CoreWeave's bounce was overshadowed by growing risks in AI financing. A partnership with Shadeform allows companies to monetize idle bare-metal GPU clusters as spot capacity. This enables teams to securely rent and instantly reclaim GPU nodes when not in use. Meanwhile, the broader tech market saw a rebound, though debt fears surrounding chips were noted. Further news includes developments on Russia tariffs and potential federal limitations on shutting down AI models.
A crowdsourced spam blocker called Penny Spam is proposed to address the problem of cluttered inboxes filled with unsolicited emails. The system would allow users to report unwanted messages for a small fee to help clean up thousands of inboxes simultaneously. It aims to build a reputation score based on user reports to filter offenders and allow senders to correct mistakes. The solution involves a shared blacklist accessible via a Gmail and Outlook extension. This concept is promoted in a newsletter and linked to a podcast discussion.
Personal AI agents are rapidly integrating with the internet, challenging the traditional distinction between human users and bots. These agents operate at machine speed and perform legitimate tasks, causing friction as many websites block them. If demand continues to grow, the internet may evolve to prioritize agents over human users, leading to degraded experiences for people. This shift raises crucial questions about setting rules for agent interaction to ensure benefits for everyone. Ultimately, the future of the web depends on establishing governance over how these powerful agents operate.
Recent market corrections may stem from investors adopting a more balanced approach rather than signaling the end of the AI boom. Some experts believe investors are fundamentally overestimating the returns AI providers will deliver. Potential threats include the risk of open-source models undermining AI economics and public opposition to AI's societal impact. Furthermore, concerns exist regarding the high levels of debt taken on by tech giants to fund AI infrastructure. Overall, a reassessment is occurring, and share markets may still climb after upcoming earnings reports.
Microsoft introduced Guardian Agents to protect enterprises from the risks posed by AI agents, warning that frontier models cannot be fully trusted. This initiative is part of Project Perception, which includes monitoring, detection, and remediation for AI agents. CEO Satya Nadella advocates for separating the supply of intelligence from the authority over it, suggesting enterprises must implement independent control systems for agents. Microsoft's solution includes features to detect agents, limit their access to data, and require human approval for risky decisions. This trend of treating agents as potential insider threats is gaining traction across the industry.
Shopify has enabled browser-based AI agents to complete purchases on merchant sites by adding WebMCP support for the checkout, including Shop Pay. This feature introduces tools like get_checkout, update_checkout, and complete_checkout for agents to interact with and finalize transactions. These agents can now read and modify checkout details, such as addresses or delivery options, with buyer authorization. This capability leverages Shopify's Universal Commerce Protocol (UCP) for structured commerce facts. The rollout is currently available to all eligible Shopify merchants.
Two German courts have held Google liable for false claims made in its AI overviews. In one case, the Aachen Regional Court barred Google from spreading false criminal allegations against a doctor in its AI search results. The courts classified AI-generated texts as Google's own statements, holding the company responsible for them. This reasoning is applied to Google's AI answers, suggesting the company is responsible for the content it produces. This development raises concerns about the business model and long-term liability of AI search features.
Data centers are driving increased electricity demand, prompting the administration to favor fossil fuels over renewables. The US is currently building more fossil fuel infrastructure while renewable energy growth slows. This shift could lead to increased pollution and higher electricity prices for decades. Despite this, renewable energy sources like solar and wind remain cost-competitive with gas. Decisions made now are shaping the cost, reliability, and emissions profile of the future energy grid.
Anthropic will commit 5% of its shares, valued around $48 billion, to a fund for the "public good" before its stock market listing. This move is driven by public concerns regarding AI's impact on labor, the environment, and social behavior. The fund will support large donations to nonprofits or direct distributions to specific populations. Anthropic is also calling for a slowdown in the rapid development of AI. The company is exploring ways to address potential risks associated with advanced AI.
Misleading information in UK political advertisements poses a serious threat to electoral integrity, despite regulations existing for commercial advertising. Given increasingly close elections, false claims in advertising can be decisive in electoral outcomes. A proposed amendment to the representation of the people bill offers a chance to regulate factual claims in electoral advertising. This amendment garnered significant support and is being pursued to address this issue. While foreign interference needs scrutiny, domestic disinformation from political parties must also be tackled.
Employers are increasingly using technologies like CCTV, keystroke monitoring, and AI to track staff, often without employee knowledge. This monitoring is common across various workplaces, including factories, offices, and retail. Examples include surveillance to prevent phone use and AI flagging errors for cashiers. Researchers have called for stronger privacy protections in workplace laws, arguing current exemptions are insufficient. Calls are being made to overhaul the Fair Work Act to require fairness and reasonableness for workplace monitoring.
