Apple has tested DRAM memory from ChangXin Memory Technologies, or CXMT, across multiple iPhone and MacBook lines, and has discussed using the components in devices sold in China. The WSJ reports that this is still an early testing and discussion stage, not a confirmed commercial rollout, and Apple has not confirmed the report.

CXMT is emerging as an option just as AI-server demand pulls memory supply away from PCs and smartphones. Citing Counterpoint, the Journal says CXMT held 7% of global DRAM revenue in the second quarter, while PC shipments fell 4% and smartphone shipments fell 11% year over year; CXMT's capacity is reportedly full for 2026. If accurate, the story puts Apple at the intersection of memory shortages, export controls, and market-specific product decisions.

According to TechCrunch, the NextSlide team has joined OpenAI. The startup turns prompts, documents, and research into editable presentations; financial terms were not disclosed, and the transaction took place earlier in 2026.

The important signal is product surface. OpenAI is moving beyond models and APIs into concrete workflows such as visual communication and knowledge work. Once a presentation can start from a prompt, a document, and research instead of a blank page, the competitive question becomes how much complete work AI can handle around the model.

The Verge reports that 43% of surveyed US teachers in grades 6 through 12 said they regularly used AI detectors during 2024 and 2025. Tools such as GPTZero, Pangram, and Turnitin try to infer whether a text was written by AI from wording, rhythm, and structure, but the article points to studies and lawsuits showing how false positives can affect non-native English speakers and neurodivergent writers.

Schools including Yale, Johns Hopkins, Vanderbilt, Georgetown, and MIT have restricted or disabled AI detectors, encouraging reflection-based assignments, in-person assessment, and disclosure of AI use instead. When “sounds like AI” becomes an accusation, AI changes not only how we write but how we decide whether to trust a piece of writing. The next market may sell not only content generation, but proof of human authorship.

Amazon data center could become one of the largest sources of emissions in the US — TechCrunch cites a New York Times estimate that Amazon's planned data center in Pecos County, Texas, could be permitted to emit up to 33 million tons of carbon dioxide a year; Amazon has offered its own response on power costs and sustainability. AI infrastructure is now a local question of energy, climate, permitting, and public infrastructure.

Cloudflare launches Kitesurf, a web runtime for AI agents — Cloudflare introduced Kitesurf as a beta runtime designed for agents interacting with the web, extending the push toward dedicated web infrastructure for agent execution. For agents, the web interface must handle authentication, page actions, and environmental limits, making it an infrastructure layer that shapes how reliably agents can work.

Apple opens a path for Qwen on Macs in China — Reuters reports that Apple says eligible Mac users in mainland China can connect Siri and Writing Tools to Alibaba's Qwen. This extends the Apple-Qwen arrangement across Apple's China ecosystem and shows Apple Intelligence taking different product forms in different markets.

Google Threat Intelligence changes how it names hacker groups — Google's new convention combines a first name with a second word suggesting national origin, making activity clusters easier to identify and discuss. As Google Threat Intelligence tracks more than 5,000 groups, naming becomes part of incident response and pattern recognition, not just a way to make threat reports memorable.

Google has two AI stories: DeepMind under pressure, GCP benefiting from AI infrastructure — SemiAnalysis argues that GCP growth and compute demand from outside AI labs may be giving Google Cloud a commercial advantage while DeepMind loses momentum at the frontier; SemiAnalysis cites GCP revenue growth above 100% year over year. Google has also moved Demis Hassabis into Chair and Alphabet Chief Scientist roles, with Koray Kavukcuoglu taking day-to-day responsibility at DeepMind, suggesting that lab performance and infrastructure monetization can move in opposite directions.

ChipAgents raises more money to build agents for chip design — ChipAgents says its Series A2 expansion brings total funding to $134 million, recurring revenue grew sixfold in the first half of the year, and its platform has reached more than 120 semiconductor companies. The company describes agents supporting design, verification, and debugging, showing agents entering technical workflows with concrete bottlenecks rather than sitting as a chat layer outside the work.

Séb Krier: AI is still advancing, but public AI discourse runs on leaky abstractions — Séb Krier argues that AI can keep improving and creating value even in a bubble-like market, while the same progress can lead one group to see recursive self-improvement and another to see diminishing returns and incremental gains. The useful reminder is to examine the assumptions and uncertainty behind an AI narrative, not only argue within the narrative itself.

a16z: Kimi K3 hype is translating into real users — a16z cites Sensor Tower data showing Kimi app downloads nearly quintupled and daily active users rose about 40% after the attention around K3. That suggests the hype produced real consumer usage, but it does not yet show long-term retention or regular workflow adoption.

With vet costs rising 60% over the past decade, pets and pet parents both deserve a little relief.

Nibbles is the first credit card that comes with pet insurance and 3x rewards at pet stores, pet services, and the vet. Plus, use the card anywhere else Mastercard is accepted for 1x rewards. Protect your pet and your wallet in one move. All those rewards mean more treats too!

Reward Your Pet With The Nibbles Pet Rewards Credit Card!

*Disclaimer: Nibbles is not a bank. The Nibbles Card is issued by Lead Bank. Other fees and T&Cs apply. Rewards & pet insurance offered by Nibbles, not Lead Bank. Base pet insurance policy paid for one eligible pet by Nibbles. See nibbles website for details. Nibbles does not operate in CT, NY, KY, MN, WA, CO, HI, MA, or ME.

A coding agent can move quickly, but that speed can create product drift: one prototype, another real interface. My default rule for user-facing changes is to lock the interaction in a prototype before letting the agent implement it.

The workflow can live in a repo's AGENTS.md or CLAUDE.md as the default interaction spec. In 30 minutes, I would turn that principle into a default rule:

After each feature, I ask four questions: Was the prototype updated? Were all affected states covered? Was approval received? Does production match the prototype? If the final answer is no, fix the drift; if the behavior must change, update the prototype first.

The article argues that rollout and access numbers can overstate an enterprise's real AI capability: a small group of power users can drive most usage while the majority barely changes its workflow. The author's 5-10-20-70 distribution and 10% of users consuming 90% of tokens are attributed observations; McKinsey and MIT NANDA provide separate context on the gap between pilots, adoption, and P&L impact. The useful shift is from “How many people have the tool?” to “How much work has actually changed?”

McKinsey argues that enterprise infrastructure must change as agents orchestrate, govern, and scale work, creating new demands across the data layer, integrations, observability, and control. The lens places launches such as Kitesurf in a wider context: the advantage will come from connecting agents to existing systems reliably, not only from model access.

The survey separates risks from agents operating across the web, software, and physical environments from traditional model safety: agents can call tools, plan across multiple steps, and create unexpected side effects. It argues for evaluating tools, permissions, environments, monitoring, and recovery, not just the model benchmark.