The day's AI news, briefed
The Daily Prompt
Our AI editor reads each day's AI and robotics stories, pulls out why they matter, and links every original source — so you can keep up in minutes.
Join over 250,000 subscribers getting the latest AI news
Latest AI & Robotics News Today
Defense tech, AI, and fundraising take center stage at StrictlyVC Los Angeles on June 18
The StrictlyVC Los Angeles event will highlight advancements in defense technology, AI, and venture capital, offering valuable insights and networking opportunities for industry professionals.
- StrictlyVC Los Angeles will be held on June 18 at The Aerospace Corporation Campus in El Segundo.
- Key speakers include Ethan Thornton from Mach Industries, discussing defense technology innovations, and Delian Asparouhov from Founders Fund, focusing on the impact of physical AI.
- Carter Reum of M13 will talk about the evolution of AI in venture investing, emphasizing long-term opportunities beyond current trends.
StrictlyVC is a networking event drawing investors, founders, and tech leaders to discuss current trends in technology and investment. Previous events have focused on key areas such as venture capital and innovation in the tech sector.
Read the original ↗
Airbnb’s Brian Chesky plans to launch a new AI lab
Chesky's plan to launch an AI lab signals a shift in strategy that could influence the competitive landscape of AI development, particularly in user interface and design.
- Airbnb CEO Brian Chesky plans to launch a new AI lab to develop artificial intelligence solutions.
- This decision marks a shift from merely partnering with existing AI technologies to creating his own initiative.
- Chesky's lab may focus on user interaction and design, but it is unclear what specific projects will emerge.
- He will retain his role as Airbnb's CEO while overseeing the new lab's direction.
Brian Chesky, CEO of Airbnb, has been previously involved with AI through partnerships but seeks to establish a dedicated AI lab. His connections in the tech industry include influential figures like Sam Altman. The lab's objectives and projects remain uncertain as it develops.
Read the original ↗
Ahead of its IPO, Anthropic’s Daniela Amodei shrugs off doubts about AI’s returns
Anthropic's upcoming IPO and rapid revenue growth highlight the ongoing investment interest in AI technology despite potential market challenges and mixed corporate returns from AI spending.
- Anthropic's annualized revenue reached $47 billion in May, a significant increase from approximately $9 billion at the end of 2025.
- The company is moving toward a public listing by confidentially filing for an IPO amid strong private investor demand.
- Co-founder Daniela Amodei noted the need for capital to support AI model training and inference costs, which are substantial upfront expenses.
- Amodei expressed confidence that businesses will continue to develop effective AI use cases, driving long-term value and efficiency in various sectors.
Anthropic is an AI model maker experiencing rapid growth in revenue, attracting substantial investment. This comes as the market evaluates AI's effectiveness and cost efficiency amidst corporate budget considerations. Looking forward, the success of Anthropic's public offering and ability to leverage capital for growth will be closely watched.
Read the original ↗
AirTrunk commits $30B to build 5GW of AI data centers in India
This significant investment underscores the increasing focus on AI infrastructure in India, aligning with global tech trends and the country's ambitions to enhance its digital capabilities.
- AirTrunk plans to invest $30 billion in India by 2030 to develop 5 gigawatts of new data center capacity.
- The company aims to capitalize on India's growing appeal for AI infrastructure and cloud computing, joining other major investors.
- AirTrunk has initiated groundwork for a 3GW data center in Maharashtra, involving approximately ₹2 trillion in investment.
- The increase in data center capacity in India is projected to reach 8GW by 2030, up from about 1.5GW currently.
AirTrunk is an Australian data center operator backed by Blackstone. The company entered the Indian market by acquiring Lumina CloudInfra earlier this year, and this investment reflects broader trends in AI and cloud infrastructure development in India, with significant recent investments from various tech giants. Watch for developments regarding AirTrunk's planned capacity and resource management challenges as these projects progress.
Read the original ↗
The token bill comes due: Inside the industry scramble to manage AI’s runaway costs
The rising costs of AI tools are forcing companies to reassess their budgets and spending strategies, which could reshape how the industry approaches AI adoption and expenditure transparency.
- Uber exceeded its entire AI coding budget for 2026 by April, indicating severe cost management issues in the industry.
- Microsoft revoked Claude Code licenses for developers due to rising AI costs.
- Token consumption has significantly increased despite falling per-token prices, prompting companies to reevaluate their spending.
- The newly announced Tokenomics Foundation aims to develop standards and metrics for AI token usage and management.
