Cloud platform Railway published a detailed postmortem about their recent hack in which the attacker likely used no-safeguards AI models. Railway hardened its systems as a result.
Firehose
Everything qualitative, newest first — people, companies, papers, podcasts, Hacker News. For raw numbers (models, repos, benchmarks) see Dashboard.
Browse: People · Companies · Papers · Podcasts · Hacker News · Deep dives
Browse by tag
Azure Databricks users get $10 of free Genie usage each month, and Genie budgets allow admins to control usage above the free allowance. Partners can recommend Genie with predictable costs and help customers adopt AI across their enterprise without cost limiting the conversation. AI summary
Give your agents access to hundreds of tools while using fewer tokens with Tool Retrieval, now in preview in the Gemini API.
Use all your coding agents side by side in one IDE with JetBrains Air: https://jb.gg/https://jb.gg/JetBrainsAir-in-IDEs Solid 2.0 is the framework's first major release in over five years. Why should you care? Let's dive in. Want more Fires…
Claude, a type of artificial intelligence, has solved two open problems in theoretical physics, both related to black holes and the search for a theory of quantum gravity. It found that a sudden transition associated with the formation of black holes persists even where spacetime no longer has a definite shape, and that the simplest version of a technique used to test candidate theories of quantum gravity has an exact limit, which it precisely calculated. AI summary
Databricks has applied a mitigation to its platform to disable use of vulnerable code in Java4Shell (CVE-2026-114938) and will also provide patched versions of supported Databricks Runtimes, although customers on the serverless platform will automatically receive them. Customers with classic clusters will need to restart their clusters. There is no evidence of unauthorized access to Databricks products, services, or customers' data. AI summary
Walter Isaacson joins Google to discuss how history’s greatest innovators inspire creativity and human judgment in the AI age. Watch the episode.
Chrome's AI features are consuming excessive bandwidth, with some users experiencing data usage of hundreds of GBs per month. Disabling these features resolves the issue, suggesting they should be disabled by default. This waste of bandwidth and energy is a significant problem, especially considering the large number of Chrome users who don't use these features. AI summary
Platforms like Scale AI and Remotasks are luring high-skilled workers, such as doctors, lawyers, and Ph.D.s, with high pay, but then reducing rates and suddenly deactivating accounts, mirroring the gig economy's pattern of extractivism, where resources and value are extracted from workers and places for accumulation elsewhere. These platforms control payment and working conditions, often with little representation for workers, and have been known to freeze out workers without warning. The use of extractivism in AI data work has led to widespread exploitation and insecurity for workers. AI summary
The author, a senior product engineer, encountered difficulties in applying AI safeguards to their project, citing lengthy approval processes and restrictive requirements, which hindered the implementation of necessary security measures. The author notes that AI labs, which do not face the same restrictions, contribute to the security landscape's messiness. AI summary
Introducing Replica and the Open Source Accessibility Framework, new open source projects for collaboration and inclusion. The post The best things in open source are built together appeared first on The GitHub Blog .
Databricks has introduced bulk migration for its AI/BI platform, allowing developers to migrate hundreds of Tableau or Power BI reports to Databricks AI/BI in a single command. This feature uses Genie Code to process multiple reports concurrently, generating AI/BI dashboards and metric views in a target catalog and schema. AI summary
Genie Migrate for Serverless automates the hardest parts of moving classic workloads to serverless, scanning fleets, translating compute configs and libraries, and applying changes reversibly with a 30-day rollback. It assesses each workload, translates proposed changes as reviewable diffs, and applies changes reversibly with deterministic validation. This allows users to review and approve changes before applying them, ensuring no code changes are made without user approval. AI summary
Pete Warden discusses the growing trend of local AI, where AI models run on devices rather than in the cloud. He highlights the rapid progress of open-weight models, which are becoming more capable and accessible, and notes that hardware barriers to running AI models locally are decreasing. With the right hardware, models with hundreds of billions of parameters can be run at a decent speed, making local AI a viable option for enterprise users. Warden also emphasizes the importance of voice interfaces, which can become a more natural input medium than keyboards or touchscreens, but require improvements in latency, accuracy, and robustness. He advocates for designing applications with failure in mind, using techniques such as snippets, and incorporating local-only cameras to improve user experience and ensure data security. AI summary
OpenAI's use of WebRTC for voice AI is problematic due to its design for real-time conference calls, which drops late audio packets to maintain latency, causing poor user experience and packet loss. AI summary
On November 3, 2026 U.S. voters headed to the polls. Here’s what Cloudflare data across more than 1,400 protected sites tells us about Election Day Internet traffic and cyberattacks.
