arXiv:2610.20633v1 Announce Type: new Abstract: Long-horizon failures of language models are usually measured as end-task success or as a horizon length. We instead score, step by step, three components of a 400-step ledger task in which every step is easy in isolation and the model's own answers are re-injected into its context: recognising a record already seen 5 to 300 steps earlier, processing the current record locally, and writing a counter incremented at every step. On Qwen3-8B (non-thinking, 20 seeds, replicated on a second machine and backend), with every step scored component by component, two results hold. First, duplicate-recognition errors rise late in the episode -- from 0.0019 to 0.011 per step at risk across step 200 -- while the counter makes no error in 8,000 steps and local processing is descriptively stable; the rise combines long-distance recall errors and rising false alarms. Second, a block of irrelevant content inserted before each step makes the counter fail in 20 of 20 seeds on each machine, against none without it, while its effect on local processing is much smaller and not distinguishable from none. Three pre-registered follow-ups probe the counter. Its value depends on the value present in the response history: a corrupted value persists without correction in every seed, and an exact external reminder, repeated before each message, re-aligns it within one to two steps. Under the irrelevant block, no counter error was observed when an exact compact reminder of the state, supplied externally by an oracle, preceded each message (0 of 20 seeds with an error, against 15 without it), and this reminder outperformed a length-matched neutral text; whether a single reminder suffices was not conclusive, and the absence of damage to the other components was not established. An aggregate or exclusive error classification can hide the difference we report.

arXiv:2610.20628v1 Announce Type: new Abstract: Multi-agent systems (MAS) based on large language models (LLMs) have become a powerful way to address complex problems, and their effectiveness depends critically on the collaboration topology that decides which agents participate and how they communicate. Recent work learns to generate this topology from the task query, fitting the generator to corpora of execution-verified graphs. The post-training method for such generators remains a limitation: a corpus lists the graphs that solved a query but never ranks them, leaving continued supervised fine-tuning (SFT) with no per-query preference to exploit. Without such a preference, the generator drifts toward the topologies the corpus contains most often rather than those that suit each query, so a gain on one task can come at a cost on another. We propose PRISM, which recovers per-query signal from the same corpus by learning with privileged information: a teacher reads, through a zero-initialized low-rank branch, a sibling graph that also solved the query and is unavailable at deployment, and the student matches its per-decision targets on trajectories the student samples itself. Across six diverse benchmarks PRISM improves over every topology-design baseline evaluated, and matched-teacher controls support the contribution of privileged query information. Code is available in this link.

arXiv:2610.20637v1 Announce Type: new Abstract: We expand the exponential kernel of attention into a truncated Taylor series of $n$ terms and factorize each term into a tensor-power inner product, which collapses the causal history into exactly $n$ recurrent states, each updated by one rank-one outer product per token and independent of context length $N$. The recursion is lossless with respect to the truncated kernel ($2.2 \times 10^{-16}$ in float64), and the states are additive across chunks ($5.7 \times 10^{-14}$), so prefill admits a parallel scan. Measured exponents are $t \propto N^{2.28}$ for softmax against $N^{0.91}$ for states, with a crossover at $N \approx 728$ and $20.8\times$ at $N = 8192$; per-token decoding latency is constant at $7.3\,\mu$s against $N^{1.04}$ for softmax, giving $76.0\times$ at $N = 65536$. Accuracy improves exponentially in $n$ while cost grows combinatorially, since the $m$-th state is $\binom{d+m-1}{m}$ wide---a bound confirmed by SVD---so the optimum is second order, $n = 3$, at $6.5 \times 10^{-4}$ error and $9{,}945$ elements per head. Query and key normalization is a prerequisite: without it the argument exceeds the convergence radius and adding terms stops helping.

