Ian Liu 劉以恆 @duckiesfloat

Causal graph neural networks for healthcare go.nature.com/4g24Usr

Moderna/Merck just ran a 1,137-patient Phase 3 trial where every single dose was unique to that patient's tumor. It worked. The pipeline: surgical resection → whole exome + RNA sequencing → ML neoantigen ranking → mRNA encoding up to 34 patient-specific targets → manufactured and shipped in 8 weeks. One drug, different sequence for every patient. The ML step is worth to note: the algorithm ingests WES + RNA-seq to identify somatic mutations, then predicts which of those will actually be immunogenic, ie, displayed on tumor cell surface and trigger a T-cell response. It's designed to keep learning from accumulated clinical and immunogenicity data across patients, not just per-patient. INTerpath-001 (Stage IIB-IV resected melanoma, 2:1 randomized): combination with pembrolizumab beat Keytruda alone on both primary (RFS) and key secondary (DMFS) at interim. Phase 2b at ASCO 2026 showed 49% reduction in recurrence/death, 59% in distant metastasis/death at 5 years. Phase 3 confirmed both. What this validates: — tumor-specific neoantigen prediction by ML works in a blinded trial at scale — 8-week personalized mRNA manufacturing is operationally real — effect is additive on PD-1 blockade, not redundant This is first positive Ph3 for individualized neoantigen therapy. First positive Ph3 for any mRNA cancer therapeutic. What a great time to live in!! This is the best time for AI & biotech!

Moderna @moderna_tx

4 days ago

Moderna and @Merck today announced positive topline results from the Phase 3 INTerpath-001 trial evaluating adjuvant treatment with intismeran autogene, a novel investigational mRNA-based individualized neoantigen therapy jointly developed by Merck and Moderna, in combination

0 1K 7K 1.4M 744

We can finally talk about it: We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company. We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried.

I am generally a very optimistic, positive person. I do not like to hate on other products. But Workday is, by a wide margin, the worst piece of software I've ever used in my career. I cannot believe that this is a $34 billion company. I pray the AGI gods fix this hot mess.

Hot take… isn’t it kinda crazy that nobody is really using AI Agents? I don’t mean software engineers or AI early adopters. I mean “college friends talking about it in group chat,” the feeling you got when everyone started using Instagram or TikTok. These frontier AI models are *insane* (as are the harnesses & tool calls & the like). And every large tech co has an AI agents platform, not to mention all the YC startups doing vertical agents. Yet all of your friends and family outside of tech — who spend all day staring at their iPhones and get paid to work in browser tabs — don’t really care or find themselves using any AI agents yet. Yes ChatGPT, Claude, etc. are extremely popular… but if you look at the engagement data the vast majority of people are still using these aI chat tools like a glorified Google + Grammarly. That’s why the AGI labs are all pushing desktop apps for Codex, Cowork, etc. so hard to non-technical ppl. And yes exceptions for lawyers and customer service but even those have some asterisks and exceptions to rule. Look I’m not saying the ChatGPT moment for AI Agents is not coming… it most definitely is! Remember we pivoted from Arc to Dia precisely because we believe computing is going to be radically reimagined around these AI primitives. No doubt. But that’s my point: it’s just so surprising it hasn’t happened yet because all of the tech you’d need is there. Again if you stop for a second and think about it… for all the press and money and hype and models and crazy ARR numbers… this “AI Agent” moment does not *feel* like the other breakthrough tech moments we’ve lived through (e.g. think the shift to Stories via Snapchat & Instagram, or shift to on-demand via Uber/Airbnb/Doordash). Which is a long way of saying: if you can figure out the answer to “why” most people don’t care about AI agents yet (and have no enduring interest in using them) — especially since the models and harnesses are here and ready — the answer to that question will allow you to capture a lot of marketshare and make a lot of money in 2027. Theoretically, the tech is ready for AI Agents to totally transform how we work and live our lives… but alas the general public dgaf… that’s the generational puzzle to solve for the next 12 months for anyone not working on the models themselves.