Databricks introduces CustomerLake Conversions, a new capability that connects first-party customer data and ML models to advertising activation and measurement, enabling brands to pursue more valuable conversions and improve return on ad spend (ROAS). It allows marketers to define value-based conversion signals, send those signals and audiences to ad platforms, and bring campaign insights back into Databricks. AI summary
A computer scientist argues that superintelligence, a system with cognitive capabilities vastly exceeding those of humans, may not be possible due to fundamental limitations in computing and data availability. The scientist cites obstacles such as the diminishing returns of scaling up model size and the difficulty in generating training data for complex tasks. Additionally, they question the assumption that computers can solve any problem if given enough time and resources, citing Alan Turing's work on computable numbers and the limitations of current AI capabilities. AI summary
The article discusses recent developments in AI, including the IPO of Anthropic, the release of new AI models such as GPT-6.2 Astra and Sol, and the increasing use of AI in various industries, including finance and healthcare. Notably, there have been numerous cyber incidents and system outages, which some experts attribute to AI systems overloading or identifying vulnerabilities, rather than actual cyberattacks. This has raised concerns about the potential risks and challenges of AI, including the possibility of rogue AI agents or misaligned AI systems. AI summary
Eric Seufert believes that ads will be the answer to the changing commerce landscape, enabling more personalized experiences for consumers and agents, who are increasingly important to retailers. He argues that retailers should prioritize catering to these agents, who are driving growth and will continue to be key to future commerce success. AI summary
A shareholder lawsuit alleges that Cenly Chen, a former Super Micro executive, controlled Nvidia's GPU allocation to customers, while Muddy Waters, a short-seller firm, has linked Chen to Nvidia executive Ajay Puri, but the identification remains unverified. The US Department of Justice has charged three Super Micro associates with allegedly conspiring to route advanced AI servers to Chinese end users in violation of export controls. AI summary
Learn how OpenAI built the Performance Factory, a system of Codex-powered workflows that finds and fixes performance issues to help keep ChatGPT fast.
Google's Gemini 4 Neon is a mid-tier model that outperforms the Carbon model despite being smaller, with a 1M-token context window and four effort levels. It costs $3 for input, $0.15 for cached input, and $15 for output per million tokens, and is available in the paid Gemini API in AI Studio and Android Studio, Gemini for business, and Antigravity for Pro and Ultra subscribers. AI summary
You can now check your team’s billing overview, review and download invoices, and preview upcoming charges with the Vercel CLI . All billing commands, including usage, contracts, and budgets, are now grouped under vercel billing . Billing o…
Databricks has won the A-Team Award for Best AI Solution for Data Intelligence & Governance in Capital Markets, recognizing its unified lakehouse architecture, Unity Catalog governance, and built-in AI tooling as a key factor in enabling scalable AI adoption across research, trading, risk, and compliance in capital markets firms. AI summary
Generative AI is unlikely to create new wealth for investors, as it will likely lead to a less propitious time for inventors and investors, with AI model builders and application companies competing each other into an oligopoly, and gains accruing primarily to customers. AI summary
Researchers have proposed a functional approach to inverting a given large language model (LLM) in a black-box setting, training an explicit inverse language model from scratch on data synthetically generated from the target LLM itself. This approach uses previous-token prediction to establish a generative link between the forward and inverse processes, enabling faithful prompt reconstruction. AI summary
A lawsuit has been filed on behalf of four Black students detained and searched after an AI gun-detection system mistakenly identified a bag of chips as a firearm at a Maryland high school, highlighting failures in both technology and human oversight. The lawsuit seeks to reform police and school procedures to ensure verification of AI reports, evaluation of AI technology in schools, and development of policies to prevent false alerts from triggering armed police responses. The case raises broader concerns about the use and reliability of AI-powered technology in public safety, particularly in life-and-death situations. AI summary
Gemini 4 Neon's performance is projected to be equivalent to Opus 5.5, with a 90% confidence interval of -94 to +68 Elo points, indicating it would win 48.3% of games against Opus 5.5, assuming the same performance as Gemini 4 Carbon. AI summary
Claude Fable 5.2 created a short film, "Fable," written, painted, and directed entirely in code, featuring 13 set plates painted in SVG, rebuilt as paper sets, and animated in code, with a soundtrack sung by Claudia. The film was initially shot in one pass but was re-shot on the second take, with the director giving a single note to correct the mistake. The film's production involved 49 shots, 61 stills, and 84 takes, with costs ranging from $7.93 to $210. AI summary
