Samuel Müller

This might be the first time after 10 years that boosted trees are not the best default choice when working with data in tables. Instead a pre-trained neural network is, the new TabPFN, as we just published in Nature 🎉

TabPFN 2.6 just went live and claimed the #1 spot on TabArena. This version outperforms TabPFN-2.5 with a 76.5% overall win rate, and 84.6% on regression tasks. But this release is really an early look at where we're headed... More soon!

@puhsuuu @eremeev_d42 @FrankRHutter @LeoGrint @LennartPurucker @DHolzmueller @JingangQu @MarineLeMorvan @GaelVaroquaux @PengCui18 I am very happy about this paper. In Freiburg, we had a master thesis to try and build a graph PFN before, but we didn’t make it work, so even happier to see they figured out how to do it.

@FrankRHutter @SamuelMullr @LeoGrint @LennartPurucker @DHolzmueller @JingangQu @MarineLeMorvan @GaelVaroquaux @PengCui18 continued training of tabular PFNs is sota on graph tasks (representative of actual graph data in industry, not citation networks). Thought you may be interested

Graph foundation model with SOTA results on real-world graphs! Our “GraphPFN: A Prior-Data Fitted Graph Foundation Model” paper recently got a major update, with better ICL performance, new ablations, code improvements and more! 🧵1/11

New NanoTabPFN training speed record: Beating RF in 10.10 minutes Previous record: 54.41 minutes Changelog: - Scaled Dot-Product Attention rewrite with explicit QKV - Pre-norm transformer blocks - bfloat16 autocast - Increase learning rate - Increase embedding size - Reduce attention heads This record is by @carterprince03

Much love to Alex Pretti and Renee Good—but remember—ICE has killed 9 people in 2026. You know the names of the 2 white people they've killed. ICE has also killed a Black man named Keith Porter, a Cambodian named Parady La, and five Latinos named Heber Sanchaz Domínguez, Victor Manuel Diaz, Luis Beltran Yanez-Cruz, Luis Gustavo Nunez Caceres, and Geraldo Lunas Campos. ICE is on pace to kill more than 100 people this year. Abolish ICE. Impeach Noem. Prosecute those who committed these crimes.

As tens of thousands across America protest the violence that ICE sows with impunity, federal agents shot and killed another person in Minneapolis today.

ICE terrorizes our cities. ICE puts us all in danger. Abolish ICE.

Big news: Prior Labs is coming to the US.🇺🇸 We’re opening offices in New York and San Francisco.

To everyone in tech who supported Trump in 2024: now would be a good time to reconsider. The parallels to the early stages of Germany’s darkest history are becoming hard to ignore. To be clear, Trump is NOT Hitler. But look at the intimidation, normalization of violence, centralization of power, threats against other nations,... The fact that Trump is less evil and less racist than Hitler shouldn't reassure us. Time to call out the emperor’s clothes. I refuse to believe that anyone in their right mind can deny that Trump is a narcissistic, egomaniacal, incompetent, corrupt, and deeply immoral person.

A trade war over Greenland? Greenland? It would be funny if it weren't tragic. But I suppose it keeps the Epstein files out of the news. Man the stuff about Trump in there must be bad...

Honored to announce that Yann LeCun @ylecun is joining Prior Labs’ Scientific Advisory Board.

Question to all those who said Trump would be better for the economy: is attacking the Fed’s independence what you had in mind?

The stuff these ICE guys are saying to the protestors is wild. In any functioning society, a cop saying to a protestor "I'm going to find when you live and give you a visit" or "you don't want to end up like Renee do you?" would be the end of your career.

This is not okay. Regardless of her citizenship (and witnesses indicated she is a U.S. citizen), this kind of thuggish behavior by ICE agents is completely over the top. Choking her & gouging her eyes is excessive force. The government is supposed to be in service of all of us.

This is completely not okay, and we can't become numb to repeated instances of illegal and unconstitutional action by government agencies. The recent days have been horrific.

