arXiv:2605.03147v1 Announce Type: new Abstract: Earnings calls are a key source of financial information about public companies. However, extracting information from these calls is difficult. Unlike the templatic filings required by the U.S. Securities and Exchange Commission (SEC) to report a company's financial situation, earnings conference calls have no built-in labels, are unstructured, and feature conversational language. We explore this challenging domain by assessing the information captured by models trained on SEC filings and in-context learning methods. To establish a baseline, we first evaluate the generalization capabilities of SEC-trained models across established SEC datasets. To support our investigation, we introduce three novel benchmarks: (1) SEC Filings Benchmark (SECB), (2) Earnings Calls Benchmark (ECB), and ECB-A, a subset with 2,460 expert annotation groups to support our qualitative analysis. We find that encoder-based models struggle with the domain shift. Finally, we propose a system utilizing LLMs to perform open-ended extraction from unstructured call transcripts, verified by human evaluation (79.7% precision), providing a baseline for this valuable domain through the consistent tracking of emergent KPIs.

arXiv:2605.02915v1 Announce Type: new Abstract: Same-model self-verification, prompting a model to audit its own predicted answer, is a plausible confidence signal for selective prediction, but its practical value remains unclear once strong likelihood-based baselines are taken seriously. We evaluate self-verification against two such baselines, LL-AVG and LL-SUM, on ARC-Challenge and TruthfulQA-MC across multiple model families, scales, and prompt variants. We measure not only correctness ranking, but also abstention quality through AURC and operating-point analyses. The results are sharply task- and model-dependent. On ARC-Challenge, self-verification substantially improves over LL-AVG for Phi-2 and the Qwen models, with the largest gains appearing in Qwen-7B. On TruthfulQA-MC, however, the signal is less reliable: smaller models can become prompt-sensitive, DeepSeek-R1-Distill-8B degrades relative to LL-AVG, and LL-SUM often remains the stronger practical baseline. We therefore do not treat self-verification as a general-purpose uncertainty estimator. In this setting, it is better understood as a conditional confidence signal whose value depends on task type, model family, prompt formulation, and, crucially, the baseline it must beat.

arXiv:2605.03103v1 Announce Type: new Abstract: Semi-structured information extraction (IE) from OCR-derived clinical reports is crucial for efficiently reconstructing patients' longitudinal medical histories. In practice, this scenario commonly involves three tasks: (i) field-header (key) discovery, (ii) key-conditioned question answering (QA), and (iii) end-to-end key-value pair extraction. However, existing evaluations often under-model two factors: heterogeneous and incompletely known key representations, and OCR-induced noise. This makes it difficult to assess model robustness in real-world settings. We present MedStruct-S, a benchmark specifically designed to evaluate these tasks under unknown keys and OCR noise. MedStruct-S contains 3,582 annotated real-world clinical report pages. Using MedStruct-S, we benchmark two representative paradigms: encoder-only sequence labeling with post-processing and decoder-only structured generation, covering four encoder-only and five decoder-only models spanning 0.11B to 103B parameters. Our results show that encoder-only models achieve the best performance for non-null-value key-conditioned QA despite being substantially smaller than decoder-only models. When comparing models of similar order of magnitude, encoder-only models still perform better overall. Without controlling for model scale, fine-tuned decoder-only models deliver the strongest overall results. These findings show that the benchmark provides a reliable and practical basis for selecting and comparing models across different semi-structured IE settings.

Discussion | Link

The Korea Herald: Coupang reports Q1 revenue up 8% YoY to $8.5B and a net loss of $266M following a 2025 Korean customer data breach, its largest loss since Q4 2021  —  US-listed e-commerce giant Coupang Inc. said Wednesday it swung to a net loss in the first quarter amid fallout from a massive customer data breach in South Korea.

Anhata Rooprai / Reuters: CRM software company Freshworks said it plans to cut 11% of its workforce, or ~500 jobs, as it deals with AI's impact; FRSH has dropped ~26% in 2026  —  Freshworks (FRSH.O) said on Tuesday it would cut 11% of its workforce, or about 500 jobs, as the business-software company navigates …

New York Times: US labor experts and economists warn US unemployment benefits are unlikely to protect against AI job losses, and the GOP-led Congress is doing little to prepare  —  Economists aren't sure if or when artificial intelligence will cause widespread job losses.  But they do agree on one thing …

Sangmi Cha / Bloomberg: Samsung reaches a $1T valuation, making it only the second Asian company after TSMC to hit the mark, after its stock rose 4x+ over the past year on the AI boom  —  Samsung Electronics Co. has reached a $1 trillion market valuation after booming demand for chips used in artificial intelligence saw …

Creating complex molecules usually requires years of experience and countless decisions, but a new AI system is changing that. Synthegy lets chemists guide synthesis and reaction planning using simple language, while powerful algorithms generate and evaluate possible solutions. The AI doesn’t just compute—it reasons, scoring pathways and explaining which ones make the most sense.

