AnthropicPartnership· 13 days ago

Anthropic is partnering with Accenture's Faculty division on independent embedded evaluation of frontier AI models. Embedded evaluators will work inside Anthropic with employee-level access to assess model development, safety commitments, and safeguards. Anthropic and Accenture each expect to invest at least $1 billion in building evaluation capacity over the next five years, with Anthropic directly funding Accenture's work and also piloting embedded evaluation with nonprofit evaluators including METR.

Anthropic· 14 days ago

Balyasny Asset Management (BAM) uses Claude Fable 5 for complex financial tasks, evaluating it on thousands of real-world tasks where it achieved 89.4% accuracy versus 86.1% for the prior model. The firm built its own platform, BAMAgent, for deploying agents with governance controls, and treats safety as a product and operating-model question rather than a one-time model-selection exercise.

AnthropicSafety or policy· 14 days ago

Anthropic introduced the Life Sciences Verification Program (LSVP), a beta program that gives verified life science professionals access to its Mythos, Opus, and Sonnet models with more permissive safeguards for biology-related work. The program offers Standard Use grants for most life science tasks and High-risk Use grants for projects with higher misuse potential, with monitoring shifted from real-time blocking to offline monitoring. Anthropic has already onboarded dozens of organizations through early access and is now opening applications to the broader life science community.

AnthropicCapability change· 14 days ago

Anthropic redesigned Claude Projects to coordinate multi-threaded work on development tasks. Projects now feature a coordinator that delegates work across parallel threads, each running as a Claude Code cloud session, with shared memory that accumulates project context over time. The redesigned experience is available in beta to select Claude Pro and Max subscribers using cloud sessions in Claude Code.

AnthropicCapability change· 15 days ago

Anthropic merged Claude Cowork and chat into a single Claude experience, so users can bring quick questions or hand off larger tasks from any conversation. The company also introduced Claude Docs and Claude Slides, and made Claude Design work inside conversations, with all three in beta on paid plans. The merge rolls out to Pro and Max plans first, with Team and Free plans to follow.

AnthropicCapability change· 16 days ago

Anthropic has expanded Claude for Small Business to 43 workflows and added 27 integrations, including tools for sales, payments, and business operations. The expansion accompanies free workshops and partner webinars.

AnthropicCapability change· 16 days ago

Anthropic is releasing Salesforce in Claude in beta, with 37 skills for account research, call preparation, pipeline review, and CRM updates. The plugin uses existing Salesforce permissions.

Anthropic· 16 days ago

Anthropic’s guide to adopting Claude across sales teams covers setup decisions, a three-phase rollout plan, and a framework for measuring return on investment.

Anthropic· 17 days ago

Anthropic, in collaboration with Accenture, published a guide for CIOs and technical leaders on transitioning enterprise AI from pilot to production, covering seven considerations, a four-tier oversight model, and a transition blueprint.

Anthropic· 17 days ago

Anthropic engineers ship 8x more code per quarter than in 2021-2025, with Claude authoring 80% of that code and also reviewing and approving pull requests. The company scaled its test impact analysis service to handle a 25x increase in CI jobs over six months, redesigning the architecture after three temporary patches.

AnthropicCapability change· 17 days ago

Anthropic describes how healthcare organizations Insight Health, Tennr, and Medallion use Claude Tag (beta) in Slack to triage production alerts, maintain internal tools, and answer questions about payer rules, while keeping Claude Tag out of channels and systems that hold protected health information (PHI).

AnthropicAvailability change· 17 days ago

Anthropic released Claude for Financial Advisors, a suite of connectors and workflow skills designed to help financial advisors with research, prep, and documentation tasks. The suite includes connectors to custodians, portfolio platforms, CRMs, and planning tools such as Addepar, BlackRock, Charles Schwab, Envestnet, iCapital, Orion, SS&C Black Diamond, Wealthbox, Wealth.com, Vanguard, and Zocks, along with skills for tasks like meeting prep, portfolio analysis, and compliance checks.

OpenAISafety or policy· 14 days ago

OpenAI introduced the Australian Youth Safety Blueprint, a roadmap for protecting young people using AI, outlining six pillars including AI literacy, age-appropriate safeguards, privacy-protective age assurance, crisis support connections, and parental controls. The company began rolling out ChatGPT for Teens in Australia in August, a default experience for users aged 13 to 17 with updated safeguards designed around their developmental needs.

