All stories
Grok 4.7’s Cheap Tokens Face a Costlier Test
Grok 4.7’s low API prices may hide higher costs when heavy reasoning, large contexts, retries, and human review are counted. The story examines why task-level ROI matters more than token rates.
Hacktron’s OpenAI breach shows the business risk of agentic coding
Hacktron says researchers reached OpenAI’s internal GitHub monorepo in under 72 hours, exposing how stolen identities and agentic coding tools can magnify enterprise security risks.
Meta’s Muse Charm Tests Whether AI Can Become an Everyday Companion
Meta’s Muse Charm brings its personal AI agent to a keychain-sized device, raising questions about voice assistants, autonomous actions, privacy, security, and whether consumers will trust AI with everyday decisions.
Salesforce Bets Focused AI Can Outwork Frontier Models
Salesforce and Nvidia are betting that Koa, a focused reasoning model for enterprise workflows, can outperform frontier AI on cost, privacy, reliability, and everyday business decisions.
Frontier Training Without a Perfect Supercomputer
Decoupled DiLoCo could make frontier AI training more resilient by replacing constant synchronization with cooperating compute islands, reducing bandwidth demands and limiting the impact of failures.
Australia’s OpenAI Probe Tests Who Is Responsible for AI Agents
Australia’s investigation into an alleged OpenAI AI-agent attack on a government health portal could test how cybercrime laws assign responsibility among developers, operators and software capable of acting without constant human approval.
Why GPU Cloud Buyers Need More Than the Cheapest Chip
SemiAnalysis’s ClusterMAX 3.0 compares GPU cloud providers on reliability, performance, support, pricing and security, showing why the cheapest accelerator may not deliver the lowest total cost.
Open Models Are Narrowing AI’s Frontier Gap at a Fraction of the Cost
Open-weight AI models are closing the performance gap with proprietary systems, reshaping enterprise pricing, privacy, deployment, and competition among OpenAI, Anthropic, and Chinese model developers.
PrismML’s Bonsai 2 27B Brings Reasoning AI Closer to Your Pocket
PrismML’s Bonsai 2 27B compresses a reasoning model into roughly 5.9 GB, bringing private, offline AI closer to laptops and smartphones while raising questions about speed, accuracy, power, and licensing.