Things were not looking great for OpenAI at the end of last year. The company had been struggling with major delays on its long-awaited GPT-5 and hemorrhaging key talent--notably, Chief Scientist Ilya Sutskever, Chief Technology Officer Mira Murati, and Alec Radford, the researcher who'd set the company on the path of developing GPTs in the first place. Several people who left either joined OpenAI competitors or launched new ones. The start-up's relationship with Microsoft, its biggest backer and a crucial provider of the computing infrastructure needed to train and deploy its AI models, was being investigated by the Federal Trade Commission. And then there was Elon Musk.

The president's comments on the oil price came after he spoke to Saudi Crown Prince Mohammed bin Salman on Wednesday. According to Saudi State media Bin Salman pledged to invest as much as 600bn in the US over the next four years, however this figure was not mentioned in the White House statement after the call. Despite the cordial exchange, Trump said he would be asking "the Crown Prince, who's a fantastic guy, to round it out to around 1tn". The price of crude fell by 1% following Trump's comments. According to David Oxley, Chief Climate and Commodities Economist at Capital Economics these comments are in keeping with Trump's desire for lower gasoline prices.

Scientists have unveiled the world's fastest computer in California that will be used to secure America's nuclear weapons stockpile. The 600 million exascale supercomputer, called'El Capitan,' is only the third of its kind in the world. That's equivalent to the processing power of about one million high-end smartphones working simultaneously, researchers said. El Capitan launched at the Livermore National Laboratory (LNNL) in November 2024 and was officially announced to the public on January 9. It will primarily focus on national security, including nuclear data and weapon testing, high-energy-density physics, materials discovery and other sensitive or classified tasks.

Crown Prince Mohammed bin Salman of Saudi Arabia told President Trump on Wednesday that the kingdom intends to increase its investment and trade with the United States by at least 600 billion over the next four years, according to the official Saudi Press Agency. The crown prince, the de facto leader of Saudi Arabia, told Mr. Trump that his new administration had the ability to create "unprecedented economic prosperity" in the United States, and that the kingdom wanted to participate, a statement from the Saudi Press Agency said. There was no immediate confirmation of the call from the White House. Mr. Trump has promised to accelerate investment in the United States, particularly to help revive manufacturing. While campaigning last year, he said he would use a mix of tax cuts and tariffs to force companies to invest in the United States.

The Israeli military's reliance on Microsoft's cloud technology and artificial intelligence systems surged during the most intensive phase of its bombardment of Gaza, leaked documents reveal. The files offer an inside view of how Microsoft deepened its relationship with Israel's defence establishment after 7 October 2023, supplying the military with greater computing and storage services and striking at least 10m in deals to provide thousands of hours of technical support. Microsoft's deep ties with Israel's military are revealed in an investigation by the Guardian with the Israeli-Palestinian publication 972 Magazine and a Hebrew-language outlet, Local Call. It is based in part on documents obtained by Drop Site News, which has published its own story. The investigation, which also draws on interviews with sources from across Israel's defence and intelligence establishment, sheds new light on how the Israel Defense Forces (IDF) turned to major US tech companies to meet the technological demands of war. After launching its offensive in Gaza in October 2023, the IDF faced a sudden rush in demand for storage and computing power, leading it to swiftly expand its computing infrastructure and embrace what one commander described as "the wonderful world of cloud providers".

Thank you for your submission to NeurIPS. After the author response and discussion, the reviewers and I are in agreement that this work presents an interesting and substantial contribution to the work on provably robust adversarial learning. The extension of such methods from the typical NN setting to one of boosted decision stumps is an interesting one, and certainly worthy of publication. The author response in particular was good at addressing the points of one of the initially most negative reviewer, and it would be good to include these points into the final version.

Weaknesses: - The theoretical foundation builds off a number of assumptions, specifically assumption 5, that may limit its applicability in real world scenarios. But there is no discussion on how this attack set is generated, its properties, or its effect on the proposed methodology.

The paper proposes a restricted black-box attach for GNN, which is claimed to be more applicable in real world scenarios. After extensive discussion and having read the reviews and the rebuttal it is clear that the novelty of the approach is acknowledged across the board. This leaves the main weakness of the work in the experimental validation. Although a fair comparison with other approaches is hard to make due to the more challenging setting of this work, an in depth analysis of the method and the various settings would have been desirable. Some of the choices seem not very realistic as well, i.e. assuming to be able to perturb 1% of the nodes.

Text classifiers suffer from small perturbations, that if chosen adversarially, can dramatically change the output of the model. Verification methods can provide robustness certificates against such adversarial perturbations, by computing a sound lower bound on the robust accuracy. Nevertheless, existing verification methods incur in prohibitive costs and cannot practically handle Levenshtein distance constraints. We propose the first method for computing the Lipschitz constant of convolutional classifiers with respect to the Levenshtein distance. We use these Lipschitz constant estimates for training 1-Lipschitz classifiers. This enables computing the certified radius of a classifier in a single forward pass. Our method, LipsLev, is able to obtain $38.80$% and $13.93$% verified accuracy at distance $1$ and $2$ respectively in the AG-News dataset, while being $4$ orders of magnitude faster than existing approaches. We believe our work can open the door to more efficient verification in the text domain.

Modern neural networks (NNs) often do not generalize well in the presence of a "covariate shift"; that is, in situations where the training and test data distributions differ, but the conditional distribution of classification labels remains unchanged. In such cases, NN generalization can be reduced to a problem of learning more domain-invariant features. Domain adaptation (DA) methods include a range of techniques aimed at achieving this; however, these methods have struggled with the need for extensive hyperparameter tuning, which then incurs significant computational costs. In this work, we introduce SIDDA, an out-of-the-box DA training algorithm built upon the Sinkhorn divergence, that can achieve effective domain alignment with minimal hyperparameter tuning and computational overhead. We demonstrate the efficacy of our method on multiple simulated and real datasets of varying complexity, including simple shapes, handwritten digits, and real astronomical observations. SIDDA is compatible with a variety of NN architectures, and it works particularly well in improving classification accuracy and model calibration when paired with equivariant neural networks (ENNs). We find that SIDDA enhances the generalization capabilities of NNs, achieving up to a $\approx40\%$ improvement in classification accuracy on unlabeled target data. We also study the efficacy of DA on ENNs with respect to the varying group orders of the dihedral group $D_N$, and find that the model performance improves as the degree of equivariance increases. Finally, we find that SIDDA enhances model calibration on both source and target data--achieving over an order of magnitude improvement in the ECE and Brier score. SIDDA's versatility, combined with its automated approach to domain alignment, has the potential to advance multi-dataset studies by enabling the development of highly generalizable models.