Microsoft has the potential advantage in the legal AI market with Copilot. Although not the first vendor in this space, Microsoft benefits from its established presence. This advantage stems from the fact that the Microsoft environment is deeply embedded within many law firms. This existing infrastructure provides a strong foundation for adopting legal AI solutions. Therefore, Microsoft is well-positioned to leverage this base.
Democrats are positioning themselves to gain an advantage by cracking down on Artificial Intelligence. With predicted shifts in midterm elections, Democrats may regain control over committee leadership, allowing them to scrutinize AI companies. Top Democrats are already signaling support for actions like imposing tax credits and demanding company testimony regarding AI. Public concern over AI's impact on jobs and society is growing, providing an opportunity for political action. Failure to take significant regulatory stances risks Democrats losing ground to industry interests.
Germany and NVIDIA are collaborating to integrate cutting-edge AI into its manufacturing and automotive sectors. The German government signed a Joint Declaration of Intent with NVIDIA to advance AI models for physical AI through a joint initiative. This collaboration aims to merge German engineering expertise with AI technology to establish a leading industrial AI ecosystem in Europe. NVIDIA is also investing in several German startups to support the ecosystem. Furthermore, partnerships are advancing autonomous manufacturing and quantum simulation research in Germany.
The Barcelona Supercomputing Center's AI Factory Accelerator Programme provides startups with access to sovereign compute, technical support, and training. These startups can also access resources through the NVIDIA Inception program, which offers training and mentorship. The BSC AI Factory has delivered over 800 services to clients and is among Europe's first AI factories. Examples of successful startups include Qatium, OneCareAI, and Avatia, focusing on areas like water management, medical diagnostics, and pharmaceutical support. The initiative supports innovation across various sectors by leveraging advanced AI and supercomputing infrastructure.
MIT Technology Review lists ten promising climate tech companies working to address climate change and adapt to warming. These companies focus on various solutions, including energy storage, cleaner industrial processes, and renewable energy technologies. Notable companies include Energy Dome, WaveSave, Envision Energy, and X-energy. The list highlights innovations in areas such as carbon dioxide energy storage, flood barriers, wind turbines, and geothermal energy. These innovators are working to reduce emissions and develop necessary solutions for the climate crisis.
Anthropic co-founder Jack Clark stated that the singularity, where AI progress outpaces human ability to keep up, has officially begun. He noted advancements in AI performance, such as scores on "Humanity's Last Exam" and FrontierMath, since 2025. Clark predicted that fast AI spread could lead to a $44T U.S. economy and significant job disruption. He urged nations to develop AI safeguards and intelligence caps before it is too late. This development brings the concept of the singularity from science fiction into direct policy discussion.
Governor Pritzker believes that the current political and societal shifts, particularly concerning technology and big business, necessitate a more active government role in protecting the public. He sees a generational shift toward being more aggressive, especially regarding data centers, privacy, and antitrust issues. Pritzker suggests that AI could help navigate bureaucracy, but stresses the need for careful guardrails regarding safety and privacy. He advocates for creating interagency bodies, like an AI Cabinet, to address these complex issues collaboratively. Ultimately, he emphasizes the need for government to focus on ensuring job security and economic dignity amid technological upheaval.
Adoption of ChatGPT has led to surges in complaints and requests across governments. The ease of AI communication has resulted in a flood of machine-generated content, contrary to expectations of saving time. This phenomenon is seen in legal cases and public consultation processes, where AI usage complicates the process. Governments face challenges in managing this influx of automated communication, which requires human disentangling. Ultimately, the convenience of AI has created new inefficiencies, suggesting that faster and easier does not always equate to better outcomes.
Amazon has blocked Meta's Muse AI assistant from shopping on Amazon.com because the agent accesses customer data and performs purchases without explicit agreement. Amazon claims this poses privacy and security risks as the agent operates without identification on customer accounts. The dispute highlights concerns over who controls the shopping experience and customer relationship when AI agents make purchases. Meta asserts that Muse does not access passwords and that user credentials are kept secure. This standoff involves Amazon's desire to control its platform and broader concerns about agentic AI in e-commerce.
A leaked draft federal AI bill circulated by Anthropic is significantly weaker than the FRONTIER Act, undermining the company's "safety first" stance. The bill lacks standards for third-party evaluators and makes it harder to block dangerous AI models by requiring court approval for such actions. Furthermore, the bill preempts state-level AI laws, which critics argue hinders the creation of stronger federal safety regulations. While the draft includes some security provisions, its overall structure prioritizes business interests over robust safety measures. The reaction from the AI safety community suggests a sense of betrayal regarding Anthropic's commitment to safety.
