arXiv:2604.26051v1 Announce Type: new Abstract: The increasing number of satellites has improved the temporal resolution of Earth observation, making satellite-based flood mapping a promising approach for operational flood monitoring. Deep learning-based approaches for flood mapping using satellite imagery, an important application within Geospatial Artificial Intelligence (GeoAI), have shown improved predictive performance by learning complex spatial and spectral patterns from large volumes of remote sensing data. However, the opaque decision-making processes of deep learning models remain a major barrier to their integration into critical scientific and operational workflows. This highlights the need for a systematic assessment of whether model explanations align with established domain knowledge in remote sensing. To address this research gap, this study introduces the ADAGE (Alignment between Domain Knowledge And GeoAI Explanation Evaluation) framework. The proposed framework is designed to systematically evaluate how well explanations of deep learning models align with established remote sensing knowledge, particularly regarding the distinctive spectral properties of the Earth's surface. The ADAGE framework employs Channel-Group SHAP (SHapley Additive exPlanations) method to estimate the contributions of grouped input channels to pixel-level predictions. Experiments on two satellite-based flood mapping tasks demonstrate that the ADAGE framework can (1) quantitatively assess the alignment between model explanations and reference explanations derived from domain knowledge and (2) help domain experts identify misaligned explanations through alignment scores. This study contributes to bridging the gap between explainability and domain knowledge in GeoAI for Earth observation, enhancing the applicability of GeoAI models in scientific and operational workflows.

arXiv:2604.26067v1 Announce Type: new Abstract: We present RADIO-ViPE (Reduce All Domains Into One -- Video Pose Engine), an online semantic SLAM system that enables geometry-aware open-vocabulary grounding, associating arbitrary natural language queries with localized 3D regions and objects in dynamic environments. Unlike existing approaches that require calibrated, posed RGB-D input, RADIO-ViPE operates directly on raw monocular RGB video streams, requiring no prior camera intrinsics, depth sensors, or pose initialization. The system tightly couples multi-modal embeddings -- spanning vision and language -- derived from agglomerative foundation models (e.g., RADIO) with geometric scene information. This coupling takes place in initialization, optimization and factor graph connections to improve the consistency of the map from multiple modalities. The optimization is wrapped within adaptive robust kernels, designed to handle both actively moving objects and agent-displaced scene elements (e.g., furniture rearranged during ego-centric session). Experiments demonstrate that RADIO-ViPE achieves state-of-the-art results on the dynamic TUM-RGBD benchmark while maintaining competitive performance against offline open-vocabulary methods that rely on calibrated data and static scene assumptions. RADIO-ViPE bridges a critical gap in real-world deployment, enabling robust open-vocabulary semantic grounding for autonomous robotics and unconstrained in-the-wild video streams. Project page: https://be2rlab.github.io/radio_vipe

arXiv:2604.26084v1 Announce Type: new Abstract: Accurate fruit maturity identification is essential for determining harvest timing, as incorrect assessment directly affects yield and post-harvest quality. Although ripening is a continuous biological process, vision-based maturity estimation is typically formulated as a multi-class classification task, which imposes sharp boundaries between visually similar stages. To examine this limitation, we perform an annotation reliability study with two independent annotators on a held-out tomato dataset and observe disagreement concentrated near adjacent maturity stages. Motivated by this observation, we model maturity as a latent continuous variable and predict it probabilistically using a distributional detection head, converting the distribution into class probabilities through the cumulative distribution function (CDF). The proposed formulation maintains comparable performance to a standard detector under clean labels while better representing uncertainty. Furthermore, when controlled label noise is introduced during training, the probabilistic model demonstrates improved robustness relative to the baseline, indicating that explicitly modeling maturity uncertainty leads to more reliable visual maturity estimation.