Getty Images Around the world, the boom in artificial intelligence is driving a parallel race between major tech companies to secure the physical infrastructure that enables it. This race has now reached New Zealand, with Singapore-based Datagrid rolling ahead with plans to open the country’s first “AI factory” near Invercargill. The scale of this major data centre, designed to power energy-hungry global AI systems, is substantial – both in its 78,000 square-metre size and its multi-billion-dollar cost. It is not the only major investment by global players. Just last year, Amazon Web Services announced plans for a large new cluster of data centres in Auckland, billed as a NZ$7.5 billion investment in New Zealand’s digital infrastructure. For smaller countries such as New Zealand, projects like these are often presented as wins – bringing jobs, investment and a pathway into the global AI economy. But they also carry less visible consequences, many of which are more complicated than they first appear. The hidden economics of AI infrastructure Behind these projects sits a rapidly expanding global demand for computing infrastructure, driven by the growth of AI systems such as ChatGPT and Claude. Increasingly, governments are leaning into this shift by competing to attract data centres as part of their approach to AI and the digital economy. In New Zealand’s case, officials have pointed to the country’s renewable energy, cool climate, available land and political stability as reasons it could appeal to foreign investors. New Zealand Trade and Enterprise has even suggested the country could become an “international data centre hub”. Datagrid’s project has been presented as an early example of that opportunity. The company has described it as “the most significant upgrade to New Zealand’s digital infrastructure in a generation”. Some of the potential downsides are straightforward. These systems rely on powerful computer chips and so use far more energy than typical digital services. The Datagrid facility, for instance, is expected to draw up to 280 megawatts of electricity, or around 6% of national demand – making it New Zealand’s second-largest electricity user after the Tiwai Point aluminium smelter. But other drawbacks are less obvious. My research on data centre markets suggests that, while the infrastructure is built locally, the systems it supports usually operate across multiple countries. Datagrid has indicated it expects to serve international AI and cloud providers, positioning New Zealand as a potential host for global AI workloads. These workloads are likely to come from AI markets already dominated by a handful of big tech companies. This, in turn, raises questions about where value is created – and how much of it is retained locally. Who builds the infrastructure – and who controls it? For smaller economies like New Zealand, relying on large international investors is often seen as a practical necessity, given the scale of capital and market reach required to build and run these facilities. But large data centre projects can also have significant impacts on the shape of local cloud service providers – companies that typically offer data storage, computing services and IT support to businesses and government. These smaller domestic players often find it difficult to compete directly with global firms on price and influence. As such, they are often forced to find new opportunities. As AI increases demand for land, power and connectivity, local firms are taking on roles in building and running the infrastructure that supports them. This tends to shift them toward the physical side of the industry, securing sites, brokering access to energy supply, building facilities to overseas market specifications and meeting both local and international legal requirements. Global companies such as Amazon Web Services, Microsoft and Google, meanwhile, provide the platforms and software that run on top, often by leasing capacity from local operators. Recent developments in New Zealand’s own cloud sector illustrate how these arrangements can work in practice. Amazon Web Services, for example, recently disclosed it had shifted away from plans for a large standalone Auckland data centre build, while continuing to expand its local cloud operations through co-location agreements with local data centre providers. For countries hosting this infrastructure, the relationship can become uneven, with hosts supplying the land, energy and networks while key decisions about how these systems operate are often made elsewhere. This raises broader questions for governments. If domestic firms are concentrated in infrastructure roles, how does that affect the development of higher-value capability over time? And how should countries weigh immediate gains – jobs, investment and connectivity – against longer-term control and positioning in the digital economy, including what this means for where value is created and retained? Projects like Datagrid’s will still deliver clear benefits. But their wider impact will depend on how trade-offs are managed, and how much influence countries retain over the systems they host. Angus Dowell does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

