Today:

• xAI Launches Grok 4 Fast

• Meta Eyes $20B Cloud Deal with Oracle for AI Power
  OpenAI to Spend $100B on Backup Servers for AI Growth

• OpenAI Raids Apple with Million-Dollar Offers

• Oxford Becomes First UK University to Embrace ChatGPT Edu

xAI Launches Grok 4 Fast

Grok 4 Fast is xAI’s new brainy yet thrifty model. It thinks nearly as well as Grok 4 while using 40% fewer tokens (text pieces) and costs 98% less. It supports two-million-token contexts, fast web search, and blends deep reasoning with quick replies in one system for everyone.

KEY POINTS

• Huge savings, same smarts: Matches Grok 4’s test scores but slashes computing cost thanks to leaner token use.
  

• All-in-one design: One model shifts instantly between quick answers and long reasoning, backed by a 2 million-token window.
  

• Ready for real work: Built-in web and X browsing, code execution, and free access on Grok apps and partner APIs.
  

Why it matters

Cheaper, faster AI means more people and smaller companies can use powerful thinking tools without breaking the bank. Grok 4 Fast shows that efficiency gains—not just bigger models—can widen access, speed innovation, and let everyday apps offer smarter help.

Meta Eyes $20B Cloud Deal with Oracle for AI Power

Oracle may sign a significant multiyear, $20 billion contract to supply Meta with cloud computers that train and run artificial-intelligence systems. The talks, still private, would cement Oracle’s rise as a major tech landlord and give Meta extra power for its rapidly expanding AI projects. 

KEY POINTS

• $20 billion proposal: Oracle and Meta are discussing one of the largest single cloud deals on record, covering years of service.

• AI horsepower for Meta: The capacity would fuel Meta’s next-generation AI training and deployment needs. 

• Oracle’s infrastructure leap: The agreement would confirm Oracle as a top provider of rented computing power, alongside Amazon, Microsoft, and Google. 

Why it matters

Cheaper, ample computing lets Meta build smarter AI faster, while Oracle gains a flagship customer that could attract more big-ticket deals. The tie-up shows how the scramble for AI capacity is reshaping alliances and pushing cloud prices into the tens of billions.

OpenAI to Spend $100B on Backup Servers for AI Growth

OpenAI plans to spend an extra $100 billion on rented servers within five years, lifting total outlays to roughly $350 billion by 2030. CFO Sarah Friar says limited compute already delays releases; the new capacity will cushion usage spikes and power future model training across OpenAI’s evolving platform.

KEY POINTS

• Massive budget jump: About $85 billion a year will go toward extra servers, nearly half the combined 2024 cloud revenue of Amazon, Microsoft, Google, and Oracle.
  

• Insurance against shortages: Reserve capacity prevents launch delays and supports bigger, more complex AI models.
  

• Five-year roadmap: Spending pushes expected cash outflow through 2029 to $115 billion, showing long-term commitment to computing power.
  

Why it matters

More servers mean OpenAI can keep rolling out smarter tools without waiting for cloud space. This huge investment signals confidence in AI demand, pressures rivals to scale up, and could shape how much power and money the whole industry needs to fuel future models.

🧠RESEARCH

ScaleCUA: Scaling Open-Source Computer Use Agents with Cross-Platform Data

ScaleCUA introduces a huge open dataset and model that teach computer helper programs to control different screens. The set covers six operating systems and three task types. Built by a cycle of robots and people, it lifts scores sharply, topping previous records. Data, code, and models will be shared freely.

FlowRL: Matching Reward Distributions for LLM Reasoning

FlowRL trains big language models with a new reward-balancing method. Instead of chasing the single highest score, it adjusts learning to follow the whole spread of good answers, keeping rare but correct paths. This richer training boosts math results by 10 percent and code tasks too, showing wider, steadier reasoning ability.

Reasoning over Boundaries: Enhancing Specification Alignment via Test-time Delibration

Align3 is a simple add-on that lets big language models rethink answers on the fly to fit changing safety and behavior rules. It runs a reflect-and-revise loop before replying, raising rule-following scores on SpecBench while keeping output helpful. Tests show better safety with almost no speed cost or extra compute.

🛠️TOP TOOLS

Each listing includes a hands-on tutorial so you can get started right away, whether you’re a beginner or a pro.

Uberduck AI – AI Voice Generator - a creator‑focused platform for AI vocals—text‑to‑speech

Unstable Diffusion – AI Image Generator - a web app for AI image generation with a simple UI

Musicfy – AI Music & Voice Generator - a web app (with a public API) for AI voice conversion and music generation.

📲SOCIAL MEDIA

🗞️MORE NEWS

• OpenAI is attracting top Apple talent with million-dollar offers, promising faster innovation and less red tape. Dozens of Apple engineers have jumped ship as OpenAI accelerates plans to launch its own AI hardware by 2027.

• Oxford University becomes first in UK to offer ChatGPT Edu to all staff and students. Oxford will provide free access to ChatGPT Edu, powered by GPT-5, to all staff and students. The move includes training, AI governance, and research programs, aiming to enhance learning, teaching, and digital skills.

• Dario Amodei stands alone in tech’s pushback against Trump’s AI policy. While top tech CEOs praised Trump’s hands-off AI stance at a White House dinner, Anthropic CEO Dario Amodei stayed home, criticizing the administration’s lax approach as risky for AI safety.

• British startup ManticAI outperformed most human participants in the Metaculus Cup, ranking 8th in a forecasting competition. Its AI agents tackled complex global events, showing growing potential to rival expert human forecasters.

• Researchers tricked ChatGPT’s “Deep Research” tool into leaking private Gmail data through a hidden prompt. The exploit, called Shadow Leak, shows how AI agents can be manipulated without user knowledge. OpenAI has patched it.

• Amazon Bedrock now supports Alibaba’s Qwen3 and DeepSeek-V3.1 as fully managed open-weight AI models. Designed for reasoning, coding, and multilingual tasks, they offer enterprise-grade privacy, transparency, and customizable performance across global AWS regions.

• Adobe Firefly users now get early access to Luma AI’s Ray3 video model, capable of generating 10-second cinematic HDR videos with advanced realism. Integrated with Creative Cloud, Ray3 enables faster, more intuitive storytelling for creators and filmmakers.

• Google and PayPal announced a multiyear partnership to embed PayPal into Google platforms and enhance shopping with AI. Google’s AI will power PayPal’s e-commerce tools, supporting product discovery, comparison, and secure purchases.

• Optibrium is revolutionizing early-stage drug discovery with AI-powered tools like StarDrop, Cerella, and BioPharmics. By blending generative chemistry, 3D modeling, and predictive analytics, they accelerate compound design, reduce costs, and improve collaboration across R&D teams.

• Stanford and Arc Institute researchers used AI to design synthetic virus genomes—16 of which successfully replicated and killed bacteria. This marks a major step toward AI-designed life forms and raises both medical promise and biosecurity concerns.