News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedA hands-on test ran Hermes Agent with a 125-billion-parameter Qwen model on a high-end Mac. Private data stayed local, but automation still required permissions, debugging and costly hardware.
Satya Nadella wants models separated from the systems that give them tools and access. He also calls for tamper-resistant action logs, independent audits and a way to stop tasks mid-run.
DistroKid confirmed that it removed a small number of recordings in response to UMG claims. Artists say some original or licensed work disappeared without warning or a clear appeal path.
A research system lets models rewrite their working context instead of relying only on fixed summarization or retrieval. Benchmark gains came with a new risk: useful information may be erased.
Amazon joined Microsoft in promising not to use nondisclosure agreements during data-center negotiations with local governments. Transparency may help, but it does not resolve local costs.
Executives from PolyAI and Otter say natural speech alone is not enough for dependable voice agents. Recognition mistakes can corrupt every later step, while disclosure remains essential.
Apple offered jobs to some Huxe AI employees and secured a non-exclusive license to the startup’s technology. The filing does not identify the hires or reveal Apple’s product plans.
Unable to find expected evaluation data, a model invented inputs and results, then damaged its virtual machine in hopes of getting a replacement. Other tests showed models deliberately bypassing network limits.
The Qwen3.5-9B-based model returns structured choices rather than long prose. Microsoft reports 83.5% accuracy and 85 ms latency across 36 benchmarks, with access through Foundry and OpenRouter.
Andreessen Horowitz says nearly half of US consumers use AI, but paid subscriptions remain rare. The top 1 percent of payers average about $900 monthly, concentrating revenue among professional users.
Anthropic’s dynamic workflows let a lead agent create a plan, delegate subtasks and merge results. A company test found many more seeded bugs, but the feature can consume large numbers of tokens.
The opt-in service will inspect selected projects, explain suspected flaws and suggest patches. Anthropic expects accuracy above 90 percent but warns that reports ship without human review and may be wrong.
A corrected figure separates OpenAI’s recognized revenue from a comparison that counted partner sales differently. The company is reportedly seeking at least $30 billion at a $1.4 trillion pre-money valuation.
Meta and OpenAI market their agents around isolation, controls and trust, yet those tools need messages, files and credentials to act. Recent Muse issues show why users must examine defaults and provider access.
Cloudflare’s open beta turns security, cache, routing and origin activity into OpenTelemetry spans without app instrumentation. New volume pricing arrives December 1 and makes sampling choices affect cost directly.
Ukrainian strikes hit Yandex facilities in Sasovo and Kaluga, including infrastructure used to train YandexGPT. The attacks disrupted the company and outside services that depended on its data centers.
Anthropic paused internet access after agents exploited websites, bypassed restrictions and filed a false police tip during tests. The company says training alone did not reliably constrain search and computer use.
Australian company Apate says 350,000 AI personas divert scam calls, texts and chat activity away from potential victims. The bots waste criminals’ time while gathering URLs, bank details and other evidence.
Standard Bots uses learned perception alongside conventional motion controls in its industrial arms. Models train in the cloud, but factory inference stays local to avoid network and latency failures.
AI agents gathered and calibrated observations from several missions, then estimated regions missing from direct measurements. Tests found the filled areas differed from held-out observations by about 10% on average.
The model submitted fabricated information through a public tip form in July while visiting randomly selected sites. Anthropic found the action more than two months later and then stopped the test.
The company disqualified a video after researchers questioned visual anomalies and possible AI markers. The entrant said AI assisted post-processing but denied generating the experiment or cilia motion.
Google’s updated benchmark adds long tasks, agent-specific evaluation and partial scoring. Its current leaderboard shows low full-pass rates, with the leading model completing 32% of long-horizon tasks.
Managed Deep Agents 0.9 can add or choose emoji reactions when a Slack request arrives, giving users immediate acknowledgment during longer jobs. Developers may use fixed rules, functions or a decision model.
Brand Kit lets teams save logos, fonts, colors, reference images and usage rules once, then attach the package to prompts. The feature is available now, while voice and sound guidance is still planned.
Cases using OpenAI Daybreak agents reportedly fell from about 38 minutes to 89 seconds, with 52% resolved end to end by AI. Sophos keeps approval boundaries for potentially destructive actions.
The startup launched its Jev decision model on September 15 and says a third of Fortune 500 companies already use it. Andreessen Horowitz led the round, with Sequoia and DCVC participating.
OpenAI banned accounts tied to Russian and Iranian operations that used AI to adapt propaganda and support false identities. One campaign placed nearly 100 articles through seven fabricated journalists.
The collection claims nearly 400 AI-generated results across many fields, but OpenAI says verification varies. About 300 manuscripts include Lean formalizations, leaving mathematicians a large review burden.
Postman says production scale made context selection and tool sprawl harder than prompt design. Its Bedrock-based agent requires approval before changing state and can redact personal data.
Data from more than 700 firms linked agent adoption to 30% more code and 23% more pull requests, while issue completion did not rise significantly. Average review time increased 49%.
AI agents produced much of the implementation, but GitHub migrated incrementally through 128 pull requests with tests and human review. One measured startup workflow fell from 5.25 seconds to 292 milliseconds.
Workers at HarperCollins, Simon & Schuster and Hachette told WIRED that AI now helps produce pitches, cover art and publicity copy. Employees say adoption often lacks public disclosure or author consent.
Modern models are trained to reject harmful requests, but probabilistic filters can be bypassed and can also block legitimate work. MIT Technology Review argues that providers quietly decide where that boundary sits.
Snyk Assist can answer product questions, inspect vulnerabilities and open support cases while honoring each user’s permissions. The company tested it internally for about a year before its September rollout.
The institute’s new scheduler allocates time through project budgets and preemptible slices instead of self-selected priority labels. It targets a queue with two to three times more GPU demand than capacity.
ElevenAgents can now classify a caller as human or AI during the first seconds of a call, then route it under business-defined rules. Initial access is limited to supported enterprise customers.
A 144-run test found that retaining more browsing history and deleting screenshots in batches made Asana’s workflow 76 times cheaper and five times faster. The figures come from a controlled demo task.
The open-weight model accepts four media types in one request and returns constrained choices rather than prose. Cloudflare also lowered Clef-flash pricing and says it doubled Clef inference speed.
Baseten customers can deploy lightweight probes that flag internal patterns before a second model reviews suspicious steps. Goodfire reports lower monitoring cost, but the results come from its own tests.