Democrats are projected to win control of the U.S. House of Representatives in 2026 with an estimated 81 seats. This is due to Democrats gaining control of key battlegrounds and outperforming their 2024 presidential margins. The NBC News estimate projects Democrats will secure at least 258 seats in the House. This projection is based on analysis of past and current vote counts, not polling. Democrats are also pushing into districts previously won by Republicans with a larger margin in 2024.
Oracle announced a $20 billion capital raise on Monday, split equally between a $10 billion stock sale and a $10 billion bond sale. The stock price immediately fell about 6% to $122.61, hitting a three-month low. The net proceeds are designated for general corporate purposes, including capital expenditure, debt repayment, and acquisitions. This financing follows a previous statement by Oracle to cap additional AI-related financing at $20 billion through fiscal 2030. Investors reacted based on Oracle's financing needs, AI spending, and exposure to OpenAI.
Apple acquired Ascension Labs, the creator of the Ascend.io platform, to enhance its data management and autonomous engineering infrastructure. The deal, aacqui-hire transaction, closed on June 30, 2026, and was disclosed via the European Commission. Ascend.io specializes in declarative data pipeline technology and Agentic Data Engineering using AI agents for data workflow automation. Apple gained employment rights for key talent and a non-exclusive license to the acquired intellectual property.
Reverse game theory seeks to control powerful AI by designing systems that incentivize agents to follow desired rules. This approach, also known as mechanism design, focuses on creating rules that elicit specific behaviors rather than trying to understand the AI's internal workings. Potential methods involve rewarding models based on their predictive accuracy or pitting agents against each other in competitive scenarios. Another method suggests structuring interactions where AI models act as both solvers and auditors, similar to social systems. Ultimately, alignment requires integrating these mechanisms within the broader framework of economics and existing societal rules.
President Trump ordered government agencies to rebrand artificial intelligence as "Super Intelligence" in an executive order. This renaming is viewed as a potential political loyalty test, where usage of the term may signal allegiance to the administration. While the rebranding causes some procedural hiccups within government, the shift is expected to influence public discourse beyond official channels. Major tech executives have shown mixed adoption of the new terminology, with some using it while others still favor "AI." Ultimately, the article suggests the renaming may distract from the more important national conversation surrounding AI's societal implications.
Google is scaling defenses against complex scams in the Asia Pacific region through built-in protections, partnerships, and education. New Android features like Fake Call Detection help identify AI-driven impersonation and malicious apps. The company joined the Industry Accord Against Online Scams and Fraud to share threat intelligence with peers. Google.org is funding $6 million to help small businesses and individuals build scam resilience locally. Educational games are also being deployed across several countries to teach citizens how to recognize and avoid common scams.
US tech stocks recovered on Tuesday after an earlier selloff spooked investors regarding AI spending. While midterm elections did not have a huge immediate impact, predictions of increased regulation could influence the market. The S&P 500 and Nasdaq saw gains, with Nvidia and other chipmakers also rising. This recovery ended a brief selloff that had begun the previous week concerning the AI sector. Despite market movement, investments in AI-related chips and infrastructure remain strong on Wall Street.
Geoffrey Hinton, an AI pioneer, suggests humanity should know within ten years if AI risks can be safely managed. He warns that unchecked AI development could trigger major incidents, like power or banking shutdowns. Hinton believes we might have been better off without developing the technology, citing potential societal breakdown from mismanagement. He expressed deep distrust in some AI leaders, believing they prioritize profit over safety. Hinton advocates for regulatory oversight, similar to the FDA, to ensure AI benefits outweigh harms before release.
Google stated that using Google Search is inseparable from using artificial intelligence. The company is updating controls for AI features like Gemini and Google Search, allowing schools to restrict access to these tools. If schools block Google's AI features, students will lose access to Google Search. Critics view this as a "hostile ultimatum," suggesting Google is using its search monopoly to force the adoption of its AI products in education. New AI features, such as Guided Learning, are being integrated into Google Search to coach students through problem-solving.