EGRL: Edge generation-guided relation-aware learning for RNA-protein interaction prediction 1. The paper introduces EGRL, a heterogeneous GNN framework for RNA–protein interaction prediction that is designed to work well not only in standard cross-validation, but also in cold-start settings where RNAs/proteins have no observed edges during training. 2. A key idea is to avoid hand-crafted meta-paths: EGRL uses implicit meta-path learning to estimate the importance of each relation type, producing a relation-level semantic vector that is injected into node features via a residual connection (feature enhancement without manual path design). 3. EGRL explicitly models multiple edge types rather than collapsing everything into one graph: RNA–RNA, protein–protein, similarity-based edges, known RPI edges, and an extra relation type for generator-produced soft edges. Each relation is processed with its own GAT layer, then fused (relation-aware message passing). 4. The cold-start module is a graph generator that predicts probabilistic “soft edges” between unseen nodes and existing nodes using only sequence-derived embeddings; these soft edges are added bidirectionally and treated as another relation type during GNN propagation. 5. The generator is trained jointly using a pseudo cold-start strategy: during training, some nodes are masked as “unseen,” and the generator learns to reconstruct their interaction probabilities with a BCE auxiliary loss, combined with the main link-prediction loss via a weighted objective. 6. For final interaction scoring, EGRL uses a multi-feature fusion predictor that concatenates four pairwise features: [h_r, h_p, h_r ⊙ h_p, |h_r − h_p|]. Ablations show this fusion is a major contributor; using only concatenation substantially degrades performance. 7. Input node features come from foundation sequence models: RNA-FM embeddings for RNAs and ESM2 embeddings for proteins, projected into a shared 128D space. This makes the framework naturally suited to sequence-only cold-start inference. 8. Across four benchmarks (RPI369, RPI1807, RPI2241, NPInter2), EGRL is competitive in standard 5-fold CV, with particularly strong results on RPI369 and RPI2241 (e.g., best ACC/MCC reported in the table for those datasets). 9. The main differentiator is cold-start generalization on NPInter2: molecule hold-out achieves AUROC 0.867 and AUPR 0.861, reported as +8.6% AUROC and +5.0% AUPR over the prior best baseline (ZHMolGraph). Under stricter sequence-cluster splits (CD-HIT 80% RNA / 40% protein), performance remains reasonable (AUROC 0.801, AUPR 0.822), suggesting reduced leakage and a harder but more realistic evaluation. 📜Paper: arxiv.org/abs/2608.12906 #Bioinformatics #ComputationalBiology #GraphNeuralNetworks #GNN #RNA #Protein #RPI #LinkPrediction #ColdStart #MachineLearning

Gemini 3.7 Flash is here. It’s stronger for coding, knowledge work, and web development. 🧵

The human brain may be biased toward seeing structure that isn’t there When humans learn a probability distribution from limited data, they do not simply recover its shape. They seem to impose a preferred level of structural complexity on it. Tianyuan Teng and coauthors show this across eight behavioral experiments spanning visual and numerical tasks. Participants estimated the overall probability density reasonably well, but their internal representation was biased toward moderate complexity: they invented multiple clusters when the true distribution was a single Gaussian, yet under-represented structure when it contained many clusters. The computational idea is particularly interesting. The authors model human learning as approximate Bayesian inference with a Distorted Economical Expansion (DEE) process. As observations arrive sequentially, the learner either updates an existing latent cluster or creates a new one. Crucially, the probability of adding structure decreases as the internal model becomes more complex, imposing a memory-like computational cost. DEE reproduces several aspects of human behavior that standard alternatives miss. It outperforms both a conventional Chinese Restaurant Process and a batch variational Gaussian-mixture model, while also predicting the tendency to form several weakly overlapping clusters from data generated by a smooth Gaussian distribution. The behavioral effect is striking. When the true distribution contained one Gaussian cluster, participants reported on average about 3 clusters. And more data did not eliminate the illusion: increasing observations from 10 to 70 actually increased the number of reported clusters. The authors then tested whether this was an artifact of explicitly asking people to report latent structure. It was not. In a preregistered recognition experiment, participants preferred a multimodal distribution generated by DEE over the true Gaussian distribution that produced the observations, even when both had approximately equal likelihood for the observed data. The broader AI lesson goes beyond cognitive science. Inductive bias does not merely help a learner generalize under limited data and computation, it also determines which structures the learner is capable of seeing. Human cognition appears to trade fidelity for a representation whose complexity remains manageable. For AI for Science, that matters: when we design priors, architectures or approximate inference schemes, we may also be deciding which scientific structures a model will systematically invent, ignore or compress. Paper: Teng et al., Nature Communications (2026), CC BY 4.0 | doi.org/10.1038/s41467…

Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I'm proud of these releases. Congrats to @alexandr_wang and the MSL team for all your great work on these models.

Pretraining a Q-function often doesn’t actually help RL finetuning, compared to initializing Q from scratch. We find that pretraining Q-functions on data from diverse policies is critical to see improvements from pretraining. Paper: arxiv.org/abs/2607.27203

Perry Dong @perryadong

3 weeks ago

Pretraining has worked remarkably well across domains We show this doesn’t hold for Q-functions in online RL from a pretrained policy — and propose IPE, a more effective way to learn Q-functions for online RL fine-tuning (1/6)

5 29 303 82K 264

Tomorrow will be my last day at Google after 27 years, and watching it grow from 25 people to 190,000+ has been an amazing journey. Below is a note I shared with many people internally at Google today. An excerpt is: It has been an absolute pleasure to work with you and to help build some of the most widely used and impactful products of all time. As a kid, I dreamed of helping build software that would be used by many people, and Google now has thirteen products used by more than a billion people (amazing!). Our work has had a tremendous impact in the world, and I have been lucky enough to collaborate and form friendships with many colleagues that I deeply admire, respect, and enjoy. It still brings me joy every time I see people out in the world using our products to find information, handle email, translate documents, watch videos, learn new things, navigate and understand the physical world, browse the web, use their phone, run large-scale computations on our infrastructure, ride in an autonomous vehicle, or perform complex tasks with the help of our AI systems. I hope you all share this sense of joy, because it is a shared accomplishment! Thank you to all of my colleagues at Google over many years! Now I'm excited to go start @DiscoLoopAI with my longtime friends and colleagues @Sanjay_Ghemawat, @OriolVinyalsML, and @quocleix. (Updated post: slightly redacted to not have some personal info)

keyboard_arrow_left Previous keyboard_arrow_right Next

NeurlPS 2026这次同行评审的问题，实际上这是一个纯数学结果。 可以发现，这次投稿量以指数增长，而合格审稿人池只能线性增长，且增量几乎全部来自入场不到两年的博士生。 要求一个自己刚被拒过三次、手上压着四篇 deadline 的人，去认真读五篇跟他方向无关的论文，无报酬、无署名、无惩罚。 系统按设计，就会产出这个结果。 顶级会议卖的一直都是合法性凭证。 签证、毕业要求、招聘 headcount、国内的帽子评审，这些才是维持 NeurIPS 存在的事。 所以Zhiping说的直接取消会议是情绪表达，但我也非常能理解他的心情。 据我的观察，学术传播的功能其实某种程度被 arXiv、社区口碑、引用接管了，剩下的壳之所以拆不掉，或许是因为拆掉它之后，没人知道怎么给博士生发学位罢！ 现在相当比例的审稿意见是 LLM 写的，rebuttal 也是 LLM 写的。全流程LLM。 现在觉得致命是这样的一个问题。 这个环节的信息量正在趋近于零，而成本不变。 我相信，等到某个会议第一个公开承认这点、改成 arXiv 预筛，然后强制公开审稿人身份，审稿记录纳入评价体系，它会瞬间吃掉所有声誉。这也是发展的结果。

Zhipeng Wang 🇺🇦 @PKUWZP

3 weeks ago

AI conference peer-reviews are completely broken. For NeurIPS 2026, my batch has very low reviewers responses rate (probably less than 5%), given that I have sent reminders twice. Maybe folks are all busy with their own rebuttals; Also for the @emnlpmeeting ARR submissions, the

22 19 318 111K 48

Are we heading to a reproducibility crisis in medical AI? My latest in @Nature on the (shaky) science of evaluation of medical AI: Medical AI has a measurement problem

keyboard_arrow_left Previous keyboard_arrow_right Next

I ran PageRank, the original Google search algorithm, on all of Wikipedia to find the most important people to humanity. Here’s the top 100. PageRank works by counting the pages that link to a page, and weighing each link by the importance of that page.