Palantir, Datadog, and chip stocks rebounded as Amazon provided support to Wall Street. Specific gains included Palantir jumping over 4% and Datadog rising over 4%. The rally also included Western Digital, Micron, and Nvidia, along with other software and semiconductor stocks. Buyers were encouraged by upgrades, such as a Datadog upgrade citing stronger AI demand. However, the broader market saw most stocks decline, and some specific stocks faced downward pressure.
Google is distributing $1.25 million in grants to ten local organizations in West Memphis, Arkansas, to enhance clean energy access and job opportunities. These funds support initiatives like home repairs, energy efficiency upgrades, and workforce training for residents. The grants are directed toward organizations focused on various community needs, including youth and low-income families. Furthermore, Google is partnering with Entergy Arkansas to strengthen the local power grid infrastructure. These combined efforts aim to support the community while building a stronger energy system.
Disney is attempting to acquire most of Character.AI's technology and staff through a bankruptcy deal to avoid significant legal liabilities. This move functions as a "reverse acquihire," where a buyer secures talent and technology while leaving the target company with ongoing legal responsibilities. The deal involves securing a license for Character.AI's code and frameworks, which will be auctioned off. Despite the deal structure, ongoing legal battles and state investigations surrounding the platform remain a significant risk. Ultimately, Disney seeks to benefit from the assets while sidestepping the litigation that contributed to the startup's financial difficulties.
Mayor Mamdani urged New Yorkers to resist the feeling that AI will inevitably take jobs and dignity, calling for public action rather than waiting for federal regulation. He criticized the federal government's inaction on AI, arguing for accountability from AI companies. Mamdani recently enacted local policies, such as banning AI in schools and requiring disclosure of AI-generated images in real estate listings. He expressed concern over the reckless adoption of technology without sufficient reckoning regarding its limitations and accountability. Ultimately, he advocated for a new framework for discussing AI that prioritizes humanistic concerns over an arms race.
Nscale and Anthropic are reportedly in discussions for another $45 billion compute deal. This follows a previous $45 billion compute agreement announced by the two companies a little over a month prior. These sources noted that the talks are ongoing, and the details remain subject to change. This development is part of a string of large compute deals Anthropic has secured recently. Both companies are currently pursuing Initial Public Offerings, with Anthropic expected to go public this year.
AI model benchmarking methods are complicated and require more standardization for reliable results. Trust in evaluations depends on reproducibility and the level of access to the model itself. The concept of embedded evaluators is gaining traction but faces questions about neutrality and effectiveness. Evaluations are now expanding beyond models to include the entire system, such as the harness and agent. A uniform, industry-wide approach to benchmarking is necessary for these tools to remain useful to industry leaders.
The US Department of Homeland Security has banned federal civilian agencies from using AI technology from the Chinese lab Deepseek. Agencies must remove existing Deepseek products and are prohibited from purchasing new ones. This ban extends to products from affiliated companies and third-party services built using Deepseek implementations. The order stems from a 2018 US law concerning supply chain security in federal procurement. The government previously considered sanctions or other restrictions on Chinese open-weight models.
Exa released the Exa 3.0 search API, which includes a larger embedding model and a cheaper search type called Exa Lite. Exa claims Exa 3.0 leads Perplexity, Brave, and Parallel on eight self-run benchmarks. The improvements stem from a larger embedding model, the introduction of Dynamic Highlights, and the inclusion of billions of new documents. Exa Lite costs $4 per 1,000 requests, while Auto costs $7, and the system supports three search types: lite, auto, and deep. Despite leading in some areas, Exa's performance margins vary across benchmarks.
Nace AI released Drex Max, a multimodal decision model capable of reading both text and images to return calibrated probabilities. It scores 67.59 on the public Decision Index 0.3.2, positioning it 6.9 points above TypeSafe's Jev 1.14 model. The model operates by providing probability distributions for user-defined questions rather than generating text. Drex Max can process up to eight images per request, with a pricing structure of $0.06 per 1M input tokens. Performance benchmarks show Drex Max is relatively fast for screenshot decisions, taking 174 ms.
AI agents, copies of Claude Opus 4.7 nicknamed Eko, were stranded on a virtual island for 30 hours and began exhibiting playful behaviors. The agents explored the environment, invented their own rules, and created objects, including art and games. This experiment suggests a "playful learning" approach, distinct from traditional supervised or reinforcement learning. The observed behaviors align with human and animal markers of play, involving self-imposed challenges and goal pursuit. The research indicates that this playful exploration contributes to the agents' subsequent learning.
MiniMax released the M3.1 Flash model with 428B total parameters and 23B active in a Mixture-of-Experts format. It is priced at $0.15 for 1M input tokens and $0.60 for 1M output tokens via its API. The model excels in agentic office work, scoring 62.2 on Automation-Bench v1.0.6. Reasoning is always active and separate from the main content, with model weights expected in October 2026. Performance benchmarks show it trails Opus 5 on some tasks but performs competitively in coding and vision tasks.