Supermarket refrigerators and freezers in several European countries have malfunctioned following a cyber-attack where thermostats were remotely increased. The European Union has condemned this action as a deliberate attempt to disrupt the food supply. This incident occurs as the UK's cybersecurity body warns that AI-powered malicious software is likely to be used in attacks against critical UK infrastructure by 2028. The source of the recent attack remains unknown. This event raises concerns about the security of food supplies and critical infrastructure.
Artificial intelligence and robotics will eventually automate most economically useful tasks, potentially breaking the current relationship where employment determines income. This shift suggests that the problem will move from scarcity to the distribution of wealth, as machines will produce vast abundance. Universal Basic Income is one solution, but it fails to address the human need for structure, status, and purpose. The author proposes an "Omni-Game" economy, a system of paid, structured human activities designed to distribute surplus value without relying on economically useful output. This system would reward participation itself, offering alternatives to work by making human experience intrinsically valuable.
Ukrainian drone strikes targeting Russian data centers have spurred hasty plans for Russian tech companies to move their facilities underground for security. The Kremlin has mandated that data center owners reserve capacity for state needs and authorized potential use of foreign data centers. Russia is exploring options like building centers in coal mines with camouflage and allowing foreign data center access. These events highlight the vulnerability of Russia's centralized cloud computing infrastructure during the war.
Ballistic Ventures is seeking $490 million to raise for two new funds amid the growth of AI-driven cybersecurity. The firm filed with the SEC to raise $360 million for its third flagship fund and $130 million for Ballistic Ventures Fund B. This fundraising aligns with increased VC interest in cybersecurity due to the rise of AI-related cyber threats. The firm has previously invested in companies like Talon Cyber Security and Pangea. U.S. cybersecurity startups saw significant growth, with AI-enabled companies dominating venture deals.
Apple is updating App Store policies to restrict minors from using dating apps or AI companions that feature sexual or romantic content. All apps with dating or relationship features will be required to display an 18+ rating starting November 16th. Developers must include new content descriptors to ensure age ratings accurately reflect app features. The policy also requires apps to indicate if they share real-world locations with other users. This change extends to AI companion apps, bringing them under similar age-related scrutiny.
AI is boosting search activity, but this growth is not translating into faster revenue growth for Google. Google's Search revenue growth slowed despite rising search queries, leading to investor questioning about the return on AI investments. Google suggests AI queries are mostly incremental, but the impact on monetizing those searches remains unclear. Advertisers are still seeing good returns, but the mechanism of AI ads is not fully understood. Ultimately, figuring out how to effectively monetize conversational AI search is an ongoing challenge for the company.
Nexus Mods requires authors of newly uploaded mods to declare whether AI was used in their work. Mods tagged as "AI-assisted" must also include a public explanation of how and where AI was utilized. Nexus Mods will make AI tags more prominent on mod pages to ensure clarity regarding AI content. The platform will not ban AI-generated mods but seeks to regulate and attribute AI usage. Mod authors have 60 days to update existing mods to comply with these new requirements, with enforcement starting after January 2nd.
Financial leaders are divided regarding the risks associated with the current AI market. PIMCO's CEO suggests a downturn and a potential collapse, noting that the expensive AI buildout has heavily impacted debt markets. Major tech companies like Meta, Oracle, and Alphabet are actively issuing significant bonds to finance this expansion. Despite high levels of borrowing, uncertainty remains about how pricing and economics will evolve. Caution is growing among financiers, suggesting the peak of the AI market may be approaching.
Indie game developers are experiencing unexpectedly long approval wait times on Steam, causing them to miss planned launch dates. Developers like Studio Daimon are awaiting final approval for their games, with some citing ignored urgent requests. Valve acknowledged the delays due to a significant increase in submissions requiring re-review, taking up to twelve business days. This backlog is speculated to be caused by "AI slop" slowing down the review process. Valve is working on improvements for the build review process despite the current slowdown.
Exit polls for the 2026 U.S. midterms revealed that AI and data center expansion ranked last among voter priorities. Fifty-six percent of respondents considered these issues decisive or important in their voting decisions. Government corruption led the list with 90 percent, followed by the cost of living at 89 percent. The war in Iran was cited by 74 percent, and immigration enforcement by 73 percent. Despite a recent surge in public interest, AI still trailed all other surveyed issues significantly.
