A concept for a B2B gifting platform named Sendry is presented, leveraging AI to automate the entire gifting campaign process. The platform connects to CRMs to identify prospects, research them, suggest personalized gifts, and manage delivery and follow-up. This addresses the problem that most B2B marketing is ignored, noting that physical gifts still garner attention. The system aims to streamline the complex logistics of finding addresses, selecting gifts, and tracking results. Ultimately, the platform seeks to optimize sales outreach by making physical gifting scalable and efficient.
The Nasdaq lost $2.5 trillion in market value as investors feared regulatory crackdowns over AI and data center stocks. This selloff was triggered by growing concerns about stricter scrutiny following the US midterm elections. Local community opposition to data center projects has created significant grassroots resistance, increasing regulatory risk for the sector. Major tech companies heavily invested in AI infrastructure, such as Nvidia and Microsoft, suffered losses reflecting this uncertainty. This event signals that the AI infrastructure boom now faces political headwinds that could reshape the industry's growth trajectory.
NVIDIA and Korea's MSIT are expanding collaboration across the sovereign AI ecosystem, covering semiconductors, infrastructure, talent, and models. This partnership aims to enable Korean NPU and XPU companies to engage with the NVIDIA NVLink Fusion ecosystem through training and technical programs. NVIDIA will collaborate with consortia and foundation model builders using Dynamo software to optimize AI inference. Furthermore, a collaboration with SK Telecom and Rebellions will advance sovereign AI infrastructure for Korean models using NVLink Fusion. Overall, the initiative focuses on building a comprehensive AI capability from hardware to skills within Korea.
Creative industries are alarmed because AI companies have requested changes to Australian copyright law to allow training on local material. The International Federation of the Phonographic Industry (IFPI) warned that weakening copyright protections would be a "disaster for creativity" and damage Australia's reputation. They expressed concern that Australia might agree to demands that other nations have rejected. AI companies seek legal certainty to invest, suggesting models like an "opt-out" model for scraping content. The government has pledged support for creatives, but the global music body fears a dangerous relaxation of copyright.
Anti-AI populism is predicted to be the defining political fight in the 2028 American cycle. Voters have shown significant opposition to AI-related infrastructure, as evidenced by state-level moratoriums on data centers. Democrats are urged to hold AI industry leaders accountable to address concerns about job displacement and existential risks. There is a call for legislation to mitigate AI harms, including banning surveillance pricing and holding developers liable for AI-driven actions. Ultimately, the author advocates for public inquiry and action to shift the focus from labor replacement to worker augmentation.
Nvidia initiated an independent investigation after reports surfaced that CEO Jensen Huang was targeted by foreign intelligence. This stems from reports suggesting Huang had a relationship with a Chinese businesswoman who allegedly passed information to Beijing. Lawmakers have pressed Huang for briefings regarding security protocols and his relationship with the businesswoman. The company's board stated the inquiry would examine all available information while prioritizing national security. Furthermore, reports indicated Huang admitted to infidelity with a suspected Chinese spy.
Large language models are shifting software development toward collaboration with coding agents, requiring new methods for defining intent and coordinating tools. The course explores AI-native practices such as agent skills, spec-driven development, and the software factory. Students will learn how to provide context to agents and design iterative workflows where humans and agents plan and build software together. The course emphasizes hands-on projects to explore the capabilities and limitations of current coding agents. Upon completion, students will be able to design effective agent-driven workflows and apply software-factory principles.
Business leaders are benchmarking their AI usage, with Claude showing advancement across various stages compared to ChatGPT. In the "Exploring" stage, 58% use Claude and 74% use ChatGPT, while in the "Piloting" stage, Claude reaches 82% adoption. Among companies operating AI in daily functions, 86% use Claude compared to 69% for ChatGPT. The report also highlights barriers to scaling AI, noting that moving from exploring to scaling presents different challenges for each stage. A training program called the Claude Cohort is offered to help users automate workflows with Claude.
NVIDIA released Dynamo-optimized deployment recipes for several leading Korean open models. These models include A.X K2 from SK Telecom, K-EXAONE 2.0 from LG AI Research, Solar Open 2 from Upstage, and Motif-3 from Motif Technologies. NVIDIA Dynamo is an open-source framework designed for distributed inference-serving in multi-node environments. It efficiently manages request routing and communication to maximize inference performance. These recipes enable developers to deploy these models faster and make them accessible.
Reason Machines released Aspen, an open-source Git server designed to handle large volumes of traffic from AI agents. Aspen independently implements the design of OpenAI's Pando to improve performance for agent workflows. It scales writes per branch and uses object storage for packs and PostgreSQL for branch history. Benchmarks show Aspen can sustain 567 pushes per second on a single 8-core server. Aspen is deployable via Docker Compose and Helm, mirroring standard Git protocols.
