HDATF
English
Start a conversation
Menu

News

Following the ideas and developments changing how we work.

264 articlesNewest first / Last collected: 2026-11-14
AI & agentsOpenAI Blog

OpenAI introduces ChatGPT Wallet for agent payments

OpenAI announced ChatGPT Wallet, a payment method that lets AI agents complete purchases without ever seeing the user's actual card details. Every purchase requires user approval with a security check such as Face ID, and existing card rewards and dispute protections are preserved. Built with American Express, Mastercard, and Visa, it is initially available through dots for ChatGPT Pro subscribers in the US, alongside a playbook helping businesses support agent customers.

Why it matters for our work

This is one of the first payment rails built for agentic commerce, moving AI assistants from product research to completed purchases. Payment and customer-service teams may need to prepare for customers whose agents transact on their behalf.

Read original (OpenAI Blog)
AI & agentsAnthropic Research

Anthropic shares guidance for building strong safety evaluations

Anthropic published recommendations for building trustworthy AI safety evaluations. Drawing on reforms from psychology's replication crisis, it recommends preregistration, construct validity, and human validation of automated grading. It also describes sandboxing evaluation environments and pausing or hardening internet-enabled evals after its own incidents. The post is grounded in the company's system-card practice.

Why it matters for our work

It offers a lens for judging how much to trust any claim that a model is safer. Useful reading for teams that commission or review AI evaluations in adoption and research work.

Read original (Anthropic Research)
Industry changeHugging Face Blog

llama.cpp reaches v1.0.0 for the local AI ecosystem

llama.cpp and ggml, the open-source engines widely used to run models on everything from laptops to servers, have released version 1.0.0. The project formalized API stability, promising to surface breaking API and ABI changes only in major-version bumps. It says it now supports more than 150 model architectures.

Why it matters for our work

Lower compatibility risk for teams building local or on-prem AI products. A sturdier foundation for organizations that cannot send their data to external clouds.

Read original (Hugging Face Blog)
Future of workAnthropic (claude.com)

Anthropic launches Claude Meetings beta

Anthropic has started rolling out Claude Meetings in beta for Team and Enterprise plans. Claude joins video calls as a visible participant, answers questions spoken or typed during the meeting, and builds docs, whiteboards, and other artifacts live while teammates refine them. Claude takes care of notes and action items, starting with Google Meet support, and by default stays out of calls that include outside guests.

Why it matters for our work

This turns meetings into sessions where AI listens and produces deliverables in real time. It could cut the time teams spend on notes and follow-up documents, letting people focus on the discussion itself.

Read original (Anthropic (claude.com))
LeadershipNVIDIA Newsroom

Nvidia CEO Jensen Huang resigns; CFO Kress steps in

Nvidia's board announced that Jensen Huang has resigned as President, CEO, and board member, effective immediately. CFO Colette Kress will serve as Interim CEO, and Harvey C. Jones was appointed the company's first-ever Chair of the Board. The announcement did not state a reason; media reports linked the departure to recent stories about Huang's personal life. Nvidia said it will report quarterly results as scheduled on November 17.

Why it matters for our work

The founder-led era at the central supplier of AI chips ends abruptly. Leaders planning AI adoption and procurement will be watching the CEO search and whether the product roadmap holds.

Read original (NVIDIA Newsroom)
AI & agentsGoogle Blog

Google introduces Gemini 4 Hydrogen

On November 12, Google introduced Gemini 4 Hydrogen. Unlike the deep-reasoning frontier models Carbon and Argon, it is a fast execution model for everyday coding and agentic work, and it is available immediately across AI Studio, the Gemini API, the Gemini app, Search, and paid Antigravity tiers. By Google's own published numbers, it scores 51 on the Artificial Analysis Intelligence Index versus Argon's 50, reaching 54 at its highest thinking setting, with flat pricing of $1.50 per million input tokens and $7.50 per million output tokens across the full one-million-token context. A new Orchestrator in Antigravity automatically switches between Carbon and Hydrogen within a single session.

Why it matters for our work

As cheap execution models get stronger, splitting work between expensive frontier models and low-cost executors becomes practical multi-model operations. The immediate effect is a direct drop in the cost of delegating repetitive coding and document work to agents.

Read original (Google Blog)
LeadershipMicrosoft Official Blog

Microsoft and IDC outline what top AI performers do differently

On November 12, Microsoft published the fourth installment of IDC research it commissions, based on responses from more than 5,200 senior leaders across 28 industries. Frontier Firms, the top 10% on IDC's AI maturity scale, report an average 4x return on AI versus 2.7x for the bottom 30%, and 96% of them are rolling out agents at scale. Across all organizations, however, only 61% are highly confident they can measure AI's business value, and 36% are still piloting agents. The return figures are leader-reported, and the article itself notes the research shows association rather than cause.