🚨这篇论文是近年来 AI 领域少有的具有深刻哲学与科学史厚度的 Position Paper。 2026 年上半年，Demis Hassabis 公开提出Einstein Test，即用 1911 年之前的知识训练 AI，看它能不能独立发现广义相对论。 这篇论文几乎是对该测试的哲学-技术回应。 论文指出，大模型只是个极其优秀的已知知识逻辑插值器，而人类顶级科学家做的是跨越认知断层，即Unexplained Jump。 以 Schmidhuber 等人为代表的派系长期宣称科学发现本质上就是数据压缩，寻找能解释观察数据的最简算法。 这是流行的谬论。 论文认为，如果科学发现全靠数据压缩，爱因斯坦根本不可能发明广义相对论。因为当时根本没有足够的异常观测数据供他去压缩。爱因斯坦是凭空做了一个思想实验，是那个升降机里的失重与引力等效的实验，完成了一次溯因跃迁。 大模型靠概率分布玩归纳注定画不出这个 Jump。 作者在附注里极其刻薄且优雅地嘲讽道：“讽刺的是，AI 对公理符号的这种幻觉，恰恰暴露出它根本无法自动化完成这一Jump的困境。” 我认为，不谈任何私货，这篇论文算是从根本上、从认知科学Peirce 的哲学框架和科学史多个维度证明Jump是一个具身世界模型问题。 你给文本模型堆再多算力，它也只是在符号的影子里打转，永远出不了洞穴。

x.com/i/article/2082…

Spatium: A Protein Language Foundation Model for Spatial Proteomics 1. Spatium is a protein-language foundation model for spatial proteomics that targets a core pain point in the field: datasets measure small, heterogeneous, and only partially overlapping protein panels, making cross-study representation learning and transfer difficult. 2. The model is pretrained on a large compendium: 51.8 million single cells from 70 human spatial proteomics studies, spanning 4 major platforms (CODEX, IMC, CyCIF, MIBI), 18 tissue types, and 18 disease contexts, harmonized into a unified vocabulary of 239 protein markers. 3. Key design choice: rank-based protein tokenization. Instead of using absolute intensities (highly platform- and antibody-dependent), Spatium ranks proteins within each cell by relative abundance, forming a “protein sequence” that is more robust to measurement scale shifts and panel variability. 4. Architecture and training follow a BERT-style masked language modeling objective on protein tokens (plus metadata tokens like technology/tissue/disease), enabling self-supervised learning of contextual protein dependencies and co-expression hierarchies that reflect cell identity. 5. Few-shot cell type annotation: across a CODEX intestine dataset (57 markers) and an IMC CRC dataset (14 markers), Spatium outperforms PCA-based baselines, CytoVI, and an imaging-oriented foundation model (Nicheformer), especially under extreme label scarcity (1–100 labeled cells per class). 6. Beyond accuracy, Spatium reduces spatially coherent “error hotspots”: misclassifications are more spatially isolated compared to baselines, suggesting the learned representations remain stable in complex tissue architectures rather than failing systematically in specific compartments. 7. The remaining confusions are biologically structured: major misclassifications concentrate between related/overlapping protein states (e.g., NK vs CD66+ or CD57+ enterocyte-like states). Simulated marker interpolation shows prediction probabilities shift continuously, indicating Spatium captures gradual identity transitions rather than only discrete classes. 8. Neighborhood-aware embeddings enable spatial niche discovery: by training a head to predict local neighborhood cell-type composition from embeddings, Spatium better preserves local tissue organization than PCA/CytoVI, and clustering these neighborhood embeddings yields interpretable niches with coherent marker enrichment. 9. In CRC IMC, Spatium identifies stromal and tumor niches with distinct programs (e.g., ACTA2-enriched activated CAF-like niche vs collagen-enriched ECM-remodeling niche). Niche labels transfer across cohorts via embedding similarity while preserving marker signatures; notably, a collagen-rich ECM-remodeling niche is preferentially enriched in recurrent CRC and appears as compact substructures within broader ACTA2+ stromal networks. 10. Protein imputation: Spatium reconstructs masked proteins from incomplete panels while preserving spatial organization (e.g., Pan-cytokeratin, CD8, Podoplanin in HCC CODEX). Many markers reach high Spearman correlation (~0.85+), performance degrades gracefully as more proteins are masked, and shuffling within-cell protein order largely abolishes accuracy—supporting the importance of learned protein-state structure. 11. Unsupervised disease-state discovery: in AML vs normal bone marrow CODEX (50-marker panel), embeddings (adapted only with MLM, no labels) recover structured cellular landscapes, detect disease-associated composition shifts, and reveal HSPC expansion and altered HSPC neighborhood contexts; niche analysis further identifies organized hematopoietic microenvironments and rare HSPC-like states with reproducible marker programs (e.g., SPINK2-associated patterns). 12. Multimodal transfer: Spatium supports cross-modal label transfer from scRNA-seq to spatial proteomics by mapping genes to protein tokens (tonsil data shows ~0.787 accuracy), and can be fused with histology embeddings (Prov-GigaPath on H&E) to produce joint representations that preserve tissue architecture while refining cellular organization. 📜Paper: biorxiv.org/content/10.648… #SpatialProteomics #FoundationModels #ProteinLanguageModels #SelfSupervisedLearning #Transformers #TumorMicroenvironment #ComputationalBiology #SpatialOmics #SingleCell #Bioinformatics