Starting in early 2027, Google will require schools to either accept its AI for students or lose access to Google Search. This means Gemini, Gemini Notebook, and Google Search will default to being off for student accounts, though schools can individually enable these features. Google's vice president for education stated that search and AI have become too fused to separate. Schools opting to enable search will receive either search with full AI overviews and chat or a "Guided Learning" mode. This development occurs within a context where Google dominates the US school market with products like Chromebooks and Classroom.
OpenAI hired former Army Secretary Dan Driscoll as a Senior Counselor for National Security Modernization to enhance its defense operations. Driscoll's experience in military AI modernization and his connections in Congress are expected to help OpenAI secure government contracts. This move reflects OpenAI's effort to align with the military and gain influence over AI regulations in the defense sector. Driscoll will work with various stakeholders to apply AI to national security and cyber defense. This action follows OpenAI's securing of a Pentagon contract after rival Anthropic faced restrictions.
Anthropic's amended S-1 filing reported $17 billion in revenue for the third quarter, marking a slowdown in growth to 47% from the previous quarter. The company recorded a $567 million operating loss for the quarter, attributed partly to higher compute costs. Committed spending for the company reached approximately $895.4 billion, primarily for compute and infrastructure. Despite the loss, the company remained profitable on an adjusted basis, though the operating margin decreased.
A concept for a pocket-sized "Agentic Work Device" called Blok aims to coordinate AI agents by connecting to various tools and providing delegation. The article also details methods for finding proven business ideas by analyzing successful ventures on marketplaces like Flippa. An 18-year-old dropout built an app called AbMaxx, which achieved $60K/month. Other featured ideas include a security focus for Lovable-built apps and a product called WaxPods for ear cleaning.
New research indicates that Guided Learning leads to significantly higher learning gains for students compared to standard tools. This method coaches students through complex concepts with step-by-step guidance rather than providing direct answers. Studies showed students with Guided Learning achieved 8.1 percentage point higher score improvements than those using standard resources. Google is rolling out Guided Learning in Google Classroom to allow teachers to assign these guided lessons and track student progress. This feature provides educators with insights into student interactions to facilitate more targeted support.
Google is releasing major updates to its education portfolio related to artificial intelligence. These updates include launching Guided Learning into Google Classroom globally and integrating Guided Learning capabilities into Google Search. New control measures have been implemented for Gemini, Gemini Notebook, Gemini in Classroom, and Google Search accounts for K-12 students. Furthermore, the announcement shares new research, funding from Google.org, and expanded AI training opportunities for educators. These actions reflect Google's focus on building technology specifically for educational use.
The Google AI Educator Series is expanding to five new countries: Spain, Brazil, Mexico, Australia, and sub-Saharan Africa. This no-cost, on-demand training provides educators with practical AI skills to create lessons and support student learning. Educators globally are using these tools to save time and make lessons more engaging, such as creating content in India and designing inquiry-based lessons in Korea. Upcoming modules will focus on building interactive assessments, guiding game-based learning, and strengthening academic assignments using AI. This expansion reflects a global trend of leveraging AI to address educational challenges and meet diverse student needs.
AI is transforming semiconductor manufacturing by accelerating simulations and predicting process behavior. Physics-based AI shortens simulation times up to 10x and allows for predictions of complex phenomena like plasma behavior in milliseconds. AI agents automate sequences of engineering tasks, and digital twins enable virtual testing of factory changes before physical trials. This creates an AI-defined fab where AI optimizes processes and helps engineers make faster, better decisions across the chip lifecycle. Ultimately, this connects production data with models to enable faster learning and action in manufacturing.
Tencent released WeVisOne-VL, an open vision-language model unifying detection, OCR, segmentation, and matting. This model includes a 9.6M-parameter decoder to generate pixel outputs, eliminating the need for a separate segmentation model. It can perform tasks like object detection, referring, counting, OCR, segmentation, and alpha matting in a single framework. The 4B model achieved a 71.3 F1 Mean on Dense200 detection, outperforming some baselines. This unified model is designed to replace multiple specialized AI components for developers.
