The people racing to build the strongest AI are asking for a brake.

A statement called Pacing the Frontier currently lists 1,224 employees from frontier AI companies. Its signatories include OpenAI chief scientist Jakub Pachocki, Anthropic CEO Dario Amodei, Meta AI chief scientist Shengjia Zhao and senior people from Google DeepMind.

They signed in their personal capacities. Their companies did not jointly ask for a pause.

The statement asks the US government to support an international effort that could deliberately slow automated frontier AI development.

These researchers are worried about AI becoming good enough to help design, test and build the next generation of AI. If that loop accelerates, capability may improve faster than companies can understand the systems or governments can respond.

I am Alex, welcome to ShortCu8 by Innov8.

Lets Dive Deep 🐰

When someone says "pause AI," check four things:

Without those answers, "pause AI" is only a feeling.

AI already contributes to AI development. Anthropic says that, Claude authored more than 80% of the code merged into its codebase. Its engineers were merging roughly eight times as much code per day as they did in 2024.

Claude still cannot build its own successor. Humans choose the research goals, provide compute, review the work and decide what enters a model. Judgment remains the large gap.

Current agents can run experiments against a goal chosen by a person. A more capable system could choose the experiments, interpret the results and decide what should be built next.

Anthropic calls the completed version recursive self-improvement: an AI system autonomously designing and developing a stronger successor. We are not there. The labs preparing for it do not want the first serious discussion to begin after the loop has closed.

Imagine OpenAI pauses its strongest training run for six months while Anthropic, Google, Meta and Chinese labs continue.

OpenAI has surrendered six months to everyone else without creating a global safety pause.

The same logic applies to each company. Even leaders who believe a slowdown may become necessary have a reason to keep building.

Anthropic learned this inside its own safety policy. TIME reported in February that the company dropped an earlier promise that could have blocked training when safety measures were insufficient. Anthropic chief science officer Jared Kaplan said unilateral commitments stopped making sense while competitors were "blazing ahead."

That experience explains Anthropic's newer proposal. Participating labs would need evidence that rivals had also slowed down and that another actor was not secretly using the pause to jump ahead.

That requires monitoring, international agreement and rules for restarting development. Nothing operates at the required scale today.

The petition does not provide a precise threshold, which is its largest missing piece.

"AI is getting scary" cannot be an enforcement rule. A credible trigger needs observable evidence, such as an AI system independently improving model research or materially accelerating the creation of a stronger successor.

The threshold also needs an independent judge. Companies should not grade the danger of their own models while releases, investments and competitive leads are at stake.

This argument becomes harder when model weights are released publicly.

NVIDIA, Microsoft, Meta, Google, OpenAI and many other companies signed a separate open-weights letter on July 24. They argued that downloadable models increase competition and let organisations inspect and run AI on their own infrastructure.

Anthropic did not sign. In his response, Dario Amodei said Anthropic has never supported a blanket ban. He called open-weight models without dangerous capabilities a public good.

His concern begins at the frontier. Providers can update, restrict or withdraw closed models. Released weights can be copied and modified, so their developer cannot reliably recall them or keep one safety layer attached.

Anthropic's proposed rule is capability testing for sufficiently powerful models, whether they are open or closed.

Frontier labs also have commercial incentives. Anthropic generates revenue through subscriptions and API access. Open weights create competitors that other companies can host without paying the original lab for every request.

Its safety case deserves scrutiny because the same rules could also protect that revenue. Testing methods, thresholds and evidence need to be public enough for outside researchers and governments to challenge.

A slowdown involving US labs changes little if a capable model continues training elsewhere in secret. Verification would need some participation from China while both countries treat advanced AI as a national-security advantage.

No. Stopping all AI development would include useful systems far below the risk described in the petition.

There is no verified trigger today that every major lab and government has accepted.

The work we need now is more specific: define the capability threshold, decide who can call a slowdown, agree on what activity stops and build a way to verify that competitors have also stopped.

If the plans are written only after automated AI research begins accelerating the frontier, every participant will have a stronger reason to keep racing.

Now go and vibecode a todo app 😅

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Adam spent 20 minutes looking for a $36 receipt. His finance team sent three Slack messages. Someone made a sticky note.

Ramp would have matched it automatically the moment he swiped. Auto-coded, in-policy, synced. Nobody had to ask Adam for anything.

This is what finance looks like when it runs itself.

Your team can be Adam. Or they can not be Adam.

Don’t be like Adam

#AD2

Skip the sales call. Walk through Gladly's interface yourself — the AI suggestions, the unified customer view, the full conversation thread. 15 minutes, no installation, no commitment.

See it yourself