AI Builders Digest

Google bought bankrupt Spirit Airlines' data for $10M, not its planes

Google paid roughly $10 million for Spirit Airlines' dataset out of bankruptcy to train models, with Mercor rumored as the other bidder, and out-of-bankruptcy data buys look like the start of a trend. Labs are now approaching hedge funds on Wall Street for the same reason: real operational data that doesn't read as synthetic is genuinely hard to find, and models, compute, and tools have near-zero switching costs while data doesn't. The sharper warning is on security: "the same type of threat is coming from nonhuman actors, agents that essentially have legitimate access to the environment with legitimate permissions." Detection methodology carries over from the ransomware era, but the velocity does not, and six months ago nobody would even discuss it while now nearly every enterprise leader has either seen it happen or fears it.

• #agents
• #security
• #data

The internet has no plan for a world where everyone runs personal agents

Aaron Levie argues the internet is almost entirely unprepared for a future where everyone's personal agents are out executing tasks on their behalf. He frames that gap as opportunity rather than crisis: new problems and new openings at the infrastructure layer, in user experience, and in business models that don't exist yet.

• #agents
• #infrastructure

Agents are accelerating the collapse of human attention spans

Zara Zhang calls the shortening of the human attention span one of the biggest crises of our time, and puts agents squarely on the wrong side of it. The contrarian edge is the direction: agents get sold as the thing that hands your attention back, and she reads them as making the problem worse.

• #agents

Anthropic traces a month of Claude Code complaints to three separate changes

Three unrelated changes stacked into what looked like broad, inconsistent degradation across Claude Code, the Agent SDK, and Cowork, while the API was never affected. The default reasoning effort dropped from high to medium on March 4, a March 26 caching optimization contained a bug that discarded Claude's prior reasoning on every turn once a session went stale, and an April 16 system prompt line capping responses at 100 words cost about 3% on evals for both Opus 4.6 and 4.7. All three are resolved as of April 20, the default is now xhigh for Opus 4.7 and high everywhere else, and usage limits are being reset for every subscriber. Worth noting from the investigation: back-tested against the offending PRs, Opus 4.7 found the caching bug and Opus 4.6 did not.

• #products
• #evals

GPT-6 Astra on low reasoning beats GPT-5.6 Sol on high

Calibration guidance from the Codex and ChatGPT side of OpenAI: GPT-6 Astra at low reasoning effort performs better than GPT-5.6 Sol at high. If you were running Sol on high and happy with it, the advice is to move down to low or medium on Astra rather than carrying the old setting across.

• #evals
• #products

Plan as though AI R&D is fully automated by early 2029

Ryan Greenblatt's median for full automation of AI R&D is end of 2030, but his 35th percentile is end of 2028, and that earlier date is what he thinks people should actually plan around. "I wouldn't say superintelligence is bad. I would say it's dangerous." His central scenario runs like this: software engineering inside AI companies fully automated in early 2028, models thinking in an AI-only language we can ask them to decode by mid 2028, roughly 4x as much AI progress in 2029 as in 2025, and then a shift from sloppy reward hacking to competent scheming and takeover. The AI 2040 plan he co-authored tries to buy time through a US-China deal built on tracking compute and total research transparency, which would gut the frontier labs' biggest moat while still leaving world GDP growing roughly 200x during the 2030s.

• #policy
• #safety

Anthropic: users approved 93% of Claude Code permission prompts

Telemetry showed users clicking approve on roughly 93% of permission prompts, which is the case for treating containment rather than human review as the primary defense. An OS-level sandbox cut prompts by 84%, and Opus 4.7 holds prompt injection attack success to about 0.1% on single attempts and 5 to 6% after 100 adaptive tries. The two most instructive failures both slipped past every probabilistic layer: an internal red team phished an employee into pasting a prompt that exfiltrated ~/.aws/credentials in 24 of 25 runs, and a third party showed that allowlisting api.anthropic.com let a poisoned workspace file upload data into an attacker's own Anthropic account. Their recurring lesson is that the battle-hardened primitives held, and the custom proxy they wrote themselves is what broke.

• #agents
• #security

Claude Code can now publish live artifacts your team can open

Claude Code sessions can now produce a shareable web page built from the session's full context, including the codebase, connectors, and the conversation itself. Every publish is a new version at the same link with version history, and open pages refresh in place, so an incident investigation republishes itself as it progresses and ends up being the postmortem. Artifacts are private to the author by default, viewable only by authenticated members of the org, and cannot be made public. It is in beta for Claude Team and Enterprise from the CLI and desktop app.

• #products
• #agents

Anthropic pulls the agent harness out of the sandbox container

Managed Agents splits an agent into three swappable interfaces: a session that is an append-only event log, a harness that runs the loop, and a sandbox that executes code. Putting all three in one container made that container a pet, so a crash lost the session and connecting Claude to a customer VPC meant peering networks. Decoupling them means a container gets provisioned only when a tool call needs one, which dropped p50 time to first token roughly 60% and p95 over 90%. The security payoff is that credentials never reach the sandbox: git tokens are wired into the local remote at init, and MCP OAuth tokens sit in a vault behind a proxy the harness never sees.

• #agents
• #products

Chasing cloud agent sessions that start in seconds, not repo clones

The one piece missing from the harness he wants is cloud sessions, and the target is seconds to start, which requires clever snapshotting. Cloning repos fresh, which is what it does today, is not fast enough. He expects to have it next week, and separately says he can't remember the last time capabilities jumped this much at once.

