postdoc to study gene–environment interplay in health & social inequalities 🧬
You'll analyze large genomic datasets as part of a 3-node consortium with Uppsala & Oslo. Based at @amsterdamumc 🇳🇱
Apply by May 4 below, please RT for karma points👇🏽
2
22
32
5,119
Philipp Koellinger retweeted
Mar 19
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.
31
132
2,053
85,917
Mar 12
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.
2
232
Mar 12
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...
Mar 11
More Claude code fan fiction, if you want to call it that. In this essay, I imagine that human researchers no longer have the comparative advantage in research.
causalinf.substack.com/p/cla…
2
37
5,729