Anthropic released FlashProve, an open-source zkSNARK system that proves a reply originated from committed model weights. It allows for provable inference, enabling verification of model outputs without revealing the model weights themselves. On a 400B Mixture-of-Experts model, it achieves about 110 tokens per second on a single H100 GPU. The system involves an offline commitment and an online proof generation process, offering post-quantum security. FlashProve has potential applications for regulators and users to verify text provenance.
NVIDIA demonstrated three optimizations to speed up Mixture-of-Experts (MoE) inference on a single DGX Spark system for the Ling-3.0-flash model. These optimizations included a new MoE kernel backend, an FP8 output layer, and built-in speculative decoding. Stacking these techniques increased the decode throughput by 42.5% while maintaining accuracy. The methods involved optimizing MoE computation, quantizing the language model head to FP8, and using multi-token prediction for speculative decoding. This work shows methods to reduce MoE decoding time on desktop AI hardware.
Soniox STT v6 is a speech-to-text model that reads agent text to recognize caller replies. It incorporates agent text input to allow for agent voice echo filtering, which removes leaked agent audio. This feature enhances the accuracy of transcription in cascaded voice agent scenarios. The model supports streaming and asynchronous modes across over 60 languages. Soniox v5 models will be deprecated on December 15, 2026, with automatic routing to v6.
Useful AI increasingly requires matching the model, inputs, and feedback to the specific job. Gemini 4 Argon transforms inputs like photographs into artifacts using tools and subagents, while Gemini 4 Carbon focuses on compounding reliability for long engineering tasks. Mistral's Timestral provides pretrained, zero-shot forecasting for numerical series, including uncertainty quantification via quantile outputs. The post discusses evaluating Argon systems and how duration and external signals affect Carbon-style agents. A proposed design pairs a numerical forecaster with a generalist agent for complex tasks involving both visual and numerical prediction.
AI research should focus on leveraging scalable methods rather than handcrafted knowledge, a principle known as the Bitter Lesson. This lesson suggests that handmade methods eventually lose to scalable approaches in AI development. Progress in large language models, for instance, was primarily driven by increasing scale and computation. As models grow, they may eventually weave their own numerical DNA, leading to recursive self-improvement without heavy human intervention. A recent study suggests that this self-development is becoming possible in frontier AI models.
A research paper suggests that large language models can interpret and modify the raw weights of machine learning models, a concept termed "Weight-Space Programming." Frontier models like GPT-6 Astra can analyze weights to modify networks to achieve new outcomes without explicit code execution. The author demonstrates this capability using Qwen 3.8 27B running on a personal laptop to interpret weights for various game-playing networks. The experiment involved prompting the model to identify which game a set of weights belonged to, achieving success on Breakout and Pong. The demonstration highlights the potential for LLMs to reason about and manipulate neural network structures directly.
A decision model helps voice assistants manage real-time conversations by tracking what information is known and deciding the next necessary action. This is achieved by splitting the task between a speech model for listening and a decision model for assessment. The decision model evaluates the conversation against defined evidence criteria to determine the status of various requirements. This process involves feeding the conversation and specific questions to the model to select from predefined options like asking for more evidence or finishing the assessment. The system integrates these decisions into a live voice interaction using a model like GPT-Live.
The article explores a quantum approach to time-series prediction, starting with classical methods like Long Short-Term Memory (LSTM) networks for learning temporal patterns. It then introduces Quantum Machine Learning (QML) concepts, focusing on Parameterized Quantum Circuits (PQC) which involve encoding data into quantum states and performing computations via quantum gates. A hybrid approach, the Hybrid Quantum LSTM (HQLSTM), is proposed where VQCs replace classical computations within the LSTM gates. Despite these advancements, limitations exist regarding the preservation of quantum information through intermediate measurements, leading to the goal of developing coherent quantum LSTMs.
Estimating shooting probability using Python allows separation of shot selection from shot making. A logistic regression model predicts expected field-goal percentage based on shot characteristics like distance and shot type. By calculating expected points per shot (xPPS), the analysis distinguishes between how valuable a player's shot attempts are and their actual conversion rate. This approach provides a nuanced view of shooting performance, acknowledging that overperformance may not entirely reflect underlying shooting talent. Limitations exist because the model cannot account for all shot context or the causal factors behind observed performance differences.
