Sections
00:00 What we're covering today
00:32 Anthropic moves its main code repository off Git as daily merges top 30,000
01:31 AI labs dispute who can control outside safety reports
02:26 British MPs ask who gets early model access and who can stop a release
03:28 IMF says weak AI profits could spread losses through finance
04:26 Microsoft's AI network agent stops for approval before making changes
05:22 Visit Hot Tea

Disclosure
Narration uses an AI-generated voice.

Transcript
Welcome to Hot Tea for Wednesday, October 14, 2026. This briefing uses an A I-generated voice.

Anthropic's move away from Git leads today's briefing. We'll also cover a dispute over outside safety reports and Britain's questions about early model access. Then we'll look at the IMF's financial stress test and the limits Microsoft placed on a network agent. In each case, the question is who can test, challenge, or stop an A I system before the consequences spread.

Anthropic moves its main code repository off Git as daily merges top 30,000. Anthropic said on October 13 that it moved the repository from Git to Scion in five months. The company reports more than 30,000 merged changes a day in September, up from about 2,500 in May. Scion handles history searches and merges on servers, while each client fetches only the files it needs. Anthropic says one global lock on the merge queue no longer worked at that volume. It now runs more tests after changes merge and before deployment. These figures come from Anthropic, and the workload is unusually large. The company says Scion shouldn't replace Git in general. Its internal coding agents also get a token budget that most businesses wouldn't choose to spend. The useful test is whether all that speed produces better software. Failure rates, recovery times, and total operating costs would help answer that.

A I labs dispute who can control outside safety reports. Reuters reported on October 13, citing two people, that Anthropic and OpenAI sought control over what evaluators could publish during talks. Anthropic disputes that this describes all of its work. It says established evaluators control their findings, while Anthropic keeps approval rights for new firms. OpenAI says it doesn't seek approval over full reports. These terms can change during negotiations. Reuters says some of Anthropic's discussions had changed by October 8. Both companies also distinguish confidential information from findings. Amodei's September proposal allows narrow redactions, but reviewers can disclose important omissions. Reuters did not report that labs control every outside review. The signed terms should show who can block a conclusion, how the parties settle disputes, and what reviewers may disclose about redactions.

British MPs ask who gets early model access and who can stop a release. The UK Parliament scheduled former researchers, industry policy leaders, and government officials for an A I session on October 13. The agenda asks whether testers get access early enough and who can stop a model release. September correspondence explains the concern. MPs asked about access to Claude Mythos 5.1. The A I Security Institute, or AISI, said no organisation outside the U S had access at launch. But AISI's September reply also says it tested GPT-6 Astra before release, so access depends on the model. This agenda shows what MPs planned to examine, not what the witnesses said. It did not grant new powers, establish a safety finding, or enact a reporting rule. The transcript and written responses should make the key distinction clearer. Developers can promise early access, but that isn't the same as legal power to demand access or stop a release.

IMF says weak A I profits could spread losses through finance. The IMF's October 13 financial stability briefing says markets could fall if A I investment or profits disappoint. It also flags rising bond yields and cyber incidents. Its severe stress test combines shocks involving A I, bond markets, and energy. In that scenario, the IMF says banks remain resilient overall, though weaker banks and non-bank firms remain vulnerable. This was a risk assessment and a stress test, not a prediction of a crash. Pension funds, private credit firms, and leveraged hedge funds can spread stress in different ways. The report's earlier analytical chapters add context, but they aren't separate findings from October 13. A I earnings, borrowing costs, and demands for cash would show where disappointment starts to hurt. They could force spending cuts or asset sales. A falling share price alone wouldn't prove damage across the financial system.

Microsoft's A I network agent stops for approval before making changes. Microsoft engineer Arun Malik described the agent on October 13. It helps with production-network incidents by collecting device state and telemetry. Ordinary software checks the model's output before anything happens. The agent only has the permissions of the person who started it. Each tool has separate authorisation and logs, and a human must approve production changes. Microsoft also grants trust separately for each procedure and type of action. This account comes from Microsoft, not an independent reliability study. A plausible diagnosis still doesn't let the agent change customer traffic. Microsoft gives no fleet-wide failure rate, and these safeguards don't prove autonomous operation is safe. Incident results and failed-action records would show how the controls handle bad input. Microsoft would also need evidence before replacing fresh model reasoning with routine procedures.

That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.