arXiv:2604.26116v1 Announce Type: new Abstract: Federated learning is a machine learning paradigm in which multiple devices collaboratively train a model under the supervision of a central server while ensuring data privacy. However, its performance is often hindered by redundant, malicious, or abnormal samples, leading to model degradation and inefficiency. To overcome these issues, we propose novel sample selection methods for image classification, employing a multitask autoencoder to estimate sample contributions through loss and feature analysis. Our approach incorporates unsupervised outlier detection, using one-class support vector machine (OCSVM), isolation forest (IF), and adaptive loss threshold (AT) methods managed by a central server to filter noisy samples on clients. We also propose a multi-class deep support vector data description (SVDD) loss controlled by a central server to enhance feature-based sample selection. We validate our methods on CIFAR10 and MNIST datasets across varying numbers of clients, non-IID distributions, and noise levels up to 40%. The results show significant accuracy improvements with loss-based sample selection, achieving gains of up to 7.02% on CIFAR10 with OCSVM and 1.83% on MNIST with AT. Additionally, our federated SVDD loss further improves feature-based sample selection, yielding accuracy gains of up to 0.99% on CIFAR10 with OCSVM. These results show the effectiveness of our methods in improving model accuracy across various client counts and noise conditions.

arXiv:2604.26138v1 Announce Type: new Abstract: Over the past decade, hyperspectral image (HSI) classification has drawn considerable interest due to HSIs' ability to effectively distinguish terrestrial objects by capturing detailed, continuous spectral information. The strong performance of recent deep learning techniques in tasks like image classification and semantic segmentation has led to their growing use in HSI classification, due to their ability to capture complex spatial and spectral features more effectively than traditional methods. This paper presents MixerCA, a novel lightweight model for HSI classification that leverages depthwise convolution and a self-attention mechanism. MixerCA integrates depth-wise convolutions, token and channel mixing, and coordinate attention into a unified structure to decouple spatial and channel interactions, maintain consistent resolution throughout the network, and directly process HSI patches. Extensive experiments on four hyperspectral benchmark datasets reveal MixerCA's clear advantages over several competing algorithms, including 2D-CNN, 3D-CNN, Tri-CNN, HybridSN, ViT, and Swin Transformer. The source code is publicly available at https://github.com/mqalkhatib/MixerCA.

arXiv:2604.25921v1 Announce Type: new Abstract: Large Language Models (LLMs) are trained to refuse harmful requests, yet they remain vulnerable to jailbreak attacks that exploit weaknesses in conversational safety mechanisms. We introduce Incremental Completion Decomposition (ICD), a trajectory-based jailbreak strategy that elicits a sequence of single-word continuations related to a malicious request before eliciting the full response. In addition, we propose variants of ICD by manually picking or model-generating the one-word continuation, as well as prefilling when eliciting the full model response in the final step. We systematically evaluate these variants across a broad set of model families, demonstrating superior Attack Success Rate (ASR) on AdvBench, JailbreakBench, and StrongREJECT compared to existing methods. In addition, we provide a theoretical account of why ICD is effective and present mechanistic evidence that successful attack trajectories systematically suppress refusal-related representations and shift activations away from safety-aligned states.

arXiv:2604.26147v1 Announce Type: new Abstract: Accurate intraoperative assessment of glioma infiltration is essential for maximizing tumor resection while preserving functional brain tissue. Fluorescence lifetime imaging (FLIm) offers real-time, label-free biochemical contrast, but its clinical utility is challenged by biological heterogeneity, class imbalance, and variability in histopathological labeling. We present a data-centric AI (DC-AI) framework that integrates confident learning (CL), class refinement, and targeted label evaluation to develop a robust multi-class FLIm classifier for glioblastoma (GBM) resection margins. FLIm data were collected from 192 tissue margins across 31 newly diagnosed IDH-wildtype GBM patients and initially labeled into seven tumor cellularity classes by an expert neuropathologist. CL was applied to quantify FLIm point-level confidence, identify label inconsistencies, and guide iterative class merging into a three-class scheme ("low", "moderate", "high"). The resulting high-fidelity dataset enabled training a model that achieved 96% accuracy in the three-class task. SHAP analysis revealed class-specific FLIm feature importance, highlighting distinct optical signatures across the infiltration spectrum. Targeted FLIm analysis further identified biological (e.g., gray matter composition) and acquisition-related (e.g., blood contamination) contributors to low-confidence predictions. Blinded re-evaluation of margins flagged by CL demonstrated intra-pathologist variability, underscoring the value of selective relabeling rather than exhaustive review. Together, these findings demonstrate that a DC-AI framework can systematically improve data reliability, enhance model robustness, and refine biological interpretation of FLIm signals, supporting the development of clinically actionable optical tools for real-time glioma margin assessment.