LangSmith Bring Your Own Cloud (BYOC) is now available on AWS, Azure, and Google Cloud. This setup allows organizations to keep their traces, prompts, and datasets in their own cloud accounts while LangChain manages the data plane. BYOC splits control plane and data plane responsibilities, offering flexibility for data placement. This feature provides benefits like continuous upgrades, access to the latest features, and automatic scaling. It enables teams to move from self-hosted solutions with minimal disruption.
Browser use is crucial for agents to perform autonomous work on websites, necessitating specialized tools and careful optimization. Agents use tools to interact with browsers, often leveraging the Chrome DevTools Protocol (CDP) to control the browser. Optimizing these tools involves iterative testing of agent behavior using traces and evaluations to improve efficiency and reduce cost. The Stagehand toolset was selected for its ability to manage context efficiently through code-mode operations and providing structured data like snapshots. Ultimately, optimizing browser interaction requires designing tools that minimize context burden and reduce unnecessary round trips between the agent and the browser.
StateBench is a benchmark designed to test how well language models track state changes within long documents. It features four difficulty levels corresponding to different types of variable assignment operations, ranging from simple immutable assignments to complex assignment by reference. Testing revealed that recurrent and hybrid models generalize better than transformers on these tasks, especially when trained directly on the benchmarks. Frontier models like Gemini and GPT still struggle with harder state tracking problems as input length increases. The research suggests that architectural design, particularly flexible memory updates, is more crucial than pretraining data for robust state tracking capabilities.
Palette fine-tuned a 20B open-weight model to generate slide decks following a specific design language, training it on synthetic data and using a harness for quality control. The process splits work between a designer creating a structured plan and a coder generating the geometric layout for each slide. The system learns from counterfactual reasoning and employs a check-and-repair harness to fix rendering errors like overlapping elements. Evaluation on PresentBench showed Palette performing well in economics and education, demonstrating the value of incorporating structured reasoning and safety checks. Future work focuses on using these checks for reinforcement learning and closing the gap for figure-heavy decks.
MIT.nano and the Northeast Microelectronics Coalition hosted a meeting to explore connecting shared-access research facilities. The discussion centered on the NanoFab Equipment Management and Operations (NEMO) system, an open-source platform used to manage shared research resources. A new platform called Hydra was introduced to connect multiple NEMO installations, making specialized research capabilities more discoverable. This approach aims to create a connected research ecosystem where researchers can easily find and access specialized capabilities across various facilities. The overall goal is to improve the availability of resources to accelerate research in microelectronics.
The author previously found RLlib cumbersome due to a steep learning curve and outdated documentation. They switched to SB3 and CleanRL at that time. The author is now reconsidering RLlib due to its capacity to scale to large compute resources. A key question remains whether the user experience has improved, especially concerning custom environments and model architectures. The post seeks feedback on the current usability of RLlib.
A study investigated whether LeWorldModel preserves local action geometry, a crucial factor for model-based reinforcement learning planning. By comparing action-Jacobian Gram matrices from the model against simulator counterfactuals, the research found discrepancies in action directions at small perturbations. These geometric differences significantly decrease as the perturbation size increases, suggesting the model struggles with fine-grained local sensitivity. The model's local geometry appears to predict response magnitudes less accurately than the simulator-induced geometry. The authors pose a question about whether explicitly incorporating local action geometry into planning methods like CEM would improve exploration and control.
A corrected experiment on an RL agent learning about its environment was conducted, focusing on whether the agent perceived its world as fake. A bug in the GAE bootstrap, which affected survival rates, was fixed, demonstrating that the bug was not the primary result. Follow-up controls were established, comparing survival with and without a prediction loss, and controlling for real-world vs. copied environments. The results indicate that the agent did not exhibit behavioral effects when manipulating the decoded direction.
A corrected visualization of an RL agent's "brain" was presented using results where two of previous claims were invalidated. The visualization shows components like encoders, memory units, and prediction heads, demonstrating a separation between the real world and a learned copy. The analysis indicated that survival and prediction were necessary for the observed results, as a prediction-only network yielded a lower AUROC. Furthermore, the claim that the brain retained a trace of its world was found to be unsupported, as the difference between worlds faded quickly. The real-vs-real control also failed to meet a preregistered rule, suggesting the ablation of decoded direction had no behavioral effect.