Greg Brockman in court with Sam Altman (R) last week Bloomberg/Bloomberg via Getty Images OpenAI's compute costs have surged to $50 billion in 2026 from $30 million in 2017, Brockman said. Brockman said Musk sought a big win for OpenAI before considering any for-profit ventures. Musk demanded 51% of OpenAI but faced resistance from founders, leading to heated confrontations. Greg Brockman's second day of testimony revealed many insights into Elon Musk's interactions with OpenAI's founders. Brockman, OpenAI's president, returned to the witness stand on Tuesday, seeking to add "context" to journal entries about how it would be "morally bankrupt" to "steal" the OpenAI nonprofit from Musk. Musk is seeking more than $100 billion in damages from OpenAI in the high-stakes trial, where he said the company deceived him and that its founders, including Brockman, unjustly enriched themselves. Here are the key revelations from Brockman's latest testimony. Musk wanted a big win for OpenAI before entertaining a for-profit In a bid to compete, Musk told the OpenAI team they needed a big win, Brockman told the jury. "Once we have a big result, then and only then can we entertain the for-profit," Brockman said Musk told them. Brockman says the team finally got their win in 2017 when an OpenAI bot defeated a top player in the video game Dota 2 on an international stage. Twenty thousand people watched the win, which Musk crowed about on Twitter, before it was renamed X. Celebrating at Musk's haunted mansion with Amber Heard After the big win, the OpenAI team was invited to Musk's "haunted mansion" in San Francisco to discuss the company's next steps, Brockman said. In his email, Musk warned of "party carnage" from the night before. "There was confetti and cups and the whole thing all around," Brockman said of the scene. Actor Amber Heard, Musk's girlfriend at the time, was also there. "I remember Amber served some nice whiskey," Brockman said. "Elon asked her to be part of the conversation. She didn't really want to," he said, adding, "But we had a conversation about the for-profit, and it was a very celebratory moment." Musk said he deserved more because he had more money Musk wanted a controlling, 51% equity stake in OpenAI and to be its CEO, Brockman told the jury. "He said he deserved more because he had started the most multi-billion-dollar companies in history, that he had zero failures," Brockman testified. "Look, you guys are great," Brockman said, quoting Musk. "But I can start another AI company tomorrow, like in one tweet." This was something the rest of the founders — Brockman, Sutskever, and Altman — objected to. The breakup fight with Musk After that, things went downhill. The founders agreed to structural changes to the company, but they fought over the details, Brockman said. Brockman recalled a meeting in August 2017 where they told Musk they would each be getting founding shares. If Musk wanted additional shares, he could pay the market price. "Then the conversation turned to equity, and something really changed," said Brockman of how Musk reacted. "Something just shifted in him. It was like, you could just sense it. And he was angry." "He just sat quiet for several minutes, just thinking," Brockman added. "And he said, 'I decline.'" Brockman said later in the testimony that he thought Musk was going to "physically attack him" and that Musk "stormed around the table." Musk said he wanted to compete with the "wolves" at Google Around the time Musk resigned from the board, he met with OpenAI's employees, Brockman said. Brockman said Musk told the team at the 2018 meeting that he was resigning because OpenAI would require "billions of dollars per year" and "the only path" he saw to make it work was to have Tesla run it. Musk told the group that in pursuing artificial intelligence at Tesla, "he would not work on safety," Brockman said. "He said the most important thing would be to catch up to DeepMind," Brockman said, speaking of Google's AI lab. "If the sheep are dictating safety and the wolves are not, there would be no purpose," Brockman recalled Musk saying. OpenAI's compute costs have skyrocketed OpenAI expects to spend about $50 billion on computing power in 2026, Brockman said when addressing the amount of funding OpenAI would need to stay competitive. According to Reuters, Brockman said that the ChatGPT maker's computing costs have surged from roughly $30 million in 2017 to tens of billions of dollars this year. "We went down the list of the Forbes 500, the wealthiest people in America, and tried to find people excited about AI," said Brockman as he detailed the difficulty of OpenAI's fundraising efforts back when it was still a nonprofit. Read the original article on Business Insider

Politico: Sources: the WH is mulling EOs to address security risks from advanced AI, including barring companies from “interfering” with the government's use of models  —  A White House spokesperson said any official policy announcement would come directly from President Donald Trump …

The Silicon Valley AV startup has not started driverless testing yet.

Hugh Langley / Business Insider: Doc and sources: Google is testing an agent in the Gemini app, internally named Remy, that can integrate with Google services to take actions on a user's behalf  —  - Google is working on an AI agent codenamed “Remy,” according to an internal document.  — Remy is described as a …

For the long-established tech vendor, the software development lifecycle platform is an accessible entry point for enterprises into the realm of AI coding.

OpenAI’s B2B Signals research shows how frontier enterprises deepen AI adoption, scale Codex-powered agentic workflows, and build durable competitive advantage.

Uber uses OpenAI to power AI assistants and voice features that help drivers earn smarter and riders book faster across a global real-time marketplace.

Meet the ChatGPT Futures Class of 2026—26 student innovators using AI to build, research, and drive real-world impact. Discover how this generation is redefining learning, creativity, and opportunity with ChatGPT.

Singular Bank built Singularity, an internal assistant using ChatGPT and Codex to help bankers save 60–90 minutes daily on meeting prep, portfolio analysis, and follow-up.