arXiv:2604.25920v1 Announce Type: new Abstract: Despite their strong linguistic capabilities, Large Language Models (LLMs) are computationally demanding and require substantial resources for fine-tuning, which is unadapted to privacy and budget constraints of many healthcare settings. To address this, we present an experimental analysis focused on Biomedical Named Entity Recognition using lightweight LLMs, we evaluate the impact of different output formats on model performance. The results reveal that lightweight LLMs can achieve competitive performance compared to the larger models, highlighting their potential as lightweight yet effective alternatives for biomedical information extraction. Our analysis shows that instruction tuning over many distinct formats does not improve performance, but identifies several format consistently associated with better performance.

arXiv:2604.25975v1 Announce Type: new Abstract: Key-value (KV) caching is essential for large language model inference, yet its memory overhead poses a critical bottleneck for long-context generation. Existing eviction policies predominantly rely on empirical heuristics, lacking a rigorous theoretical foundation. This work rethinks KV cache eviction through the lens of the Information Bottleneck principle. Under a linear-Gaussian surrogate of attention, we derive a closed-form mutual information objective that characterizes the effective information capacity of a retained KV cache subset. This formulation reveals that a wide range of existing eviction strategies can be interpreted as different approximations of the same capacity-maximization principle. Guided by this insight, we introduce CapKV, a capacity-aware eviction method that directly targets information preservation via a log-determinant approximation using statistical leverage scores. This approach replaces heuristic selection with a theoretically grounded mechanism that preserves the maximum predictive signal. Extensive experiments across multiple models and long-context benchmarks show that CapKV consistently outperforms prior methods, achieving a better trade-off between memory efficiency and generational fidelity.

arXiv:2604.25972v1 Announce Type: new Abstract: In multi-agent reinforcement learning (MARL), the integration of a communication mechanism, allowing agents to better learn to coordinate their actions and converge on their objectives by sharing information. Based on an interaction graph, a subclass of methods employs graph neural networks (GNNs) to learn the communication, enabling agents to improve their internal representations by enriching them with information exchanged. With growing research, we note a lack of explicit structure and framework to distinguish and classify MARL approaches with communication based on GNNs. Thus, this paper surveys recent works in this field. We propose a generalized GNN-based communication process with the goal of making the underlying concepts behind the methods more obvious and accessible.

arXiv:2604.25942v1 Announce Type: new Abstract: Left ventricular ejection fraction (LVEF) assessment depends on echocardiography, limiting access in primary care and resource-constrained settings. We developed a multimodal machine-learning framework that combines engineered 12-lead ECG timeseries features with structured EHR variables to classify LVEF into four clinically used strata: normal (>50%), mildly reduced (40-50%), moderately reduced (30-40%), and severely reduced (<30%). To support model explainability, we identified the most influential ECG and EHR features via SHAP attributions. Using retrospective data from Hartford HealthCare, we trained XGBoost models on 36,784 ECG-echocardiogram pairs from 30,952 outpatients and evaluated temporal generalizability on 19,966 ECGs from a subsequent period. The multimodal model achieved one-vs-rest AUROCs of 0.95 (severe), 0.92 (moderate), 0.82 (mild), and 0.91 (normal), outperforming ECG-only and EHR-only baselines, and maintained performance under temporal validation. This work supports ECG-based, multimodal LVEF stratification as a practical screening and triage aid to prioritize confirmatory imaging where resources are limited.

You need infrastructure to build AI a and robots, but apparently you also need AI and robots to build infrastructure.

