Machine Learning

Runway

2026-10-01 14:30:53

Runway says Praxis-1 brings internet video pretraining to robots with results close to specialist datasets

Runway has introduced Praxis-1, a robotics model that applies techniques developed for AI video training to real-world robots. The system learns from large volumes of ordinary internet video, using those clips to model how objects move and how people complete tasks, then converts that knowledge into robot actions. In an object placement test, Runway said robots pretrained on internet video posted an average error of 16.1 centimeters. When trained instead on specially collected robot teleoperation video, the average error was 16.0 centimeters. The company said the benchmark covered 93 evaluations, with the two training approaches finishing nearly even. Runway argued that if this approach scales, the amount of usable training data for robotics could expand by orders of magnitude, since teleoperated robot data requires hardware, physical space, and human labor, while ordinary video is already abundant online. Praxis-1 has been tested on multiple robots from Noble Machines, Standard Bots, and Ultra, including robotic arms, dual-arm systems, and mobile robots. The company also showed the same policy continuing a task after moving from a studio setting to a kitchen. Praxis-1 is currently available only through an early access application, while model weights are planned for release in the coming months. Parameter count and fuller independent evaluations have not yet been disclosed.

740

NVIDIA

2026-10-01 07:03:53

NVIDIA launches Kumo Tabular, a tabular foundation model that predicts new rows in one forward pass

NVIDIA has introduced Kumo Tabular, a family of tabular foundation models built for classification and regression tasks. The company said the models use labeled rows as context and can predict a new row in a single forward pass, without training, hyperparameter tuning, or feature engineering. The lineup includes small, medium, and large versions, with parameter counts ranging from about 28 million to 215 million. Kumo Tabular runs through NVIDIA’s open-source Structured Data Models, or SDM, library, and its weights are released under the OpenMDW-1.1 license, which allows commercial use. NVIDIA said the models were fully pre-trained on synthetic tables sampled from structural causal models, while also incorporating real-world data patterns such as missing values and high-cardinality categories. According to the NVIDIA team, Kumo Tabular ranked first on the TabArena benchmark with an Elo score of 1950 under default settings. The team also reported that it runs 17 times faster than LimiX-2 on a single RTX 6000 Pro. The item was cited by MarkTechPost in the Techub News report.

590

NVIDIA

2026-09-30 17:17:22

NVIDIA opens applications for its 2027-2028 graduate fellowship program

NVIDIA has opened applications for its 2027-2028 Graduate Fellowship Program, according to Techub News, citing the NVIDIA Blog. The program is open to PhD students worldwide and supports research in AI, machine learning, autonomous driving, computer graphics, robotics, healthcare, and high-performance computing. Selected students can receive up to $60,000 in funding, along with mentorship and technical support from NVIDIA. The application deadline is Oct. 30, 2026. Winners are also required to complete an internship at an NVIDIA research office in the summer of 2027. NVIDIA said the fellowship program has been running since 2002 and has awarded more than 220 grants with a total value of about $8 million.

580

McDonald's

2026-09-30 11:19:32

McDonald’s uses AI pricing engine to adjust menu prices by willingness to pay

McDonald’s is using an artificial intelligence pricing engine to dynamically adjust the prices of menu items including the Big Mac, according to a Techub report citing Decrypt. The system assigns an "optimal price" to each individual item at each restaurant rather than applying a uniform price across locations. As a result, price gaps have widened even between stores that are only two miles apart. The report said the machine learning model is designed to maximize revenue and updates menu prices in real time by analyzing local market data. The move points to a more granular pricing approach at the store level, with item pricing tied to local demand signals and what customers are willing to pay.

600

TMLR

2026-09-26 04:49:11

TMLR editor tested whether authors understood their own papers. All 10 flagged submissions were rejected

Transactions on Machine Learning Research, or TMLR, has published a firsthand account of a small editorial experiment that speaks to a broader problem facing academic publishing in the AI era. In a Sept. 16 blog post, co-editor-in-chief Nihar B. Shah of Carnegie Mellon University said that during his Aug. 14-28, 2026 rotation, he pulled 10 submissions that were already headed for desk rejection and invited the authors to discuss their papers before any external review. The result was stark. One paper was withdrawn, one author declined the call, one no-showed, and Shah ultimately spoke with authors of seven papers. Only one author answered all questions. Three could explain the high-level idea but struggled on technical details. Three others could not answer even basic questions about the problem setup, notation, or where the abstract’s claims were supported in the paper. Those three were all single-author submissions. Even the one strongest case was rejected after Shah found a major error in a central conclusion, though resubmission was allowed after correction. The post also outlines how TMLR is responding to a surge in submissions: author quotas, AI-generated reliability reviews, and a stronger requirement that papers clearly communicate their findings. Shah argues the issue is not whether AI was used, but whether named authors can stand behind the work.

600

Pluralis

2026-09-25 19:53:07

Pluralis team has three papers accepted to NeurIPS 2026

Techub News reported that @coinfund said in a post on X that the team at AI research company Pluralis had three papers accepted to NeurIPS 2026. CoinFund also congratulated Pluralis co-founder Alexander Long and the broader team. NeurIPS is widely regarded as a top academic conference in artificial intelligence and machine learning. The update did not disclose the paper titles or research topics, and the announcement was limited to the acceptance result and the congratulatory message shared by @coinfund.

510

TypeSafe AI

2026-09-21 11:23:00

TypeSafe AI’s Jev spread to about 13% of Vercel’s paid teams within 24 hours of launch

TypeSafe AI came out of stealth on Sept. 15, 2026 and introduced Jev, a model built for structured decisions rather than text generation. Within 24 hours, about 13% of Vercel’s paid teams had started using it, and Vercel described the rollout as one of the fastest-adopted launches in its history. Native support from Cloudflare, LangChain, and Langfuse followed within days. Jev is positioned as a “System One Model,” designed to return probabilities, scores, and confidence values instead of natural-language answers. TypeSafe says the model is 40x to 200x faster than comparable large language models on classification-style tasks, with end-to-end latency of 70 to 500 milliseconds and costs that are 40x to 400x lower. Its pricing is unusual as well: input tokens cost $0.042 per million, while output tokens are free. The company says Jev is not a replacement for large language models. It is aimed at narrow, high-frequency decision tasks such as spam detection, moderation, and routing. Third-party tests cited by developers and platform engineers point in the same direction on speed and cost, though long-term production reliability, retention, failure rates, and the durability of its business model remain unclear. TypeSafe has also not disclosed Jev’s underlying architecture.

1370

FLock

2026-09-13 09:13:39

FLock Research gets three papers accepted to ECML PKDD 2026 on complex LLM reasoning

FLock said three papers from its research team have been accepted to ECML PKDD 2026, a major international conference in machine learning, data mining, and knowledge discovery in Europe. The company said the papers focus on recurring problems in complex reasoning by large models, including shortcut-taking, reasoning errors, and drift away from the original objective during long multi-step chains of thought. According to FLock, one of the studies helps models stay closer to their original line of reasoning during extended multi-step inference. It reduced reasoning drift by 63% and improved average accuracy by 7.4 percentage points across models of different sizes. The other two papers target reasoning on unfamiliar problems and the efficiency of partial correction after a model detects an error in its reasoning process. FLock Research said it continues to conduct foundational work around large-model training, inference, and reliability, with the stated goal of improving accuracy, stability, and generalization on complex tasks so models can move from answering questions to reliably solving more complex problems.

1350