Foundation models are brilliant reasoning engines, "the Einsteins of the engineering world," but they "know zero facts about your actual business." SVP and Technical Fellow Ashish Kumar of @MongoDB tells @JasonTAndersen that real-time signals + historical records + business rules = context AI can trust. $MDB

Foundation models are reshaping computational biology. Adapting models to a specific task is nontrivial, so to reduce the difficulty of building these workflows, @nvidia BioNeMo Recipes provide step-by-step training recipes built on familiar PyTorch, Hugging Face, and other patterns. This post walks through two case studies that show how the same parameter-efficient and readable recipe applies across biological modalities. Read the full post:

"Foundation models are now better technically than most people." "OpenAI and Anthropic can subsume any business that only relies on technical talent." "Regulatory moats take years and a lot of trust. There's a lot involved beyond just technology that cannot be vibe-coded." @neerajprasad on why @MoonPay acquired Helio, Iron, Sodot, and Dawn in 18 months: "There are four moats in launching any product today. Technical talent, capital, regulatory licensing, and distribution. At MoonPay we have all four." "We genuinely are playing to win." Neeraj's full thesis on the un-vibe-codeable moats in crypto infrastructure:

Tabular foundation models have mostly just been predicting a single target variable. A new paradigm learns the relationships across an entire table instead. Title: LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence URL: 📝 Overview Contextual Mechanism Networks (CMN) learn the full joint distribution p(x, y) across a table, not just the target y, letting a single model handle classification, regression, missing-value imputation, and causal discovery. ❗ Problem it solves Existing PFN-based methods specialize in conditional prediction of the target only, without explicitly modeling the dependency structure between variables. ⚙️ Methodology Cell-level embeddings feed a 24-layer dual-axis transformer, trained via "Context-Conditional Masked Modeling" that masks part of the query rows and jointly predicts both features and targets. It's pretrained at scale on synthetic data generated from structural causal models. 📊 Results On TabArena (51 datasets), it hits an Elo of 1935 vs. 1818 for the baseline; on BCCO (156 datasets), cumulative wins are 2.06x higher. It also took 1st place on F1 score for causal skeleton recovery on all 6 tested datasets, all while using 4x fewer parameters than TabFM. 🔬 Use cases Since the same model handles classification, regression, and causal discovery, it could replace much of a data analysis pipeline with a single foundation model. #TabularData# #MachineLearning#

Robot foundation models look great in simulation, but fail the moment the camera angle or lighting changes. Turns out they might just be cheating on looks. Title: Breaking the Vision-Action Shortcut: Latent Interface Training for Generalizable Robotics Foundation Models URL: 📝 Overview This paper proposes Latent Interface Training (LIT), a method that breaks the "vision-action shortcut" where models exploit task-irrelevant visual cues that just happened to correlate with actions in the training distribution. ❗ Problem it solves Robot foundation models perform well in-distribution but degrade sharply under visual distribution shift — different camera viewpoints, lighting, or sensor noise. ⚙️ Methodology It's a two-stage recipe: first pretrain the action expert using only pose goals, no images at all, then route all visual information through 100 latent tokens supervised with a pose-reconstruction loss. 🤖 Use cases It applies to 4 different robot foundation models — π0.5, MolmoAct2, FAST-WAM, and ImageWAM — making it a genuinely framework-agnostic technique. 📊 Results On LIBERO-Plus out-of-distribution evaluation, every model improved by 3.87 to 10.70 points. In real-world robot manipulation, it gained +16.7 points under lighting changes, and up to +60 points on a specific task. Relying on real spatial information instead of surface appearance feels like a direct path to more trustworthy robots in the field. #Robotics# #VLA#

World Foundation Models (WFM) are the future of Physical AI, and we believe this future should be open-sourced. We are partnering with @Valeo_Group to build one of the largest open-source, multi-camera WFMs — redefining how Physical AI interacts with the real world.

Crazy how far robot foundation models have come in terms of true fine motor skills: Without task-specific training, no sim-to-real transfer, and... fully autonomous ❗ The dual-arm robot places an O-ring onto a hydraulic rod and removes it again in a fully autonomous and repeatable manner, with frequent contact, precision, and without task-specific training. The model uses short physical prompts (a demonstration lasting just a few seconds) for one-shot learning and adapts to new manipulations in just a few gradient steps. No sim-to-real transfer is necessary; instead, pretraining is performed on real interaction data. It is precisely these types of tasks (O-rings, seals, small components) that are classic bottlenecks in assembly. If this can be reliably scaled up, the cost-effectiveness of robots in manufacturing will change significantly. Source: @GeneralistAI / Congrats to @peteflorence, @andyzengineer and the team on the GEN-1.5 demo!

“MindTopo: Can Foundation Models Reason in Topological Space?” Current models can recognize topological structure, but struggle to preserve it through sequential actions. And MindTopo exposes this reasoning–planning gap across 11K procedurally generated examples. They found that even image and video world models often reach plausible endpoints through topologically invalid trajectories.

GPT-3 level robot foundation models already exist It’s just not obvious because GPT-3 was only partially successful most of the time. That’s ok for some knowledge work but not for the work we need robots to do GPT-5 level intelligence for robots we could see within a year because research compounds We won’t necessarily have something that looks exactly like the ChatGPT moment for robotics, but the FigureAI livestream was close to that for a lot of people