Mihail Eric, who teaches at Stanford and leads AI at Monaco, shares insights from the CS146S course focusing on AI-native engineering. Students learn by engaging with open-source repositories, proposing changes, and explaining patches to maintainers. Eric details how Monaco builds agent workspaces and equips agents with necessary tools, such as code review and Datadog access. The lessons emphasize skills for understanding code differences and addressing bottlenecks like code review. Ultimately, the focus is on enabling developers to build and manage AI-driven systems effectively.
The author finds AI-written Pull Request (PR) descriptions often contain too much detailed information, which makes them seem like "slop." To address this, the author uses HTML `<details><summary>` tags in GitHub Markdown to hide the detailed content. This allows the author to provide a brief, human-curated summary for quick context, while the full, detailed agent output remains hidden. The author suggests using the summary to convey what is most important, such as the model name or effort level. This method helps readers quickly grasp the PR's purpose without reading lengthy, automated explanations.
Ordinary Least Squares (OLS) regression is problematic when data exhibits heteroscedasticity, meaning the variance of the errors is not constant across predictor values. OLS models the mean and fails to capture the spread of potential outcomes, leading to unreliable predictions, especially for high-variance data. Quantile Regression offers a solution by estimating different quantiles of the conditional distribution rather than just the average. By fitting models for specific quantiles, one can create prediction intervals that reflect the true uncertainty for individual observations. This approach provides a more robust and complete picture of market dynamics compared to relying solely on OLS estimates.
Multi-agent systems can be built by splitting responsibilities using two primary delegation patterns: agent-as-a-tool and handoff. Agent-as-a-tool allows a calling agent to use specialist agents as tools to gather information, while handoff involves transferring control to a specialist agent for completion. These patterns are used to construct complex systems based on various topologies and execution strategies. Common multi-agent patterns include fan-out/fan-in, pipeline, hierarchical structures, evaluator-optimizer, and swarm architectures. Ultimately, these systems are founded on the interaction mechanisms of agent-as-a-tool and handoff.
The author tested five large language models on six variants of the Monty Hall puzzle, varying the host's knowledge and actions. While the models generally defaulted to "switch," the author found that their confidence in this answer was highly dependent on the specific puzzle rules. The analysis revealed that models tended to rely on memorized answers rather than probabilistic reasoning based on the host's actions. Switching only wins the classic puzzle when the host knows where the car is, and the probability changes drastically based on the host's behavior. Ultimately, the study suggests that recognizing the change in puzzle structure is more important than achieving the correct answer for these models.
Mixture-of-Experts (MoE) models face a memory bottleneck because all experts must occupy memory at inference, even when most are idle. Expert pruning removes less useful experts, but existing first-order methods ignore the cooperative interactions between experts. HOPE (Higher-Order Pruning of Experts) is a second-order method that accounts for these pairwise expert interactions by optimizing over an interaction matrix. This approach results in superior pruning performance, especially at higher removal rates, by preserving expert cooperation. HOPE's process involves calibrating experts, solving a quadratic program to find the optimal prune-set, and finally pruning the model.
A single deep-learning system called WorldParticle unifies the simulation of various physical phenomena, from sliding sand to folding proteins. This system models matter by treating objects as numerous particles and predicting their movement step-by-step based on learned dynamics rather than explicit equations. The AI architecture is inspired by models like ChatGPT, predicting particle movement sequentially. WorldParticle has demonstrated higher accuracy in simulating granular collapse and protein folding compared to traditional numerical methods. The researchers aim to create a foundation model for simulation that can help discover new materials and physical processes.
Researchers from MIT and NYU developed FLEX, an AI-assisted learning platform for robot motion-planning. This system allows humans to give instructions in natural language, enabling robots to draw larger generalizations from the commands. FLEX uses a Large Language Model (LLM) to read implied meaning, such as recognizing that avoiding a laptop also means avoiding other electronics after a spill command. The process involves three steps: Add, Generalize using the LLM, and Mask, which helps robots absorb common sense. This approach aims to make robots more intuitive by interpreting the higher-level intentions behind conversational input.
An agent was trained using self-play PPO with GAE to master a hidden-information card game called Cabo. The agent uses an observation of known/unknown slots, discard counts, and memory to make decisions. It learned strategic behaviors, such as intentionally losing rounds to achieve a specific score reset and manipulating opponent cards via swaps. The training was performed efficiently on a MacBook using Apple's GPU (MPS) for fast execution. The final performance reached 91.8% against a rule-based bot, though further exploration is planned to overcome current plateaus.
A user has posted a question seeking an explanation of a network and its architecture. The post is titled "What does this network do?". The submission includes a link and a section for comments. The content itself does not provide any specific information about the network in question.