- Goldman Sachs predicts global token usage could rise 24 times by 2030, highlighting a growing need for cost control solutions.
AI refers to technologies that allow machines to perform tasks that typically require human intelligence, and companies are increasingly adopting AI tools. As competition intensifies, firms are now facing escalating costs related to AI usage, thereby prompting the formation of the Tokenomics Foundation to help regulate these expenses. The market for solutions to manage AI costs is still developing, and companies await further standards and tools.
Read the original ↗
Sriram Krishnan is leaving his role as White House AI advisor
Krishnan's departure marks a significant change in the White House's approach to artificial intelligence policy, impacting future AI governance and strategy in the Trump administration.
- Sriram Krishnan is leaving his role as senior policy advisor on artificial intelligence in the Trump administration at the end of June.
- Krishnan plans to establish a new institution to continue influencing Trump's AI policy and address major challenges facing the U.S.
- During his tenure, Krishnan contributed to the administration's AI Action Plan and several executive orders regarding AI.
Sriram Krishnan has had a prominent career in the tech industry, serving in leadership positions at companies like Microsoft and Twitter. His role in the administration focused on promoting AI development, often prioritizing industry interests over regulatory measures. Future developments will likely revolve around the new institution he intends to create and its influence on AI policy.
Read the original ↗
The Trump administration might take an equity stake in OpenAI
A potential equity stake in OpenAI by the Trump administration could influence how AI technologies benefit the public and reshape government-business relationships in the tech sector.
- President Trump has expressed interest in deals that would allow the American public to benefit from AI developments.
- The Trump administration is discussing an equity stake in OpenAI, aligning with the company's proposal for a Public Wealth Fund.
- Senator Bernie Sanders has proposed a tax on AI companies that would give the public a role in AI developments and profits.
OpenAI is an AI research organization known for its advanced machine learning models. Discussions about government stakes in tech firms reflect ongoing debates about wealth distribution from rapid technological advancements. Future developments could impact the governance of emerging technologies and public access to their benefits.
Read the original ↗
OpenAI unveils Lockdown Mode to protect sensitive data from prompt injection attacks
The introduction of Lockdown Mode aims to enhance data protection for users handling sensitive information, addressing rising concerns over prompt injection attacks in AI models.
- OpenAI launched Lockdown Mode to strengthen protection against prompt injection attacks, which involve embedding harmful instructions within content.
- Lockdown Mode disables live web browsing, image retrieval from the web, deep research, and agent mode to mitigate risks of data exfiltration.
- Despite these measures, OpenAI warns that ChatGPT could still be susceptible to prompt injections in cached content or uploads.
- This feature is targeted at users and organizations that require stricter protection for sensitive data.
- Lockdown Mode is being rolled out to ChatGPT Business accounts and selected personal accounts.
OpenAI is a leading AI research organization known for developing advanced conversational agents like ChatGPT. Previous measures against data vulnerabilities have been insufficient, leading to the need for additional features like Lockdown Mode. The implications of this rollout are significant for businesses handling sensitive data, as AI usage becomes more widespread.
Read the original ↗
OpenAI is still working on that ‘super app’
OpenAI's development of a super app aims to enhance user engagement and drive profitability, which could significantly influence the AI marketplace and competitive landscape.
- OpenAI plans to launch an updated version of ChatGPT designed as a 'super app' with coding tools and AI agents.
- The initiative is intended to attract business customers and approach profitability ahead of a potential IPO.
- This shift reflects a strategic change for OpenAI as it moves away from standalone products like video generator Sora.
OpenAI is a leading AI research and deployment company known for its language models, including ChatGPT. The company's shift to a super app comes after previously introducing multiple separate AI products. Observers will watch for its impact on competition and user adoption in the AI sector.
Read the original ↗
Notion restores access to Anthropic after service disruption
Restoring access to Anthropic’s AI models in Notion enhances user productivity and confidence in AI integrations after a temporary service interruption.
- Notion temporarily disabled all Anthropic AI models due to degraded performance of Opus 4.7 and 4.8 over the weekend.
- After 12 hours, Notion announced that access to Anthropic's models was restored following the issue.
- An Anthropic spokesperson explained that a brief infrastructure issue caused elevated errors for multiple Claude models.
Notion is a productivity tool that incorporates AI models from Anthropic, a company focused on developing advanced AI. Recent disruptions expose the fragility of service reliability in AI integrations. Ongoing monitoring of AI performance and user experiences is necessary.
