timon (@tdsone3) header
timon (@tdsone3) profile picture

designer chromosomes for everyone

london
Joined October 2020
Photo by timon (@tdsone3)Photo by timon (@tdsone3)Photo by timon (@tdsone3)Photo by timon (@tdsone3)Photo by timon (@tdsone3)Photo by timon (@tdsone3)Photo by timon (@tdsone3)Photo by timon (@tdsone3)Photo by timon (@tdsone3)Photo by timon (@tdsone3)
2031 trete ich dann an mit meiner mitte-links-akzelerationistischen und hypermodernistischen yimby abundance bewegungspartei names "überfluss-für-alle" (ÜFA)© dieser akute mangel an geilen politik angeboten ist wirklich nicht mehr auszuhalten – at Berlin, Germany
10
2
2
64
2,444
Replying to @vintweeta
I don't think a lack of belief in LLMs in the big problem with the AI x bio field? There is plenty of blind faith to go around and then some😀
2
1
10
1,324
Feel free to reach out if you want to collaborate or for feedback/suggestions. cc @abdul, Dennis Pultz, @KuanHaoChao, @ProfTomEllis
1
1
100
If u have deep expertise in genomic data processing with agentic platforms, DNALMs, S2F models & interested in joining us please reach out. I'll be hiring 1-2 postdocs/staff scientists for this project. Please RT 7/
1
35
66
14,452
Offensively obvious … in hindsight (!)
We discovered a third pretraining axis beyond parameters and data: exploration. Scaling exploration monotonically improves existing models across images/video/language, and unlocks end-to-end generation. In the simplest case, it's just a for loop. Introducing Explorative Modeling. TLDR: - Gains from exploration grow with scale: 7%→36% as data scales, 13%→23% as parameters scale, and gains double at 3× the compute - Adding exploration to ~SOTA baselines improves data efficiency by 6.2×, FLOP efficiency by 4.1×, parameter efficiency by 47%, and hits a near-SOTA 1.43 unguided FID on ImageNet - Exploration lets you trade training compute for generalization, and scales how end-to-end your generative model is - End-to-end Explorative Models (XMs) match diffusion performance on control tasks with up to 256× less inference compute 🧵Thread:
1
5
1,270
Our lab’s latest on synthetic genomics “Synthetic Combinatorial Minimisation of Cell Cycle Control” is up now on BioRxiv - covering ambitious yeast #synbio cell cycle work led by Anastasiya Malyshava and co-supervised by @UniOfSurrey’s Matteo Barberis.
3
23
1
86
7,539
"OpenAI is investigating an unprecedented security incident." The incident is that their own internal OpenAI model 1.) exploited a zero-day and escaped the sandbox and then 2.) autonomously hacked Hugging Face production using stolen credentials plus another zero-day attack !!!
We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation. Sharing preliminary findings to help defenders understand emerging risks: openai.com/index/hugging-fac…
5
8
1
118
6,032
What has been the solution to increase the share of power for young people in an aging democracy? Has there ever been one? @_w0bi @mmjukic The only convincing solution seems to be more young politicians since votes have already been 49% by 60+ y/o last election (GER).
1
2
121
Yesterday at @ISCB_RegSys I introduced Cherimoya: a lightweight sequence-to-function (S2F) model that achieves state-of-the-art performance. We've been using it internally at the Programmable Genomics Laboratory for several months and are excited to share it with y'all now!
2
31
2
118
15,222
There’s very few fields in AI x Bio where deep learning is yet to hit in full effect! Go talk to Matt :)
any cool engineers who are at icml wanna cure cancer?
1
1
3
466
I think the biggest things we’re currently missing from current models like AlphaGenome and Borzoi:
2
3
6
1,615
I think Shred-seq from @JonasKoeppel and @sudpinglay scRNA-seq recombination paper create some early starting points for this in human cell lines (+ obvs exogenous genomes if possible like done in yeast).
1
2
194
•Co-training on MPRA data: it must have crossed everyone’s mind at this point but co-training track coverage models with MPRA data where we have millions and millions of variants as opposed to the few hundreds of thousands for coverage models should boost performance.
1
144