Kenya insurtechs Sh8.5bn, EU AI Act labels, SQLite fake CVEs

Cloudflare introduced @cloudflare/computer, an agent runtime it argues gives agents a full computer rather than just a container. The runtime dynamically moves work between fast, efficient V8 isolates and full Linux containers, so an agent gets whatever execution environment a task actually needs without the developer pinning it down upfront. That matters because agents today do not just run one function; they manage state, spin up tools and interact with the wider platform, which a plain container lifecycle handles badly. For developers building agentic applications on Workers, it reduces the friction of managing separate environments and hands the orchestration decision to the runtime.

AWS has trimmed prices for GPT models offered through Amazon Bedrock, the latest round of its weekly platform updates, which also added managed CloudWatch collectors for Prometheus metrics. The Bedrock price cut lowers the per-token cost of calling OpenAI models through the AWS gateway, which matters for teams that have standardised on Bedrock as their single LLM front end. It is also a signal about margins in model reselling: as frontier models commoditise, platforms pass savings through to win inference volume. The practical consequence for builders is to re-check Bedrock pricing before locking in a per-token budget, since the cheapest route to a given model shifts frequently.

Superblocks, a 50-person vibe-coding startup that has raised $60 million, signed a multiyear joint marketing deal with AWS that lets enterprises run its tool inside their own private clouds. Apps built that way spin up Amazon Aurora databases and integrate with Bedrock inside the customer's account, so data is not sent to external model providers or throwaway databases, and the apps fall under IT's management and security rather than living as rogue applications. AWS does not yet have its own business-user vibe-coding agent, which is why it is backing a partner. For builders, the signal is that enterprise vibe coding is moving toward regulated, private-cloud deployments where the app layer stays inside governance boundaries.

Meta says its Generative Ads Recommendation Model (GEM), the foundation model behind Instagram and Facebook ad recommendations, now trains at LLM scale on several thousand of its newest GPUs. Engineers doubled end-to-end training efficiency to 20 to 25 percent model FLOPs utilisation while scaling training fourfold over 12 months, by co-designing kernels, low-precision training, parallelism, networking and memory together. Why it matters beyond Meta: recommendation workloads do not fit generic LLM training stacks, so the team built custom kernels and topology-aware parallelism rather than waiting for a tool to arrive. The practical lesson for teams scaling ML systems is that gains come from co-designing across the whole stack, not swapping in a bigger GPU.

The startup behind Design Arena, called Intelligence, raised a $7.9 million seed round led by Index Ventures, with Conviction, A-star and Valkyrie joining. Design Arena is a tool used by 5.3 million people to rank AI outputs in head-to-head A/B matchups across websites, images and other visual formats. Frontier model labs pay for that river of human judgment because it is the missing bottleneck in making design and creative models produce genuinely good results. The round is a reminder that taste and human preference data are becoming a priced input to AI development, and that companies supplying scalable evaluation are finding a willing market among the labs.

Audi CEO Gernot Döllner told The Drive that the new Q9 three-row SUV nearly shipped with flush electronic door handles. A final test drive during validation changed that: the handles were innovative but worse for customers, so engineers re-tooled the car about 45 weeks before the media reveal, a change that normally takes two years. It is a quietly practical decision. Electronic handles are a recurring reliability and emergency-access headache, and regulators are already pressing on egress in collisions. The consequence for buyers is that the Q9 keeps handles that work when you need them, and it signals that even premium automakers are pulling back from form-over-function door hardware as ownership reliability reasserts itself.