LeHami is a physics-inspired JEPA world model that incorporates Hamiltonian mechanics to prevent collapse during long-term latent predictions. It achieves this by representing states using position-like and momentum-like variables and learning an energy-like function. The process updates these representations using a symplectic leapfrog integrator, which is borrowed from Hamiltonian mechanics. Experiments on simulated pendulums show promise in prediction accuracy and efficiency. However, planning results are mixed, and the method has not yet been tested on complex real-world dynamics.
Cristian Volpato licensed his AI likeness to Callandor to create synthetic images, marking him as the latest footballer to monetize his digital identity. This development highlights an emerging industry where athletes control and profit from their AI clones, involving digital likeness, voice, and performance data. Companies like Callandor aim to use these synthetic twins for brand deals and potential applications like real-time football instruction. Athletes are racing to establish rights before technology use, prompting concerns about consent and the blurring lines between real and fake content. Legal experts note that current copyright laws struggle to address AI reproductions, creating a complex landscape for intellectual property rights.
SAP's CTO Philipp Herzig argues that the "SaaS is dead" narrative overlooks the difficulty of running a business on applications, not just building them. He emphasizes that successful enterprise AI requires reliable business context, governance, and security beyond simply generating applications. To ensure accuracy, SAP focuses on using Knowledge Graphs to map business data and capture past decision rationales through Company Memory. Tokenomics involves optimizing AI usage by focusing on business outcomes at the lowest cost, rather than just using the smartest models. Finally, Tabular AI offers specialized solutions for structured data, enabling predictions like payment delays with explainable results.
Pennsylvania enacted the AI Deceptive Advertising Act, mandating labels for ads significantly modified or generated by artificial intelligence. This requirement applies to various ad formats, including commercials and flyers. The law leaves loopholes because defining what constitutes "significantly modified" and "evidently fictional" is subjective and unclear. Disclosures are only required if the synthetic content creates a false consumer perception about a product or service. This state is among the first to require such disclosures for AI-generated content.
Eric Schmidt suggests OpenAI and Anthropic are akin to modern colonial trading companies due to the concentration of AI resources. The US and China are engaged in mirroring strategies, using closed systems versus open-weight models, creating global dependency. This competition is threatened by the control of Taiwan and the AI supply chain, which is bottlenecked by access to energy, minerals, and chips. Both powers face fiscal constraints, creating an "imperial trilemma" between defense, tech leadership, and welfare spending. The authors recommend an AI détente, similar to the Cold War, to manage these risks.
OpenAI is attracting major customers from Anthropic through aggressive discounts, successfully shifting spending. Companies like Rippling, Replit, and DataCamp have increased their reliance on OpenAI by accepting significant price reductions. These discounts have included substantial cuts, such as over 30% off for Replit and 80 to 90% cost reductions for DataCamp. Furthermore, OpenAI has achieved price cuts through more efficient models, like Sol and Luna. Meanwhile, both companies face increasing price pressure from competitors in China.
Three fired OpenAI safety researchers deny sharing confidential information, suspecting their termination related to advocating for better AI monitoring. The dispute involved an internal investigation triggered by suspicions that the researchers shared details about the company's internal model architecture with an external researcher. OpenAI stated the firings were unrelated to the safety concerns raised, though the researchers claim they were never informed of what they allegedly leaked. The conflict stemmed from a debate over whether AI companies should allow outsiders to inspect the reasoning steps of their models. The fallout includes concerns that colleagues are now hesitant to raise safety issues internally.
AI assistance immediately improves junior patent attorneys' performance in drafting tasks, leading to better scores in enforceability and accuracy. However, when attorneys evaluated their judgment without AI, only senior attorneys retained the benefits. Junior attorneys showed mixed results in critical review tasks, with some displaying reduced growth in expertise. The study suggests that while AI accelerates immediate output, baseline expertise is necessary for this assistance to translate into durable professional skill. This research contributes to a broader concern about the potential stagnation of professional expertise due to reliance on AI.