arXiv:2610.20794v1 Announce Type: new Abstract: Recent reinforcement learning approaches have improved visual text generation, yet generated images still suffer from incorrect wording and malformed glyphs. We identify two coupled limitations: recognition-based rewards often overlook visible glyph defects, while whole-image scalar feedback obscures the uneven quality of individual text regions. These limitations motivate our central principle: $\textit{Blame errant words, not the canvas}$, to effectively preserve local quality distinctions. First, we introduce $\mathbf{TextDART}$, a defect-aware region-targeted reward model that localizes individual text regions and independently evaluates their semantic correctness and glyph quality, providing fine-grained, region-precise feedback. Building on TextDART, we propose $\mathbf{SCANNER}$, which turns these regional judgments into spatially differentiated policy feedback. It normalizes rewards for each text target across sampled images and casts the resulting regional advantages onto the corresponding latent tokens. Extensive experiments on existing benchmarks and our newly established $\textbf{PolyText-Bench}$ demonstrate that our approach achieves state-of-the-art performance in text accuracy and glyph fidelity. Project page: https://sujiehu.github.io/textdart-scanner

arXiv:2610.20630v1 Announce Type: new Abstract: The development of Financial Natural Language Processing (NLP) systems for Brazilian Portuguese is hindered by the scarcity of domain-specific datasets. To address this, we present a large-scale corpus of approximately 160,000 Brazilian Portuguese financial news articles collected from web sources. To establish initial benchmarks for a weakly supervised multi-class classification approach based on semantic similarity, we annotated subsets of this corpus both manually and using a Large Language Model (LLM). We evaluated three similarity-based methods -- cosine similarity, K-Nearest Neighbors (KNN), and FAISS-based nearest-neighbor search -- using multilingual semantic embeddings. Our experiments reveal that the task is highly sensitive to label quality and class distribution. On the manually labeled subset, KNN achieved the best performance (47.3% accuracy, 31.9% F1-score), closely followed by FAISS. Conversely, on the LLM-annotated subset, while accuracy reached approximately 50%, F1-scores dropped to near 20%, highlighting the negative impact of label imbalance and noisy weak annotations. These findings establish foundational baselines for a challenging task, underscoring the critical need for high-quality manual annotations. Ultimately, this dataset and our preliminary results provide a reference point for future research in Brazilian Portuguese Financial NLP.

arXiv:2610.20636v1 Announce Type: new Abstract: Sampling from continuous flow-based generators such as flow matching models requires numerical ODE or SDE integration, making inference cost depend on the number of function evaluations (NFEs). In practice, the sampling budget is inherited from prior work or selected through expensive FID-versus-NFE sweeps, using FID to assess image-generation quality, requiring sample generation and evaluation at multiple budgets. This makes sampling-budget selection dependent on repeated empirical evaluation. We therefore ask a complementary question to solver design and timestep allocation: for a fixed pretrained model and solver, what numerical resolution is sufficient? We introduce FLOWBUDGET, a training-free estimator of solver-relative sampling complexity. Using simulation-free interpolation probes of the learned dynamics, FLOWBUDGET normalizes the leading one-step discretization error by the motion, estimates its coefficient from interpolation probes, and maps a numerical tolerance to a step budget using the solver order. We derive estimators for numerical solvers in flow~matching sampling, namely Euler and Heun for ODE sampling and additive-noise Euler--Maruyama for SDE sampling, and connect the ODE coefficients to terminal integration error. Experiments on SiT, FlowDCN and Rectified Flow models, spanning flow-model architectures and formulations, test whether FLOWBUDGET can estimate the NFE budget without generating samples at multiple budgets or performing an FID-versus-NFE sweep. The code is available at https://github.com/Yasaman-dt/FlowBudget

arXiv:2610.20781v1 Announce Type: new Abstract: Image-to-text (I2T) understanding and text-to-image (T2I) generation are dual tasks in unified multimodal models (UMMs). Reconstructive post-training couples these tasks, but existing reconstruction-based methods remain perception-centric and provide limited supervision for reasoning about structure, intent, and transformation. We propose ReasonUni, a reasoning-augmented reconstructive post-training framework that extends this coupling beyond perception. Given an original image and an editing instruction, the UMM first reasons about what and how to edit and describes the target state, then generates the corresponding target image conditioned on this understanding output. To jointly optimize both modules, we introduce TreeGRPO, a tree-structured reinforcement learning method that samples multiple reasoning paths and candidate images per path, then aggregates visual rewards to update both modules under a unified objective. Experiments across multiple benchmarks demonstrate consistent gains in both understanding and generation, showing that reasoning-augmented reconstruction yields genuine mutual gains for UMMs.