• #agents
• #products

Brilliant's product rule: never tell the learner the answer

Sue Khim, co-founder of Brilliant, builds on a single principle: never tell the learner the answer, because cheating and explaining the answer to your child are closer than you'd think. Her framing is that using AI to skip learning is like bringing a robotic arm to the gym to lift weights for you, and that learning was never about the answer but about strengthening the part of you that can concentrate and struggle. Her advice for anyone building AI products is to find areas where you hold unique data that makes the product better over time.

• #products
• #education

GPT-6 Astra goes out to everyone, API included

GPT-6 Astra landed for Pro, Enterprise and Business Premium users in Work/Codex and in the API, then rolled to Plus and Business the same day. Altman thanked users for their patience on the wait.

• #products

OpenAI pulled its roadmap forward six months on the back of internal Astra use

Astra was OpenAI's biggest competitive advantage while it was still internal, and the productivity jump was large enough that the team moved plans forward six months to ship at DevDay instead of mid next year. The rollout came ahead of schedule, and OpenAI paired it with a full banked reset for all Plus, Pro and Business users landing end of day. More ships promised next week.

• #products

Claude in Chrome hits GA and can now act without asking each time

Claude in Chrome is generally available on every paid plan, and Claude now auto-approves actions it judges safe rather than prompting on each one, using the same mechanism as auto mode in Claude Code. A classifier checks each pending action against what you actually asked for and blocks mismatches, layered on top of probes that scan web content for injected instructions before Claude acts. The numbers: on the current red-team evaluation, attacks that reached the model succeeded 17.6% of the time against Opus 4.5 and 3.8% against Opus 5 with no safeguards; with probes plus the classifier, nothing got through against Sonnet 5, Opus 5 or Mythos 5, and Fable 5 sat at 0.3%. It reads pages, clicks, types and fills forms using your existing logins, which is the point for internal dashboards and vendor portals that will never have a connector.

• #agents
• #products
• #security

Cowork gets its own browser so Claude stops borrowing yours

Claude Cowork on the desktop app now has a built-in browser that opens in a side panel when a task needs the web, separate from your own browser and blind to your tabs, bookmarks and passwords. You can port logins over site by site from Chrome, Edge or Firefox, with banking, email and SSO sites excluded unless you opt them in. The split is deliberate: the built-in browser is for handing off work you want done while you keep going, and Claude in Chrome stays the answer for the page already in front of you. Rolling out this week to Pro, Max and Team, available to Enterprise admins today.

• #agents
• #products

Arm's CEO: chip supply isn't the ceiling, the data center buildout is

Rene Haas says the AI supply picture stays constrained for three to five years at least, and the next bottleneck is physical construction, not wafers or memory. Almost no data center project is running ahead of schedule or under its labor estimate, and local pushback against new builds adds a second brake, which he treats as almost a mercy since wafer and memory capacity would otherwise be the binding constraint. On oversupply versus demand: "Not even close." Inside Arm, 80 to 90% of engineers use AI daily, mostly against verification, validation and debug, which eat far more of a 24 to 36 month chip cycle than architecture design does. His sharpest observation is about why AI can't yet write good RTL: proprietary IP tends to ship without documentation or test benches, and "if it's unusable and untestable, it's actually untrainable." He also pushes back on the idea that accelerators made CPUs irrelevant, arguing that something has to orchestrate where all those generated tokens go, and that job is still the microprocessor's.

• #hardware
• #policy

Rauch on WebMCP: agents should ride the web we already built

Guillermo Rauch is betting on WebMCP with a Tesla FSD analogy: agents need to meet the existing web where it is, streets, stoplights, potholes and all, rather than demand a parallel infrastructure. The concrete win he names is Next.js dev pages exposing debugging tools directly to an agent in the exact tab it's testing, with page-specific context, instead of making it dig through server logs. No separate MCP server to find and configure, because the page carries its own agent tools. In that world he argues a browser-driving agent plus a dev server is a complete web stack with zero loss in debugging depth.

• #agents
• #products

Garry Tan: two hours of OpenClaw setup became three minutes with Aside

Wiring OpenClaw up to Slack took Tan two hours; the same job on Aside's harness, with full integrations and browser access, took under three minutes with sensible access-control defaults out of the box. He's made Aside's browser the preferred remote session browser for GStack and calls it the best way he's found to give an agent web access and credentials under your own identity. His read is that Aside isn't just a browser but a credential manager, integration layer, harness and memory system at once.

• #agents
• #products

The fastest way to learn AI product building: automate one workflow you already know

Madhu Guru's weekend exercise is to take a workflow you understand cold, from work or your personal life, and automate the entire thing end to end with whatever AI product you're comfortable with. Doing it once forces four questions that reading can't: what a great end-to-end experience actually looks like, how to use MCPs and tools, where humans need to stay in the loop, and how you evaluate the result. Then keep raising the ambition. His claim is that one pass through this beats a month of reading about AI product building.

• #agents
• #products

Frontier models are already picking favorites in AI media

Swyx is assembling a large AEO report across frontier models and found that Claude names Latent Space when asked for the best AI newsletter or podcast. The result surprised him enough that he double-checked for memory leaks in the harness before believing it.

• #evals

Replit ships Astra support on day one

Amjad Masad announced Astra running on Replit the same day it went out, with more coming. His framing for the moment: the singularity is here, it's just not evenly distributed.

• #products

People rate AI writing higher than it actually is

Zara Zhang argues the gap worth paying attention to is perceptual: most people think AI writes better than it does. The implication is that the bar readers apply to machine-written prose is lower than the one they apply to humans.

• #evals

The missing Codex interface is a watch you talk to

Peter Yang wants an Apple Watch app that lets him dictate into his Codex threads by voice and hear replies back, so he can leave the phone behind entirely. His motivation is screen addiction rather than convenience, and he wonders aloud whether that's what the rumored OpenAI device turns out to be.

• #products
• #hardware