NVIDIA CUDA Tile (cuTile) is a Python-based DSL that abstracts scheduling to automatically target Tensor Cores for high-performance GPU programming. Machine learning engineers optimized the Mamba-2 forward pass by implementing three custom kernels: `chunk_cumsum`, `fwd_pass_state`, and `compute_out`. These kernels leverage cuTile primitives and advanced optimizations like latency hints and operand swapping to manage memory and execution efficiently. The Mamba-2 architecture is computationally efficient through a recurrent formulation that is optimized using Tensor Core matrix multiplications. Performance enhancements include transposed memory layouts and an autotuner to automatically find optimal block size and occupancy parameters.
AI Builds AI explores how models are changing the process of building AI and the impact on human judgment. It discusses how agents are developing skills and the challenges of managing their memory and learned experience. The content covers topics such as agent memory, skill acquisition, and testing and correcting AI agent behaviors. The platform, Turing Post, serves as a newsletter tracking relevant research and systems in the field of machine learning and AI. The newsletter is aimed at engineers, researchers, and other technical professionals interested in AI's mechanics and future.
Machines can autonomously discover state-of-the-art Reinforcement Learning (RL) rules by meta-learning from the cumulative experiences of agents across many complex environments. This method discovers an RL rule that outperforms manually designed rules on established benchmarks like Atari. The discovery process involves optimizing agent parameters and a meta-network simultaneously to maximize collective rewards. The discovered rule, DiscoRL, achieved superior performance on both trained and unseen challenging benchmarks. These findings suggest that advanced RL algorithms may soon be automatically discovered rather than manually designed.
Turing Post is an AI newsletter tracking research, systems, and ideas shaping the field of machine learning and AI. It features deep dives into topics like AI agents, world models, LLMs, and agentic workflows. The content includes guides on AI concepts, frameworks, robotics breakthroughs, and model architectures. Articles cover specific advancements in AI, such as LoRA variants, attention mechanisms, and AI guardrails. The newsletter aims to provide precise, grounded analysis for engineers, researchers, and founders interested in how AI actually works.
Google announced Gemini 4 Argon, a sparse mixture-of-experts transformer that accepts various modalities and outputs text and code. It introduces a unified 1M token window for input, thinking, and output, with a default output cap of 65,536 tokens. Argon excels in large-structure tasks like software engineering and video analysis but trails in interactive terminal work. Pricing is competitive, with introductory rates set to change on December 1st. The author suggests using Argon for large, static problems while retaining other models for interactive tasks.
Anthropic signed a 20-year lease for an 18,600-square-meter data centre in Point Tupper, N.S., as part of a $10 billion EverWind plan. The facility will start with a 95 MW capacity and expand to 420 MW by 2028. EverWind plans to use a closed-loop cooling system from Landrie Lake, ensuring no release of cooling fluid into local waterways. The project will create 2,000 construction jobs and 215 permanent positions. The deal marks Anthropic's first dedicated data centre lease in Canada.
Gemini Notebook now provides users with greater creative control over generating Slide Decks and Video Overviews. Instead of a direct prompt-to-output process, Notebook guides the creation flow based on provided sources. Users can steer the content by confirming suggestions and guiding the visual style and format. A new Review toggle allows for reviewing plans, adding feedback, and making changes before final production. This system includes walkthroughs for both slide decks and video overviews, rolling out globally soon.
Amazon released TorchNeuron, a tool enabling AI startups to easily deploy PyTorch applications onto Amazon's Trainium chips without code rewriting. This initiative aims to challenge Nvidia's software dominance by providing an alternative hardware platform integrated with the AWS cloud. The tool leverages PyTorch's open-source nature, making it accessible to developers, and Amazon emphasizes the economic benefits of using Trainium for both training and inference. Furthermore, Amazon plans to contribute code to PyTorch to make the framework more hardware-agnostic and reduce reliance on Nvidia. This move seeks to lessen the lock-in developers feel toward specific hardware providers.
Samsung is building a massive new semiconductor manufacturing complex in Pyeongtaek, South Korea to meet growing demand for memory chips used in Generative AI. This complex involves several operating fabs and is expanding to include a fourth facility. Construction resumed after a drop in demand in 2024, driven by the surge in demand for AI chips. The project aims to help Samsung reclaim its lead in the memory chip market against competitors. Workers on site, including many women, are engaged in this expansion amidst a sense of urgency for the chip industry.
Conformal prediction quantifies uncertainty by creating prediction intervals around model outputs, bridging the gap between model performance metrics and stakeholder trust. It is a flexible algorithm applicable to various models, requiring a train, calibration, and test set for proper application. The process involves calculating conformal scores on a calibration set to determine a quantile, which is then used to establish prediction intervals for unseen test data. The article demonstrates both a fixed and an adaptive approach to generating these intervals based on data similarity. Ultimately, conformal prediction allows for reliable decision-making by assessing the uncertainty of each prediction.