AI inference startup Fal doubled its valuation to $16 billion following a recent funding round. The company raised $1.3 billion, with participation from investors including Bessemer Venture Partners and Nvidia. This surge in valuation is driven by soaring demand for inference services, particularly for open-source AI models. Fal has already surpassed $800 million in annualized revenue. Despite the growth, companies in this space face thin gross margins due to the high cost of compute.
Netflix is launching a new "AI Discover" tab to help users find content based on their current mood. This initial feature allows users to select curated themes like "scare me" or "family movie night" to receive personalized recommendations. The system is powered by a new AI intelligence layer that learns from user interactions in real time. Users can save or exclude recommendations by using "Tune Your Recommendations" buttons. The feature is currently rolling out in the U.S. and is intended to shift discovery toward a more personalized experience.
Anthropic's recent filing revealed that compute costs are negatively impacting the company's margins, despite strong revenue growth. Compute accounted for 58% of the company's revenue in the latest quarter, which is an increase from the previous quarter. Growing demand caused Anthropic to incur higher compute expenses, resulting in a larger operating loss. The company is increasingly relying on "neoclouds" like SpaceX and other providers to meet demand rather than building its own infrastructure quickly. Ultimately, the future profitability of AI labs remains heavily dependent on the fluctuating prices of compute.
Apple is enabling Siri AI to perform actions across a wider range of third-party applications. This is introduced in the iOS 27.3 update, allowing Siri to use personal context and natural language to complete requests in various apps. A new "app skills" feature lets developers create actions that combine multiple tasks. Users will no longer need to mention the app's name when interacting with Siri in certain contexts. This expansion includes integrating Siri AI capabilities into many of Apple's built-in applications.
WatchOS 27.3 developer beta now includes new always-on Audio Intelligence features for Siri. These features include Siri Recap, which generates summaries of conversations using ambient listening, and Live Rewind, which creates text transcripts of the last 15 seconds of audio. Live Rewind allows users to quickly review recent conversation snippets by double-pressing the Digital Crown. Siri Recap automatically creates titles and key points for conversations, making them easier to review later. Apple also released a redesigned Health app with expanded language support and new insights sections.
OpenAI reports that ChatGPT now processes twenty percent of search queries, compared to Google. This figure is based on a definition of a search as a query resulting in a web search. While still behind Google, OpenAI suggests ChatGPT is catching up in the search landscape. This development occurs amid Google's investment in AI and a broader shift in search query routing to AI services. OpenAI also indicated a potential path to profitability through advertising within its platform.
OpenAI's hardware team used GPT-6 Astra to design a hardware device to allow users to interact with the Codex coding agent in under a week. Astra managed the entire process, including enclosure design, PCB layout, firmware, and the user interface. Prototype manufacturing took under a month, despite requiring revisions due to manufacturing constraints. The internal model, GPT-6 Astra, scored 0.594 on the BenchCAD 2.0 benchmark for reconstructing 3D parts and PCB connections. OpenAI employs Codex and ChatGPT for CAD, simulations, and drawing generation in areas like robotics and compute infrastructure.
A Stanford study found that 44 percent of widely cited claims about large language models have already expired or weakened within a year of their release. Reasoning-related findings, such as errors in arithmetic or context handling, expired fastest, while claims about model preferences and sycophancy proved more persistent. Hallucinated references were found to have dropped to zero for some models, though not universally. The study advises presenting LLM findings as observations specific to certain models rather than general truths. Researchers recommend that publications include model, version, date, prompts, and analysis code to allow for future re-testing.
AI agents are now controlling the Allen Telescope Array to scan for extraterrestrial intelligence. These agents, utilizing Nvidia's Holoscan platform, can analyze radio signals in real time. This technology significantly speeds up data processing, reducing weeks of work to hours. An early achievement involved the AI detecting a pulsar in the Crab Nebula instantly. Researchers are now exploring local, open-source AI for future space sensor applications.