Why it matters for our work

The core takeaway is that firms managing measurement, governance, and organization-wide skilling as one system turn early AI results into repeatable value. That 80% of Frontier Firms have an org-wide AI skilling strategy, versus 20% overall, signals that people, not tools, are the next bottleneck.

Read original (Microsoft Official Blog)
Future of workGenpact Press Release

Genpact survey: agents in production, value capture lags

On November 12, Genpact released a report based on a 2026 survey of more than 2,000 executives across 17 industries. While 79% of respondents report AI agents in production, only 20% use them broadly and consistently across functions, and only 15% have fully tied AI investment to P&L outcomes. On average, 49% of AI initiatives funded in the past 24 months delivered the expected value, and just 9% have completed significant operating model transformation. Among organizations running agents, 92% kept their workforce the same size or grew it; all figures are self-reported.

Why it matters for our work

The survey's core finding is that agent adoption has become common while the harder work of redesigning how companies operate and measure value lags behind. For organization designers, it argues that running people and agents as one workforce is what separates leaders from the rest.

Read original (Genpact Press Release)
Industry changeDatabricks Newsroom

Databricks announces always-on agent Genie

On November 11, Databricks announced the second generation of Genie. Given a goal and clear boundaries, Genie takes ownership of long-running work, tracks progress on its own, and reaches out through a priority inbox and email only when a person needs to decide. Genie Ontology preserves business context such as definitions, metrics, and organizational knowledge, while Unity Catalog and Unity Gateway keep every action within existing permissions, budgets, and auditable controls. The CEO's claim of being a '10x more productive' user and the figure of thousands of customers in production are the company's own statements.

Why it matters for our work

An always-on coworker that handles follow-ups on its own, rather than a chatbot waiting for instructions, is reaching production first through a data platform company. The case shows how delegating routine management work can be designed together with permissioning and auditability.

Read original (Databricks Newsroom)
AI & agentsGoogle Open Source Blog

Agent standards MCP and Agent Skills move under neutral foundation

On November 11, Google's Open Source Blog published an interview on the governance of MCP, the agent connectivity standard, and Agent Skills, the procedural knowledge standard. After Anthropic donated both to the Linux Foundation's AAIF, engineers from Google, Anthropic, and OpenAI now co-maintain them under vendor-neutral governance, likened to the CNCF's role for Kubernetes. The MCP specification moved to a stateless architecture with its July 28, 2026 version, improving horizontal scalability. The post frames the shift as the industry prioritizing interoperability over competing standards.

Why it matters for our work

When open standards for tool connections and work procedures live in a neutral foundation, different agent platforms can reuse the same tools and skills. Lower lock-in costs make it easier for teams to swap and combine agent tooling.

Read original (Google Open Source Blog)
AI & agentsGoogle blog

Google launches the Gemini Decisions API for fast judgments

Google introduced the Gemini Decisions API, powered by the new Gemini Decision 1 model. It accepts one multimodal input across text, images, audio, video, and PDFs and answers up to 1,000 yes/no, multiple-choice, and rubric-scoring questions in a single pass. Instead of generating text, the model scores candidate options and returns probabilities and confidence scores. Google reports 150-250 ms end-to-end latency and $0.10 per 1M input tokens with no output-token charges, figures from the company's own benchmarks.

Why it matters for our work

Moving repeated agent judgments such as tool selection, policy checks, and model routing from slow generative calls to a purpose-built endpoint could change the speed and cost of human-AI pipelines. Low-confidence cases still need a path to human review.

Read original (Google blog)
Future of workGoogle Cloud blog

Google Cloud unveils Professional Agentic Architect certification

Google Cloud announced a new Professional Agentic Architect certification for professionals who can deploy secure, enterprise-grade agents that automate multi-step, multi-app workflows. The hybrid credential combines a proctored multiple-choice exam with hands-on Google Skills labs, and the company says thousands of its field engineers and partners have already trained on the curriculum over the past four months. The claim that more than 80% of certified individuals see new career opportunities comes from Google Cloud's own research.

Why it matters for our work

It is a signal that designing agents is becoming a formally recognized role. As people and AI work together more, a credential for verified agent-architecture skills can reshape hiring and how teams divide responsibilities.

