Josh Woodward
Josh Woodward introduced Dreambeans, a new Google Labs experiment. A fixed number of “beans” appear each morning. They point users toward activities in the real world. The emphasis is on doing things they care about with real people. Josh said it was developing a devoted following and invited people to try it.
Nikunj Kothari
Nikunj Kothari shared a view that every small business owner could have bespoke software. The quoted example focused particularly on the trades. Differentiation would emerge in the last mile between software and the actual business. For customers, that means a distinctive experience. For employees, it means tools shaped around how their own business works.
Thariq
Thariq is redesigning Claude Code’s plan mode based on feedback. Some users do their own planning and do not need a separate mode. Others like a dedicated state for thinking and brainstorming. His plan is to make plan mode a built-in mod and let mods add modes or override Shift+Tab. Users could customize the planning prompt, create and share modes, or bind the shortcut to something else.
Peter Steinberger
Peter Steinberger argued that an agent stops too early when simply told to clean up. He suggested giving it an ambitious, verifiable goal. His example was to remove the least useful 20% of tests while keeping code coverage within 2%. In a separate dependency investigation, Daybreak found eight more longstanding memory leaks. Together, the examples emphasize explicit quality constraints and care for the open-source code a project depends on.
Guillermo Rauch
Guillermo Rauch shared two months of usage changes on Vercel AI Gateway. Anthropic still led spending, but its share fell from 69% to 40%, while OpenAI rose from 10% to 24%. He highlighted strong growth in GPT-6 Astra and GPT 5.6 Sol, with OpenAI now leading token volume. Kimi K3 and DeepSeek captured roughly half of Anthropic’s lost spending share. Opus 5.5 reached 10% of spending in two days, while OpenAI accounted for 62% of image generations. These figures describe traffic on Vercel AI Gateway.
Swyx
Swyx revisited his “Scaling without Slop” content strategy from the beginning of the year. He believes it is starting to work. The first 100,000 YouTube subscribers took three years; the next 100,000 took only 1.2 months. He reported similar momentum in AEO, SEO, and subscriber growth. He is preparing the next phase of Latent Space, AINews, and other new-media projects.
Amjad Masad
Amjad Masad announced that Muse can now create apps on Replit. His post links to a demonstration.
Official blog · Claude for Small Business launches new workflows, integrations, and training programs
Claude for Small Business added 27 integrations, bringing its workflow count to 43, with tools including Shopify, Salesforce, TikTok, Xero, Gusto, Square, Stripe, and Zapier. It has been installed more than 900,000 times since launching in May. After hearing from over 1,000 owners in 10 cities during its spring tour, Anthropic expanded beyond back-office operations into lead generation, inbound inquiries, and proposals. Roughly a third of those owners had requested help growing their businesses.
Example workflows include a morning view of cash, sales, and overdue invoices; handling and logging leads after hours; turning a voice memo into a priced, branded proposal; and preparing month-end accounting materials. Customer Garrett French said it reads the CRM at 6 a.m., prioritizes tasks, and emails the results: “We call it the daily briefing.” Free workshops return to 10 US cities in the fall. More than 150 trained organizations plan over 750 community workshops, and 14 integration partners will run webinars.
The MAD Podcast with Matt Turck | Who Feeds the GPUs? Inside AI's Hidden $30B Layer | Renen Hallak, VAST Data
Core takeaway: Competition in AI infrastructure is expanding beyond buying GPUs to managing data, models, memory, and permissions together.
VAST Data founder and CEO Renen Hallak worked in storage R&D at EMC before starting VAST in 2016. He describes its software infrastructure layer as an operating system for the AI era. One AI-cloud customer initially expected to need 500 PB over three years, then requested another 2 EB, illustrating the change in data demand.
First, DASE changes scaling through shared access. In traditional sharded systems, communication becomes more complex as nodes multiply. VAST places storage across a fast network so nodes share access to data and metadata. Hallak described a customer cluster with multiple exabytes, tens of thousands of nodes, and tens of terabytes per second of throughput.
Second, inference demands more coordination than training. Training primarily needs fast access to large datasets. Online inference also requires low latency, continuous availability, and routing requests to GPUs that already hold the appropriate model and cached context. Long-term agent memory, RAG, and fine-tuning records need granular permissions to prevent one user’s knowledge leaking to another.
Third, DataEnclave aims to protect both enterprise data and model weights. Inference can run on enterprise premises or in its VPC, with hardware capabilities from vendors including NVIDIA protecting weights across networks, memory, and GPUs. Hallak emphasized control: “Doesn’t necessarily need to be physical equipment within a building that belongs to you.”
Fourth, growth should follow customers’ actual problems being solved. VAST employees can move customer status from green to yellow or red, but recovery requires the customer to confirm resolution. Hallak advocates repeatedly finding and removing bottlenecks so infrastructure never limits customers’ AI progress. The company supports existing interfaces such as S3, SQL, and Kafka, allowing customers to start with new AI apps and expand gradually. Hallak noted that standard interfaces also let customers leave, making genuine satisfaction essential to retention.