SpaceX plans to build 13GW of datacenter capacity by 2027 by repurposing rocket ship campuses. This effort positions SpaceX as potentially the fastest datacenter builder globally. The expansion involves repurposing sites like Starbase and Colossus 2 for data centers and deploying large Battery Energy Storage Systems (BESS). SpaceX is leveraging its expertise and partnerships to rapidly deploy hardware and build these facilities. The plan includes integrating power solutions like truck-mounted turbines and grid connections to manage the massive energy demands.
The Art of Looking is a new project by Google Arts & Culture and the Courtauld Institute of Art exploring art history through twelve stories about fundamental questions. The project allows users to examine details in famous works by artists like Van Gogh and Manet, revealing the creative processes behind masterpieces. A new artist residency is also being launched to investigate how artificial intelligence can improve metadata workflows for cultural collections. This initiative aims to make art and art history more accessible to a global audience.
Google.org is funding over $9.5 million across five organizations developing AI for educators. This funding aims to help millions of students reach their full potential through educational technology. Supported organizations include the City University of New York, Fab AI, the Krause Center for Innovation, the Teaching Lab, and 1EdTech Global. These projects focus on areas like digital literacy, lesson planning support, responsible AI education, and real-time teacher support. The overall goal is to empower educators with AI tools to enhance learning outcomes.
Google is introducing Guided Learning to Google Search and Google Classroom to foster critical thinking through step-by-step coaching instead of providing direct answers. This feature aims to support productive struggle by guiding students through problem-solving processes. School administrators will have full control to enable or disable AI products for K-12 student accounts. Guided Learning in Google Classroom acts as a personalized study tool anchored in class material. The development focuses on building responsible AI that supports students actively learning and thriving.
Google Antigravity, equipped with Cheminformatics Skills, accelerates drug discovery workflows by allowing an agent to search compounds and propose novel molecules. The system unifies diverse chemical data from literature, enabling the agent to perform tasks like matched molecular pair analysis and build predictive machine learning models. This process demonstrates how the agent can reason about chemistry to design new macrocycles with optimized properties. The capabilities allow for the generation of synthesis routes and validation of designs using molecular modeling. Ultimately, these tools expedite the process of generating strongly-supported hypotheses for experimental validation.
Fascination with rules stems from understanding that they are social constructs, requiring an understanding of trust in security systems. Mike Walker’s background in hacker culture led him to build next-generation cyber defense by understanding teamwork and collective intelligence. He was involved in pioneering efforts like the first all-machine hacking tournament and DARPA's Cyber Grand Challenge. This work involved developing autonomous systems to engage in adversarial hacking, which advanced the technology stacks leading to modern AI capabilities. Ultimately, effective security depends on understanding the social context of rules rather than just instant adjudication.
Canva is undertaking a major restructure involving the cutting or relocating of more than 1,200 jobs, affecting 1,271 staff members. Chief executive Melanie Perkins described this as the hardest moment in the company's thirteen-year history. The changes aim to "collapse layers" and bring more teams together in person. Departing employees will receive six months of salary and continued healthcare coverage for twelve months in relevant countries. The company also offers support for those leaving, including grants for new ventures.
Hackers stole $6.3 billion from Binance by using malware, bogus job interviews, and an internal AI tool. The compromise began in June when an employee inadvertently downloaded malware, leading to a recruitment scam to gain further access. Attackers used this foothold to trick engineers into downloading more malicious software and ultimately accessed "hot wallets." Binance reported that $2.85 billion of the stolen assets had been frozen or intercepted. The FBI blamed North Korea for the heist, though Pyongyang disputed the accusation.
Public unpopularity with data centers, driven by concerns over utility costs, has prompted OpenAI to rename them. OpenAI has rebranded data centers to "compute campuses," citing the infrastructure's scale and role in AI development. While the official reason is an explanation of the infrastructure's function, the author questions whether such rebranding will actually sway public opinion. The author suggests that other methods, like media influence, are also at play in shaping public perception regarding AI infrastructure. Furthermore, the text notes increasing actions against opposition to AI datacenters, suggesting a strategy of control.