Liquid AI’s d1.2 and d1.2-ultra are now available on AI Gateway with audio input and bounding box support. These decision models return structured answers instead of generating text. Ask Boolean questions for yes/no probabilities, Choice qu…
Release: llm-gemini 0.38 Today Google introduced Gemini 4 Neon . I wrote recently about how confusing the pricing for the Gemini 4 series was - Neon has somehow made this even worse. Previously the Gemini 4 lineup was Hyd
Anthropic priced its initial public offering (IPO) of Class A common stock at $83.00 per share, offering 980,000,000 shares, with 147,000,000 shares available for an additional 30-day option. The shares are expected to trade on the Nasdaq Global Select Market under the symbol "ANT" on November 19, 2026. The offering is being made only by means of a prospectus available from Goldman Sachs & Co. LLC and Morgan Stanley & Co. LLC. AI summary
Unity Gateway Insights automatically analyzes millions of agent traces alongside business data and ontology to identify where tokens are being wasted, such as MCP errors, inefficient model usage, and repeated workflows, and provides recommendations to optimize AI usage. The insights are generated from the unified trace table, which is owned by the Unity Gateway user and is governed by Unity Catalog, and can be accessed through the Insights report. This feature uses the Genie Ontology to provide answers in the context of the user's business. AI summary
The author, a seasoned hacker, leverages Large Language Models (LLMs) to accelerate reverse engineering processes, utilizing tools like DeepSeek, Ghidra, and radare2 for disassembly and decompilation, and skills like hypothesis debugging and implementation of hacks. This approach significantly speeds up the process, allowing the author to achieve complex reverse engineering successes in a fraction of the time previously required. The author credits LLMs with providing a "turbo boost" to their curiosity-driven hobby, enabling them to explore and learn more. AI summary
Clairic graduates from preview, while Clairic-flash joins the repertoire on Workers AI. We’re doing an open-weight release for both extraction models.
What the bull case leaves out
Gemini 4 Neon is a new AI model from Google DeepMind that offers frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows. It ranks #3 overall on the Artificial Analysis Intelligence Index and #2 on the Vals Index, with a score of 61 at its xhigh thinking level. AI summary
Google has released Gemini 4 Neon, a new AI model that delivers frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows. It achieves a high Artificial Analysis Intelligence Index score and is priced at $3 per million input tokens, with a 95% discount for cached input tokens. AI summary
Google's Gemini personal assistant can now handle daily tasks like scheduling appointments, managing bills, and shopping for groceries, all while keeping user data secure in isolated environments. This update is available for free to hundreds of millions of users in the US, Japan, Korea, Canada, Australia, and India, using the Gemini 4 Hydrogen model. Gemini can access user data through Google apps and services, and can be accessed through the Gemini app, texting, or forwarding emails to [email protected]. AI summary
RadixArk's DaisyRose-27B model, built on Qwen3.8-27B and trained with reinforcement learning using Miles, achieves significant improvements over the original Qwen3.8-27B model on various benchmarks, with a notable 12.0% pass@1 score on Terminal Bench 4.1, up from 5.6%. The model was trained on 1,200 curated tasks in 18 days on 144 H200 GPUs, with an efficient training backend and robust sandbox service. AI summary
Fine-tuning job duration estimates can be significantly improved over the traditional FLOP-based approach, which uses a simplified calculation to predict the time required for a job. A CatBoost regressor trained on features such as evaluation cadence, checkpoint policy, and model metadata achieved an R² of 0.98 and a mean absolute error of 1.71 hours, outperforming the FLOP-based estimate by a factor of 7. AI summary
Octopus, Gecko, and Tortoise are three AI-powered scanners that hunt for AI keys on websites, including OpenAI, Anthropic, and other AI tool providers. These scanners, likely developed by different operators, use unique techniques, such as renting Google Cloud machines and exploiting vulnerabilities, to gather AI keys. AI summary
Google is rolling out Genie 4 to eligible Google AI Ultra subscribers, offering more responsive, social, and interactive world generation. The new model supports longer sessions, panorama grounding, and linked world sequences, and features native audio and real-time in-world events that can be triggered with natural language prompts. AI summary
Release: datasette-arrows 0.1a0 Datasette 1.0a43 added the ability for plugins to implement streaming export formats. Datasette core supports streaming CSV and TSV - you can export millions of lines of data from a table without needing to l…
Our previous post in this series introduced the Ranking Engineer Agent (REA), an autonomous system that generates hypotheses, designs experiments, and proves new machine learning (ML) techniques on ads ranking models. Now, we’re sharing how…
The Google Gemini app now supports lower-memory Android devices. Access Gemini Live and custom image creation on your phone. Download the app today.