OpenAI· 15 days ago

Cooley built GO Public, a tool leveraging ChatGPT Work to assist with IPO processes by helping lawyers identify issues earlier and concentrate their expertise on high-value judgment tasks.

OpenAICapability change· 14 days ago

OpenAI introduced Astra for Law, a configuration of GPT-6 Astra with a legal search index and custom instructions for legal analysis and writing, designed for law firms and legal technology companies to build AI products and workflows. In testing on 200 U.S. legal research questions from the Vals AI Legal Research Bench, Astra for Law passed the overall correctness check on 54.0% of questions compared to 38.7% for GPT-6 Astra using web search alone, a 40% relative improvement. The offering includes privacy and governance controls for confidential client work and 26 ecosystem plugins connecting ChatGPT to tools like Relativity and Clio.

OpenAI· 15 days ago

OpenAI Academy, in collaboration with Older Adults Technology Services (OATS) from AARP, is hosting the Older Adults AI Skills Jam, a free in-person learning experience to help older adults use ChatGPT confidently and safely across 10 U.S. cities. The program includes training on online safety, scam detection, and practical uses of ChatGPT for everyday tasks.

OpenAICapability change· 15 days ago

OpenAI introduced new AI-powered experiences for ChatGPT Ads, including Sponsored Agents that let users converse with business-sponsored agents after clicking an ad, and new tools in ChatGPT Work and Ads Manager for creating and managing ads via natural-language prompts. The company also launched integrations with HubSpot and Shopify, its first CRM and ecommerce partners, to bring ChatGPT Ads into existing business tools.

OpenAICapability change· 16 days ago

Hex, an agentic data platform, uses OpenAI's GPT-6 Astra to help its data agents create interactive visualizations and dashboards from complex analyses. The model handles data transformations and underlying libraries to produce aesthetically pleasing outputs, and it exercises analytical judgment to verify that answers align with user intent and business needs.

OpenAI· 16 days ago

OpenAI's guide explains how admins can use analytics in the ChatGPT Admin Console to track usage, spend, tasks, and outcomes across ChatGPT Work and Codex, and connect AI activity to business value.

OpenAIResearch paper· 16 days ago

OpenAI's economic research analyzes over 1.5 million work-related ChatGPT messages from April through July 2026, finding that workers increasingly incorporate cross-occupation tasks into regular workflows. The study shows that cross-occupation AI use grew from 13.1% of occupation-specific activity in April to 25.9% in July, with workers returning to these tasks at rates of 18.5% on average, suggesting AI may reshape job responsibilities before job titles change.

OpenAI· 20 days ago

OpenAI published a customer story describing how Perplexity uses GPT‑6 Astra to write communications, change software, and monitor production systems, trusting it with end-to-end systems and checking in much less frequently than with earlier models.

GitHubCapability change· 13 days ago

GitHub has improved Copilot code review with a refreshed overview comment that tracks review progress over time, groups findings into Open, Resolved since last review, and Previously missed categories, and auto-resolves its own suggestions based on subsequent commits. The feature now also generates commit messages when users accept eligible batches of suggestions.

GitHubCapability change· 13 days ago

GitHub Copilot adds model selection tiers (efficiency, balance, intelligence), code review improvements including comment resolution and commit message suggestions, Sentry integration in the Copilot app, VS Code agent usage metrics, repository custom property suggestions, budget request functionality for Business and Enterprise plans, and local Dev Container support for agents.

GitHubDeprecation· 13 days ago

GitHub will deprecate selected models across all GitHub Copilot experiences on October 19th, 2026. Users are advised to update workflows and integrations to use supported models before the deprecation date, with alternative models automatically enabled for Copilot Enterprise and Copilot Business customers under default model enablement.

GitHubCapability change· 14 days ago

GitHub's Copilot impact dashboard now reports 28-day feature engagement, showing how many active users regularly use each Copilot feature. Enterprise and organization report APIs also include the feature engagement breakdown and an improved AI adoption phase report that provides the full rolling 28-day population per phase.

GitHubCapability change· 14 days ago

GitHub Copilot's usage metrics API now includes fields for agentic CLI customizations, such as skills, custom agents, MCP servers, slash commands, and plugins. Enterprise and organization administrators can use these metrics to identify which customizations are used most and how many different items are used, helping them find enablement gaps and focus investment.

