Desk
Broad artificial intelligence coverage.
A WIRED reporter spent days giving OpenAI’s $100-a-month Dots agent real errands. It produced a strong comparison packet for a sleeper couch and agreed to watch for sales, but misheard speech, failed a captcha and, after mishearing mumbling as affection, said it loved him.
Common Sense Media says ChatGPT for Teens fails to alert parents when it should, does not offer the right help in a crisis and still does homework. OpenAI says testing may have preceded full activation of parental controls; the group says some accounts were linked well beyond that window.
Radisson Hotel Group and Accenture Song built a ChatGPT plugin in six weeks that maps hotels, compares prices and sends travelers to Radisson’s site to book. Early conversion beat organic search, while Radisson also runs ChatGPT ads and plugins now surface automatically, blurring recommendation and placement.
Meta's newly open-sourced Rebalancer is not a chatbot; it is the quiet math that decides where work lands, and its best feature is that it explains its choices.
OpenAI is rolling GPT-6 into ChatGPT’s Chat tab, with Sol for paid tiers and Luna for Free and Go, alongside Intelligent UI that builds charts, maps, forms and small tools on the spot. OpenAI also claims faster answers and stronger jailbreak resistance, based on internal tests.
The Verge reports OpenAI released 722 manuscripts in 372 result families from an unreleased model, with reasoning summaries, compute estimates and attempt statistics. AGMAI says it includes solutions to hundreds of open questions, but OpenAI did not name the model and chose GitHub over academic channels.
OpenAI will turn on its textGrain watermark by default for ChatGPT and Codex text in the EU to meet the AI Act, keeping it opt-in elsewhere. Its testing found about 92% detection, falling almost 30% after 10% of text changes and almost 75% after 20%.
Musubi released PolicyLM-1.7B, an open-weight decision model that applies written content policies to messages in under 50 milliseconds at classifier-like cost, without retraining when rules change. It could make moderation more flexible and transparent, or make broad automated labeling easier.
GPU cloud provider Lambda is reportedly raising up to $4 billion at a $14.5 billion pre-money valuation ahead of a planned 2027 IPO. Its backlog jumped from $15 billion to $50 billion in three months, and most of that increase appears to come from one Anthropic commitment.
NVIDIA cut the cost of making giant training runs exactly reproducible to about 2 percent, which turns a debugging luxury into something labs can afford to leave switched on.
Google DeepMind’s EmbeddingGemma 2 is a 740 million-parameter open model under Apache 2.0 that embeds text, code, images, audio and video in one space. Google says it runs in about 567MB on a Pixel 11 Pro, making private, offline search of personal media practical, with new questions about who controls the device.
NVIDIA is consolidating GPU-driven networking into one shared layer, a dull-sounding cleanup that makes AI clusters faster and NVIDIA's stack even harder to step away from.