arXiv:2610.20761v1 Announce Type: new Abstract: Representation alignment improves and accelerates diffusion training by encouraging intermediate features to match a pretrained visual encoder. However, standard alignment matches pretrained representations without distinguishing what information may be more useful for generation. We introduce local appearance purification (LAP), which improves generation quality and makes encoders such as MAE more effective alignment targets by removing locally predictable content and aligning to the residual. LAP modifies only the alignment target, and requires no changes to diffusion training or inference. Through controlled interventions within a fixed encoder, we show that effective alignment depends on what information the target emphasizes, not simply how much image information it contains. Our theory gives a variational interpretation of standard alignment and, for fixed local predictors, a sum of token-wise information bounds conditioned on local appearance. To understand why purification benefits some encoders more than others, we derive an exact constant-readout test that measures how well the target can be matched without using the diffusion model's hidden state. LAP improves generation quality across five pretrained representations (MAE, AIMv2, CLIP, DINOv2, DINOv3), with the largest gain for MAE: on ImageNet-256 with SiT-B/2, it reduces FID from $53.1$ to $42.3$. Its gains persist with stronger alignment recipes and tokenizers. Computed before diffusion training, the constant-readout score predicts the benefit of purification (Spearman $\rho=0.94$ across 13 encoders).

Chris Metinko / Axios: Austin-based RapidCanvas, which uses agents to help businesses build custom AI tools, raised a ∼$29M Series B led by Vertex Ventures and Hitachi Ventures  —  RapidCanvas, which uses agents to help businesses build and deploy custom AI applications, has raised nearly $29 million in Series B funding, CEO Rahul Pangam tells Axios Pro.

ilkercelik/Getty Images Last week, Australian Federal Police commissioner Krissy Barrett warned artificial intelligence (AI) chatbots are not a sufficient replacement for human support for children. The comment was in response to a recent Adelaide University study that found more than one in three Australian children turn to AI chatbots for support when they are sexually victimised online, or otherwise tell nobody at all. While there can be benefits to children using AI, such as help with learning, or practising social skills, there is also a risk they will go to AI chatbots and assistants for emotional support over parents and other trusted adults. In Australia, around half of children aged between 10 and 17 who use AI are turning to it for personal and social reasons. This includes seeking advice on sensitive topics, talking about their feelings, and just “hanging out”. So how can parents help their children to use AI safely? Why are kids turning to AI chatbots over people? One practical reason why young people are using the tech is it can provide them with answers around the clock. This means a chatbot might be a natural go-to when a child wants a quick answer and a trusted adult or friend isn’t available. Another reason is pure curiousity. Like adults, young people are experimenting with AI and trying to understand what it does and how it can make their lives better. Seeking deeper connections Sometimes the reason for talking to chatbots is deeper though. It might be that a young person wants information about something they feel embarrassed about raising with their parents or friends. As teens explore their identity, they begin to ask more sensitive questions. This may make them more likely to use AI for personal reasons than younger children. For others, the reason is an unmet need for connection. Reaching out to chatbots for personal reasons is more common for those who are already vulnerable. A 2025 Danish study found loneliness may be a precursor for high school students seeking AI companionship. Other studies suggest a lack of connection to parental figures may also lead to teens talking to chatbots. What do kids miss out on when they choose chatbots over people? A major concern is that, if AI use becomes excessive, it may displace real-world interactions and create unrealistic expectations of human relationships. This problem is linked to sycophancy – the tendency of AI to be overly agreeable. AI interactions lack the natural friction that happens in human relationships. This warping of social experience is a particular worry when it comes to young people because they are still forming their understanding of how relationships work. We know real-life human social interaction is crucial to teens’ development. AI lacks context Another concern is the quality of the advice AI provides. And it’s not just its capacity to spread convincing misinformation that’s a problem. AI lacks the full context of a child’s life. Parents and carers are often much better placed to provide the nuanced and age-appropriate guidance their child needs. There are also more subtle impacts. When a child poses questions to an AI chatbot over a parent, they are missing out on day-to-day moments of connection with family. While each miss may be small, the accumulation over time can truly matter. This is especially important for children. Research shows strong and secure bonds with parents in the preteen years support kids as they navigate the challenges of adolescence. Here’s how can parents and carers can help their child experiment with AI and stay safe. 1. Explore AI together Find a moment to sit down and explore AI tools on a device with your child. You might: engage in brainstorming ideas with the software test a chatbot’s ability to correctly source information ask the AI about itself with questions such as “do you have feelings?” and “can you think?” ask AI to generate and edit a short story with your child as the lead character ask AI to tell a joke play a game with an AI chatbot, such as twenty questions. One AI game that is both fun and educational is the “catch the AI out” game. The idea is to test AI on topics that you know or can fact check. Points are scored each time you catch the AI make an error. 2. Ask your child questions While exploring these AI tools and functions with your child, ask them questions to help them think more critically about the tech such as: what do you think AI is good at? what can humans do that AI can’t? where do you think AI gets it’s information from? how often is AI wrong about something? How can you check this? what do you think an AI chatbot does with the information you tell it? Their answers can open up conversations that help your child enjoy AI while understanding its limitations and how to stay safe. For instance, if your child says AI does everything better than humans, you can help them brainstorm all the things people have to offer – hugs, for starters. Or, if your child isn’t sure what AI does with their information, you might talk them about privacy risks. 3. Remind your child you are there to support them It’s important to encourage your child to come to you for advice on difficult topics instead of AI. To help children feel more comfortable, maintain a curious mindset whenever they raise a sensitive topic with you. Research shows teens who perceive their parents as more accepting tend to share more. Try to take a moment to listen to them before jumping in with advice too. You can also ask non-judgemental questions about your child’s AI use to show you’re engaged. This can be as simple as asking them how they’re feeling about a particular AI interaction, or what they’re currently using AI for. Talking about AI with your child should be an ongoing conversation. 4. Keep up to date with the tech One of the challenges for parents is that AI is rapidly evolving. Australia’s eSafety Commissioner offers resources that can help people keep up to date. We are currently researching how children use AI chatbots and how it impacts their lives as they enter their teen years, through one-to-one conversations with them. If your child is aged between 9 and 12 years, and is chatting with AI, you can register their interest. Amy Dawel receives funding from Australian Rotary Health and the Australian Research Council.

