Latest / HotTea: The AI Morning Briefing / Aleph Alpha releases 78-billion-parameter Kolibri, but recommends less than its one-million-token maximum for complex work
Transcript
- Sections 00:00 What we're covering today 00:22 1. Aleph Alpha releases 78-billion-parameter Kolibri, but recommends less than its one-million-token maximum for complex work 02:15 2. Trump puts Jay Clayton in charge of a new federal AI task force 03:13 3. OpenAI safety employee resigns and criticizes its release culture 04:16 4. Amazon says it has ended government NDAs on data-center projects 05:23 5. Reuters asks how AI could pay for a buildout projected to cost more than $30 trillion 06:35 Visit Hot Tea Disclosure Narration uses an AI-generated voice. Transcript Welcome to Hot Tea for Sunday, October 4, 2026. This briefing uses an A I-generated voice. Trump puts Jay Clayton in charge of a new federal A I task force. OpenAI safety employee resigns and criticizes its release culture. Amazon says it has ended government NDAs on data center projects. Aleph Alpha releases 78-billion-parameter Kolibri, but recommends less than its one-million-token maximum for complex work. Aleph Alpha released Kolibri on October 3. It published the German-English model’s weights under Apache 2.0. The company built Kolibri for document analysis, reasoning and tool use. The model has 78.1 billion parameters, but it uses about 3.46 billion for each token. It selects a small part of the model at each step, which cuts the computation needed. The full weights still need substantial memory. Aleph Alpha advertises a maximum context of one million, forty-eight thousand, five hundred and seventy-six tokens. The model’s native training length is two hundred and sixty-two thousand, one hundred and forty-four tokens. It recommends staying at or below that length for efficient serving and complex tasks. The company says it tested the longer context, but no independent evaluation is cited here. The published eight-bit weights take up about 78 GB. The minimum hardware is two A one hundred GPUs, each with eighty gigabytes. One H two hundred, B two hundred or B three hundred GPU also meets the minimum. Downloadable weights give organizations another way to deploy the model. They do not eliminate hardware or operating costs. Aleph Alpha supplied these claims and has an interest in the result. Its benchmark and efficiency comparisons have not been independently verified here. Open weights and an Apache 2.0 licence do not prove that the model gives reliable answers. They also do not prove that it lowers total costs or complies with the rules for a particular application. Independent tests should show how Kolibri handles German and English documents, especially when evidence is missing or the task uses a long context. Memory use and the cost of each completed task matter too. Trump puts Jay Clayton in charge of a new federal A I task force. President Donald Trump announced the Super Intelligence Force on Sunday, with Director of National Intelligence Jay Clayton leading it. The other leaders are FTC chair Andrew Ferguson, Pentagon technology chief Emil Michael and Office of Personnel Management director Scott Kupor. The group will coordinate federal engagement with consumers, public-interest and religious groups, infrastructure providers and A I companies. AP reports that it will answer to Trump and chief of staff Susie Wiles. The announcement follows the previous week's voluntary industry safety accord. The group now has leaders and a job to do. That doesn't show that companies have kept their commitments, or that any new binding safety rules are in force. The useful evidence will be published recommendations, plans for reporting incidents, and a clear account of how the government will judge those commitments. OpenAI safety employee resigns and criticizes its release culture. David Robinson said in an October 3 Atlantic essay that he'd resigned from OpenAI. He argued that A I developers need stronger safety expertise and precautions before releasing more capable systems. Reuters reports that Robinson spent three and a half years at the company. He said he helped draft its preparedness framework and oversaw safety reports for 12 frontier-model launches. He criticized both the pace of those launches and the practice of strengthening safeguards after problems emerge. OpenAI told Reuters that it keeps capabilities within what it can safely manage and secure. The company said it pauses training or holds back models when necessary. Robinson is speaking as a former employee, and OpenAI disputes the suggestion that it's moving beyond its safeguards. His resignation doesn't establish whether any particular model is safe. Stronger evidence would come from published release assessments and containment tests, especially tests showing that required safeguards stop unauthorized behavior before wider deployment. Amazon says it has ended government NDAs on data center projects. Amazon Web Services CEO Matt Garman says the company no longer uses nondisclosure agreements with government agencies working on its projects. The statement appeared in an October second post and received further coverage on Saturday. Amazon also announced Built Together, promising more than $1 billion over five years for data center communities. The company says communities will help choose priorities such as education, workforce training, energy affordability and water conservation. It separates that future commitment from the more than $1 billion it says it contributed over the past three years. These are Amazon's claims about its policy and spending. Ending an NDA doesn't prove that residents can inspect every record for a project. It also doesn't show that anyone has resolved local utility costs or the effects on local resources. The useful evidence will be government records released before project decisions, project-specific utility agreements, and public reports on energy and water use. The amount of community funding actually paid out will matter too. Reuters asks how A I could pay for a buildout projected to cost more than $30 trillion. An October 3 Reuters analysis looks at how companies could finance A I infrastructure. It cites a PwC projection that cumulative global data center spending could exceed $30 trillion by 2050. PwC published the underlying outlook on September second. Its central scenario puts the total at $31.6 trillion. The forecast depends on how quickly organizations adopt A I. It also counts repeated replacements of computing equipment, not only new buildings. PwC says servers and other equipment typically need replacing every four to six years. PwC sells advisory services and has an interest in the forecast, which has not been independently validated here. The figure is a long-range scenario, not money already spent or contracted. Even useful technology can struggle to earn revenue fast enough to cover equipment, electricity and financing. The analysis does not prove that the investment will fail. Watch the revenue companies collect, whether customers renew, how much infrastructure they use and what replacement equipment costs. Those records will show whether cash generation keeps pace with the commitments. 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.