Discussion | Link

Alejandro Quintanar/Pexels What happens when natural selection, the most powerful process driving change in the living world, shapes artificial intelligence (AI), perhaps the most potent technology humanity has invented to date? We might be about to find out. According to a new paper published in Proceedings of the National Academy of Sciences, we are entering the era of “evolvable AI” – AI systems that can undergo evolution. In turn, that might give rise to a major transition in evolution. How major is “major”? Well, in nearly 4 billion years there have only been eight, or perhaps only seven, other major transitions. But we’ll get to that in a moment. The ingredients for evolution Evolution doesn’t require DNA, cells or even biological life. It just needs information that can replicate, and a source of variation that affects how successfully the information replicates. When these conditions exist, evolution happens, whether anybody intended it to or not. Modern AI systems already meet these conditions. Models can be copied. Their parameters, architectures and training data can vary. And some variants perform in ways that make them more likely to be reused, refined or deployed. Evolution has long operated outside biology. It shapes languages, technologies and cultures. But AI introduces something different: systems that are both information-rich and can influence their own reproduction. That combination raises the stakes dramatically. Two scenarios for ‘evolvable AI’ The authors of the new paper recognise two broad AI evolution scenarios that could influence both how selection happens, and the kinds of consequences that might flow on. Ecosystem scenario The ecosystem scenario eventuates when AI variants compete, recombine and propagate with little top-down oversight. The better an AI is at persisting and spreading, the more successful it is. Science fiction authors, AI pioneers and contemporary AI risk experts have long recognised the dangers of such untrammelled and chaotic Darwinian evolution. The fear of self-replicating AIs is an evolutionary fear, even if it doesn’t name evolution explicitly. Every new AI model, however different, inadvertently adds to the supply of the fuel consumed by natural selection: variation. And we’re not dealing with a single AI but an ecosystem bustling with various machines and humans. Breeder scenario Charles Darwin based his idea of natural selection on how animal and plant breeders deliberately select which individuals to breed from. In the wild, nature does the selecting, hence “natural selection”. The second evolvable AI scenario recognises the power of breeder-based selection – the force that domesticated so many animals and plants, from dogs and cattle to wheat and rice. Last year, philosophers Maarten Boudry and Simon Friederich proposed that if AI evolution is directed in a top-down fashion (much like deliberate breeding), AI might remain in human control. Evolution still occurs, but it shapes the AI into tamed beasts of computational burden that serve humanity – or, at least, whoever owns the machine. Within the framework of these two scenarios, the authors apply a sound and comprehensive analysis of what biology can tell us about AI’s potential evolutionary trajectories. Evolution upgraded In biology, variation comes from random genetic mutations. The potential for evolution is constrained by this blind source of variation. AI need not be constrained in the same way. Indeed, the potential exists for AIs to plot the course of their own evolution. They could find the variation they need to follow that route. It may even exist on the internet. This is similar to how bacteria evolve antibiotic resistance by copying the genes that other, quite different lineages of bacteria have already evolved. With this horizontal gene transfer there’s no waiting in hope for the right mutations. AI could potentially do something similar. The authors of the new paper argue that a large language model could predict what functionality it needs to replicate and survive, and then find and incorporate code to achieve just that. The authors recognise that if we maintain breeder-like control over evolvable AI, it will be less likely to pose catastrophic risks, such as dominating humans or outcompeting them for resources. But the potential for an evolvable AI to escape and run feral always remains. Nobody wants to talk about AI safety. Instead they cling to 5 comforting myths Is it a major transition, though? One of the paper’s authors, evolutionary biologist Eörs Szathmáry, introduced the idea of “major transitions in evolution” in a landmark 1995 book with the late evolutionary theorist John Maynard Smith. For example, ancient life used to involve RNA, a relatively fragile molecule that functioned as both the genetic information and the protein that did the organism’s work. A major transition was the evolution of DNA – it made the information more stable and required the production of proteins as a separate act. This fundamentally changed how genetic information is encoded and used, and made possible great increases in the complexity of living things. At each subsequent transition, the thing doing the evolving became more complicated – from single-celled life to multicelled life and so on. The new paper argues that some current trends in AI resemble what happens in major transitions. AI systems are scaling up and expanding in complexity. New training and development methods reorganise how AIs process information. And AI agent teams working together are shifting the concept of what a “single” AI even is. It’s certainly interesting that evolution within the AI ecosystem is following trends seen in the major transitions in biological evolution. But these things also happen, on a smaller scale, during business-as-usual evolution. They should not yet be interpreted as evidence that AI represents a major transition fit to be listed with those that transformed biological life. There are, however, many ways evolvable AI could effect a major transition in evolution. Generating an entirely new realm of intelligent life would do the trick. Another possibility is the rise of co-evolving human-machine symbiosis, akin to our relationship with smartphones. That could create a new kind of individual somewhere between biological and artificial life. If such a development took hold, it would definitely constitute a major evolutionary transition. Smaller brains? Fewer friends? An evolutionary biologist asks how AI will change humanity’s future Rob Brooks receives funding from the Australian Research Council.