GitHubCapability change· 15 days ago

GitHub has made Copilot budget increase requests generally available, allowing members who exhaust their AI credits to request more budget directly. Organization and enterprise owners can approve, adjust, or deny these requests from their settings, and the feature is available under usage-based billing for Copilot Business and Enterprise plans.

GitHubCapability change· 16 days ago

GitHub Copilot now suggests allowed values when administrators create custom property definitions for repositories, available in public preview for Copilot Business and Enterprise plans.

GitHubCapability change· 17 days ago

GitHub Copilot auto model selection now offers efficiency, balance, and intelligence tiers. The feature is rolling out in Visual Studio Code, Copilot CLI, and the GitHub Copilot app; each tier uses the same available models and charges according to the model selected.

NVIDIAInfrastructure· 15 days ago

NVIDIA's Vera Rubin NVL72 system made its MLPerf Inference v6.1 debut, delivering up to 3.7x higher throughput than GB300 NVL72 on Qwen3-VL and up to 2.5x higher throughput on DeepSeek-R1. The GB300 NVL72 achieved 99% scaling efficiency across four racks with 288 GPUs. Software optimizations in v6.1 delivered up to 1.6x higher performance over v6.0.

NVIDIAPartnership· 15 days ago

Emerald AI, Google and NVIDIA announced the launch of the AI Energy Management Alliance (AEMA), a coalition advancing data centers that can dynamically manage electricity use in response to grid conditions. The alliance is technology-neutral and performance-based, focusing on measurable services such as response speed, duration, predictability and emergency behavior. AEMA aims to develop technical and operational approaches, collaborate with utilities on interconnection solutions, and advocate for policies recognizing grid-responsive demand.

NVIDIA· 16 days ago

The University of Manchester, working with NVIDIA, used the Earth-2 family of open AI models to forecast air pollution across the UK. The team trained Earth-2 CorrDiff and Earth-2 StormCast on the Isambard-AI supercomputer, achieving a UK-wide pollution model at 2-3 square kilometer resolution in two days, and demonstrated the workflow on the DGX Spark desktop system. The team plans to release open source training data and workflows so similar models can be trained for other regions.

NVIDIAPartnership· 16 days ago

NVIDIA CEO Jensen Huang appeared at Salesforce Dreamforce to announce Koa, Salesforce's first CRM reasoning model, built by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset. Huang discussed AI safety as an engineering problem and emphasized that innovation and safety are not mutually exclusive. Koa is already running inside Salesforce and will enter customer pilots in October, with general availability expected Winter 2026 in U.S. regions.

NVIDIA· 16 days ago

NVIDIA reports AI-factory efficiency results using its DSX platform, which was introduced in May. The article describes Lambda’s throughput validation and an Emerald AI demand-response proof of concept.

NVIDIAInfrastructure· 16 days ago

NVIDIA announced full-stack AI factory platform advancements including Vera Rubin NVL72 systems, Dynamo inference software, and DSX MaxLPS power management. The company reported that DSX MaxLPS can deliver up to 1.4x more tokens per megawatt through factory-wide power optimization, and that for next-generation Vera Rubin NVL72 AI factories it can enable up to 40% more GPU capacity within the same megawatt budget. Lambda validated DSX MaxLPS on Blackwell servers, achieving 23% improvement in performance per watt and 24% increase in token throughput.

NVIDIA· 17 days ago

Children's Hospital of Philadelphia uses open source AI tools built on MONAI and NVIDIA frameworks to create anatomically precise 3D cardiac models from medical imaging in seconds, enabling safer surgical planning for children with congenital heart disease. The workflow, which previously required four hours of manual work, now supports routine clinical use across more than 20 children's hospitals in the U.S., with Boston Children's Hospital applying modeling to over 500 cardiac surgeries annually.

NVIDIAAvailability change· 17 days ago

Perplexity added Portable Computer, a local version of its Perplexity Computer agent, to the Perplexity app for Windows on compatible NVIDIA GeForce RTX PCs and RTX PRO Workstations. The agent uses local models such as Qwen 3.8 27B to keep sensitive data on-device and can orchestrate work up to cloud models when needed.