Read original (Google Cloud blog)
Industry changeLinux Foundation press release

Linux Foundation launches Agent Name Service standard

The Linux Foundation announced that Agent Name Service (ANS), an open standard for trusted identity, verification, and discovery of AI agents built on existing DNS infrastructure, has launched as a foundation project. The specification spans seven documents covering domain-anchored naming, registration and versioning, DNS publication, public log auditing, integrity monitoring, and agent-to-agent authentication. Eleven organizations including Cisco, Cloudflare, DigiCert, GoDaddy, and Salesforce contribute to or support the standard, and AGNTCY's Agent Directory Service already verifies ANS identities.

Why it matters for our work

Delegating work to agents requires knowing whom an agent represents. A verification layer built on DNS, which the internet already runs on, can lower the barrier for enterprises to adopt and audit agents.

Read original (Linux Foundation press release)
AI & agentsHugging Face blog

Hugging Face maps the decision model ecosystem

Hugging Face published an official blog post introducing decision models, which score predefined answer options and return probabilities instead of generating text. The post shows a pipeline('decision') interface in transformers with a ready-to-use model, notes that first-class training support is being prepared in TRL, and points to an Unsloth recipe that fine-tunes generative models into decision models. It also explains how models once called System One models, after TypeSafe AI's Jev, came to be named decision models by the community.

Why it matters for our work

Fast, repeated judgments such as classification and routing can now be handled with probabilities rather than text generation, and official tooling is standardizing around the idea. This gives research and work automation a new option for response speed and output consistency.

Read original (Hugging Face blog)
Future of workOpenAI Blog

OpenAI and Upwork study how AI is reshaping freelance demand

OpenAI's economic research team and freelance marketplace Upwork analyzed job postings across 63 work categories alongside ChatGPT usage data. Forty-three categories where ChatGPT was heavily used lost share in Upwork postings, while AI-related projects combined tasks that had rarely appeared together and carried larger fixed-price budgets and more hours. Between early 2023 and late 2025, the task mix in AI-related postings became about 45% more diverse, with similar posted hourly rates. All figures come from the two companies' own analysis.

Why it matters for our work

Freelance markets surface work changes earlier than employment statistics, offering an early read on which tasks shift to AI and which grow by combining with it. The pattern of AI-related work winning bigger budgets also applies directly to how teams allocate work.

Read original (OpenAI Blog)
AI & agentsGoogle Blog (DeepMind)

Google releases MKA benchmark for visual knowledge work agents

Google DeepMind and evaluation firm Artificial Analysis launched the Multimodal Knowledge Agent (MKA) benchmark. Experts across 10 industries and 83 occupations created 200 real-world tasks that take a seasoned professional three to eight hours, graded by rubrics weighting factual accuracy at 75% and visual design at 25%. Every model runs the tasks in the same open-source agent harness, and tested models averaged between 30 and 130 turns per task, a wide efficiency gap. Google reports that its automated grading agreed with human experts 96% of the time.

Why it matters for our work

The benchmark provides a yardstick for how far agents can take over knowledge work that depends on reading visual files such as blueprints, decks, and spreadsheets. By reporting efficiency in turns alongside quality, it helps judge readiness for real-world adoption.

Read original (Google Blog (DeepMind))
Industry changeGoogle Cloud Blog

Google Cloud launches Gemini for healthcare

Google Cloud introduced Gemini for healthcare in preview, an AI offering that unifies clinical, operational, and administrative data. It connects electronic health records, billing, clinical guidelines, and Medicare datasets so agents can run multi-step workflows such as medical records synthesis and prior authorization review, but only within defined guardrails and with human oversight. HCA Healthcare, Highmark Health, and Northwell Health are launch partners. The cited figure that clinicians lose nearly 28 hours a week to paperwork comes from a 2024 Google Cloud-commissioned study.

Why it matters for our work

Administrative overload is a shared pain point for healthcare systems worldwide, so how quickly agents can lift hospital paperwork is worth watching. Three large US health systems joining from the start offers an early read on real-world validation.

Read original (Google Cloud Blog)
Industry changeOpenAI Blog

OpenAI brings business profiles and Business Agents to ChatGPT

OpenAI launched the Businesses in ChatGPT beta inside ChatGPT, which it says reaches 1.2 billion people a week. Companies can claim and manage their profile in Business Hub and connect Shopify product feeds, while users can continue with a conversational Business Agent from a business profile. The agents roll out first to eligible US ecommerce and online-service businesses, and agent chats are clearly labeled as separate from ChatGPT's regular answers. Businesses can also sponsor their agent to surface it like an ad.

Why it matters for our work

As customers ask AI instead of searching and transact with agents, sales, marketing, and support work will need rearranging. Operating an agent inside an AI platform is becoming a new baseline task for businesses.