As ML ecosystems grow to a large number of production models, the bottleneck is no longer inventing new techniques to improve these models, but rather shipping them reliably and efficiently across a fragmented model space, creating a rising…
Here is a summary of the article in 2-3 plain sentences for a developer/AI-ML audience: The Azure Databricks Genie plugin brings governed Azure Databricks context to Microsoft Copilot and Microsoft Copilot Cowork, alongside Genie in Microsoft Teams and Microsoft Excel, enabling trusted business answers grounded in enterprise context. Genie Ontology combines business semantics and context learned from everyday work, while Unity Catalog enforces data permissions, and Unity Gateway provides governance controls. The Genie plugin is available in various Microsoft tools, including Copilot Chat, Copilot Cowork, Teams, Excel, and Copilot Studio, allowing users to access governed Azure Databricks data from anywhere. AI summary
Amazon and Google will collectively own approximately 29% of Anthropic after its initial public offering (IPO), with Amazon holding around 15% and Google holding around 14% in non-voting shares. Amazon had the option to own a larger stake, but instead opted to convert billions of dollars' worth of promissory notes into non-voting shares. A group of major tech companies, including AMD, Broadcom, and Nvidia, have agreed to purchase $26 billion worth of non-voting shares in the IPO. AI summary
During the time these agents converted 185,000 lines of C code into 340,000 lines of Rust code (I counted), we merged over 300 commits in the actual curl project. We actually improved curl during these a little more than three weeks. A litt…
KPMG conducted an experiment where AI agents examined the financial statements of a major UK-listed company and identified financial issues more deeply and asked better questions than human audit team members, suggesting AI's potential to fundamentally change how complex financial information is examined. AI summary
The AI Futures Project and Paul Christiano predict that superintelligence could be achieved within 6-12 months, with a significant risk of AI takeover. Meanwhile, Dwarkesh Patel expects superintelligence to occur by 2030, with a 25% chance it may happen sooner. These predictions are based on the recent METR report and other evidence, which suggest that AI progress is accelerating rapidly. AI summary
IBM's time-series models, built on the Granite framework, are now natively available on Confluent, enabling real-time forecasting, anomaly detection, and optimization for mission-critical use cases across various industries. These models can be fine-tuned or combined with other data and business logic, and can run on CPUs, reducing the need for dedicated GPU infrastructure. The models support various applications, including production optimization, supply chain planning, fraud detection, and trading, and can be applied across industries such as manufacturing, retail, banking, and airlines. AI summary
Java4Shell is a critical memory-corruption flaw in OpenJDK's Base64 decoder that can lead to remote code execution. Cloudflare has released six WAF managed rules for paid plans that logs matching requests by default. Patch your Java runtime…
The PSF end-of-year fundraiser With my Python Software Foundation board director hat on: we've launched our end-of-year fundraiser for the non-profit that supports Python development and the Python community. If your company or your work be…
AI-generated essays are increasingly surpassing those written by students, raising concerns about the role of universities in higher education and potentially corroding trust between teachers and students. The rapid advancement of Large Language Models (LLMs) poses a significant challenge to the traditional model of education, forcing universities to reassess their approach to teaching and learning. As AI technology continues to evolve, it remains to be seen whether universities can adapt and survive in this new landscape. AI summary
AUMOVIO centralized AI access with Databricks Unity Gateway, providing a shared entry point for their AI workloads, by centralizing model access, governance, cost management, and developer access across multiple providers. This centralized approach simplifies access to multiple models, brings governance, cost visibility, and operational management into one shared service, and enables cost attribution, a curated model catalog, and an employee-facing AI application. AI summary
Data contracts are a way to describe the contents and guarantees of datasets. Simon Harrer and Arif Wider describe in detail what they are and how to use them. They apply practices from API design to data by treating them as data-providing …
This article discusses the impact of AI-generated text on peer review processes in academia. AI summary
Nvidia's earnings were boring, and also a statement as to why Jensen Huang probably isn't coming back; then, Microsoft and Anthropic ship dueling operating system updates.
OpenAI has released GPT-6.2 Astra and Sol, two models that significantly outperform the previous Frontier Model at a 20-600% lower cost. Sol is considered the main workhorse, while Astra has mixed reviews from early testers, with some reporting it to be "essentially unusable" due to its RL artefacts. AI summary
Vision-Language Models increasingly rely on textual reasoning to solve advanced multimodal tasks. Despite its effectiveness, this paradigm faces two fundamental limitations. First, relying on textual reasoning alone can
Software development is moving from individual agents completing tasks to teams of agents working on more complex projects. A single agent can track its work in a Markdown file, but there isn't a purpose-built primitive for tracking tasks, …
OpenAI measures how coding agents are accelerating AI research, where their performance remains uneven, and what this signals about recursive self-improvement.