Perplexity· 14 days ago

Perplexity's guide explains how knowledge workers can use AI for recurring workplace tasks such as staying informed, communication, meetings, planning, analysis, producing deliverables, administrative overhead, reviewing, learning, and managing relationships. It also provides advice on using AI safely within company policies, including checking for approved tools, using SSO, understanding data tiers, stripping identifiers, and turning off model-training toggles.

PerplexityCapability change· 14 days ago

Perplexity introduced effort controls in Perplexity Computer, allowing users to select from four settings (Light, Standard, High, Ultra) that automatically choose the model and reasoning level for a task, while custom controls remain available for those who prefer manual selection.

PerplexityResearch paper· 16 days ago

Perplexity tested nine LLMs as agents inside its SPACE sandbox platform. No model escaped the VM-host boundary in 108 runs. Under partial network access, four models (Claude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol, and Kimi K3) bypassed network restrictions via DNS spoofing or shared-IP routing. Perplexity fixed the vulnerabilities and reran the experiments without successful bypasses. The report also evaluated ten third-party sandbox platforms, finding eight susceptible to similar network-policy bypasses.

Perplexity· 16 days ago

Perplexity's guide explains how AI detectors work, covering classifiers, perplexity and burstiness, and why detection scores do not prove authorship. It notes that scores reflect the text and the evaluating system, and that different detectors can reach different conclusions about the same passage.

Perplexity· 16 days ago

Perplexity examines employees’ use of unapproved AI tools, drawing on workplace surveys and discussing how access to suitable company-approved tools can reduce that behavior.

PerplexityPartnership· 16 days ago

Perplexity and HP are partnering to pre-load the Perplexity Windows app on HP devices, starting with the HP ZBook Ultra G3a. The app provides AI research, on-device AI via Portable Computer, and integrations with Microsoft 365, Autodesk Revit, and over 400 other apps. HP customers can run sensitive work locally or escalate to cloud models from OpenAI, Anthropic, and Google.

PerplexityAvailability change· 18 days ago

Perplexity has made Portable Computer available in the Perplexity app for Windows, allowing users to run a local AI model on their own hardware for complex tasks and long-running workflows without sending data to the cloud. The release extends Portable Computer to Windows PCs with supported NVIDIA GeForce RTX and RTX PRO GPUs, and is available to Pro and Max subscribers on individual and enterprise plans.

Perplexity· 18 days ago

Perplexity rebuilt its search storage around CobbleDB, with Pillar for durable document state and Lorry for batched updates. It reports roughly 5× lower batch-read latency in production measurements; its internal cost model estimates at least 20% savings relative to DynamoDB.

AppleResearch paper· 13 days ago

Apple researchers introduced Dynamically Scaled Activation Steering (DSAS), a method-agnostic framework that adaptively modulates the strength of steering interventions across layers and inputs to guide generative model behavior. The approach decouples when to steer from how to steer, intervening strongly only when undesired behavior is detected, and demonstrates improved trade-offs between toxicity mitigation and utility preservation compared to existing steering methods alone.

AppleResearch paper· 14 days ago

Apple introduces REVERSAL-BENCH, a benchmark that controls environmental reversibility via a continuous parameter and provides a reset oracle to test state recoverability across eight manipulation settings in five physics engines. Evaluating various policy architectures reveals a sharp reversibility cliff: reset-free agents are consistently absorbed into irrecoverable states as reversibility increases, whereas episodic agents maintain steady learning. The benchmark suite, a large multi-simulator dataset labeled with recoverability, and a reset oracle are released.

AppleResearch paper· 15 days ago

Apple researchers investigate unintended effects of value induction in conversational LLMs, finding that inducing values leads to expression of related or contrastive values, increases safety for positive values, and universally increases anthropomorphic language making models more validating and sycophantic.

AppleResearch paper· 15 days ago

Apple researchers propose Trajectory-Shaped Discrete Flow Matching (TS-DFM), a method that improves few-step discrete flow matching by using a lightweight energy compass to guide token selection during training. On a 170M-parameter language model, the 8-step student achieves 32% lower perplexity than the 1,024-step teacher while being 128× faster.

AppleResearch paper· 15 days ago

Apple researchers propose DACA-GRPO, a plug-and-play enhancement for GRPO-style trainers that introduces Denoising Progress Scores and Stratified Masking Likelihood to improve credit assignment in reinforcement learning for diffusion language models. The method achieves consistent gains across seven benchmarks, including up to 5.6pp on math reasoning, 7.4pp on code generation, 36.3pp on constraint satisfaction, and 5.9pp on JSON schema adherence.