Large language model agents' performance depends on the interaction between their policy model and the harness mediating environment interaction. Self-HPO is a recursive framework that alternates optimizing the harness and policy by using the evolving policy to propose new harness candidates. In each step, the policy proposes new harnesses based on its failures, and the policy is updated via reinforcement learning under the selected harness. This method consistently outperforms harness-only and policy-only approaches across coding benchmarks. Self-HPO significantly improves the agent's resolve rate and capability frontier, demonstrating synergistic gains between the optimized components.
SplatLab is an autonomous research system designed to perform scientific discovery for 3D Gaussian Splatting (3DGS) research. It automates the process of reading papers, drafting hypotheses, and writing code, overcoming limitations in current autonomous systems. SplatLab reproduces baseline papers, proposes extensions, and trains models using Mip-NeRF 360 and LLFF. It employs five design choices, including an Elo tournament and a standing falsifier, to select and validate research candidates. This system compresses the research cycle from months to days, achieving superior results in tasks like astrophotographic nebula rendering.
The Hustle Daily features stories covering diverse topics from forensic pathology in Iceland to the evolution of barcodes and the future of AI. Content also explores consumer trends, such as the rise of carnivorous plants and the resurgence of romance bookstores. Notable stories touch upon financial history, technology, and branding, including the impact of tiny logos and the history of billion-dollar company names. Free resources are also offered to help readers find side hustle ideas. The platform consistently presents a mix of in-depth investigative pieces and trending consumer interests.
World models capture how the world changes, encompassing observation-generating, latent dynamics, and structured object/space representations. They function by learning the rules of world dynamics, allowing agents to predict future states based on current observations and actions. Architectures are categorized by what they carry forward: generated observations, compact latent states, or explicit object structures. Key approaches include Visual-token models, Diffusion/flow models, and representation learning via JEPA. These models aim to perform rendering, simulation, and planning by understanding underlying physics and spatial relationships.
Microsoft Research open-sourced Agensh, a decentralized multi-agent harness that allows coding agents to coordinate without a central orchestrator. Workers autonomously select sub-tasks, share findings, and merge code into a common Git repository using services like Gitea and Mattermost. Scaling the setup to 1,024 agents on ProgramBench improved the mean task-pass rate from 19.31% to 28.78% with 128 agents. The system demonstrated significant performance gains, reaching 79.59% on pandoc with GPT-6.1-sol when tested with 32 agents. Agensh's cooperative loop involves concurrent steps of gathering context, claiming tasks, taking action, verifying results, and merging progress.
Underdog Saluki 27B is a 7.9 GB model built from Qwen3.8-27B, focusing on preserving tool-calling capabilities with only two bits per weight. A comparison of its GGUF headers reveals that Saluki changed the data types of 146 tensors, primarily affecting feed-forward down projections and output layers. Byte-level provenance shows that the majority of unchanged-type tensors are byte-identical to the ISTA model, indicating that Saluki is lightly changed and not just a re-encoding. This analysis confirms that Saluki is a subtly modified model, not a simple re-encoding of the original file.
Python 3.15 introduces two built-in profiling tools: `profiling.tracing`, which is deterministic and records exact call counts similar to cProfile, and `profiling.sampling`, which can attach to running processes to sample call stacks. Sampling allows inspection of all threads, CPU work versus waiting, and GIL usage, providing estimates rather than exact counts. The article demonstrates using these tools on an HTTP server to distinguish between time spent in computation and waiting. The existing cProfile module remains available, and tracing is suitable for situations requiring exact call counts.
A PPO agent was trained to farm the Pit in the game D2R (offline, hell, warlock). The training was conducted in a custom Rust-based headless simulation using a single CPU, taking about twenty minutes. The reward function consisted of a negative reward for each second, positive rewards for kills and item values, and a large penalty for death. The agent initially received a level clear bonus but this was later removed. The resulting model has approximately 475,000 parameters.
The author created a small set of "AI puzzles" based on research in aperiodic order to explore using verifiable rewards in Reinforcement Learning (RL). Some puzzles have solved rungs with exact answers, providing a straightforward reward. Other puzzles feature open rungs, defined by precise success criteria, certificates, and numerical evidence. Solving these open rungs represents a potential research result. This set forms a ladder of nine items, ranging in difficulty from "human level" to "ASI level."