Google Deepmind argues that LLMs can never make real scientific discoveries. They published a paper breaking down Albert Einstein’s private view of scientific discovery. In a famous letter to his friend Maurice Solovine, Einstein drew a diagram of how science actually happens. It is a cyclical loop. First, you experience raw sensory data. Then, through a mysterious, non-logical act of intuition, you make an intuitive "jump" to abstract axioms. Finally, you use strict logical deduction to derive consequences from those axioms. Generative AI has completely mastered two-thirds of this loop. • Induction: Statistical pattern matching across billions of tokens. • Deduction: Formal proof generation, like AlphaProof solving complex math Olympiads. AI can crunch data and it can prove theorems. But it cannot make the jump. The paper argues that AI completely lacks Abduction, the generation of novel explanatory hypotheses when observational data is scarce. The prevailing tech myth says that "creativity is just data compression." That if you feed an LLM enough text, scientific breakthroughs will naturally pop out. Einstein’s formulation of General Relativity proves that is a delusion. When Einstein formulated relativity, the observational data didn't demand a new physics framework; classical mechanics was still massively successful. The breakthrough required a conceptual rupture. An intuitive leap from physical reality to a brand-new set of foundational axioms. An LLM can execute the math once the axioms are given. But it is structurally incapable of formulating those premises on its own. It can interpolate inside existing human thought, but it cannot transcend it. The translation of physical reality into formal axioms remains the absolute, hard bottleneck of artificial scientific invention. We can build models with trillions of parameters. We can scale compute into the stratosphere. We can make the calculator infinitely fast. But until we solve grounding, the machine can process all the data in the universe. It still can't make the jump.

𝗡𝗲𝘂𝗿𝗼𝗻 𝗿𝗲𝘀𝗽𝗼𝗻𝘀𝗲𝘀 𝘄𝗶𝘁𝗵𝗶𝗻 𝗿𝗲𝗴𝗶𝗼𝗻𝘀 𝗮𝗿𝗲 𝗻𝗼𝘁 𝗰𝗮𝘁𝗲𝗴𝗼𝗿𝗶𝗰𝗮𝗹, 𝗺𝘂𝗹𝘁𝗶-𝘀𝗲𝗹𝗲𝗰𝘁𝗶𝘃𝗶𝘁𝘆 𝗶𝘀 𝘁𝗵𝗲 𝗻𝗼𝗿𝗺 doi.org/10.1038/s41586…

x.com/i/article/2077…