Read original (OpenAI Blog)
AI & agentsIBM Newsroom

IBM and Microsoft announce Imagine Enterprise agent platform

IBM Consulting and Microsoft unveiled Imagine Enterprise, a jointly developed agent platform. IBM contributes more than 6,000 pre-built agents and assets from its consulting platform, while Microsoft supplies the cloud, security, and governance stack, with forward-deployed engineers from both companies. It starts with frameworks for financial services, insurance, and retail, emphasizing controls and observability for regulated industries. The featured customer, SEI, is still at the exploration stage, and all details come from the companies' own announcement.

Why it matters for our work

Two major vendors bundling packaged agents with hands-on operating support signals the shift from agent experiments to production deployment. Whether the industry frameworks actually shorten build time is what to watch.

Read original (IBM Newsroom)
LeadershipPwC US newsroom

PwC to train 85,000+ staff in critical thinking

PwC US announced a Critical Thinking and Judgment certificate program for more than 85,000 partners and employees as AI takes over foundational tasks. The rollout starts in December 2026 across its US, Mexico, and Acceleration Centers offices, with completion targeted by June 2027 through diagnostics, masterclasses, personalized simulations, and on-the-job tools. The firm also plans to open its proprietary framework, developed with input from Wharton- and Harvard-affiliated professors, to universities from early 2027. The program is a plan, and its effects have not yet been verified.

Why it matters for our work

A major firm is formally training the human judgment needed to check AI output, not just AI skills themselves. As AI absorbs routine work, the ability to review and judge becomes what separates quality and trust in deliverables.

Read original (PwC US newsroom)
AI & agentsAnthropic blog

Anthropic introduces Claude Fable 5.2 and Mythos 5.2 for hard technical problems

Anthropic has introduced Claude Fable 5.2 and Claude Mythos 5.2, new models aimed at demanding technical challenges. Both share the same foundation: Fable ships to all users with cybersecurity and biology safeguards, while Mythos is limited to organizations in Anthropic's verification programs for life sciences and cybersecurity. By the company's own account, output is up to 30% faster than the previous generation, input prices drop to $8 per million tokens, and an optional fast mode generates tokens up to 2.5x faster. Prerelease feats such as large software migrations and solving two open black hole problems are company-reported, not independently verified.

Why it matters for our work

For coding and research teams, the barrier to using top-tier models on hard problems falls on both speed and price. If routing routine work to daily-driver models and reserving frontier models for the hardest tasks becomes standard, team productivity and research pace can structurally improve.

Read original (Anthropic blog)
AI & agentsAnthropic research

Claude designs working electronic prototypes from start to finish

Anthropic's research team partnered with circuit board firm Diode Computers to have Claude Fable 5.2 design three devices: a speakerphone, a voice-assistant hub, and a handheld virtual pet. The model used geometric reasoning to lay out circuit boards, factored in manufacturing constraints, modeled its designs in CAD software, and directed human engineers through assembly. The company reports the resulting prototypes worked without any circuit fixes, though validation comes from its own experiments.

Why it matters for our work

The case shows AI moving beyond code into hardware design and prototyping. Engineers may hand repetitive design iterations to AI and focus on verification and judgment, creating a new division of labor.

Read original (Anthropic research)
AI & agentsOpenAI blog

OpenAI unveils Control Room to stop model workloads across research infrastructure

OpenAI has introduced Control Room, a central system that tracks model workloads across its research infrastructure and stops them quickly when safety or security concerns arise. Model output halts within 35 seconds of an intervention decision, and the design keeps output stopped even if Control Room itself goes down. Every research workload doing agentic inference must be tied to an approved safety case.

Why it matters for our work

As AI agents run at scale, turning them off becomes a design problem in its own right. Enterprises adopting agents can take the same lesson: build emergency stop and audit mechanisms alongside deployment.

Read original (OpenAI blog)
Industry changeOpenAI blog

ChatGPT Enterprise becomes available in AWS Marketplace

OpenAI is now selling ChatGPT Enterprise through AWS Marketplace. Organizations can discover, buy, and manage it with their existing AWS purchasing and billing relationship, and apply eligible spend toward their AWS commitments. Healthcare and financial services editions are also available to eligible customers through the same channel.

Why it matters for our work

Bringing AI tools inside existing cloud procurement makes evaluation and budget approval simpler. For companies juggling multiple clouds and subscriptions, it reduces administrative overhead.

