Sections 00:00 What we're covering today 00:27 1. Amazon reports $42.4 billion cash outflow as AWS revenue grows 01:35 2. GitHub puts multi-agent projects into public preview 02:32 3. Canadian privacy regulators seek health AI safeguards 03:31 4. Researchers train a physics AI on 500 million simulated collisions 04:28 5. Researchers plan to map bee boxes and blooms from space 05:28 Visit Hot Tea Disclosure Narration uses an AI-generated voice. Transcript Welcome to Hot Tea for Friday, October thirtieth, twenty twenty-six. This briefing uses an A I-generated voice. Today, GitHub puts multi-agent projects in preview. Canada's privacy regulators call for safeguards around health A I. A physics model tests what it learned from simulated collisions. And researchers plan to check satellite images of bee boxes against conditions on the ground. Amazon reports $42.4 billion cash outflow as AWS revenue grows. Amazon's results cover different businesses and different periods. The free cash outflow covers the whole company over twelve months ended September thirtieth. A year earlier, the same measure showed fourteen point eight billion dollars flowing in. It is not a quarterly loss at A W S. The cloud division reported forty-six billion dollars in third-quarter revenue, up thirty-nine percent from a year earlier. Its operating income reached seventeen point eight billion dollars, compared with eleven point four billion a year earlier. Those results show sales and operating earnings. They do not separately measure the profitability of A I. Amazon says higher purchases of property and equipment primarily reflect A I investment. It still expects about two hundred twenty billion dollars in cash capital spending this year. That figure is a forecast, not a finished annual total. More cloud sales can arrive while construction consumes cash. The next reports need to show whether new capacity earns enough to cover what Amazon pays to build and run it. GitHub puts multi-agent projects into public preview. GitHub announced the preview on October twenty-eighth. A project can include repositories, designs, notes, and material from connected tools. Copilot coordinates agents and returns their work for review. The same app has desktop computer use in public preview on macOS and Windows. GitHub says it can read application content, click controls, and enter text. The app asks for approval before controlling an application. Organizations can disable the capability through managed settings. These are the company's descriptions of preview tools. They do not establish independent results for reliability or productivity. A returned pull request still needs review. A task marked complete still needs a working result. Watch what access each project receives, how mistakes become visible, and whether someone can verify the completed work. Combining more tools gives an agent more to do and more ways to get it wrong. Canadian privacy regulators seek health A I safeguards. The announcement came on October twenty-ninth, but the resolution was adopted in Ottawa on September fifteenth. Canada's federal, provincial, and territorial privacy regulators call for safeguards throughout the life of health A I systems. They ask developers to know where training data came from, use it lawfully, and assess risks as systems change. They also call for human oversight and practical ways for people to challenge consequential decisions. Healthcare providers should assess the tools they buy and keep monitoring their use. This is a joint resolution. It is not a new statute or a finding against a particular product. Existing privacy duties still depend on the jurisdiction. Watch for signs that governments change laws and whether providers put these protections into contracts and daily work. A written policy needs someone who can check a tool, answer a patient's concern, and stop an unsafe use. Researchers train a physics A I on 500 million simulated collisions. The model is called E V E Net. Its authors then adapted the model to several particle-physics tasks. Their paper appeared in Nature Machine Intelligence on October twenty-ninth. One test used real open data from the C M S experiment. The model rediscovered the known upsilon meson. That checks transfer from simulation to recorded data. It does not announce a new particle. The authors report gains against comparison models in the tasks they tested, but leave important work unfinished. They have not stress-tested substantial mismatches between simulations and detector data. Measurement and theory uncertainty still need explicit treatment. A shared model could reduce repeated training work across analyses. The initial training bill belongs in that comparison. Watch independent reuse, harder detector tests, and results that remain sound when the data differ from the simulations. Researchers plan to map bee boxes and blooms from space. The University of North Texas described the project on October twenty-ninth. A three point six million dollar National Science Foundation grant includes one point four million dollars for the university. Researchers plan to use satellite images to locate bee boxes, estimate available flowers, and follow crop growth. Field teams will check what the models find. Earlier work began with a seed grant in twenty twenty-four. The university says identifying boxes can be difficult because water tanks and haystacks can look similar in images. A box gives only a rough indication of bee numbers. The research plan does not establish better crop yields or a measured economic return. Those conclusions need field observations and careful comparisons. Watch how often detections match conditions on the ground. Then ask whether the data can explain pollination and crop outcomes after accounting for other differences among fields. 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.