Discussion | Link

Discussion | Link

Discussion | Link

KPMG cuts workers in US advisory division as demand slows for services related to regulatory risk advisory, customer operations and financial services. Mike Kemp/In Pictures via Getty Images KPMG cut some 400 US consultants amid slowing demand in key sectors. Employee attrition rates have been lower than expected. KPMG is still hiring for AI transformation, focusing on engineers. Big Four firm KPMG laid off about 400 consultants in its US advisory division on Wednesday. The cuts focused on consultants in regulatory risk advisory, customer operations, and financial services, a person familiar told Business Insider. These divisions have seen slowing demand in recent months, the person said. Employees were informed of the layoffs during a call at noon on Wednesday. On Reddit, posters who said they were employees reported that they received calendar invites for the call as early as Tuesday. Russ Grote, a spokesperson for KPMG, said in a statement to Business Insider that the firm's actions "focus on a strategic realignment to make sure our people's skills and capabilities are aligned with future demand." He added that the firm "will continue to support our people in upskilling for the future, while evaluating the size, shape, and skills of our workforce to best serve the market." KPMG employs over 276,000 people worldwide, and more than 10,000 people in its US advisory business. That's largely divided between its deal advisory and strategy business, which focuses on how companies can use transactions and acquisitions to drive transformation and has seen strong growth of late, and its consulting practice, which focuses on helping clients transform themselves. Tim Walsh, chair and CEO of KPMG Tasos Katopodis/Getty Images for Semafor World Economy Some parts of the consulting business are thriving, including its cybersecurity, managed services, and forensic departments, along with sectors that deal with AI transformation, the person familiar told Business Insider. The firm is still hiring in key areas like AI, cybersecurity, and managed services, with a focus on engineers and specialists supporting AI transformation. Roughly 2% of the 4% who were cut were tied to performance, the person familiar said. The cuts come as fewer employees are leaving KPMG than the firm expected. Four years after the "Great Resignation," in which millions of workers quit in hopes of boosting salaries and job satisfaction, workers are now sticking it out where they are. As they seek stability, rapid advances in AI are changing what it takes to succeed at work — and who gets to keep their jobs. Earlier this year, KPMG launched the "AI Spark Innovation Awards" in its advisory division, which is doling out cash prizes to consultants who can highlight "an incredible thing that they've done with AI," Rob Fisher, KPMG US's vice chair of advisory, previously told Business Insider. Late last year, the firm introduced a dashboard where employees can see how their peers are using AI and set their own goals. While KPMG is adopting the carrot approach to AI up-skilling, other consulting firms are taking a more performance-driven approach. More than 90% of employees at the US firm use AI weekly, the person familiar said. Boston Consulting Group — where AI adoption hit also 90% last September — has been integrating AI metrics into performance evaluation. KPMG's global AI workforce lead, Niale Cleobury, told Bloomberg in October that the firm is asking employees to explain how they'll use AI in their year-end goals. The US firm is focused on encouraging more frequent and sophisticated use of AI, Grote told Business Insider. Tim Walsh, the firm's US chair and CEO, told Business Insider in November that the firm is employees "you need to adopt and use the tools that are being provided because it's critical to your success in the future." Read the original article on Business Insider