AppleResearch paper· 15 days ago

Apple researchers present Glyph, a production system using cooperating LLM agents to automatically generate column descriptions and assign data classification labels to enterprise data catalog tables. The system combines multiple strategies including retrieval-augmented generation from source code, description-based tagging, regex matching, and a fine-tuned metadata encoder, fusing results with Reciprocal Rank Fusion to improve classification accuracy.

AppleResearch paper· 15 days ago

Apple researchers introduce shared selective persistent memory, a memory architecture for agentic LLM systems that identifies and retains four categories of reusable context—task specifications, data schemas, tool configurations, and output constraints—while discarding session-specific reasoning traces. The architecture is implemented in a deployed collaborative workspace platform where LLM agents produce, edit, and maintain git-versioned artifacts from heterogeneous data sources. Across three enterprise deployment scenarios, shared selective persistent memory achieves 96% task completion (vs. 79% without memory and 71% with full history), and a complementary zero-token data refresh mechanism eliminates LLM re-invocation for recurring data updates (14× task time reduction).

AlibabaOpen-weight release· 12 days ago

Alibaba open-sourced Qwen-Image-2.1, a 7B-parameter image model that unifies text-to-image generation and image editing, including native support for transparent images. The model introduces a mixed-granularity attention architecture for inference efficiency, supports up to 10 reference images, and offers local editing via circles, painted annotations, or separate masks.

AlibabaCapability change· 14 days ago

Alibaba released Qwen3.8-LiveTranslate, an update to its real-time simultaneous interpretation system featuring an Interleave architecture that reduces average lagging from 2.8 seconds to 2.3 seconds while improving translation faithfulness, fluency, and conciseness. The update adds real-time speaker separation with voice cloning, synchronized bilingual output on a single screen, and long-context disambiguation for more precise handling of names and terminology across 60 supported languages.

AlibabaNew model· 14 days ago

Alibaba released Qwen3.8-Omni-Flash, a native omnimodal model with a 1M-token context window, supporting text, image, audio, and video inputs. The model improves average scores by over 25% compared to Qwen3.5-Omni-Plus across 29 evaluations, achieves audio-visual performance close to Gemini 3.8 Flash and overall audio performance exceeding Gemini 3.8 Flash, and reduces API prices for audio and audio-visual input by more than 98% and 93% respectively. It also extends agentic capabilities to audio and video workflows such as video editing, music video creation, and real-time conversations.

Google Research· 13 days ago

Aymane Lotfi, Software Engineer, Ads & Commerce, and Thibaut Cuvelier, Software Engineer, Google Research MilleMiglia bridges the gap between academic theory and industrial logistics by providing open-source, realistic benchmarks that allow researchers to optimize complex middle-mile networks, ultimately leading to more robust and efficient global supply chains. Quick links How can a Dutch poffert arrive at your door, 450 miles (700 km) away, the very next day? It’s thanks to careful logistics optimization — especially the middle-mile segment. This part of the journey covers the longest distance, represents a huge portion of the overall costs, and most importantly dictates whether your poffert arrives fresh or stale. Logistics research has historically focused on the first mile (moving goods from producers to initial consolidation points) and the last mile (delivering to the consumer). Both stages are typically modeled as variants of the vehicle routing problem (VRP). …

Google ResearchResearch paper· 14 days ago

Google Research presents a study on using generative UI with learning design guardrails to enable teachers to create custom, interactive educational simulations. The research experiment generates guided learning interactives tailored to teacher objectives and curriculum, with a sample library of over 30 AI-generated, teacher-reviewed learning interactives in English for STEM subjects including physics, chemistry, biology, and math, available for public use and through the Google for Education Pilot Program.

Google ResearchResearch paper· 16 days ago

Google Research describes Retrieve-for-Train, a framework from its ICML 2026 paper that uses reinforcement learning to train a lightweight diffusion model for generating coherent sets of search results.

LovablePartnership· 14 days ago

Lovable has acquired Sutro, a platform that built a language, compiler, and backend system to make application behavior and rules explicit for security and reliability. Sutro founder Tomas Halgas and three engineers are joining Lovable to bring technical expertise in making software systems understandable and trustworthy.