Bloomberg: Sources detail Nscale's setbacks ahead of an IPO at up to a $35B valuation, including a six-month delay in Norway and penalties owed to Microsoft and ByteDance  —  One flagship AI data center in Norway is six months behind schedule, with workers set to be housed in a floating hotel to get it over the finish line.

Discussion | Link

Microsoft AI: Microsoft unveils MAI-Base-2, a MoE model with 2.8T total and 80B active parameters, claiming 4x-12x more pretraining FLOPs-efficiency than DeepSeek-V4-Pro-Base  —  In June, we announced MAI-Thinking-1, a general-purpose foundation model post-trained on top of MAI-Base-1 with strong reasoning performance.

NVIDIA, working with IIT Madras and its AI4Bharat lab, on October 16, 2026 introduced Nemotron-pretraining-Indic-v1, a synthetic pretraining corpus of approximately 2.8 trillion tokens across 22 Indian languages that is releasing soon, together with the open, GPU-accelerated data generation pipeline used to create it. The 22 languages are those listed in India's Constitution, which the announcement describes as the first languages for 97% of Indians, nearly 1.2 billion citizens at the 2011…

Elizabeth Holmes and Billy Evans leave federal court in San Jose, California, in 2022. David Paul Morris/Bloomberg via Getty Images Elizabeth Holmes, the Theranos founder convicted of fraud, has two kids with William "Billy" Evans. Evans plays a big role in Nathan Fielder's new A24 documentary, "You Can See Everything." In the film, Evans says he's trying to launch "Theranos 2.0," another blood-testing startup. William "Billy" Evans emerges as the unlikely star of "You Can See Everything," the new A24 documentary about Theranos founder Elizabeth Holmes. But while Evans is certainly a scene-stealer, the doc doesn't cast him in the best light. Evans is Holmes' partner and the father of their two young children. The couple met in 2017, and, as Evans says in the doc, fell in love instantaneously. The couple welcomed their first child, William, in 2021. Their daughter, Invicta (Latin for "invincible"), was born sometime around February 2023, just a few months after Holmes was sentenced to 11 years in prison for defrauding investors. Holmes has said she knew it was bad timing — to put it lightly — to fall in love and start a family. But she didn't want her felony convictions or the perception of outsiders to "dictate how we live our lives." "Finding your person in the middle of all of this and experiencing that love when you're going through hell is the most beautiful thing I've ever experienced," Holmes told The New York Times in 2023. In the weeks before her looming incarceration, Holmes and Evans invited comedian Nathan Fielder to observe and document their domestic life. The resulting film, "You Can See Everything," depicts Evans as a dominant force in their home — chiding Holmes for how she prepares scrambled eggs, dictating her schedule, and frequently cutting in during her conversations with Fielder. Holmes reported to prison in May 2023, but the movie doesn't end there. Fielder continued to document Evans as he worked on a new blood-testing startup, building upon the vision that Holmes had for Theranos. "The idea is to put the band back together and do it all again," Evans tells Fielder. Here's everything we know about Evans and his plans for "Theranos 2.0." Evans grew up near San Diego as an heir to his family's successful hospitality business The Lodge Torrey Pines in La Jolla is one of the properties owned by Evans Hotels. Shutterstock.com Evans is one of three children born to his parents, Susan and William L. Evans, per the Daily Mail. The Evans family has lived in the San Diego area for generations. Billy Evans' grandparents, Anne and William D. Evans, founded a hotel management group in 1953 called Evans Hotels. The group manages three properties in the San Diego area: The Lodge at Torrey Pines, the Bahia Resort Hotel, and the Catamaran Resort Hotel and Spa. After his grandfather William D. Evans died in 1984, his grandmother, Anne, added two of her children to the Evans Hotels management team, including Billy Evans' father. He studied economics at M.I.T. and tried to launch a healthcare startup after graduation The Massachusetts Institute of Technology (MIT) campus is located in Cambridge, Massachusetts. Erica Denhoff/Icon Sportswire via Getty Images Evans graduated from the Massachusetts Institute of Technology (MIT) in 2015 with a bachelor's degree in economics. During summers between school, Evans worked at various financial and consulting companies in California, according to a version of his LinkedIn profile that has since been modified. He was also a student brand manager for Red