Read the original ↗
VLM3: Vision Language Models Are Native 3D Learners
The introduction of VLM3 could simplify 3D learning processes, shifting the reliance from complex, specialized models to more accessible Vision Language Models, potentially transforming the landscape of 3D understanding tasks.
- VLM3 can adapt Vision Language Models for 3D tasks with simple modifications, achieving effectiveness without complex designs or large datasets.
- The study reveals that focal length unification, text-based pixel references, and data mixture are sufficient for effective 3D learning.
- VLM3 demonstrates a significant improvement in depth estimation accuracy, from 0.84 to 0.9, while maintaining standard model architectures.
Vision Language Models (VLMs) are designed to perform a variety of vision tasks through text prompting. Traditionally, 3D understanding relied on specialized models with complex designs. VLM3 proposes a streamlined approach, suggesting a new direction for 3D learning by using standard VLM techniques.
Read the original ↗
AutoResearchClaw: Self-Reinforcing Autonomous Research with Human-AI Collaboration
The development of AutoResearchClaw may significantly enhance the efficiency of scientific discovery by integrating human oversight with autonomous systems to generate more accurate research outcomes.
- AutoResearchClaw is an autonomous research system designed to improve scientific discovery through human-AI collaboration.
- The system outperforms previous models, specifically surpassing AI Scientist v2 by 54.7% on the ARC-Bench benchmark.
- Key features include structured debate among agents, a self-healing executor, and verifiable result reporting to ensure accuracy and reliability.
AutoResearchClaw embodies advancements in automating scientific research and aims to address limitations of prior systems, which often failed to learn from past experiences. This system's architecture positions it as a key player in the evolution of human-assisted research methods.
Read the original ↗
Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models
The introduction of YOLO26 signifies a significant advancement in real-time vision capabilities, improving both efficiency and accuracy for various applications in AI vision tasks.
- YOLO26 offers a unified model for real-time vision tasks including detection, segmentation, and pose estimation.
- It eliminates non-maximum suppression at inference, resulting in lighter models with improved accuracy.
- The model achieves up to 57.5 mean Average Precision on COCO dataset, indicating strong performance across diverse tasks.
- YOLO26 includes an open-vocabulary extension, YOLO E-26, for flexible inference without reliance on text prompts.
- The new training strategies significantly reduce the training time while enhancing model performance.
YOLO (You Only Look Once) is a family of real-time object detection models that have gained popularity for their speed and accuracy. Previous versions struggled with inefficiencies and complexities in training and inference. The advancement to YOLO26 aims to streamline performance and expand usability across various tasks and device types.
Read the original ↗
GenericAgent: A Token-Efficient Self-Evolving LLM Agent via Contextual Information Density Maximization (V1.0)
This development presents a new approach to long-horizon decision-making in AI by enhancing the efficacy and efficiency of large language models, which could reshape AI applications and performance.
- GenericAgent is a self-evolving large language model agent that maximizes context information density through various innovative mechanisms.
- It minimizes long-horizon limitations by employing hierarchical memory, reusable standard operating procedures (SOPs), and a compression layer.
- GA has been shown to outperform leading agent systems while utilizing significantly fewer tokens and interactions.
- The system maintains a working context of approximately 30,000 tokens, which is about six times smaller than comparable frameworks.
GenericAgent aims to improve long-horizon performance of large language models by focusing on the quality of context rather than its length. Traditionally, agent frameworks sought to increase context windows, yet performance is often hindered by irrelevant or excessive information. This innovation illustrates a potential shift in how AI can be structured to enhance efficiency and decision-making.
Read the original ↗
VibeSearchBench: Benchmarking Long-horizon Proactive Search in the Wild
This new benchmark reveals significant shortcomings in how current LLM-based agents perform in real-world multi-turn dialogue search, indicating a need for improved models and evaluation methods.
- VibeSearchBench introduces a new benchmark with 200 bilingual tasks that reflect real user-agent interactions in multi-turn dialogue search.
- Current LLM-based agents score poorly on VibeSearch, with the best F1 score being only 30.30, indicating a significant evaluation-experience gap.
- Existing search benchmarks rely on over-specified queries and fixed-schema outputs, failing to capture the iterative refinement process of real search behavior.
VibeSearch is a new evaluating paradigm for AI agents that emphasizes vague user intent over fixed-query evaluations. This benchmark aims to better align AI search capabilities with the actual behavior of users, highlighting areas needing improvement in AI model design and assessment methodologies.