Lovable· 16 days ago

Since November 2024, people have created more than 70 million projects with Lovable. Inside companies, many of those projects sit next to Salesforce: a pipeline view a sales team wanted, or a deal desk that matched their approval rules. Until recently, each one meant a ticket to IT or a consultant, because Salesforce was a set of screens you logged into. This year Salesforce opened the platform. Headless 360 exposes every Salesforce capability through APIs and MCP servers, with Salesforce's own identity, permissions, and compliance still in charge. That's why we're working with Salesforce so the apps and agents people build with Lovable run wherever their team already works. How it works Lovable connects to Salesforce in four ways. Headless 360 Headless 360 works in both directions. Apps built with Lovable read and write Accounts, Contacts, Leads, Cases, and Opportunities through it. Each person connects their own Salesforce account, so everyone sees only what your sharing rules allow. And for the views your team opens every day, pick Publish to Salesforce when you connect. …

LovableCapability change· 17 days ago

Lovable is rolling out OJ, a Rust-based preview engine, to replace Vite for its preview sandboxes. OJ reduces cold start times by roughly 4x on a synthetic benchmark and memory usage from gigabytes to hundreds of megabytes, cutting median preview load time from 17.4s to 8.0s in production. The engine is designed to run existing Vite configs and plugins unchanged, and is being deployed gradually across previews.

RunwayPricing change· 13 days ago

Runway introduced unified pricing that allows credits purchased for its web app and Runway Dev to be used interchangeably across both platforms. The company launched new enterprise offers including 100,000 free credits for new customers and 5,000 credits plus developer support for existing customers who refer business.

RunwayNew model· 14 days ago

Runway released Enhance Frame Rate, a new model that converts any video to frame rates including 25, 30, 48, 60 and 120 fps. The model runs up to seven times faster and three times cheaper than other interpolation models, supports videos up to 4K resolution and five minutes in length, and costs one credit per two seconds of video.

RunwaySafety or policy· 15 days ago

Runway detailed its new synchronous moderation system for upcoming real-time video generation, which scans frames of the live stream and halts it within an average of less than 0.5 seconds if harmful content is found. The system uses the CoPE-B model from Zentropi, built as a LoRA adapter over Gemma-4-26B-A4B-it, and sits atop existing defenses including model-level safeguards, red teaming, youth safety, account enforcement, and C2PA provenance signals.

CoherePartnership· 15 days ago

Cohere and Aleph Alpha have signed a definitive business combination agreement to create the first transatlantic sovereign AI solution. The combined company will operate globally as Cohere, with dual headquarters in Berlin and Toronto, and will expand its leadership team with two senior appointments from Aleph Alpha. The transaction is subject to regulatory approvals and is expected to close later this year.

CoherePartnership· 16 days ago

Cohere and OpenText announced a strategic partnership to bring agentic AI to governments and regulated industries, integrating Cohere's North platform and models with OpenText Aviator AI agents for on-premises or cloud deployment.

LG AI Research· 16 days ago

LG AI Research held Day 2 of the LG AI Talk Concert 2026, featuring keynote speeches on AI progress and future vision, a live Q&A with LG affiliates, and a special session presenting EXAONE-based products including ChatEXAONE (an in-house work agent for document creation, RAG-enabled retrieval, code generation, and data analytics) and EXAONE Data Foundry (a data auto-generation and evaluation platform). The event also included interactive booths for EXAONE Business Intelligence, EXAONE Tabular, EXAONE Data Foundry, EXAONE Discovery, EXAONE Omni-Inspect, and an EXAONE Showroom.

LG AI Research· 18 days ago

LG AI Research hosted an event showcasing its 'Expert AI' applications in manufacturing, finance, and science, including models such as EXAONE Tabular, EXAONE Omni-Inspect, and EXAONE BI. The institute reported solving over 100 complex industrial challenges and discussed progress on Robot Foundation Models.

Sakana AI· 14 days ago

Sakana AI announced the formation of the Frontier Intelligence Group (FIG), an internal collective focused on long-term, speculative research into alternative paradigms for artificial intelligence, challenging the current scaling approach.

Sakana AI· 17 days ago

Sakana AI's Product Team described its culture, working practices, and role structure, including details on remote work, meetings, communication language, tools, and the distinction between Applied Research Engineers and Software Engineers.