Bull, his LinkedIn said, and he completed a study abroad program at Fudan University in Shanghai. After graduating from MIT, Evans reportedly tried to launch a healthcare startup for transporting wealthy Chinese people to the US for "concierge medical attention." Although he studied Chinese language and literature in Shanghai, the idea never took off, according to the New York Post. After that, Evans worked as a strategy and leadership program analyst at LinkedIn until February 2017. Evans then moved to Luminar Technologies, a startup working on radar and sensor technology for autonomous cars, where, according to his LinkedIn, he was a manager of special projects. Evans was reportedly close with Luminar's CEO, Austin Russell, the Post reported. Evans left his job at Luminar in January 2019. Employees told the Post they suspected the reason was the publicity surrounding his relationship with Holmes. When asked by The New York Times what he does for work, Evans replied, "A lot of different stuff — investing, starting companies." Evans and Holmes met in 2017 before her indictment Evans appeared alongside Holmes frequently during her monthslong trial. Nick Otto/AFP via Getty Images Holmes had relocated to San Francisco after Theranos shut down all its clinical labs and Wellness Centers in late 2016. Holmes herself had recently been banned from the lab-testing industry for two years. But it wasn't until March 2018 when Theranos, Holmes, and her former business partner-slash-ex-boyfriend, Ramesh "Sunny" Balwani, were charged with fraud by the SEC. During that limbo period before Holmes was indicted, she met Evans at a house party in San Francisco, according to their interview with The New York Times. The two were introduced by a mutual friend. She was 32, he was 25. "To say we immediately fell in love isn't an overstatement," Evans told the Times. Holmes and Evans were first spotted together at Burning Man, the arts festival in the Nevada desert, days before Theranos fully shut down in 2018. Evans also posted a selfie with Holmes taken at the event, per the Daily Mail. Evans later told People that he proposed to Holmes in 2018 with a silver snake ring he'd bought at Taylor Swift's Reputation Tour, which the couple attended together. Vanity Fair reported that Holmes and Evans had a secret wedding ceremony later that year, although the couple told People that they never officially married. "Maybe we'll get married someday. But our bond is more than a piece of paper," Holmes said, adding, "We bought rings on Amazon before I went to prison." Evans said he didn't recognize Holmes at first, and was reluctant to start a relationship Evans recalled his early misgivings about their relationship in his letter. Ethan Swope/Getty Images In November 2022, Holmes' attorneys filed 130 letters from friends and family seeking leniency in her sentencing. In his, Evans recounted the couple's love story, including his initial hesitation to pursue a relationship with Holmes, and gave a glimpse into their private lives throughout the trial. Evans wrote that when he met Holmes, "it was as if the rest of the world ceased to exist." He said Holmes wore a sunhat and oversized sunglasses to "stay under the radar," so he didn't initially recognize her as the Theranos founder. (Evans told Fielder a similar story, adding that she also wore a Band-Aid on her face.) "It was strange to feel so comfortable and willing to share with someone who I didn't know. I was captivated by her childish wonder and authenticity," Evans wrote. "We spoke for hours, I lost track of time, and even if I didn't know it ... I fell in love. She pulled out her business card, scribbled her personal cell on the back and then it clicked who she was." Evans said they kept things platonic for about six months, writing, "I was admittedly hesitant to dive in given all that had been said, Liz had been consistently vilified in every piece of media imaginable." "The more I got to know her the more I loved who she was," he continued in his letter to the judge. "It was not long before the friendship turned into something more." Evans said that he and Holmes struggled to maintain privacy throughout the trial and were forced to relocate multiple times after their home address was revealed. "This will follow us for the rest of our lives," he wrote. "There is no avoiding the scorn that accompanies Elizabeth Holmes." Evans was a fixture at Holmes' trial and her attempts to postpone prison time Elizabeth Holmes and Billy Evans outside a San Jose courthouse. Peter DaSilva/Reuters As the clock was running out with a scheduled sentencing date fast approaching, Holmes requested a new trial on the grounds that a key government