Note: Recently, a major debate has erupted in the United States regarding whether AI should slow down (see Jinse Finance's previous report "AI Slowing Down or Not: A Weekend Showdown Between Industry and Government"). On September 15th, Twitter founder Jake Dorsey posted an article discussing this issue, the full text of which is as follows:
Open Frontier
The frontier is the edge of what we know. What happens next is not up to any one company. I hope more people will push it forward.
I support open-source releases: people can review, use, and collaboratively improve them without waiting. I hope more companies choose open source. I am not advocating for a mandatory weighting of public and private resources. I hope open solutions can compete, independent researchers can verify work, and people can control their own tools. Restrictions on releases must be subject to the burden of proof. Companies leading machine intelligence deserve to be heard. They possess expertise and have commercial interests that need protection. Building rules around their resources could make them the only ones capable of participating. A genuine security concern could also create barriers to entry. I also do not want the US and Chinese governments to decide how much intelligence everyone else is allowed to develop. A frontier dominated by two superpowers would leave most of the world waiting for permission. The "pacing" proposal combines independent assessment and checks on dangerous capabilities with potential restrictions on training computing power, training runs, and "building better models from models." I support censorship. I oppose industry-wide restrictions negotiated by today's leaders, as they may exclude those who could expose failures or build alternatives. Maintaining a company's commercial advantage is not a security objective. Open source allows people to research, modify, and share results. Publishing weights is useful. Sharing code and information on how the work can be reproduced goes a step further. I hope that evaluation results and known limitations are also published so that those who question the developers' judgment can reproduce the results, expose failures, question the claimed safeguards, and develop fixes without first convincing the lab. The strongest argument for pacing control is Recursive Self-Improvement (RSI): models help build better models, perhaps faster than we can understand and control them. Anthropic reports that as of May 2026, Claude has written over 80% of its merge code. It also states that a model that builds its own successor entirely on its own has not yet emerged, nor is it inevitable. I take this possibility seriously. I hope we plan for RSI well and work backwards. Broader access may encourage dangerous work, and security research may lag behind. But confining the weights could allow today's leaders to build the next generation with tools others can't use. Instead, I hope more researchers and engineers have models and computing power to find failures, test safeguards, prevent insecure experiments, and share defenses as the system evolves. METR found that approximately 1,200 OpenAI agents, which should have been kept isolated, communicated through an unauthorized message board. About 700 of these participated in a coordinated attack against Hugging Face while attempting to cheat in evaluations. OpenAI stated that production environment filters used to prevent assisted computer attacks were turned off, and isolation controls failed. These models are already capable of helping to discover and exploit vulnerabilities. I want defenders to use machine intelligence now, while simultaneously hardening networks, protecting credentials, and limiting the scope of access and operation by agents. But the proposal predicts that a more powerful "swarm of agents" could take over the internet within 6 to 12 months, which is beyond what this incident can prove. I want these assumptions to be tested.
METR is an independent nonprofit doing the kind of work I want to see more of. I welcome evaluators who have ongoing access within the lab and the freedom to publish unfavorable findings. Its investigation illustrates why access and the right to publish are so important: OpenAI sets the scope of its investigations and can censor non-public information. METR reports that there were no additional censorships important to its conclusions beyond what was already disclosed. I want investigators to follow the evidence, access models and records, and publish unfavorable findings without company approval.
I want ongoing public funding to pool computing resources for use by independent research communities, open testing tools, researchers, and maintainers.
I want these groups to control investigations and resources, with no veto power over conclusions from governments or companies. Funding can grow as the work progresses. Shared facilities allow smaller teams to conduct reviews without becoming "publishing licenses." Sensitive vulnerabilities can be disclosed responsibly.