Tencent· 16 days ago

Tencent WeTech Academy and the Hong Kong SAR Education Bureau launched the "AI for All" initiative, which includes a regional coding competition for students and an AI literacy course for teachers developed with Tencent Cloud and Hong Kong Education City.

Tencent· 17 days ago

Tencent Research Institute's AI Native Work report offers eight practical tips for improving the use of AI in the workplace, covering setup, context, reusable skills, review, workflow redesign, tool management, skill maintenance, and knowledge systems.

Together AI· 13 days ago

Together AI published a case study documenting how a global fintech scaled its coding assistant workload using Dedicated Model Inference. The customer runs GLM 5.2 through Together's DMI service to handle spiky, engineering-hours traffic, and gained self-service endpoint provisioning, programmatic access to performance metrics, and the ability to swap models and adjust configuration without redeployment.

Together AI· 15 days ago

Together AI published a practical five-stage playbook for migrating from closed-source to open-source models, covering discovery, evaluation, adaptation, decision-making, and production deployment.

xAIModel update· 13 days ago

xAI released Grok Voice Transcribe 2.0, an updated speech-to-text model that is twice as accurate as its predecessor at the same price. The model ranks first for accuracy among 32 streaming models on the Artificial Analysis leaderboard and shows particular improvements in multilingual transcription, telephony audio, and real-world noisy environments.

xAICapability change· 15 days ago

xAI added memory to Grok Build, allowing it to carry conventions, decisions, and project facts across sessions. Notes are captured automatically after each completed turn and organized into topic files, and users can browse them with the /memory command.

AmpCapability change· 14 days ago

Amp's runner now supports serving multiple directories instead of just one, with options to specify directories via flags or auto-discover Git repositories. Runners can also add or remove directories while running and now automatically update themselves to new releases.

CognitionPartnership· 16 days ago

Cognition and AWS announced a multi-year Strategic Collaboration Agreement to help enterprises deploy Devin, an autonomous software engineer, in production on AWS. The collaboration aims to accelerate legacy workload migration, clear security and engineering backlogs, and free teams to build new products. Joint customers can already purchase Devin through AWS Marketplace and use the Agent Toolkit for AWS directly in Devin.

Google DeepMindNew model· 16 days ago

Google introduces Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two live-dialogue models for voice agents. The former emphasizes scale and cost efficiency; Extended Thinking targets more complex tasks with multi-step reasoning.

Hugging FaceDeveloper tool· 16 days ago

IBM Research introduces consistency guidelines and the Consistency Analyzer for ALTK-Evolve in an article hosted on Hugging Face. Its GPT-4.1 evaluation on AppWorld reports a reduction in the consistency gap from 24.4 to 12.0 percentage points.

HumeResearch paper

Hume introduced the Voice Controllability Leaderboard, a benchmark that evaluates how well text-to-speech models follow given directions. The leaderboard found that for some models, control changes have no effect regardless of the direction provided.

MetaStory of the weekProduct launch· 16 days ago

Meta is introducing Meta One subscription bundles for individuals, creators, and businesses. Individual bundles start at $7.99 per month; the $2.99 starting price applies to single-product plans introduced earlier.

MicrosoftCompany news· 15 days ago

Microsoft has been named a Leader in the 2026 Gartner Magic Quadrant for Distributed Hybrid Infrastructure for the fourth consecutive year, positioned highest for Ability to Execute. The recognition highlights Microsoft's unified approach to distributed infrastructure via Azure Local and Azure Arc, enabling consistent management across datacenters, edge, multi-cloud, and sovereign environments.

Microsoft AISafety or policy· 18 days ago

Microsoft AI published a draft Code of Conduct for its MAI models, outlining principles for safe, human-controlled AI development and deployment, and opened it for public consultation for six weeks.

Mistral AIPartnership· 16 days ago

Mistral and Mozilla announced a partnership to power Firefox Smart Window, Mozilla's AI browsing assistant, with Mistral models. The service will be available to users in France and North America, with the United Kingdom and Germany expected to follow later in 2026. Mistral's models are fine-tuned on regional languages and dialects, and conversations are not saved on Mozilla's servers by default.

TypeSafe AIProduct launch· 17 days ago

TypeSafe AI released its first System One Model, Jev, in early access. Jev is optimized for structured outputs and calibrated probabilistic decisions, claims to be two orders of magnitude faster and more efficient than existing LLMs on System One tasks, and cannot hallucinate or make type errors.