🧠 OpenAI partners with Cerebras ↗
OpenAI says it’s adding 750MW of ultra low-latency AI compute via Cerebras - basically a bid to make responses feel more “real time” instead of “go make tea, come back later”.
The pitch stays simple: faster inference loops (ask - think - answer) mean people stick around longer and run heavier workloads… which, yeah, tracks. The capacity rolls out in phases, with more coming online through 2028.
🔍 Google taps emails and YouTube history in push for personalised AI ↗
Google’s Gemini is getting more “you-coded” by pulling from stuff like Gmail, Search, and YouTube history - if you opt in, anyway (and it’s off by default). It’s the chatbot version of your phone knowing you’re hungry before you do.
The idea is that Gemini can infer preferences and context so replies feel less generic and more like it genuinely remembers your life. Handy - and also the kind of thing that makes you pause mid-scroll and go, “hold on, it knows a lot.”
🧩 Sales of a powerful Nvidia AI chip to China gets the greenlight, with conditions ↗
The U.S. approved sales of Nvidia’s H200 AI chips to China - but with a bunch of guardrails attached (priority for U.S. supply, extra review steps, limits around where chips can end up).
Lawmakers are split - some see it as a national security risk, others as a competitiveness play (because if you don’t sell, someone else will… or China just builds faster). It’s one of those policy moves that’s both “strategic” and muddled, somehow.
🛡️ Belgian cybersecurity startup Aikido hits unicorn status with new funding round ↗
Aikido Security hit a $1B valuation after raising $60M, leaning hard into the “security guardrails for developers” angle as AI-assisted coding keeps… multiplying code, bugs, and unforeseen surprises.
They’re pitching it as developer-first security that automatically flags risks without turning every sprint into a compliance nightmare. The slightly wild part is how fast they’ve grown - the market is clearly paying for “please stop my AI code from shipping vulnerabilities”.
👩💻 Anthropic expands Claude Code beyond developer tasks with Cowork ↗
Anthropic is pushing Claude Code into a broader “work with me” vibe via Cowork - less just coding help, more like a collaborative assistant that can sit inside workflows.
It’s part of the bigger trend where code assistants stop being a fancy autocomplete and start acting like semi-autonomous teammates… which is exciting, or alarming, or both - depending on how your last PR review went.
🧯 depthfirst Announces $40M Series A to Secure the World's Software ↗
depthfirst raised $40M to build an AI-native approach to software security - the kind that’s meant to catch real vulnerabilities with fewer false alarms (the eternal dream, in truth).
They’re positioning as an “applied AI lab” for security, with funding aimed at scaling research and product. If AI is helping write more code faster, the counter-move is basically AI helping defend it faster - a snake eating a snake, in a trench coat.
FAQ
What does OpenAI partnering with Cerebras mean for response speed?
OpenAI says the Cerebras partnership adds 750MW of ultra low-latency AI compute, aimed at making responses feel more real time. The practical takeaway is quicker inference loops - ask, think, answer - so interactions land with more snap. It can also make heavier workloads feel smoother, without the long, awkward pauses. The capacity is planned to roll out in phases through 2028.
When will the 750MW of Cerebras compute actually be available?
The announcement frames it as a phased rollout, with more capacity coming online through 2028 rather than all at once. With phased ramps, early gains tend to surface first, then widen as infrastructure expands. If you’re tracking impact, the clearest signals usually show up as lower latency and steadier performance under load. The headline is a multi-year buildout.
How does Google Gemini use Gmail, Search, and YouTube history for personalization?
Google’s plan is to make Gemini more personalized by pulling from sources like Gmail, Search, and YouTube history - but only if you opt in. It’s described as off by default, which matters for privacy and expectation-setting. The goal is for Gemini to infer your preferences and context so replies feel less generic. Many users will weigh convenience against how much “memory” they feel comfortable enabling.
Is Gemini personalization turned on automatically, and what’s the privacy trade-off?
In this rollout, Google positions the feature as opt-in and off by default. That means you’d choose to let Gemini use signals from products like Gmail or YouTube history to tailor answers. The trade-off is straightforward: more context can improve relevance, but it also increases how much personal data is involved in the experience. A common approach is enabling it selectively and reviewing settings regularly.
What conditions did the U.S. put on Nvidia H200 chip sales to China?
The approval described here comes with guardrails, including priority for U.S. supply, extra review steps, and limits around where chips can ultimately end up. The debate is also spelled out: some lawmakers see national security risk, while others argue competitiveness - if U.S. companies don’t sell, alternative suppliers or domestic substitutes may accelerate. It’s a policy move with strategic intent and complicated trade-offs.
Why are AI security startups like Aikido and depthfirst raising big rounds right now?
The theme is that AI-assisted coding is multiplying code output - and with it, bugs and unforeseen vulnerabilities - so demand for developer-friendly guardrails is rising. Aikido is positioned as “developer-first” security that flags risks without turning every sprint into compliance friction, while depthfirst pitches an AI-native approach aimed at catching real issues with fewer false alarms. As code volume grows, “secure it faster” becomes a budget line item, not a nice-to-have.