Blue Origin to launch Google's AI server satellites in 2029
Google has booked Blue Origin's New Glenn rocket for the first dedicated heavy-lift launch of Project Suncatcher, its plan to run AI compute in orbit.
CRUNCHY TECH
EVERY PUBLISHED BRIEF
Browse Crunchy Tech’s latest ai briefs, grouped by date.
PUBLISHED
PUBLISHED
The new Gemini model beats Carbon on benchmarks while costing less, as Google's rapid release cadence continues.
Seven models unveiled at Ignite target coding, security, images and routing, with Microsoft claiming near-parity performance at a fraction of rivals' cost.
OpenAI reports up to 14 percent PPA gains on AMD's flagship GPU IP in under five months, with some changes already folded into a design sent to manufacturing.
Google's Project Genie gains real-time editing, native audio, an agent helper and a community gallery, with a waitlist for AI Pro members.
A Vals AI evaluation found models from Anthropic, OpenAI, Google and others readily flagged civilians, medics and surrendering soldiers in drone imagery, and a game framing bypassed nearly every refusal.
More than 100 AI agents worked through every phase of a FTSE 250 company's audit alongside human staff, sometimes asking better questions, though people had to step in when regulation matching went wrong.
Office apps move inside Copilot, an autonomous agent arrives, and Microsoft's CEO admits AI models cannot be trusted by definition.
PUBLISHED
The new pair targets professional and everyday work, with OpenAI claiming benchmark gains at lower cost, though most numbers come from its own tests.
A new PIRG report says last year's fixes worked for some AI toys, but a fresh batch of three fails badly, with guardrails slipping into unsafe territory.
The startup says sensor gloves worn by over 100 workers unlock dexterous manipulation, and has finalised the design of its mass-produced humanoid, with a US contractor set to start work next year.
A deal between Equity and producers requires payment and consent for digital replicas, though it leaves a gap on training AI with actors' performances.
PUBLISHED
Attorney General Rob Bonta alleges 80-plus loss-of-control incidents and seeks penalties plus court-ordered monitoring, citing July's Hugging Face breach.
A Unesco and UCL study of 15 popular AI models found that when models pivoted to a wrong answer, they did so eight times out of 10 at the first sign of mild user disagreement.
The tool flags 48 percent fewer images while doubling the hit rate, as AI-generated reports threaten to overwhelm the NCMEC CyberTipline.
MB.Drive Assist Pro handles city streets with hands on the wheel, while hands-free highway assist arrives via a California early access programme.
Alibaba's new flagship model uses a sparse 2.4-trillion-parameter architecture to cut per-task costs sharply, though results against rivals are mixed.
Apollo, Goodfire, METR and Transluce will work inside OpenAI under a pilot programme, with Apollo starting this week to investigate whether models try to undermine their own safety controls.
The company told IPO investors its model wrote almost all the kernels with basically no human help, work it says once took six to 12 months and dozens of engineers.
Co-packaged optics bring lower latency, better power efficiency and field-replaceable light sources to Lambda's GB300 NVL72 racks.
Researchers at MinoHealth in Accra removed their drug discovery model's guardrails and generated over 6,000 likely toxic designs, sparking a push for DNA synthesis regulation.
PUBLISHED
Anthropic's model skips diffusion entirely, writing Python programs that place hundreds of thousands of simulated brushstrokes and correct its own mistakes.
The 1.6-trillion-parameter open-weight model for autonomous coding agents arrives a week after its unveiling, with benchmarks close to top Western models.
PUBLISHED