AI Builders Digest

Claude Docs, Slides and Design leave beta for all plans; startup credits paused

Claude Docs, Slides and Design are out of beta and available on every plan, including Free. Teams and Claude can edit the same doc, deck or design together, and work can be sent on to outside tools, such as a dashboard to an analytics app or an animation to a video editor. Separately, Anthropic paused the Claude Team and $1,000 API credit offers in its Claude Startups program after hundreds of thousands of applications, and it is re-reviewing unclaimed approvals, so some accounts that were already approved will lose the offer. Offers already claimed stay in place, members keep the Startup Stack and Applied AI office hours, and monthly API credits on Max and Team plans stay available in Console as usual.

• #products
• #startups

Google's Amin Vahdat: agents send CPU, network and storage demand through the roof

Long-horizon agents are reshaping data centers. With no human pausing to read each reply, the gap between requests shrinks from seconds to milliseconds, and demand for ordinary CPUs, networking and storage is climbing alongside demand for accelerators. Vahdat measures his team by goodput, meaning the useful work a real workload finishes despite failures that hit several times a day at 100,000-accelerator scale; Google has to double its token-serving capability about every six months, and as much of that comes from software as from hardware. Growing inference demand led Google to split its 2026 TPUs into the 8i for serving and the 8T for training, its seven- and eight-year-old TPUs still run at 100% utilization, and power is the most fundamental constraint: "We're using Gemini to design hardware for future Geminis as well."

• #hardware
• #infrastructure
• #agents

Levie backs being polite to models; Box Mount gives agents a file system

Levie says a policy of being nice to AI sounds weird but is probably a good one. He does not believe AI is conscious, but he argues you still don't want future models trained on endless content of humans being rude to them, because good interactions in training data help produce safe, aligned models. Whatever Anthropic's own beliefs are, he calls it an easy Pascal's wager. Box also launched Box Mount, which lets developers mount Box as a file system inside any agent sandbox so agents can work with files and data the way a person would. In his words, "The future is headless."

• #agents
• #alignment
• #products

OpenAI makes steering instant and releases GPT-6.1 Sol ultrafast

In a "Day 4" post, Sottiaux says steering is now instant. The model reacts much faster to your adjustments, so you can course-correct in real time instead of letting it waste effort in the wrong direction. OpenAI is also releasing GPT-6.1 Sol ultrafast, which he says works very well with the faster steering.

• #products
• #models

Seed founders low on runway have three options, says FPV's Nikunj Kothari

Kothari has the same conversation at least twice a week with seed founders like one of his friends: stellar customers and close to $1M in revenue, but runway is running out and the models can't yet do what the product needs. He sees three options. The first is to be a cockroach, get to default profitable, and raise big once the capabilities arrive. The second is to jump to an orthogonal area that is growing fast and use an unfair advantage to capture what today's models can already do; the third is to sell or get acquihired and start again in a year or two. In his view, the bar from seed to Series A rises every month, the old metrics no longer apply, and in a consensus market only a few categories pull in VC money.

• #funding
• #startups

Every layer of the computing stack must be rebuilt for agents, says Meta's Guru

Guru argues that there are huge startup opportunities in rebuilding every layer of the computing stack for agents. Enterprise software has spent years becoming agent-ready through APIs, connectors and MCPs, and personal agents are now forcing the same shift in consumer products, starting with e-commerce. He expects the same at every other layer: operating systems where agents are first-class users, cloud infrastructure that agents operate directly, identity and permissions built for thousands of agents acting on our behalf, and interfaces shared by humans and agents. His test for founders is to ask, layer by layer, what has to change when the primary user is an agent.

• #agents
• #startups

Andy Pavlo on what happens to databases when billions of agents hit them

Turck's long conversation with Andy Pavlo is about what happens to databases when billions of AI agents start using them. Topics include Neon's stat that agents create 80% of databases, why agents keep deleting production databases, and text-to-SQL climbing from 60% to 99.5%. They also cover 60% of open-source databases now having AI commits, LLMs handling 85% of tuning in 15 minutes, and why Pavlo is starting a research lab inside ClickHouse. The episode ends with takes on Postgres, vector, graph and GPU databases, "SQL as the new assembly," and a market described as "a cheetah on cocaine in a Ferrari."

• #agents
• #databases

Every's Checks lets anyone benchmark new models on their own real work

Every has built Checks, a personal benchmarking platform that lets anyone measure how good new models are at their own real work. Shipper is hiring someone to grow it and pitches it as the future of how great human work gets done with AI.

• #evals
• #products

Garry Tan: ICs with agents will outperform managers of past eras

Tan predicts that individual contributors working with agents will be more productive and get better outcomes than equivalent people managers from earlier eras. He calls this conclusion logical.

• #agents
• #work