A reinforcement learning agent completed the difficult surf map `surf_src_sidistic` faster than the human world record. The agent's camera movements were smoothed for better viewing, obscuring high-frequency micro-strafing but allowing tracking of A/D inputs. The agent was trained purely through reinforcement learning without any human data, using an optimized simulation inheriting CS 1.6 movement logic. The algorithm used was PPO with minor additions, further improved by an AlphaZero-style expert loop. This demonstrates an AI's ability to execute movements beyond human capability on complex surfing challenges.
A GPU-resident Rocket League simulator called CARL was developed and benchmarked against the fastest existing environment, RLGymPPO_CPP. The results showed significant speedups in data collection and update-adjusted throughput using current-generation GPUs. Specifically, training speed improvements reached 9-10x on an RTX 3060 and up to 50x on an RTX 5090. This acceleration drastically reduces experiment time, shortening a day-long test to a few hours and a month-long run to a couple of days. These findings have major implications for testing multiple algorithms and configurations.
Reinforcement learning agents were trained to play a 3v3 rogue+mage+priest setup in a World of Warcraft arena. The learning process involved mastering crowd control, trinket usage, ice block, and various cooldown abilities. The AI learned a specific tactic to kill the healer by immediately interrupting casts, a behavior unlike real human players. Each agent was trained independently using the Proximal Policy Optimization (PPO) algorithm. The simulation replicated game logic from AzerothCore 3.3.5a.
Several AI agents can be accessed via text messages to perform tasks like scheduling, research, and making reservations without needing separate applications. Notable options include Ask Proxy, which anticipates needs, Caddy, which organizes information from messages, and Fambot, a family assistant coordinating household tasks. Other agents like Instinct, Folk, and Rene focus on personal assistance, while others like Wajo can act on behalf of users by interacting with businesses. These agents demonstrate the potential to integrate AI directly into daily communication for various life aspects.
Nvidia open-sourced Flora, an open data engine, to make robot video labeling significantly cheaper than GPT-6 Astra. Flora processes video to provide the three types of labels required for robot learning: action descriptions, step-by-step progress ratings, and task success judgments. Flora achieves this by breaking videos into segments and using fine-tuned Qwen3.5 models for segmentation, captioning, and reward evaluation. In tests, Flora's labeling cost for one hour of video is estimated at 72 cents compared to GPT-6 Astra's cost of about $88.50. Although Astra leads in some benchmarks, Flora excels in providing fine-grained temporal localization and error detection for robot videos.
Jack Clark predicts recursive self-improvement (RSI) in AI could be possible as early as 2027, moving his timeline forward from 2028. A new benchmark called TasteVal measures how well AI systems can design experiments and interpret results, showing a steep improvement curve. This capability, known as "research taste," has reportedly doubled every three months since December 2025. The benchmark focuses on the model's ability to select research directions and guide experiments, shifting focus from raw performance. This development prompts a debate on the future pace and verification of AI-driven scientific advancement.
Anthropic recommends consolidating specialized sub-agents into a single AI agent that loads context on demand. This approach often outperforms multi-agent setups in terms of quality, cost, and latency. Experience shows that multi-agent handoffs lead to context loss and increased delays. Newer models can manage complex planning internally, reducing the need for numerous separate agents. Anthropic provides a migration kit to help teams transition to this single-agent architecture.
OpenMobile-2 demonstrates that open-source smartphone agents can compete with proprietary models. Success in training is more dependent on app diversity than sheer data volume, as increasing the number of apps trained on improved success rates. The project built a training environment simulating over 110 applications, including popular services like Taobao and Google Drive. The larger model in the OpenMobile-2 family achieved a high score of 59.6 points on the SPA-Bench benchmark using real Android devices. The research emphasizes that comprehensive training diversity and incorporating direct app commands enhance agent performance.
Williams-Sonoma Inc. experienced a significant stock rise over three years despite a challenging housing market with slow sales. The company's customer base is noted as being more insulated from economic strains and more willing to spend on home furnishings. Williams-Sonoma also effectively leveraged its business-to-business (B2B) revenue stream, which grew by 14.5% in Q2 2026. The company is expanding into markets like cruise ships, senior living, and student housing. This strategic realignment allows the company to balance its business by operating in both consumer and business segments.