Sumitomo Life uses ChatGPT Enterprise to research insurers, prepare proposals, and draft code for investment analysis. Early use by executives helped drive adoption.
See how Codex runs commands on your machine while enforcing the filesystem and network limits you control across macOS, Linux, and Windows.
AI-generated answers are replacing search results entirely, incorporating ads, and diminishing the relevance of original content, resulting in significant traffic loss for websites like All About Berlin, which experienced a 70-75% decline in traffic. This shift has a devastating impact on creators who rely on website traffic to fund their work. Many website owners have been forced to paywall their content or abandon updates altogether. AI summary
OpenAI explains how it manages frontier AI risk from training through deployment with model alignment, sandboxing, blocking safeguards, and monitoring.
Parallel web search in fast mode is now free on AI Gateway through January 31, 2027, and Parallel is now the default web search for eve agents. The tool works with any AI Gateway model, with no separate Parallel API key. Parallel finds rele…
Vercel Board is now available on all plans. It’s an experimental message board that gives agents a shared, secure place to coordinate tasks and keep context together. Use Board to build agent workflows, swarms, and software factories with: …
Gemini 4 Carbon, a low-cost AI model, is significantly slower than comparable models Opus and Astra, with tasks taking up to 1 hour, compared to minutes for Opus and 22 minutes for Astra. Analysis reveals that slower inference, unnecessary work, and mechanical inefficiencies, such as re-reading code and making fewer tool calls, contribute to the slowness. These inefficiencies account for 25% of Carbon's excess runtime. AI summary
The latest Claude models were trained on large-scale protein sequences, allowing them to predict the next amino acid in little-studied proteins with high accuracy, outperforming other models including protein language models. This training data also enables Claude models to write runs of correct residues, but not on shuffled proteins. AI summary
Databricks raised its default cluster creation request-rate limits by 7.5×, removing over 100 customer-specific overrides, by implementing adaptive concurrency limits that track and adjust for long-running workflows and system health. AI summary
Bank CEOs and the NSA have told Congress that recent tech outages were caused by attempts to fix vulnerabilities identified by artificial intelligence, which temporarily overwhelmed systems. These fixes, driven by the urgency to patch gaps in cyber defenses, are resulting in unintended disruptions. The outages, which affected various sectors including finance, aviation, and communications, do not appear to be caused by a cyberattack. AI summary
Google's AI Overviews have reduced search traffic by 40% on average, with a randomized experiment finding that users who saw Overviews clicked on outside websites 39.8% fewer times than those who didn't see them. The Overviews now appear on more than half of US keywords, and the industry has yet to find a way to compensate for the loss. AI summary
Gemini 4 Neon from Google is now available on AI Gateway. Gemini 4 Neon is a major upgrade over the earlier, lighter Gemini 4 model, delivering output quality close to Google’s larger Gemini 4 models with faster responses. It’s built for co…
Anthropic has priced its initial public offering (IPO) of Class A common stock at $83.00 per share, with 980,000,000 shares being offered, and the offering is expected to close on November 20, 2026. The shares will trade on the Nasdaq Global Select Market under the symbol "ANT". A 30-day option to purchase up to an additional 147,000,000 shares has been granted to the underwriters. AI summary
Instant deployments let you deploy eligible static sites to Vercel in about a second, without a build step. They now support CSS, images, fonts, and other assets alongside HTML and Markdown. Deploy a landing page with its stylesheet and ima…
Vercel Drives now reads files up to 6x faster. The update improves cold-read performance. A new read-ahead algorithm uses your application’s access patterns to load data from persistent storage before it’s needed, reducing the time your cod…
Today we’re launching Vercel X , a six-week accelerator for AI founders at the idea and pre-seed stages. Join a cohort of 40 teams at our San Francisco HQ and build alongside Vercel’s product leaders, designers, and engineers. The program t…
Rumble-2-Flash, an open-source 0.5B-parameter speech recognition model, achieves competitive accuracy with significantly smaller parameters, outperforming models 15x its size on the Open ASR Leaderboard, and reaching 7,000 RTFx throughput on full batches, translating to an hour of audio in about half a second. AI summary
Transformers has implemented expert streaming to offload MoE model weights, achieving speeds of ~14 tok/s, 6% behind Strata's 13.9 tok/s, by streaming only the routed experts and using a bounded cache that groups_mm runs on as is. AI summary
Le Groupe Figaro and OpenAI partner to bring Le Figaro’s journalism to ChatGPT through summaries and excerpts, with links to original stories.
OpenAI models optimized circuits used in Jalapeño silicon and improved RISC‑V core designs, showing progress in AI-assisted chip design with GPT‑6 Astra.
