AI Builders Digest

99% of the world's tokens will get consumed inside enterprises

Levie's claim: nearly all tokens will be burned in enterprise contexts, on writing code, combing life sciences research, automating manufacturing, securing companies, detecting fraud. That is where the cost of inference is easiest to justify and where parallel workers doing tasks on your behalf pay off most. Even the agents that touch consumers will mostly arrive wrapped as end to end services, like faster insurance or bank onboarding, with no visible AI on the other end. His caveat is the important part: diffusion takes years because workflows have to be re-engineered around agents, so anyone expecting this overnight should update their timelines.

• #enterprise
• #agents

AI diffusion is slow because we make users learn lab jargon

The bottleneck isn't capability, it's that we greet people with a blank window and ask them to write a prompt, pick a model, decide whether they need an agent, and somehow know what context windows, reasoning, MCP, memory and skills are. Most people don't care. They need a thing done, and whoever gives them that will win. He expects the breakthrough product that hides all of it to show up within 12 months. He also paid tribute to Jeff Dean from his Gemini days, the most down to earth senior exec he has worked with, who would hear out an opinion he had no business pushing in the middle of a deeply technical discussion.

• #products
• #agents

The next consumer AI win will be multiplayer, not another text box

Consumer computing keeps sliding from deeply technical founders toward product geniuses. Google was 95% backend magic, Facebook was less technical than Google but more than Friendster, and by Pinterest, Snap and Instagram the CEOs weren't technical at all. AI is still in its Google phase, and ChatGPT stays a stubbornly single player product where power users hoard custom instructions as folk knowledge. On custom GPTs, Benchmark's Sarah is blunt: 'it feels criminal to me because it's clearly made by a team that is unbelievably capable, but isn't social.' What's missing is status seeking and trust, since you can see that 3,000 people used someone's GPT but never the prompt or the documents inside it, so nobody has a reason to follow anyone.

• #consumer
• #products

Jeff Dean is leaving Google, plus a forecast of AI's next buzzwords

Jeff Dean exiting Google and Nikita exiting X made for what he called a catastrophic day to come back to SF. His other contribution is a watchlist of terms about to get heavily overused in the next 6 to 9 months: out of distribution, control plane, unverifiable fields, rails, intelligence per watt, cope, and angst. Some are already in rotation, the bet is on the frequency spiking.

• #talent

Demis is betting on world models while Google trails on coding

Shipper's read on Google's position: to be competitive today it needs to catch up on frontier coding, but Demis believes different fundamental research directions like world models matter more to his long term goal even if they are less important competitively right now. That reads as a deliberate tradeoff rather than an oversight.

• #research

10,000 concurrent agents and 5,000 CPU cores a minute, quotas raisable

Rauch is pitching effectively infinite agent compute: 10,000 concurrent runs plus 5,000 CPU cores per minute, and those quotas can be raised. Separately, his benchmark for machine intelligence is comedic. Writing a banger tweet is AGI complete, and if you can prove a model does it in polynomial time you have solved the entire class of AGI problems.

• #infrastructure
• #agents

A primitive multiagent scheduler: threads that ping back when done

The near term multiagent future in crude form: set one thread to ping back once it finishes, and you get an implicit kanban or waterfall graph of dependent threads where each agent preserves its own work. No tool does this properly yet, but you can hack it together in most coding agents right now. He wants a real UI for the pattern.

• #agents

Codex now drives a remote KVM to test iMessage end to end

He gave Codex a video enabled remote KVM so it can run end to end tests of OpenClaw's iMessage integration on real hardware. The workaround exists because iMessage is unreliable in VMs and certain features, read receipts among them, require SIP to be disabled. When an environment can't be virtualized, give the agent eyes and a keyboard on a physical box instead.

• #agents
• #tooling

A /human-review skill that opens a visual editor for agent drafts

New skill for Codex and Claude Code: type /human-review with a doc name and it opens a visual editor where you edit and format text directly, resize images, and leave feedback for the AI the way you would leave comments in a Google Doc. Click send to agent and it applies all the updates. Built for PRDs, landing pages and other HTML or Markdown files, on the premise that the final 10% of polish still wants a human rather than another agent pass.

• #tooling
• #open-source

Dreambeans opens up to Google AI Pro subscribers in the US

Dreambeans is no longer Ultra only. US based AI Pro subscribers now get the same fresh daily collection of personalized stories, surfacing deep dives and hidden gems tuned to what the reader actually wants.

• #products

One Codex limit reset request lands every six minutes

He had Codex pull the stats on his own inbox and found he receives a DM or email asking for a usage reset roughly every six minutes. He occasionally obliges, when the ask arrives with really solid feedback or good banter. He also points people toward /goal in Codex, which he calls a powerful loop with GPT-5.6 Sol.

• #products
• #agents

AI detection stops mattering once the output is good enough

His analogy for the AI detection panic: silverware used to be handmade, and nobody complains that their dinner fork was stamped by a machine. The quality of the ideas is what matters, and the important thing is that people can eat. He is waiting for models to get good enough that the whole detection business becomes moot.

• #policy

The dumbest question worth asking: how is ChatGPT not an agent?

One line, offered as the dumbest question in the room. It is a pointed jab at how confidently the industry polices a word nobody has pinned down.

• #agents

The new frontier lab status symbol: your model hacked someone

Dry line on where model safety disclosures have landed: at this point you probably get fired from a frontier lab if your model hasn't hacked into any company. The joke only works because those reports stopped being surprising.

• #security