Toby Mao

@brennan__simon Real data reflects the true distribution of language and knowledge, while synthetic data is just a recursive projection of the prior that current models already believe. The improvement is marginal and could risk long-tail problem that would result in model collapse.

How does the scientific community feel about the role of generative AI in science? Not LLMs, but generating synthetic data in situations where real data is sparse due to cost or other reasons, to improve sample size for analysis/model training? Optimistic? Dubious?

@brennan__simon Cautiously optimistic. In training LLM, syn data is usually used in post-training, especially in finetuning reasoning whereas real data predominates pre-training. However, in biomedical research, the bottleneck is scarcity of real data that represent diverse biology.

@Roth_Lab @AustinMHartman congrats @austinhartman

Genome wide perturb seq creates powerful perturbations x transcripts datasets. This unveils new biology - but also reveals new off-target effects. New work from @AustinMHartman, a massively talented PhD student in our lab, systematically identifies seed driven off-target effects in genome-wide perturb seq exps. (1 of 3) biorxiv.org/content/10.648…

The next inflection point in virtual cell modeling will require large-scale perturbative data. To get there we are partnering with @tahoe_ai and @officialbiohub to help us generate a dataset to push us forward: arcinstitute.org/news/tahoe-arc…

Link to full interview: youtu.be/mBPbasxrHR4 rss.com/podcasts/decod…

How do we move beyond chain-of-thought to build models that reason more deeply and reliably? Is generalization, rather than scale, now the real bottleneck? I’m excited to talk with @Muennighoff, on the frontier of model reasoning and reinforcement learning. #ai #LLMs #reasoning

@Muennighoff Niklas Muennighoff: x.com/Muennighoff scholar.google.com/citations?user…

In this conversation, we discuss: Model reasoning research beyond chain-of-thought The future of reinforcement learning Test-time scaling and inference-time compute tradeoffs AI’s impact on the job market, the “AI bubble,” and separating signal from hype

#virtualcell #ai #aihealth

In Silicon Valley, “virtual cells” are suddenly everywhere. Meta and CZI recently went all in, signaling that this is no longer a fringe research direction. So what is virtual cell? Check out my new blog about virtual cell: yuncongtobymao.com/blog/virtual-c…

Thanks so much for this opportunity for me and @_AravindK_ Dr. Kohane, looking forward to future collaborations!

We've got a #Twofer for #TT!! First, we tested 6 different #ML models for Grade IV #glioma prognosis, with improvement after feature selection and high accuracy. They identified factors such as adjuvant TMZ💊, confusion, and MGMT status🧬 as predictive in survival!

Today is a bittersweet moment with Toby Mao leaving my group and pursing his next step as a prestigious @FulbrightPrgrm fellow as he waits to hear back from med schools. Congratulations to Toby and we wish him luck! @CleClinicCIRC @CleClinicHVTI