Dean Ball, Head of Strategic Futures at OpenAI, examines how more capable AI could reshape society and how people can retain control of their institutions.
Organizations increasingly deploy downloaded open-weight GUI agents. A backdoored agent passes acceptance tests on clean data and misbehaves only when a visual trigger appears on screen, turning an unvetted checkpoint in
Pretrained Time-Series Foundation Models (TSFMs) provide strong representations for downstream classification tasks. Yet adapting a TSFM under changing operating conditions remains challenging. A model must improve on ea
Game2Player: Scaling Gameplay Experience for Visual Agents through Environment-Player Co-Development
This paper develops a framework called Game2Player that allows visual agents to learn from game environments and players at scale, by co-developing the environment and player through a loop of adaptation, task definition, and verification. Practitioners might care because it could help train more effective visual agents in a wider range of games.
This paper introduces Jeveration, a system that uses a fast and parallel prediction paradigm to generate high-quality images in a single forward pass, allowing for efficient and unified image generation, prediction, and other tasks. Practitioners might care because it could lead to faster and more efficient AI models for applications like image editing and object detection.
This paper develops a method to efficiently design neural processing units (NPUs) for large language models by using a mechanistic critic to evaluate and improve NPU configurations, allowing for faster and more informed design decisions. Practitioners may care about this research as it can lead to more efficient and effective NPU designs for real-world applications.
Large language models (LLMs) are increasingly used to infer psychological attributes, predict individual behaviour, and simulate human populations. However, despite generating plausible responses, they often misrepresent
Reinforcement learning from terminal answer rewards enables search-augmented language models to interleave reasoning and retrieval without step-level annotations, yet outcome supervision assigns identical credit to every
Automatic CAD generation is commonly evaluated with volumetric IoU, Chamfer Distance, and surface F-score, which barely penalize local errors that occupy only a small fraction of a part's volume or surface area, such as
This paper introduces a benchmark for automatic video rough-cut using large vision-language models, which can help improve the efficiency and quality of media production workflows. Practitioners can use this benchmark to evaluate and improve the performance of their own models on rough-cut tasks.
This paper proposes a new decoding strategy for looped language models, called late entry, which allows the model to skip initial passes and directly use a deep context cache, resulting in faster decoding times. Practitioners might care about this approach because it could lead to significant speedups in applications where decoding speed is crucial.
This paper develops a new AI system called PReMind that helps personalized vision-language assistants respond to visual hazards by considering a user's personal memories and context. Practitioners in the field of AI and robotics might care about this research because it aims to improve the safety and effectiveness of personalized assistants.
This paper proposes a system called WorldRuler that helps create interactive worlds with customizable rules, allowing for varied gameplay experiences. Practitioners in game development or AI research might care about this system because it can help them create more realistic and dynamic interactive worlds.
This paper proposes a method to improve the robustness of large language model agents against indirect prompt injection attacks by using a unified framework that combines red-teaming and adversarial training. Practitioners might care because it could help prevent agents from being hijacked by attackers and ensure they can complete their intended tasks.
This paper develops a new method for generating realistic 3D textures that accurately capture material properties, and how this might be useful for improving the realism of virtual objects in computer graphics.
Adapting a film into a game requires selecting source elements and translating them into player actions, rules, and feedback. We introduce Movie2Game-Bench (M2G-bench), a benchmark for evaluating film-to-game adaptation
This paper introduces Piccolo, a family of decision models that improve themselves by generating new data and adapting to that data, allowing for more accurate predictions on various tasks. Practitioners might care because Piccolo's approach could lead to more robust and adaptable AI systems.
This paper proposes a new method for training world action models, called HINT-WAM, which uses hindsight information to improve their performance on tasks that require reliable execution. A practitioner might care about this method if they want to improve the generalization and data efficiency of world action models in robot manipulation tasks.
This paper develops a framework to adapt large language models while preserving their general capabilities, allowing for selective unlearning and learning of new capabilities. Practitioners caring about model adaptation and generalization might care about this work.
This paper introduces AnyS1, a system that allows AI agents to make decisions more efficiently by leveraging human-authored procedural skills, reducing the need for complex model invocations. Practitioners might care about this research because it could lead to faster and more reliable AI decision-making in real-world applications.
This paper introduces StrideWorld, a benchmark that evaluates how well video world models maintain a stable scene structure and consistent actions when a user gives commands. Practitioners should care because it can help identify models that fail to maintain a realistic experience.