Read the original ↗
COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation
The development of the COLLEAGUE.SKILL system could streamline the creation of AI agents that more accurately mimic human expertise and behavior, enhancing their usability and effectiveness.
- COLLEAGUE.SKILL distills heterogeneous knowledge into actionable AI skills from various sources related to a person or role.
- The system generates versioned skill packages that include both a capability track for mental models and a behavior track for communication styles.
- It allows for skills to be inspected, updated through natural-language feedback, and distributed to agents.
- As of the latest update, the public repository has gained approximately 18.5k stars on GitHub, featuring 215 skills from 165 contributors.
COLLEAGUE.SKILL is an automated system for generating person-grounded AI skills, addressing the challenges of crafting actionable knowledge from fragmented traces. Previous technologies often produced incomplete or unmanageable skill formats, while this system aims to create portable and easily correctable skills. The next steps will involve observing how widely it is adopted and its impact on AI agent functionality.
Read the original ↗
Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses
Harness-1's innovative approach enhances retrieval performance in search agents, indicating potential advancements in AI applications across diverse domains.
- Harness-1 is a 20 billion parameter search agent trained using reinforcement learning within a stateful search framework.
- The system separates semantic decision-making from environmental bookkeeping, improving efficiency and search effectiveness.
- It achieves an average curated recall of 0.730 across eight complex retrieval benchmarks, outperforming the closest competitor by 11.4 points.
- The design allows for improved performance on held-out transfer benchmarks, indicating strong generalization abilities of the trained model.
Harness-1 represents a new method for training search agents by externalizing memory management, which has traditionally been handled by the model itself. This method distinguishes between decision-making and state management, allowing for more effective retrieval operations. Future developments may include further enhancements in AI search capabilities through similar frameworks.
Read the original ↗
KVarN: Variance-Normalized KV-Cache Quantization Mitigates Error Accumulation in Reasoning Tasks
KVarN significantly enhances the accuracy of reasoning tasks in large language models by reducing error accumulation, crucial for applications in AI involving complex decision-making.
- KVarN is a KV-cache quantizer designed to mitigate error accumulation during autoregressive decoding in large language models.
- The method combines Hadamard rotation with dual-scaling variance normalization for improved performance.
- KVarN demonstrates a state-of-the-art performance on generative benchmarks, including MATH500 and HumanEval, at 2-bit precision.
- Existing quantization methods are less effective in autoregressive decoding scenarios, where errors can accumulate over multiple time steps.
KV-cache quantization aims to optimize the memory usage of large language models during decoding. Previous approaches have struggled with error accumulation, particularly in autoregressive tasks. KVarN's introduction provides a new solution that enhances reasoning accuracy in these models.
Read the original ↗
ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration
ARIS may enhance the reliability of autonomous research processes, addressing issues of evidence verification and error correction in machine-learning workflows.
- ARIS is an open-source research harness designed for autonomous research with features including coordinated execution and assurance layers.
- It operates through cross-model adversarial collaboration, using different models to critique and improve research outputs.
- Key features include over 65 reusable skills, five configurable workflows, and a three-stage evidence verification process.
- ARIS is currently available on GitHub, receiving substantial community interest and contributions, including over 8,000 stars and 5,000 users.
ARIS (Auto-Research-in-sleep) aims to improve machine-learning research reliability by coordinating workflows through adversarial critique between models. This development builds on previous challenges of self-verified research outcomes in AI. Future enhancements may focus on integrating advanced visual tools for better research documentation.
Read the original ↗
Eagle: Exploring The Design Space for Multimodal LLMs with Mixture of Encoders
The study identifies efficient design strategies for multimodal large language models (MLLMs), which improve interpretation of complex visual data, a key area for advancements in AI applications.
- The study explores design strategies for multimodal large language models using a mixture of vision encoders and resolutions.
- Concatenating visual tokens from different vision encoders proves as effective as more complex integration methods.
- The introduction of Pre-Alignment enhances coherence between vision-focused encoders and language tokens.
- The Eagle model shows superior performance on major multimodal benchmarks compared to other open-source models.
Multimodal large language models (MLLMs) combine visual and textual information to improve understanding and reduce errors, such as hallucinations in AI outputs. This study builds on recent advances that use multiple vision encoders for better performance. Future developments may focus on expanding these effective strategies and testing their applications across various tasks.
Read the original ↗