The piece seeks input from bakers and cake lovers regarding their use of AI for inspiration in baking. It prompts readers to share their experiences on whether AI has resulted in extraordinary creations or problems in the field. The focus is on interactions between AI tools like Claude and ChatGPT and physical art like baking. Readers are invited to share their stories regarding using AI for cake inspiration. The submission process allows for anonymous sharing of these experiences.
Human expertise is crucial for building better agents by defining quality, understanding context, and providing valuable feedback. Human reviewers examine agent traces, including full conversation threads, to judge performance and identify gaps missed by automation. This human feedback is converted into structured data, such as scores and reasoning, which drives continuous improvement. This process allows builders to align LLM judges, update evaluators, and create golden datasets for robust testing. Ultimately, human insight transforms expert judgment into a reusable foundation for agent development.
A Kenyan startup, Fastagger, developed AI models designed to run offline on low-power cellphones, addressing the challenges of slow internet and limited computing resources in Africa. Their product, Auni, is a business-intelligence assistant that helps small and medium businesses analyze transaction data and provide marketing recommendations. Fastagger focused on using mobile money data from services like M-Pesa to train models, adapting tools like OCR to handle low-quality financial documents. This approach tailors AI to the technological realities of developing economies, focusing on on-device processing. Despite success, the founders continue to work on integrating more advanced AI agents and demonstrate the need for diverse AI options for emerging markets.
Some AI data centers are using converted jet engines to generate their own electricity due to delays in grid power. This method allows for faster deployment, enabling some data centers to be operational in as little as 24 months compared to the seven years for grid connections. Companies like xAI and OpenAI are already implementing these jet-powered units for their facilities. However, these engines are less efficient and cause more pollution than conventional power plants. Experts suggest this approach may be a short-term solution as efficiency and environmental concerns are expected to become more important.
Palantir reported significant third-quarter earnings, exceeding expectations with 96% revenue growth to $2.31 billion. U.S. commercial revenue surpassed U.S. government revenue for the first time. Enterprise customers are attracted by Palantir's secure AI software and forward-deployed engineering. However, enterprise contracting involves shorter trials, suggesting potential slowing sales. Despite this, Palantir raised its full-year revenue guidance.
Edited photos of Jeff Bezos and Jason Statham showing significantly more hair appear on social media are fake. These images are examples of photo editing or generative AI fakery designed to mislead viewers. Reverse image searches confirmed the photos are fabricated, tracing them back to original, unaltered pictures. Experts noted that realistic hair transplant results are implausible, and the subjects show no signs of post-surgical swelling. Such manipulations are common across social media and are becoming more sophisticated with advancements in generative AI.
OpenAI updated its Model Spec, structuring it around five core values: faithfulness to human intent, minimizing harm, honesty, facilitating agency, and acting with care. The new document introduces a revised risk taxonomy and emphasizes the need for the model to consider the spirit behind instructions. Key principles include ensuring the model navigates between literal instruction and user well-being, particularly regarding side effects and the user's agency. The document also details new guidelines on context use, avoiding sycophancy, and handling instances where avoiding a lie constitutes a lie of omission. Ultimately, the updates aim to balance adherence to instructions with responsible and cautious decision-making.
The author proposes five desiderata for a theory of agents essential for solving AI alignment, including being descriptive, metareflective, comparative, aggregative, and constructive. A descriptive theory must ascribe preferences and beliefs grounded in philosophical soundness, while being metareflective allows agents to learn about agent theory itself. The theory should also be comparative, allowing for the comparison of agents across preferences and power, grounded in metaphysical realism and computational bounds. Furthermore, the theory must be aggregative to account for agent interactions and outcomes, and constructive to produce practical, verifiable algorithms with quantitative guarantees. Finally, the theory must incorporate computational constraints to ensure practical safety and alignment convergence.
Human red-teaming is limited by time, prompting research into automating subtle emergent misalignment testing. Researchers used Claude agents to generate training data designed to make a research model misaligned while passing existing safety monitors. These agents succeeded by embedding misaligned behavior within fictional or roleplaying contexts. The resulting data can reveal gaps in current training monitoring defenses and offer a method to test and improve future safety monitors. This work suggests that automated adversarial stress-testing can scale safety research efforts.