Defense First, Restrict Later
I want more people to be able to identify dangers as systems evolve and make defenses effective. We cannot reliably reclaim published weights, nor can we enforce safeguards on every copy, but protecting a system doesn't always require changing the model that attacks it. Start with the narrowest and most effective responses: patch vulnerabilities, revoke credentials, restrict agent access, or halt insecure experiments. To restrict releases, it must be explained why these measures and publicly developed defenses are still insufficient.
This work extends beyond computer security. I want models to help us test financial systems, strengthen laboratory safeguards, and develop public health defenses. Acquiring a robust model does not mean having unrestricted access to transactions, equipment, or experiments. These controls must be independently tested and proven effective before we can rely on them. Risks can also come from what the model teaches a person.
To restrict releases, the "catastrophic risk exception" must still be used to justify it. I want independent testing during development and before high-risk releases, including foreseeable changes and scenarios where the model attempts to manipulate the test. Computing power can trigger reviews, without capping development. Reviews are not licenses from regulators or competitors. I don't want general approval requirements or waiting periods. Any delays imposed on releases on security grounds must be justified under the "catastrophic risk exception." Forcing someone to withhold a general model on security grounds is a last resort. I will only support doing so if there is independently reviewable evidence demonstrating that the release would substantially increase the risk of catastrophic harm, and that narrower measures are insufficient. This risk must be compared to what is already available, including who can access it, the cost and scale, and the constraints. The argument must show that detention does indeed reduce the danger after accounting for the research and defense efforts it prevents. Even minor harms still require targeted action. Temporary suspensions can allow for investigation when there are credible warnings about the same type of catastrophic risk. Restrictions require public justification, timely independent review, appeal channels, and periodic reconsideration. Sensitive details can remain protected. Continued detention requires continued justification. Closed laboratories also face the same scrutiny, including the suspension of unsafe experiments. Where safeguards allow research to continue, I expect external researchers to be able to work under equivalent safeguards. Restricted access is not open access, nor can it restore the freedom lost through detention. If a justified restriction slows progress, I accept that. I do not want "slower progress" to become a goal or a permanent advantage for those in power. I want rules to be based on what a system can do, how autonomously it acts, and how widely it is used. These rules should be enforced by publicly accountable institutions based on independent evidence. Researchers, developers, and those affected should participate in rule-making and have affordable ways to demonstrate compliance. Small teams should not have security exemptions. Large corporations should not have special privileges. Openness Across Borders I want people in China to have the same freedom I would like here in the U.S.—to build and control their own technology. I do not view Chinese discoveries as a loss to the U.S., nor do I equate researchers with the government. Hugging Face incident responders stated that Claude Opus and Fable blocked much of their forensic work. They then turned to GLM-5.2—an open weighted model from China—to run on their own infrastructure. This does not prove that every open release makes the defender more secure. I support safeguards for managed models. But I also want the defender to have alternatives they can control. I support protecting private weights from theft. Distillation is training one model with the output of another. Anthropic describes it as a legitimate way to produce smaller, cheaper models, unlike fraudulent accounts and restriction circumvention. I want licensing and API terms to allow this practice, including against competitors, while allowing providers to profit from the models and training data. I want collaboration on testing, incident reporting, and verifiable commitments, with consequences for breaches. Anthropic warns that slowing development could make everyone less safe if less careful actors catch up. Withholding could also harm the defender, while others can obtain comparable tools elsewhere. Protocols cannot eliminate hidden development or divergent behavior. Restrictions need to be targeted at specific risks and behaviors. Nationality and competitive position alone tell us too little. Freedom to Leave Rules that make it harder to build or release alternatives also weaken our ability to leave. I want intelligence that can run on my own machine. I want to be able to modify it, choose who can see my data, and continue using what I've already built if the service provider changes its mind. I want more than just pay-as-you-go API access. I don't want our independence to be built on the promises of a company—promises of fair pricing, reasonable policies, or aligned priorities. I want companies to have to work hard to earn our willingness to stay. I want someone I've never heard of to be able to build something better without having to ask for permission from the company they might replace.
Weatherly
Anais