Matthias Samwald
@matthiassamwald
Followers
609
Following
1K
Media
47
Statuses
269
🔬 Associate professor of AI & biomedical informatics 🌐 Recent co-chair EU General-Purpose AI Code of Practice ✨ Aligning human and machine intelligence
Vienna, Austria
Joined January 2009
The EU GPAI Code of Practice is now public. I'm proud to have helped shape the Safety & Security Chapter, and grateful for the fantastic collaboration with the co-chairs, vice-chairs, and all stakeholders! - To me, this work has been about striking an important balance: enabling
1
4
13
🙏 We thank the eight experts who participated in our elicitation: @Nik_Fortelny, Máté G. Kiss, Bernhard Kratzer, @TKrausgruber, Anna Redl, Rob ter Horst, @p_traxler & Dimitrios Tsiantoulas for sharing their valuable insights.
0
0
1
📢 Shout outs to the groups doing excellent work on biomedical research acceleration we covered in our exploratory analysis, including @FutureHouseSF @EmeraldCloudLab @daniil_boiko @vivnat @KyleWSwanson @james_y_zou @ProfBuehlerMIT @m_skarlinski @NJSzymanski @biogerontology
1
1
4
➡️ Bottom line: GPAI can compress biomedical research timelines dramatically, especially for cognitive work, but hard biological & social limits remain. Responsible deployment requires robust infrastructure, high‑quality data, ethical oversight & community uptake. (13/14)
1
0
2
Harnessing GPAI requires more than tech. We need investment in cloud labs & frontier models, reform how science is communicated and rewarded, safeguards against misuse & agile governance. Without these, acceleration may compromise quality and not reach full potential. (12/14)
1
0
2
👩🔬We surveyed 8 senior biomedical researchers. Their projects typically lasted ~72 months with ~73 % cognitive tasks. They doubt max-level gains for idea generation, design&execution, but see high potential for admin tasks & flag community assimilation as main barrier. (11/14)
2
0
2
🚧 Bottlenecks abound: building self‑driving labs needs major investment & integration; missing metadata & negative results hamper models; biological processes like cell growth & animal models have irreducible times; ethics approvals & peer review can take months. (10/14)
1
0
1
⏳ In a worked example, a 36‑month project (24 months cognitive, 12 months physical) shrinks to ~3.6 months under maximum‑level GPAI — a 10× acceleration. Note the assumed 3 months of biological “irreducible” time dominates the accelerated timeline. (9/14)
1
0
1
Biology and physics set natural limits to how fast research can go. No matter the GPAI or automation level, some processes—like cell growth or reaction kinetics—can't be accelerated. We call this "non-compressible” or “irreducible” time constant. (8/14)
1
0
1
Real‑world evidence varies: drug discovery reports 1.3–2× speedups; agent‑written articles are ~75–300× faster; self‑driving labs promise 10–100× throughput gains. Cognitive tasks often outpace physical ones. (7/14)
1
0
1
Our scoping literature review yielded acceleration factors for our two scenarios: “Next-level” GPAI could double the speed of both cognitive and physical tasks (~2×); “Maximum-level” systems might achieve ~100× for cognitive tasks & ~25× for lab work. (6/14)
1
0
1
We then integrated the GPAI capability levels with research tasks to illustrate qualitative differences. “Next-level” GPAI: assists humans but requires substantial oversight. “Maximum-level” GPAI: operates largely autonomously, significantly reducing human workload. (5/14)
1
0
1
🔬 We outlined 9 major research tasks in a typical biomedical project: knowledge synthesis, idea/hypothesis generation, experiment design, ethics & permits, experiment execution, data analysis, results interpretation, manuscript prep & publication. (4/14)
1
0
1
We drew on existing AGI & automation frameworks to map General-Purpose AI (GPAI) capabilities in science. Our two axes separate cognitive tasks (e.g. literature review, analysis, writing) from physical tasks (lab work). Combined, they define the level of GPAI autonomy. (3/14)
1
0
1
📖 Full paper on arXiv: https://t.co/pRFOwpYT0M -- highlights below (2/14)
arxiv.org
Although general-purpose artificial intelligence (GPAI) is widely expected to accelerate scientific discovery, its practical limits in biomedicine remain unclear. We assess this potential by...
1
0
2
Could AI ever 10x the speed of biomedical research? Our new preprint explores the limits to research acceleration. We reviewed evidence, modelled scenarios & asked experts to see where speedups are possible and bottlenecks emerge. 👇(1/14)
2
3
11
Great to see so many frontier AI companies support the Code!
I’ve been thrilled to see the support for the Safety & Security Chapter of the Code of Practice. Most frontier AI companies have now signed on to it: @AnthropicAI, @Google, @MistralAI, @OpenAI, @xAI Why this is important: 🧵 1/6
0
0
2
I’ve been thrilled to see the support for the Safety & Security Chapter of the Code of Practice. Most frontier AI companies have now signed on to it: @AnthropicAI, @Google, @MistralAI, @OpenAI, @xAI Why this is important: 🧵 1/6
13
51
308
Reasoning model transcripts are sometimes misleading, and it's plausible that we'll ultimately have good reason to directly train them to look a certain way. But we've gotten lucky with how informative they can be in their current form. Let's not throw that away casually.
A simple AGI safety technique: AI’s thoughts are in plain English, just read them We know it works, with OK (not perfect) transparency! The risk is fragility: RL training, new architectures, etc threaten transparency Experts from many orgs agree we should try to preserve it:
2
7
52
Great summary thread of the code of practice and Safety & Security chapter
The challenge: the AI Act imposes requirements on providers of the most advanced models – like OpenAI, Anthropic, Google, Mistral, Meta. But these requirements are very high level. It says they should “assess and mitigate systemic risk”. But what does that mean?
1
0
2
The Code of Practice is out. I co-wrote the Safety & Security Chapter, which is an implementation tool to help frontier AI companies comply with the EU AI Act in a lean but effective way. I am proud of the result! 1/3
8
31
107