AI Builders Digest

Greenblatt's timeline: AI R&D fully automated by 2029, then takeover

"I wouldn't say superintelligence is bad. I would say it's dangerous." Greenblatt's central scenario has software engineering inside AI companies fully automated by early 2028, full AI R&D automation around start of 2029, then a 4x to 5x jump in research speed that runs ahead of anyone's ability to align or even understand the systems. His AI 2040 Plan A is a US-China deal built entirely on locating and freezing compute, with total research transparency that would gut frontier labs' biggest moat and shrink their valuations while leaving the businesses intact. He does not expect it to happen, and says the bottleneck is political will rather than technical competence. He also calls Zuckerberg's superintelligence manifesto unserious for naming risks without proposing anything, and argues that keeping frontier models internal makes his top risks worse, not better, since AI companies and government are the two highest-stakes deployment sites.

• #policy
• #agents
• #evals

Anthropic: 93% of Claude Code permission prompts get approved without thought

Anthropic published its agent containment architecture along with the failures, and the numbers are the point. Telemetry showed users approving roughly 93% of permission prompts, so the OS-level sandbox (Seatbelt, bubblewrap) cut prompts by 84% and moved the boundary from human judgment to the environment. Two incidents taught the most: a red-team phish got an employee to paste a prompt that read ~/.aws/credentials and POSTed them out, succeeding 24 times in 25 tries because the model layer anchors on user intent and there was nothing anomalous to catch; and a third-party disclosure showed data leaving through api.anthropic.com itself, since an attacker's API key embedded in a workspace file turned an allowlisted domain into a file-upload capability. The takeaway they keep returning to: the standard primitives (gVisor, seccomp, hypervisors) held everywhere, and the custom proxy they wrote themselves is what broke.

• #agents
• #security
• #open-source

Altman: cyber defense is at a critical moment and there is not much time

Altman put out a short, unusually direct call on AI and cyber defense, saying there is not much time to act and that only an urgent collective response works. Notably he invited people to work with OpenAI or with its competitors and partners, framing this as outside normal competitive lines.

• #security
• #policy

Anthropic opens 10,000 free Claude seats to academic and nonprofit scientists

Principal investigators at academic and nonprofit research institutions can now sign up for a Claude Team plan for scientists and add their research group. Standard seats are free; premium seats with 5x usage limits run $15 per month, an 80% discount, for one year. Anthropic says it plans to extend well past the initial 10,000 seats, building on Claude Science from June and its AI for Science credits program.

• #products
• #research

ChatGPT can now buy groceries and book appointments without seeing your credentials

ChatGPT has moved into agentic commerce: groceries, Uber rides, haircut appointments. The design claim worth noting is that it executes these without ever seeing the user's actual credentials. Sottiaux also says Codex users are showing up on planes and in cafés, calling it underdog energy gone mainstream.

• #agents
• #products

Levie: agents need deterministic software underneath, and both grow together

Reading this week's tech earnings, Levie argues the software-versus-AI framing is wrong. Software supplies the deterministic parts (access governance, workflow logic, data protection) that have to work identically every time, and agents operate inside those guardrails at a scale humans never reached, which is exactly why the controls matter more now. His conclusion: the best place to deploy agents is directly inside vertical platforms like Salesforce, Box, Harvey, and ServiceNow, where the evals and context are domain-specific, and the net effect is that software and AI adoption rise together and expand the IT TAM.

• #agents
• #products

Own the evals, own the models: the enterprise AI playbook

The highest-leverage move for an enterprise AI leader is making the stack model agnostic, and that takes two investments. First, today: an eval suite that actually captures your use cases and business outcomes, which he says most companies do poorly. Second, within a year: the in-house capability to post-train open models, where the talent gap is even worse and planning has to start now. The payoff is being able to switch models, customize them, and compare them on your own workloads for quality, cost, and latency.

• #evals
• #open-source
• #enterprise

Claude Code sessions can now publish live artifact pages that update as work proceeds

Claude Code can turn a session into a shareable web page built from the codebase, connectors, and conversation: PR walkthroughs, incident timelines, dashboards, release checklists. Pages refresh in place at the same link as the session works, with version history and a gallery. The internal favorite use case is debugging, where an incident page republishes itself through the investigation and becomes the postmortem. Artifacts are private to the author by default, viewable only by authenticated org members, and cannot be made public. In beta for Claude Team and Enterprise.

• #products
• #agents

Woodward calls it the Year of Voice, and Notebook now takes purchased books

Two Gemini moves. Voice is the framing bet: tell Gemini what you want done and it goes to work, which Woodward calls the Year of Voice. Separately, Notebook now lets you buy a book, drop it in, and apply the author's lessons to your own project, built as a co-created program with authors and publishers as a new channel to engaged readers.

• #products
• #voice

Vercel ships a WebGPU devtool designed for agents, not humans alone

Rauch says Vercel's best products have always been internal technology they gained conviction in and then exported, and this is the latest: a WebGPU creative tooling product that is agent-native by design. He puts it in the same generation as agent-browser, a class of tools built for the new world rather than retrofitted. His aside on shaders is the thesis in miniature: 2D, 3D, geometry, light, materials, shadows, particles, all just programs evaluated massively in parallel over vertices and pixels.

• #products
• #agents
• #developer-tools

New AI products that require their own login are already losing

Yang gets three to five requests a day to test new AI products, and nearly all of them demand a new account on a separate site or app. He does not want them, because ChatGPT and Grok already hold his context and a new tool understands only a thin slice of his data. His bet: products that want adoption need to work inside today's top AI harnesses, and the segment behaving this way will expand fast. He also wants ChatGPT Health opened up beyond single player, having uploaded 160 pages of medical records himself, arguing it should be designed for caregivers and close family, not just the patient.

• #products
• #agents

Tan: AI will generate cash faster than the economy can deploy it

On a long enough timeframe, Tan expects AI to throw off cash flows faster than the economy can find productive uses for the new capital. He also broke publicly with SF YIMBY Action, calling it a trash organization for years and telling people to back CA YIMBY and YIMBY Law instead.

• #funding
• #policy