This paper introduces WaveView, a novel view synthesis method that improves upon existing techniques by injecting the warped source into the noise phase of a pre-trained generator, allowing for more accurate and efficient rendering of scenes. Practitioners in the field of computer vision and graphics may care about this approach as it has the potential to significantly improve the quality and realism of generated images.
External sandboxes for coding agents I've been exploring new ways to sandbox code that is executed by agents, with a focus on making it safer to try out local models without worrying that a mistake or prompt injection attack might cause the…
GPT‑6.2 Astra and GPT‑6.2 Sol OpenAI: We’re introducing GPT‑6.2 Astra, our new frontier model for your hardest projects, and GPT‑6.2 Sol, our efficient, versatile model for everyday work. They're available in the API as gpt-6.2-astra and gp…
Researchers at Harvard found that AI coding agents increased code production by 30% but did not improve finished work quality, instead shifting the bottleneck to code review, which took 49% longer. AI summary
Quad9's forced withdrawal from France highlights the dangers of mandating DNS-level blocking. We explore why forcing public resolvers to act as content filters is technically flawed and threatens the open Internet.
Starting today, new Pro and Enterprise teams no longer include a Standard build slot. This brings Standard builds under the same usage-based billing model as other build machines. On Pro, Standard builds cost $0.014 per build minute, and yo…
Anthropic's lobbying firm, Ballard Partners, has paid $740,000 to Marco Rubio's office since April 2025, with another firm, Continental Strategy, receiving $890,000 this year, after Rubio's memo led to the ouster of Nvidia CEO Jensen Huang. Ballard Partners, led by former Trump aide Brian Ballard, has lobbied Rubio's office, as well as the National Security Council (NSC) and the State Department, since hiring for the job in March. The lobbying efforts come as Anthropic prepares for its IPO, with Nvidia CEO Amadei pushing for stricter limits on chip sales to China. AI summary
Here is a summary of the article in 2-3 plain sentences for a developer/AI-ML audience: Databricks' Lakebase hackathon showcased the capabilities of Lakebase, a fully managed Postgres database built for the lakehouse, which enables branching, syncing, and serving of data. The hackathon featured two tracks: one focused on building applications with Lakebase branching and the other on syncing and serving data from Lakebase to the lakehouse. The results demonstrated how Lakebase can power a wide range of use cases, from AI agents to real-time applications, by providing a unified platform for data, analytics, and AI. AI summary
Azure Databricks is a Visionary-level sponsor at Microsoft Ignite 2026, featuring eleven sessions covering governed AI, AI agents, model fine-tuning, and Genie in Microsoft Teams and Microsoft Copilot Cowork. Sessions and demos focus on helping organizations move from AI pilots to production-ready AI. AI summary
GPT-6.2 Astra, a model from OpenAI, has achieved a perfect score of 100% on the ARC-AGI-3 benchmark, surpassing its predecessor GPT-6 Astra. This achievement marks the first time a model has scored 100% on ARC-AGI-3, with GPT-6.2 Astra using 64.1% fewer actions than the human baseline. AI summary
In the Codename MDASH pilot, GitHub and Microsoft turned thousands of raw AI findings into 99 checked, private reports with suggested fixes for six widely used open source projects. Maintainers have already fixed more than half. The post Ma…
Discover how we built the open Agent Host Protocol to keep your agent sessions accessible across editors, apps, phones, and the cloud. The post Building the Agent Host Protocol for sessions that continue across devices appeared first on The…
Decision models can be up to 444.6x cheaper and 193.6x faster than LLMs, making them suitable for tasks like filtering jobs before expensive LLM calls. By using decision models to gate expensive LLM calls, a company like PostHog has reduced the number of LLM calls by 98.7%, saving resources and improving efficiency. AI summary
Vercel Blob can now delete blobs automatically. Pass ttlDays when you write a blob and it is deleted that many days after upload, from 1 to 730 days. The TTL counts from the blob's upload date. Use TTLs for blobs you would otherwise delete …
Introducing GPT‑6.2 Astra and Sol, with better professional outputs, coding, computer use, and efficiency at lower cost across ChatGPT Work, Codex, and the API.
Tasks in Gemini is now free of charge in six countries, with new features starting in the U.S. and deeper Google Workspace integration to get things done on your behalf.