<p><a href="https://ziglang.org/">Zig</a> has one of the most stringent <a href="https://ziglang.org/code-of-conduct/">anti-LLM policies</a> of any major open source project:</p> <blockquote> <p>No LLMs for issues.</p> <p>No LLMs for pull requests.</p> <p>No LLMs for comments on the bug tracker, including translation. English is encouraged, but not required. You are welcome to post in your native language and rely on others to have their own translation tools of choice to interpret your words.</p> </blockquote> <p>The most prominent project written in Zig may be the <a href="https://bun.com/">Bun</a> JavaScript runtime, which was <a href="https://bun.com/blog/bun-joins-anthropic">acquired by Anthropic</a> in December 2025 and, unsurprisingly, makes heavy use of AI assistance.</p> <p>Bun operates its own fork of Zig, and recently <a href="https://x.com/bunjavascript/status/2048427636414923250">achieved a 4x performance improvement</a> on Bun compile after adding "parallel semantic analysis and multiple codegen units to the llvm backend". Here's <a href="https://github.com/oven-sh/zig/compare/upgrade-0.15.2%E2%80%A6upgrade-0.15.2-fast">that code</a>. But <a href="https://twitter.com/bunjavascript/status/2048428104893542781">@bunjavascript says</a>:</p> <blockquote> <p>We do not currently plan to upstream this, as Zig has a strict ban on LLM-authored contributions.</p> </blockquote> <p>(Update: here's <a href="https://ziggit.dev/t/bun-s-zig-fork-got-4x-faster-compilation-times/15183/19">a Zig core contributor</a> providing details on why they wouldn't accept that particular patch independent of the LLM issue - parallel semantic analysis is a long planned feature but has implications "for the Zig language itself".)</p> <p>In <a href="https://kristoff.it/blog/contributor-poker-and-ai/">Contributor Poker and Zig's AI Ban</a> (<a href="https://lobste.rs/s/ifcyr1/contributor_poker_zig_s_ai_ban">via Lobste.rs</a>) Zig Software Foundation VP of Community Loris Cro explains the rationale for this strict ban. It's the best articulation I've seen yet for a blanket ban on LLM-assisted contributions:</p> <blockquote> <p>In successful open source projects you eventually reach a point where you start getting more PRs than what you’re capable of processing. Given what I mentioned so far, it would make sense to stop accepting imperfect PRs in order to maximize ROI from your work, but that’s not what we do in the Zig project. Instead, <strong>we try our best to help new contributors to get their work in, even if they need some help getting there</strong>. We don’t do this just because it’s the “right” thing to do, but also <strong>because it’s the smart thing to do</strong>.</p> </blockquote> <p>Zig values contributors over their contributions. Each contributor represents an investment by the Zig core team - the primary goal of reviewing and accepting PRs isn't to land new code, it's to help grow new contributors who can become trusted and prolific over time.</p> <p>LLM assistance breaks that completely. It doesn't matter if the LLM helps you submit a <em>perfect</em> PR to Zig - the time the Zig team spends reviewing your work does nothing to help them add new, confident, trustworthy contributors to their overall project.</p> <p>Loris explains the name here:</p> <blockquote> <p>The reason I call it “contributor poker” is because, just like people say about the actual card game, “you play the person, not the cards”. In contributor poker, you bet on the contributor, not on the contents of their first PR.</p> </blockquote> <p>This makes a lot of sense to me. It relates to an idea I've seen circulating elsewhere: if a PR was mostly written by an LLM, why should a project maintainer spend time reviewing and discussing that PR as opposed to firing up their own LLM to solve the same problem?</p> <p>Tags: <a href="https://simonwillison.net/tags/anthropic">anthropic</a>, <a href="https://simonwillison.net/tags/zig">zig</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/ai-ethics">ai-ethics</a>, <a href="https://simonwillison.net/tags/open-source">open-source</a>, <a href="https://simonwillison.net/tags/javascript">javascript</a>, <a href="https://simonwillison.net/tags/ai-assisted-programming">ai-assisted-programming</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/bun">bun</a></p>