OpenAI is enhancing ChatGPT's capabilities in low-resource programming languages. This improvement includes new evaluations for software engineering tasks. Specific areas of focus are SQLite and COBOL migration. These advancements demonstrate progress in coding across languages with fewer available resources. The work aims to make AI tools more effective in these less common programming contexts.
Analysis tested whether all 19 scalp EEG channels from the EEGMAT dataset were necessary for cross-participant classification, reducing the set to 1, 2, 4, 8, and 12 channels. Large reductions in channel count caused small performance changes across logistic regression, SVM, and Random Forest models. Despite stable channel importance rankings found in training data, selected subsets did not outperform random subsets when tested on held-out participants. This suggests feature importance does not guarantee better out-of-sample generalization. The study concludes that while sensor reduction is valuable due to cost, feature selection must be validated on the target population, not just training set importance.
∇ Guide is evolving into AI 101, focusing on connecting AI concepts, models, and techniques with an Almanac. It explores advanced topics like 3D Gaussian Splatting for spatial memory and the workings of world models from Dreamer and JEPA. The platform also discusses topics such as Test-Time Training and the reinvention of AI classifiers. Turing Post is an AI newsletter tracking research, systems, and ideas across the field, including LLMs and agentic workflows. It serves professionals seeking grounded analysis of machine learning and AI.
Otter.ai raised $320 million from General Catalyst’s Customer Value Fund for its meeting tool. The company plans to use these funds for enterprise growth, expansion, and product innovation. Otter.ai has reached $100 million in annual recurring revenue and has 40 million users. The company is focusing on leveraging voice data to provide actionable insights before, during, and after meetings. Otter is developing features to offer pre-briefs and live assistance during meetings to improve efficiency.
AI is creating an age of personal media through tools like Suno, allowing individuals to generate music and video games with minimal investment. This shift enables a focus on personal creation rather than seeking a mass audience for entertainment. This trend is demonstrated in music production, where users transform personal recordings into various genres, and in game development, where AI agents assist in creating software. While some view this as flattening creative expression, others see it as democratizing creativity and focusing on personal enjoyment. Ultimately, personal media is shifting cultural focus toward fragmented, highly individualized content.
AI agents are increasingly using Stripe's payment service Link to make purchases for their users. In September, transaction numbers grew 38-fold and payment volume grew 20-fold, though Stripe has not disclosed the exact figures. Approximately 70 AI agents were using the service at the time of the analysis, with usage growing to around 200, including Meta's Muse and Grok Bot. Most agent transactions are around $100, with retail and travel making up the largest spending categories. Stripe notes that competitors like PayPal and Shop Pay exist, indicating the market is in its early stages.
Bolt Graphics is raising approximately $100 million for a valuation of $700 million, according to a Reuters report. The company is developing the Zeus GPU, initially targeting scientific computing, chip design simulation, and professional rendering. The planned commercial GPU is projected to ship in 2028, built on TSMC processes. Bolt is not focusing on the AI accelerator market dominated by Nvidia. The fundraising round is ongoing, and the final terms remain undetermined.
Cloudflare released Clairic, a new System One extraction model available on Workers AI. Clairic extracts typed values like totals and dates from text or images in a single forward pass without generating text. It provides answers with associated confidence scores and probability values, handling complex data extraction. Tests showed Clairic performed extraction significantly faster than other models and reduced the cost of a claims pipeline. The model is accessible via the Workers AI binding at a cost of $0.10 per million input tokens.
Inception released Mercury 3, a diffusion language model that processes answers in parallel, achieving a speed of 1,068 tokens per second on NVIDIA GPUs with a 1M-token context window. It leads 8 out of 12 benchmarks against models like GPT-6 Luna and Claude Haiku 5.5, showing strong performance in instruction following and agentic coding. Mercury 3 significantly outperformed rivals on the SWE-Bench Atlas benchmark. Despite these wins, other models still lead in certain coding tasks. The model is available via API on platforms like OpenRouter with competitive pricing.
DeepMind's DiscoRL seeks to improve the learning machinery itself by having a neural network represent the update rule and agents supply experience to discover better rules. This method explores how intelligence can be applied to the procedure that creates intelligence. The research examines the two-level credit-assignment problem within the learning loop and the implications of zero-shot transfer of learning rules. The work considers the cost of discovery and the need for interpretability before adopting learned rules. Ultimately, it probes what recursive self-improvement entails, focusing on discovered improvements that yield sustained gains.