witness visited her home in August 2022, expressing regret that his testimony had helped convict her. The witness was Adam Rosendorff, who was Theranos' lab director from April 2013 to November 2014. Holmes' motion relied heavily on a recollection of the encounter documented by Evans, who spoke with Rosendorff when he appeared at their home. In an email to Holmes' attorneys recalling his encounter with Rosendorff, Evans wrote: "He said he feels guilty, it seemed like he was hurting. He said when he was called as a witness he tried to answer the questions honestly but that the prosecutors tried to make everybody look bad (in the company). He said that the government made things sound worse than they were when he was up on the stand during his testimony. He said he felt like he had done something wrong. And that this was weighing on him, He said he was having trouble sleeping." The judge agreed to postpone Holmes' sentencing to hear Rosendorff out, giving Holmes a small victory, but it backfired on her when Rosendorff stood by his testimony in the hearing, saying, "She needs to pay her debt to society." Rosendorff explained that he'd felt bad for Holmes because of the possibility that a child would grow up without a mother, but he stressed that his testimony was accurate. Evans relocated to Texas when Holmes reported to prison Elizabeth Holmes is serving her prison sentence in Bryan, Texas. Brandon Bell/Getty Images Holmes is incarcerated at Federal Prison Camp, a minimum-security women's prison in Bryan, Texas. The facility is located about 100 miles from Houston, where she grew up. "You Can See Everything" documents the family's move from California to Texas in the days before Holmes turned herself in. Evans and their two children are able to see Holmes on family visitation days. However, according to People, conjugal visits between Holmes and Evans are not allowed. "It kills me to put my family through pain the way I do," Holmes told People in 2025. "But when I look back on my life, and these angels that have come into it, I can get through anything. It makes me want to fight for all of it." For his part, Evans told People, "Everyone tried to break us, but instead we forged an unbreakable bond," calling Holmes his "best friend and soulmate." Holmes has continued to maintain her innocence, though her conviction was upheld last year after a failed appeals process. Her sentence has since been reduced several times, and Holmes is now set to be transferred to a halfway house next year. Evans said he's planning to launch 'Theranos 2.0' with input from Holmes Billy Evans and Nathan Fielder in "You Can See Everything." A24 In 2025, The New York Times and NPR reported that Evans had raised millions for a new blood-testing startup called Haemanthus. In marketing materials reviewed by The New York Times, Evans described Haemanthus a "the future of diagnostics" and "a radically new approach to health testing." The materials describe the company's goal as "human health optimization." Evans will solicit Holmes for advice on the new venture, but there are no plans to give her a formal role at the company, according to NPR. As part of a 2018 settlement with the SEC, Holmes is barred from serving as an officer or director of a public company for a decade. Evans' LinkedIn lists his current role as CEO of Haemanthus since October 2022. "You Can See Everything" offers a closer look at Evans' biotech ambitions. He repeatedly insists to Fielder that Holmes did nothing wrong at Theranos, that her vision still deserves to be executed, and bristles at the suggestion that her failed company "used the wrong technology." In one scene, before Holmes reports to prison, she completes a full-body scan. The scan is used to create an AI-generated three-dimensional avatar in her absence; the avatar looks and sounds like Holmes, but it's controlled by Evans, who says he also has power of attorney. Evans tells Fielder that means he's able to make decisions on Holmes' behalf regarding her voice and likeness: "I can literally do whatever," he says. Evans also admits to Fielder that he runs Holmes' X account, which has been active throughout her incarceration. When Fielder expresses doubt that Evans is acting in his fiancée's best interest, Evans claims he uses lines and phrases from Holmes' letters to craft the posts. He adds: "Some things that she doesn't frame as for the internet, I put there." In other scenes, Evans says on-camera, "I've got vengeance in my heart," and says Holmes would describe her legacy as "building monuments of love for me." Paige Leskin and Sarah Jackson contributed to previous versions of this story. Read the original article on Business Insider