This paper really is groundbreaking. It solves a long-standing embarrassment in machine learning: despite all the hype around deep learning, traditional tree-based methods (XGBoost, CatBoost, random forests, etc) have dominated tabular data—the most common data format in real-world applications—for two decades. Deep learning conquered images, text, and games, but spreadsheets remained stubbornly resistant. This paper's (published in Nature by the way) main contribution is a foundation model that finally beats tree-based methods convincingly on small-to-medium datasets, and does so very fast. TabPFN in 2.8 seconds outperforms CatBoost tuned for 4 hours—a 5,000× speedup. That's not incremental; it's a different regime entirely. The training approach is also fundamentally different. GPT trains on internet text; CLIP trains on image-caption pairs. TabPFN trains on entirely synthetic data—over 100 million artificial datasets generated from causal graphs. TabPFN generates training data by randomly constructing directed acyclic graphs where each edge applies a random transformation (using neural networks, decision trees, discretization, or noise), then pushes random noise through the root nodes and lets it propagate through the graph—the intermediate values at various nodes become features, one becomes the target, and post-processing adds realistic messiness like missing values and outliers. By training on millions of these synthetic datasets with very different structures, the model learns general prediction strategies without ever seeing real data. The inference mechanism is also unusual. Rather than finetuning or prompting, TabPFN performs both "training" and prediction in a single forward pass. You feed it your labeled training data and unlabeled test points together, and it outputs predictions immediately. There's no gradient descent at inference time—the model has learned how to learn from examples during pretraining. The architecture respects tabular structure with two-way attention (across features within a row, then across samples within a column), unlike standard transformers that treat everything as a flat sequence. So, the transformer has basically learned to do supervised learning. Talk to the paper on ChapterPal: chapterpal.com/s/a1899430/acc… Download the PDF: nature.com/articles/s4158…

@karpathy But don't you think it will lead to much more demand for driving especially in urban areas? Actually making them worse for everyone by making Taxis more (too) affordable.

I am unreasonably excited about self-driving. It will be the first technology in many decades to visibly terraform outdoor physical spaces and way of life. Less parked cars. Less parking lots. Much greater safety for people in and out of cars. Less noise pollution. More space reclaimed for humans. Human brain cycles and attention capital freed up from “lane following” to other pursuits. Cheaper, faster, programmable delivery of physical items and goods. It won’t happen overnight but there will be the era before and the era after.

check out the technical report storage.googleapis.com/prior-labs-tab…

TabPFN 2.5 is out 🔥

Thrilled to share our new paper, TempoPFN! 🚀 TempoPFN is a new foundation model trained ENTIRELY on synthetic data. Most Time Series models use massive, proprietary real-world datasets. We asked: Can we compete with just a Linear RNN and 100% fake data? (Spoiler: yes)

This paper shows a simple linear RNN trained on synthetic time series can do strong 0-shot forecasting. It uses a linear block named GatedDeltaProduct that keeps a running state across steps. Training and inference happen in parallel over the full sequence, so no windowing is needed. The big deal is that a small linear RNN, trained only on fake yet diverse signals, can match or beat many real-data models. State weaving passes each layer's final state to the next to share horizon context. The input mixes past times with values and future times without values into one sequence. The model outputs quantiles in 1 pass, which are ranges rather than a single guess. Pretraining uses only synthetic series like Gaussian processes, sawtooth ramps, steps, spikes, audio rhythms, and regime switching patterns. Augmentations such as TS-Mixup, censoring, quantization, amplitude tweaks, shocks, and random convolutions add variety and robustness. On Gift-Eval it beats synthetic-only baselines and rivals many models trained on real data. It stays efficient and reproducible with parallel compute and a fully open data pipeline. So a lean linear RNN delivers strong 0-shot forecasting without touching real datasets. ---- Paper – arxiv. org/abs/2510.25502 Paper Title: "TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-shot Time Series Forecasting"