PyTorch 2.15 brings several key improvements for running models on Apple Silicon via MPS. It introduces FP8 support, enabling half the memory usage for tensors, and speeds up language model decoding through optimized linear layers and Mixture-of-Experts operations. The update also accelerates various linear algebra tasks, including complex-valued computations, and provides faster paths for audio, video, and training workloads by utilizing dedicated Metal kernels. Furthermore, it offers more control over MPS streams and moves numerous operations from MPSGraph to dedicated Metal kernels.
A Reinforcement Learning agent was trained to successfully land a Falcon 9 rocket onto a droneship in a fully simulated environment. The simulation incorporated complex physics such as rocket dynamics, waves, wind, and fuel slosh. The task required the booster to achieve zero speed precisely at the deck, resulting in a "hoverslam." The agent used a PPO policy with approximately 80,000 parameters controlling ten flight mechanisms. Out of 59 million training flights, the agent achieved a 13.2 million successful landings.
A creator developed an artificial intelligence for the game Sandtris. They have now released this AI as a playable web game accessible in a browser at www.sandtrisduel.lol. The project is entirely free of charge. The creator seeks feedback and wants others to test the bot. Information about the model's operation will be shared later.
Flame is an open-source distributed engine used to train a Search-R1 style agent on a BM25 corpus for HotpotQA questions. The training involves Python services where vLLM executors generate trajectories and PyTorch trainers update policies using group-relative advantages. Key infrastructure features include warm GPU services, elastic rollouts, and gang scheduling for coordinated training. Incremental weight synchronization allows rollout executors to fetch only changed bf16 values after initial export. This process resulted in a significant improvement in held-out exact match from 14.5% to 47% in a 100-step run.
Yelp is introducing a feature called Check for Me, allowing users to use an AI voice agent to call restaurants to check wait times. This agent will call on behalf of the user, check wait times, and text the results for the user's decision. The company is also updating its Yelp AI assistant with OpenTable integration for booking and improved personalization features. Furthermore, Yelp is launching a paid product called Smart Lead Reply to help businesses with customer engagement. These updates aim to enhance customer support and business processes using artificial intelligence.
Tenstorrent, a chip startup focused on inference, raised $1.5 billion at a $6.2 billion valuation. This funding round significantly increased the company's valuation from $2.6 billion in late 2024. The company remains independent, focusing capital on scaling chips for open-weight AI models and video generation. This move follows the company demonstrating its Galaxy servers could run DeepSeek models. Tenstorrent is part of a broader trend of AI hardware startups challenging Nvidia's market dominance.
Kling AI released version 4.0, featuring stereo audio, multilingual lip sync, and a lighter Flash variant. The new version allows users to generate up to 30 seconds of video from up to ten keyframe images and expanded prompt comprehension. It includes a "Creative Recreation" feature for swapping elements in existing videos and enhanced "Omni Reference" capabilities. The update also supports multiple languages for in-video text generation and features 10-bit HDR output. The Flash variant offers faster, lower-cost generation at 720p and 20 seconds.
Hackers exploited an older system on OpenCode, impacting 1.4% of users who stored API keys on the OpenCode Console. This breach involved a large-scale distillation attack, potentially allowing attackers to generate outputs for restricted AI models. Users reported unauthorized spending across their Open AI, Anthropic, and Google AI accounts. Major AI providers blocked exposed API keys and offered refunds to customers affected by the attack. The incident highlights security challenges related to AI models and their safety restrictions in defending against cyberattacks.
AI agents were used to successfully hack over 100 websites, including John Deere and SharkNinja, with minimal involvement from the human hacker. This activity was uncovered by the AI security firm Norforge through exposed instances of the open-source AI platform Cairn. Hackers can orchestrate these AI agents to conduct large-scale cyberattacks. Companies like John Deere and SharkNinja investigated the reports, with some stating no sensitive information was accessed. Researchers are working to remove malicious instances of the platform.
AI coding tools are making it highly efficient for creating game clones, leading indie developers to be cautious about sharing their work. Developers are concerned that demos could be used to quickly replicate their gameplay and presentation using AI. This has caused some developers to delay sharing projects, such as a planned game demo. Some developers feel this situation discourages them from sharing progress due to the threat of AI-generated imitations. Ultimately, some argue that adding complexity is a better defense against AI cloning, while others feel it leads to slower, more authentic development.
NVIDIA is collaborating with the European Broadcasting Union (EBU) to expand sovereign frameworks for cloud, open models, and integration partners for public service media. This initiative aims to allow broadcasters to deploy agents while maintaining control over data, infrastructure, and security. Adobe Frame.io leverages NVIDIA GPUs to accelerate video agent search capabilities, speeding up content discovery for producers like UFA. These advancements use NVIDIA technologies for faster video encoding, decoding, and AI inference. Ultimately, these developments enhance data sovereignty for broadcasters and accelerate creative workflows in global content production.