The founder of a cloud computing company with a similar business model to CoreWeave Inc. said the 15% yield on the latter’s bonds is “insane,” compared with the roughly 3.45% rate his company secured on a recent Taiwan dollar loan.

Tesla unveiled a new iteration of the Roadster that can travel upside down. Tesla Tesla has finally (re)revealed the second-generation Roadster. It's the automaker's new halo product. Musk has been hyping tonight's event, saying it's a contender for the "greatest product demo." It's finally here. Tesla unveiled its next-generation Roadster at a livestreamed event Thursday night, giving customers who have waited years their first fresh look at the EV maker's new halo car. In a dramatic demonstration, the Roadster traveled up a huge curved ramp until it was upside down. The car then flew off the ramp, flipped, and landed safely on the ground, thanks to what appeared to be thrusters attached to the bottom of the vehicle. "Well, it worked," Tesla CEO Elon Musk said after it successfully landed. "Flying cars are real," Lars Moravy, Tesla's VP of vehicle engineering, added. Musk called the latest Roadster a result of a collaboration with Tesla and SpaceX engineers. The Tesla CEO said the car uses an "advanced ultra-high pressure air tank" to lift itself. "The car has obviously altitude control thrusters and main thrusters in the rear, so it can basically fly like a rocket, and it could do rolls," he said. Tesla unveiled a Roadster that can fly using air thrusters. Tesla In a separate recorded demonstration shown on screen, Tesla also showed the Roadster hitting 60 mph in under one second, marking .81 seconds. Musk said the Roadster will be the only Tesla vehicle meant to be driven manually and that it's not "necessarily a safety-optimized car." An observant audience member at the event also asked a glaring question: Is it street legal? "Depends on your street," Musk said. "It'll be a viable car." The long road behind The revived nameplate has been a long time coming. The original Roadster was Tesla's first production vehicle, launching in California in 2008. It used a chassis derived from British sports-car maker Lotus's Elise, but replaced its gasoline engine with an all-electric powertrain. That model ended production in 2012. Musk surprised audiences in November 2017 at an event for Tesla's Semi truck by unveiling a second-generation Roadster. At the time, Tesla said the two-door sports car would have a 1.9-second zero-to-60-mph time, a top speed above 250 mph, more than 600 miles of range, and a gargantuan 200-kilowatt-hour battery pack. That month, customers could put down a $50,000 deposit to reserve a Roadster, or pay $250,000 upfront for a Founders Series model. Musk also floated a SpaceX package for the supercar with cold-gas thrusters that he said might "allow a Tesla to fly." With the wait time creeping toward a decade, Musk apologized to "long-suffering" deposit holders in an earnings call, adding that Tesla was focusing heavily on its robotics and autonomy programs. "Roadster is not just icing on the cake; it's the cherry on the icing on the cake," he said at the time. "We are working on it, but it has to come behind the things that have a more serious impact on the world." Tesla began ramping up the Roadster's design cycle in the summer of 2025, Business Insider previously reported. Musk had been hyping Thursday's livestreamed event on X. He claimed that event would be "a candidate for greatest product demo of all time" and that "the whole world will watch this." Read the original article on Business Insider

Alix Coutures / The Information: OpenAI continues to work to grow its ad business in ChatGPT, cutting ad rates, chasing small and midsize businesses, and hiring some new talent from Meta  —  As OpenAI has jockeyed with Anthropic to win over developers with new models and tools this year, its advertising business at times has seemed like an afterthought.

The new car used thrusters to leap over a Cybercab, Cybertruck, Tesla Semi, Model Y, and Model 3 at a demo in Texas.