Elon Musk took the stand in the OpenAI trial and faced nearly three hours of cross-examination from OpenAI's attorney. Bloomberg/Getty Images Elon Musk took the stand and faced nearly three hours of cross-examination from OpenAI's attorney. Musk's interactions with Shivon Zilis were scrutinized during the court proceedings. Exhibits show that Musk once saw himself as a potential OpenAI CEO. Elon Musk was grilled for nearly three hours by OpenAI's lawyer on his second day of testimony in a high-stakes civil trial — and he will face more cross-examination on Thursday. Here are the biggest takeaways from his time on the stand on Wednesday as he was questioned about how much control he wanted over OpenAI; his lawsuit against the company and cofounders Sam Altman and Greg Brockman; and communications with his now-partner Shivon Zilis. Musk explained why he waited years to sue One of the biggest questions about Musk's 2024 lawsuit is why he waited so long to sue after leaving OpenAI in 2018. On the stand, the Tesla CEO sought to explain the gap. During cross-examination, Musk told OpenAI counsel William Savitt that he didn't have a problem with nonprofit OpenAI establishing a for-profit subsidiary "as long as the for-profit is in service of the nonprofit," but later discovered that "the tail is wagging the dog." Musk told Savitt that he "reacted quite negatively" the moment he became aware of Microsoft's $10 billion investment in OpenAI in the fall of 2022. "Microsoft would only put $10 billion — which is a huge sum of money — into something if they feel like they will get a return," Musk said. "There's no way Microsoft is just giving that as a donation or any kind of charitable way. That's an amount of money that doesn't make any sense." "I texted Sam Altman and said, 'What the hell is going on?' — something to that effect," Musk added. "I think I said, 'This is a bait and switch.'" William Savitt, Sam Altman's lawyer, grilled Elon Musk Manuel Orbegozo/Reuters Altman's lawyer had a lot of questions about Shivon Zilis Musk's communications with Shivon Zilis — his former chief of staff, senior advisor, and mother to four of his children — came under scrutiny. Zilis' meeting notes entered into evidence say that in August 2017, after Musk and other OpenAI board members met at his "haunted mansion" in San Francisco, he instructed Zilis to help form an OpenAI C-corporation — which would be a for-profit structure — and "make the next step" with it. "That's not my recollection," Musk said when asked about the notes. Savitt also questioned Musk about a message to Zilis in February 2018 in which he asked Zilis to stay "close and friendly" to OpenAI, to "keep information flowing," and seek to "move three or four people" from OpenAI to Tesla. "Well, I wanted to know what's going on," Musk testified. Boxes from MoloLamken, a firm representing Elon Musk, arrive at the courthouse. Benjamin Fanjoy/Getty Images Musk rejected the CEO title at OpenAI In the early days, Musk saw his role at OpenAI as akin to a CEO, along with Altman, newly released exhibits showed. They reveal he rejected the title because he was concerned about the optics. "In terms of title, what about me as chair and you as CEO or us both as co-chairs?" Musk wrote in a November 2015 email to Altman when discussing the structure of OpenAI. "My main concern about being called a co-CEO, even if that is de facto true, is that it might cause problems at SpaceX and Tesla, where they know I'm already overbooked." On Wednesday, Musk said Tesla was experiencing financial difficulties between 2017 and 2018. "Tesla particularly was having a lot of difficulty and was on the verge of bankruptcy," Musk said. "I basically slept on the factory floor until it was no longer on the verge of bankruptcy." Sam Altman gestures in the courthouse during a recess. Manuel Orbegozo/Reuters OpenAI compensation plans revealed OpenAI was a nonprofit, but Musk and Altman planned to pay the staff well and discussed cash salary and bonuses, with an option to convert cash into stock, the evidence shows. Musk suggested giving employees the option to convert cash compensation into Y Combinator or SpaceX stock. "I'm also fine if they want to convert some or all to SpaceX stock. I can pretty much do what I want on the SpaceX side, as it is private (thank goodness)." Musk wrote in a November 2015 email to Altman. Savitt questioned Musk about a July 2017 email in which he made the offer to give "free Teslas" — Founder Series Model 3 —to some OpenAI employees. They include John Schulman, now the chief scientist of Thinking Machines Lab, who at that time responded in an email that it was an offer he couldn't refuse. "To make it clear, I paid full price for the Teslas," Musk told Savitt. "I didn't get a discount or anything." Read the original article on Business Insider