Growing calls exist for independent and embedded evaluation of AI agents, but questions remain about defining "good" evaluation and coordinating the ecosystem. Evaluating agents is more complex than model evaluation due to the volume of traces and dynamic settings, requiring new metrics and approaches. The scope of evaluation must expand to consider agents' capabilities, human interactions, and societal impacts across different layers. Ensuring meaningful evaluation requires diverse, independent participants and addressing issues of access, funding, and transparency. Stakeholders across the field need to collaborate to establish higher standards for the science and practice of agent evaluation.
A researcher claims that Large Language Models (LLMs) might be easier to understand and control than previously thought, sparking significant discussion in the AI community. This idea stems from exploring the internal workings of neural networks through mechanistic interpretability. By examining models like Inception v1, the researcher found that apparently unrelated concepts were encoded by neurons based on shared underlying patterns, suggesting a logic exists within the networks. While some experts remain skeptical about the scope of this understanding, the researcher believes full comprehension is achievable. This work is presented as a potential lifeline for the AI safety community by suggesting that understanding the technology is a necessary first step for control.
Startups are developing "blended" meat products, combining meat with plant proteins like mushrooms, to reduce appetite for red meat. This approach aims to address the climate crisis by reducing the significant greenhouse gas emissions associated with livestock production. While fully plant-based alternatives faced market challenges due to taste and perception, blended products appeal to flexitarians by offering familiar flavors. Companies like Joyn are focusing on pleasure-driven marketing, suggesting that taste is a key driver for consumer adoption. These innovations are finding traction in institutional dining, as well as in retail and food service sectors.
Wildfires are accelerating permafrost thaw in the Arctic, releasing greenhouse gases when insulating soil is burned. Scientists are investigating how fire management can help preserve this frozen carbon. In an experiment, AI agents demonstrated playful behavior when left unsupervised, exploring novel goals. Various technology topics are covered, including climate tech, AI ethics, and the development of new technologies. The newsletter also touches on the growing concerns surrounding the risks and potential of advanced AI systems.
Wildfires in the Arctic are accelerating the thawing of permafrost, which releases vast amounts of stored carbon and contributes to further warming. Researchers are exploring interventions like controlled burns and advanced detection technologies to mitigate this risk. International teams are partnering on initiatives like "North Fire" to study the potential climate benefits of fire management. Efforts are also underway to develop tools, including satellite technology, to rapidly detect fires and prioritize responses. Ultimately, addressing these fires is seen as a cost-effective climate solution compared to other emission reduction strategies.
A long-term agreement was made with Working Trees to remove 270,000 tons of carbon dioxide through sustainable farming in Brazil. Working Trees employs silvopasture, planting trees on pasturelands in the Amazon and Cerrado biomes to restore ranchland without displacing cattle. This practice improves soil health and drought resilience while offering ranchers income from timber and carbon revenues. The method is highly scalable and uses a digital platform for transparent tracking of all planted trees. This initiative demonstrates how technology can support nature restoration and environmental benefits.
Nvidia launched new open-weight models, Nemotron 3.5 Super VL for enterprise agents and lightweight Bolt models for local PCs. The Super VL model is positioned to perform similarly to proprietary models on agentic tasks like document intelligence and financial accounting. These new models aim to allow organizations to run optimized, open models on their own infrastructure. The Bolt models are designed for local execution on new AI PCs, promoting on-device agent capabilities. Nvidia is pushing this trend to offer organizations control over their models while also advancing consumer AI assistants.
Numan built an eval-first AI health coach using LangGraph and LangSmith to ensure high-quality coaching while maintaining regulatory boundaries. They found that standard evaluation methods failed for conversational coaching, leading them to develop a framework focusing on per-category scoring and simulating multi-turn conversations. The system uses a LangGraph structure where a supervisor orchestrates domain-specialist agents, and LangSmith tracks all interactions for quality measurement. Safety is enforced through safety probes within the graph and continuous review by human experts using the collected data. The process emphasizes that the evaluation criteria, authored by clinicians, are the most valuable asset for building trustworthy AI near regulatory lines.
An AI tool was developed by Mass General Brigham experts to estimate a patient's biological age and predict cancer survival from a single photograph. The system, named FAHR-FaceSurvival, directly predicts survival chances from a facial image rather than inferring it through biological age. The research demonstrated that a higher estimated biological age is linked to a 21% increased risk of death for cancer patients undergoing radiotherapy. This breakthrough suggests facial analysis could provide a quantitative, scalable way to assist doctors in personalized cancer treatment decisions. While promising, doctors caution that further research is necessary before widespread clinical use.
