DoorDash Launches "Ask DoorDash" AI Chatbot — Ordering by Voice and Image
On June 11, DoorDash announced "Ask DoorDash," an AI chatbot built into its app. Users launch the chatbot from the "Ask" button in the search bar, submit a request by voice, text, or image, and the bot automatically assembles restaurant or grocery orders.
Type "I'm craving something spicy" and it suggests matching menu items; snap a photo of your fridge and it adds the missing ingredients to your cart. Text responses come with direct add-to-cart buttons — a design that dramatically reduces the friction between search and purchase.
According to TechCrunch, the company is also positioning the technology as a new revenue stream for businesses. As a real-world deployment of AI in food delivery, its impact on consumer products is drawing attention.
Why it matters: Commercial AI chatbot deployments have mostly centered on customer support, which makes this a notably large case of one embedded directly into actual purchasing behavior. Its support for voice and image input also makes it a forerunner of consumer-facing multimodal AI implementations.
Workers Spend 6+ Hours a Week on AI "Botsitting" — Business Insider Survey
A June 11 Business Insider report picked up 108 points on Hacker News and drew a strong reaction. According to the survey, many workers whose employers have rolled out AI tools spend six or more hours a week reviewing and correcting AI output — a practice known as "botsitting."
The criticism: AI is supposed to boost productivity, but in practice it is driving up a hidden labor cost for humans. The verification work needed to guarantee the accuracy of AI-generated content is, in turn, adding to workers' burden and frustration.
The story also echoes the "AI and the Productivity Paradox" discussion reported by the Financial Times. While AI investment surges, the data on actual productivity gains remains less clear than expected — a fundamental question now corroborated by what workers are experiencing on the ground.
Why it matters: In debates over the ROI of AI adoption, human oversight costs are an easily overlooked factor. As the concept of "botsitting" gains wider recognition, decisions on adopting AI will increasingly demand comprehensive evaluations that include labor costs on the human side.
NTT to Raise $1 Billion for US Data Centers, While Crusoe Stumbles on Its Wyoming Project
The AI infrastructure investment push continues. According to Bloomberg, NTT Global Data Centers is raising at least $1 billion, with Citigroup's assistance, for data center development in the United States. The company may scale up depending on demand and is also reportedly considering setting up a credit facility.
Meanwhile, data center developer Crusoe has pulled out of the massive AI campus it had planned in Wyoming (which was set to consume as much power as the city of Denver). It was officially announced as a "pause," but reports say the company actually failed to secure customers including Google, and contract negotiations fell through over cost and schedule concerns.
It is a sign that while massive investment keeps flowing into AI data centers, challenges around customer acquisition and profitability are beginning to surface.
Why it matters: The rush to invest in AI infrastructure continues, but individual projects face uncertainty over profitability and customer acquisition. NTT's expansion signals Japanese companies entering the AI infrastructure market, while Crusoe's setback highlights that sheer scale does not necessarily mean success.
DiffusionGemma Tested in Real-World Workloads — Divergence from Benchmark Demos Flagged
Real-world workload test results for DiffusionGemma (the diffusion-based fast text generation model recently announced by Google DeepMind) have been reported in Reddit's LocalLLaMA community, revealing behavior that differs from the benchmark demos.
According to the user who ran the tests, the scaling behavior of H100s and A100s differs significantly from ordinary Transformer inference, and the gap between the GPUs widened beyond expectations as the number of concurrent requests increased. While the model shows remarkable speed on tidy workloads with short outputs, efficiency was confirmed to drop sharply in "realistic" environments — long-form outputs, uneven request lengths, streaming, and mixed multi-user loads.
GPU utilization patterns were also completely different from conventional token-by-token decoding, exhibiting new characteristics that do not fit the classic bottleneck structure.
Why it matters: The gap between benchmark environments and production environments is a perennial challenge when deciding whether to adopt a new architecture. As community verification progresses, this is becoming a key data point on how far innovative approaches like DiffusionGemma can maintain their edge in real-world settings.
Lovable's AI-Generated Site Wasn't "100%" Accessible After All
After AI coding tool Lovable claimed it had "built a 100% accessible site," accessibility specialists Axess Lab ran a verification and reported that multiple issues actually remained.
It is a lesson that when it comes to the quality of AI-generated code — especially specialized, standards-based requirements like accessibility — expert human review remains indispensable.
Why it matters: As AI code generation becomes widespread, AI's limits in the realm of quality assurance — particularly requirements tied to laws and standards — have been thrown into sharp relief once again. The recognition that "built by AI" is not, in itself, a guarantee of quality is spreading through the developer community.