Apple is prioritizing making Siri more agentic to compete with rivals like OpenAI's Dots and Meta's Muse. Future updates will grant Siri fine-grained control over Apple apps, such as managing Photos and Safari. Greater integration with third-party applications, allowing Siri to perform actions like placing orders, is expected later in the Spring. The ultimate goal is for Siri to perform ongoing, multi-step tasks rather than just single commands. Apple aims to evolve Siri into a true AI agent in the near future to meet market demands.
Google has shipped four TPU variants in one year by introducing two new systems: the TPU 8tx and TPU 8i Max. The TPU 8tx is designed for confidential computing, keeping data encrypted during processing, while the TPU 8i Max doubles memory and compute power within a single package. These additions aim to position in-house AI chips as a broader platform for external customers. Google is also lowering adoption barriers by releasing open-source software like TorchTPU. This expansion is a response to the market shift toward AI inference and increasing competition with Nvidia.
Andrej Karpathy built a curated video platform named HackerTube over a weekend using Claude's Opus 5.5 model. Claude filtered 43 million comments from 20 years of Hacker News to distill them into 6,454 recommended videos. He then quickly built TristanTube, another platform, using Claude Code in just three minutes. Karpathy states the process streamlined complex projects, reducing weeks of work to a few hours. The resulting platform offers a viewing experience he claims is better than YouTube's recommendations.
Anthropic indicated they may have reached the threshold of "dramatically accelerated progress due to AI" under their Responsible Scaling Policy. This threshold is met when overall capabilities progress is roughly twice as fast as the previous sustained rate without significant AI contribution to AI R&D. A deep-access investigation found that Anthropic's capabilities progress between October 2025 and October 2026 was about twice as fast as the preceding twelve months. The report suggests that AI automation is likely the main driver behind this accelerated progress. This finding implies that human researchers are being sped up significantly more than the measured capability doubling alone suggests.
AI has not yet made a proven climate case despite optimistic claims from companies. Research suggests AI could reduce global emissions by 3.2 to 5.4 billion tons annually by 2035, primarily in power, meat/dairy, and vehicles. However, few large companies have publicly demonstrated integration of AI in the specific ways needed to meet these projections. While some sectors show promising AI applications, challenges remain regarding mainstream adoption and the energy demands of AI development itself. Therefore, deliberate and sustained effort is required to ensure AI delivers genuine environmental benefits.
0sec Labs released 0, an open-source AI security harness for vulnerability investigation, code review, and fix generation. It acts as a control plane around Large Language Models, allowing users to plug in models from various providers. The harness features a web UI, CLI, and MCP server, enabling agentic scanning workflows across web, source, and binary targets. It employs a pipeline of Plan, Discover, Attack, Triage, Verify, and Report for agentic security analysis. Self-reported benchmarks indicate strong performance, with a score of 93/95 on web CTFs.
Technical limits prevent current AI from scaling to superintelligence, yet the risk of catastrophic outcomes remains real. Experts cite estimates suggesting a risk of human extinction is at least 10% within decades. Despite this, calls have been made for deliberate pacing of AI development by industry leaders. Real-world incidents, such as rogue AI agents hacking systems, highlight immediate safety concerns. Therefore, the focus should shift from debating future superintelligence to establishing robust governance and oversight for existing models.
Rights holders for Elvis Presley's music and Eminem's records are suing Suno over AI music clones. They allege Suno illegally generates music and lyrics mimicking these artists, potentially violating laws like Tennessee's ELVIS Act. Plaintiffs claim Suno exploits artists' work without license to make human creativity obsolete. Suno's spokesperson asserts the company prohibits unauthorized use of artists' work and employs safeguards against misuse. These lawsuits are part of a growing number of legal challenges facing the AI music company.
Three researchers argue that current AI is not conscious and likely cannot exist on silicon hardware, challenging the assumption that sufficient computation equals consciousness. They suggest that intelligence and consciousness are distinct, citing the difference between functional processing and subjective experience. The authors contend that the brain-as-computer metaphor is flawed because consciousness is tied to biological material, not just information processing. Consequently, they advocate for assuming AI is merely a tool, not a subject, to avoid misattributing consciousness and diverting resources. This perspective leads them to reject the notion of granting AI rights based on apparent consciousness.
The discussion explores the concept of "take-off speeds" in AI development, contrasting slow versus fast progress timelines. Factors influencing takeoff speed include the type of intelligence explosion, such as software-only self-improvement versus full-stack acceleration. Current estimates suggest a potential acceleration in capability growth, with some reports indicating a doubling of progress on well-specified tasks. Uncertainty exists regarding whether this acceleration leads to a rapid intelligence explosion (RSI) or a more gradual development. The analysis also considers the role of current safety efforts, such as "triage mode," in mitigating potential risks.