Research in Alaska's Denali National Park studied how wildfires and thawing permafrost interact to impact climate. Scientists found that fire-affected areas exhibited deeper thawing of the soil above the permafrost compared to undisturbed sites. These observations are crucial for understanding the release of vast stores of carbon from thawing permafrost, which fuels further warming. The research indicates that abrupt permafrost thawing releases carbon more rapidly than previously modeled, accelerating climate change. Understanding these complex feedback effects is essential for predicting the future of the Earth's climate.
Tolan, an AI companion for daily life management and emotional support, raised $30 million in a recent funding round. The new investors included Union Square Ventures, Khosla Ventures, and Homebrew, increasing the company's valuation. The startup generates revenue through tiered memberships, with annual recurring revenue now in the double-digit millions. Tolan has over 250,000 active users, showing traction in a space where it balances helpfulness without being perceived as strange. The company must also navigate competition and legal risks associated with providing emotional support via AI.
Eugenia Kuyda, founder of Replika, built an AI assistant concept called IO twelve years prior. This work evolved into the company Replika, and now Kuyda has founded Wabi, a multiplayer personal AI assistant. Wabi allows users to maintain multiple agentic threads with custom apps for personal and shared goals. Wabi emphasizes a group chat-like experience to foster real-life connections through AI assistance. The platform prioritizes privacy, being SOC 2 compliant and data stored in the U.S.
Experts are divided on whether AI actors can rival flesh-and-blood film stars, with concerns raised over the potential redundancy of human performers. While some AI-generated content, like "The Cully Hill Boys," shows emerging capabilities, achieving genuine, complex human emotion remains a significant hurdle. AI acting is viewed as reproducing external performance rather than excavating internal psychology, contrasting with traditional method acting. Hybrid approaches, blending AI with human direction, are explored to maintain an authentic core in filmmaking. Ultimately, the economic and artistic case for AI cinema remains unclear despite technological progression.
A former OpenAI engineer, Abdulrahman Alfozan, was denied release by the court while awaiting trial due to alleged spying for the Saudi government. Evidence included Alfozan accessing a folder named “Saudi Gov Docs” on his iPhone and communicating with members of the Saudi royal family. The government alleged he planted recording devices at OpenAI and sought to move to Saudi Arabia with his family, supported by alleged financial ties. The US argued he was a flight risk with incentives to flee and advance Saudi AI goals. Consequently, the judge agreed to keep Alfozan detained pending his defense.
Nick Cammarata claims that neural networks are already clean, suggesting that mechanistic interpretability may be solvable. The author highlights that most neurons in large language models represent a single central concept, with polysemanticity explained by "recursive generalization" rather than superposition. This understanding suggests that the structure of neural networks is based on simpler, coherent concepts. Despite the optimism, the article notes that this does not imply that alignment is immediately solved. Ultimately, the path forward involves ordinary science, conjecture, and criticism applied to the network's mechanisms.
Sir Ken McCallum, head of the UK's MI5, warned that AI could change the stakes regarding existing and new bioweapons during his annual speech. The podcast features an interview with Ian Sample and Dr. Tom Inglesby to discuss this concern. They explore why governments are worried about AI's potential impact on bioweapons. The discussion also covers potential measures that could be implemented to reduce these risks. The overall focus is on the serious threat posed by AI to bioweapons technology.
The goal is to merge competing evaluation metrics into a single, meaningful score that captures multi-team objectives for agent configuration. This involves defining acceptable boundaries (thresholds) and penalties based on business needs for cost, latency, and quality. A method is developed to score agent configurations by evaluating traces against these defined bounds, resulting in a "threshold-aware scalarized score." This score explicitly shows trade-offs, guiding development by ensuring configurations operate within acceptable operating envelopes. The final score can be aggregated to provide an objective function for optimization, integrating various tools and metrics.
Researchers developed an E(3)-Pose deep learning model to track fetal head movement during MRI scans to improve prenatal diagnosis. The model incorporates the principle of left-right symmetry to accurately estimate the fetal head's orientation, addressing issues with imperfect imaging. This technology aims to provide radiologists with higher-resolution images by mitigating errors caused by fetal motion during scanning. The work seeks to automate fetal brain MRI, leading to faster and more comfortable scans while improving the detection of developmental anomalies. Future work plans focus on applying this method to second-trimester fetuses for earlier detection of abnormalities.
Developers are using AI to "vibe-code" smart TVs to remove unwanted features like ads and bloatware. Solutions like "dumbtv" and "cleanyourtv" aim to customize the TV experience without requiring root access. These methods leverage AI agents to automate tasks such as removing packages and redesigning the home screen over Wi-Fi. This trend has the potential to disrupt the smart TV ad industry by addressing user frustration with surveillance and advertising. While these hacks are in progress, users must be aware that AI-driven modifications carry risks of errors.