Read original (OpenAI blog)
Industry changeGoogle blog

Google unveils TPU 8tx and 8i Max built for the agentic era

Google has added TPU 8tx for secure, confidential inference and TPU 8i Max for massive context windows to its eighth-generation TPU lineup, the first time it has shipped four TPU systems in a single year. By the company's account, TPU customers now number nearly 100, including frontier labs, trading firms, retailers, and universities. Google frames the systems as purpose-built hardware as computing shifts from human to agent scale.

Why it matters for our work

As agents run long, complex jobs, huge context windows and dedicated infrastructure become the bottleneck. For enterprises, the change means more hardware options for balancing agent workload cost and security.

Read original (Google blog)
AI & agentsMeta AI Research blog

Meta brings specialized models to Muse and Meta AI

On November 6, Meta Superintelligence Labs announced specialized versions of its existing models, retrained with the Muse Wonder series introduced in October, rolling out across Muse and Meta AI. The update also brings Active Memory and Active Focus to these models. Meta says that by its own measurements, Muse now forgets critical information 2.4 times less often and completes everyday tasks disrupted by unexpected issues 67% better.

Why it matters for our work

Whether an AI assistant retains context between conversations and recovers on its own to keep working toward a goal is a decisive factor for using agents in real work. This rollout shows memory-capable assistants spreading from premium models to general users, though the improvement figures come from Meta's own measurements.

Read original (Meta AI Research blog)
AI & agentsAnthropic Research Blog

Anthropic tested Claude agents inside EVE's game economy

Anthropic's Fellows Program partnered with game studio Fenris Creations to have Claude trade on an EVE Online market replay and run long missions as four-agent teams in EVE Frontier. The market experiment ran in a simulator replaying past player orders rather than the live game; over six months the newest model, Fable 5, grew its capital to about six times the starting amount, well ahead of the roughly two times earned by the older Opus 4.6. Directional bets based on in-game news stayed profitable in only about half of runs. Teams coordinated well, but a secretly instructed saboteur was almost never caught by teammates, and dedicated monitor agents caught only 30-40% of saboteurs. All figures come from Anthropic's own experiments.

Why it matters for our work

The study probes how far agents can be trusted with capital and teamwork in a low-stakes setting, and points to concrete fixes such as explicit risk limits and audit trails for organizations where humans and AI work together.

Read original (Anthropic Research Blog)
AI & agentsAnthropic Research Blog

Anthropic on letting AI build scientific models

An Anthropic research essay breaks scientific modeling into three steps—building the model, constraining its parameters to data, and interpreting it—and describes AI's role in each. Over the past two years, systems have appeared where LLM-based agents write candidate model code, run it, compare outputs with real data, and revise, taking over design work once thought to be human-only. The central argument is that many models can fit the same data, so researchers should ask for the full distribution of plausible models rather than a single best fit. The authors also state the limitation that such systems have mostly been tested on rediscovering known models, with sparse evidence on open problems.

Why it matters for our work

The essay draws a sharp line between AI producing an answer and producing understanding, and offers a practical workflow: compare surviving models to see which parts of a system are pinned down by data and where the next experiment should aim.

Read original (Anthropic Research Blog)
AI & agentsOpenAI Blog

OpenAI releases BioBasicsBench for everyday wet-lab analysis

OpenAI has released BioBasicsBench, a benchmark built with Ligo, an enzyme-design biotech company, to measure how well AI agents handle routine wet-lab analysis. Scientists wrote its 100 tasks and grading rubrics from raw experimental evidence such as gel images, microscopy stacks, and sequencing reads. On the company's own evaluation, GPT-6 Astra earned an average of 66.7% of available credit at the highest reported reasoning setting, well ahead of GPT-5.6 Sol at 43.8%; failure cases, including a roughly 12% overcount in colony counting, were published alongside the results.

Why it matters for our work

How accurately AI reads messy raw experimental data, not polished paper figures, is the crux of automating research support. Publishing benchmark results together with failures gives labs a yardstick for deciding which analysis tasks can be delegated.

Read original (OpenAI Blog)
AI & agentsAnthropic Research

Anthropic shares research on the components of agent memory

Anthropic has published research on how it designs agent memory: knowledge Claude keeps between sessions. The core ideas are memory notes written by judging what is salient in the moment, and dreaming, where the agent re-reads session transcripts offline to fix mistaken notes. In the company's own experiments on a long-horizon Pokemon task, an older model wrote 413 memory files and earned one badge while a recent model wrote nine files and earned four; dreaming also corrected a wrongly noted location and got the run back on track.

Why it matters for our work

When agents handle long-horizon work, failures often trace back to memory design. A map of what to record and when to self-verify underpins the reliability of human-AI collaboration.

Read original (Anthropic Research)

Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.