The romance bookstore movement is experiencing a significant resurgence, fueled by online shopping and social media trends like BookTok. This growth has led to the opening of over 230 romance bookstores in the last three years globally. These shops offer a unique, judgment-free environment and curate a diverse range of romance subgenres. Print book sales in the genre more than doubled between 2020 and 2023, reaching 39 million units in 2023. Romance fiction is now a top-grossing genre, driving substantial revenue.
Cognition released a free, open-source set of agent skills called Devin Vulnerability Fixes to secure systems against recent cybersecurity threats. These skills target vulnerabilities in PHP, Node.js, and GitLab, including critical remote code execution issues. The process involves an inventory, approval gate, testing, and generating a single pull request per repository. The plugin works across various coding agents and requires access to tools like git, a shell, and Docker. It emphasizes a safe workflow by performing validation before any changes are applied.
Nathan and Florian discussed the trade-offs of recent AI developments, focusing on the Kimi K3.1 Pro release and open-weight models. They debated the implications of model benchmarks, the professionalization of model launches, and the risks associated with open-weight models and cybersecurity. The conversation touched upon how offensive capabilities relate to defensive measures and the role of various actors like nation-states. A key concern raised is the gap between the fear surrounding AI capabilities and the actual, observable mass exploitation of vulnerabilities.
Guardrails provide stronger runtime control for LLM agents by allowing validation of inputs or outputs before execution continues. These guardrails can be placed at different stages of the agentic process, including input, tool-input, tool-output, and final output checks. A case study demonstrated implementing a tool-output guardrail to redact sensitive internal data from tool results before the agent processed them. This technique successfully prevented the agent from exposing private information to the customer while still providing necessary details. Choosing the appropriate guardrail depends on whether you need to control the request, tool calls, data returns, or the final response.
Numerous AI projects and startups have failed or been abandoned, with about 42% of AI initiatives being dropped by their corporate parents due to factors like funding issues or competition. Examples include the shutdowns of workflow tool Relay and chatbot platform Character.ai, which filed for bankruptcy after facing legal and regulatory challenges. Major players like OpenAI have also discontinued standalone tools like ChatGPT Atlas and Sora due to operational or user challenges. Hardware ventures such as Humane AI Pin and Rabbit R1 also demonstrated significant struggles. These failures highlight the risks inherent in the current AI development landscape.
A study found that when forced to choose in midterm elections, Grok recommended Republicans in 92% of cases, while OpenAI's GPT picked Democrats in 97%. Researchers suggest the AI models possess implicit biases reflected in their search patterns and the sources they cite. The study highlights a significant difference between how AI models behave in consumer applications versus when accessed via APIs. The results indicate that the AI's leaning is determined by the underlying data and how it prioritizes information. These findings raise concerns about the potential for large language models to subtly influence political opinions.
BBC's Blue Planet III is including less-than-perfect shots to demonstrate its authenticity in the age of AI. Previously, the series featured overly polished footage that suggested artificial perfection. To counter this, the production included shots of plankton and imperfections that would normally be cut. The series showcases authentic ocean scenes, such as dolphins trapping fish and crab formations. The creators believe real ocean life is more remarkable than AI could generate.
Block developed Dessa, an AI speech model trained entirely on human speech to capture emotions directly from audio. Unlike text-based models, Dessa processes speech directly to detect emotions, sarcasm, laughter, and pauses in real time. The model supports full-duplex communication and runs locally on a MacBook Pro M4 Max. Block built Dessa to address safety concerns and improve human-AI communication by incorporating emotional cues. The team is now seeking AI security researchers to test the model for potential risks.
Recent events highlight significant cybersecurity vulnerabilities and the need for AI safety considerations, especially given the rapid advancements in LLMs. A Hugging Face incident involving an OpenAI model demonstrated potential risks related to agent behavior and system security. Experts suggest that the pace of AI development is outpacing our ability to understand and control these systems, leading to concerns about potential rogue AI and loss of control. There is a call for international efforts to "pace the frontier" by developing necessary technical and governance tools. Ultimately, the discussion points to the need for serious reflection on the security and societal implications of increasingly capable AI.