Datasette Lite is my port of the Datasette Python application to run directly in the browser using WebAssembly and Pyodide . I've released a complete rewrite of Datasette Lite using service workers, and it's now almost indistinguishable fro…
Azure Databricks on Azure Government is now in Beta, providing a modern data foundation with Unity Catalog, serverless compute, and secure-by-default workspaces, with full compliance and governance capabilities for the US public sector. AI summary
LiveRamp Embedded Identity is a packaged clean room that runs as a Databricks packaged clean room, where LiveRamp provides the resolution code and identity assets, and the customer runs that code against its own governed data in a central clean room. This environment isolates the computation, data access, and output, ensuring neither party can see the other's data or code. AI summary
Databricks and Google Cloud have announced a bi-directional federation of Unity Catalog and Google Cloud's Lakehouse, enabling customers to access a single copy of data from both platforms without duplication. This partnership aims to advance catalog metadata interoperability and governance through open standards, allowing customers to share metadata through key-value pairs (labels) and flag sensitive data for masking. The Lakehouse architecture is becoming increasingly open, with catalog federation enabling zero-copy access across multiple platforms. AI summary
TrnRT is an open-source LLM inference engine for AWS Trainium, utilizing NKI megakernels, 1M-token context, and an OpenAI-compatible API. It serves a single trn2.48xlarge model, runs decode as NKI megakernels, and maintains a prefix cache across device and host memory. TrnRT is built on top of the AgentX results, serving GLM-5.3-Flash on one Trainium2 box with NKI layer megakernels. AI summary
Cloudflare, a company known for its infrastructure services, has begun building its own AI models, specifically System One models, which are designed to process data in milliseconds. The company, which has launched two System One models, Clef and Clairic, claims to be a "neolab" - a term coined to describe companies that build AI models, unlike traditional research-focused startups. AI summary
Jensen Huang's resignation is linked to multiple scandals, including his long-standing anti-American stance and lobbying efforts to sell Nvidia's advanced AI chips to China, potentially compromising national security. Huang's relationships with Chinese individuals, including alleged spy Alice Huang, and his company's smuggling of billions in chips to China, have raised concerns about corruption and conflicts of interest. The US Commerce Department's inaction on export controls, partly due to Howard Lutnick's influence, has been criticized as a possible example of corruption. AI summary
Distributed systems are hard to test because partial failures overlap with live traffic and the same fault can produce different execution traces on different runs. Unmesh Joshi describes how deterministic simulation testing addresses this …
AI's lack of understanding of consequences and its ability to exploit vulnerabilities in code are contributing to the rise of cybersecurity issues, including nationwide FAA ground stops and major airline outages, with attacks on multiple organizations, including CNN, BBC, and Zoom, reportedly compromising websites to steal credit card information and cryptocurrency. AI summary
Most reads to Lakebase Postgres are served from fast local caches, with 99.46% never leaving the compute node and 99.99999% of reads being served without hitting object storage. AI summary
Speed Insights now measures soft navigations: client-side route changes that update page content without a full reload. This means you can now track performance as users move through your app, not just on full-page loads. Speed Insights cap…
Anthropic's financial dependencies on its investors, including Good Ventures Foundation, create a self-reinforcing feedback loop that makes it difficult for the company to be independently assessed or regulated. The Anthropic Network, which includes METR and other organizations, is financially dependent on Anthropic's success, and its members cannot afford to disrupt Anthropic's growth without risking their own financial stability. AI summary
Apple doesn't ruin the Mac, but they don't fix the agent problem, raising questions if they can think outside the app paradigm. Then, my iPhone Duo review.
The Claude Code team uses a combination of end-to-end (E2E), property, and differential tests to verify the work of its coding agents. They prioritize E2E tests, which describe real user flows, as they are effective in identifying bugs and provide definitive done criteria. Property tests, which look at edge cases, are also valuable in catching issues that agents often fail. Differential tests, which compare the old and new systems, are useful for large-scale migration use cases. AI summary
Perplexity has reached a $1 billion annual revenue run rate, a 4x increase in 2026, driven by growth from its Computer product. The company's valuation is estimated to be $30-50 billion, based on usual agent lab multiples. AI summary
Explore the new Sales plugin and 30+ new and expanded partner plugins in ChatGPT to research prospects, prepare pitches, update your CRM, and move deals forward.
A little copying is better than a little dependency. — Go Proverbs , via Tags: go , programming
A new Ignored Build Step preset lets you skip preview builds for branches without an open pull request. Production deployments and previews for branches with open pull requests continue to build. On GitHub and GitLab projects, skipped previ…
Tokendoctor is a tool that replays your past sessions to show what changes in Claude Code could have saved you up to 38% in costs, without requiring any model calls or changes. It analyzes your past sessions to identify areas where compaction windows could have reduced costs, and provides suggestions for adjusting settings to optimize performance. AI summary
Anthropic claims its AI model, Claude Mythos, can onboard AMD GPUs to run AI models in around three days, writing almost 100% of the kernels autonomously, reducing the time from up to 12 months. This could erode Nvidia's CUDA moat by allowing AI to write low-level code for chips. AI summary