Sections 00:00 What we're covering today 00:30 1. SemiAnalysis's new AI work test tops out at 12 percent 01:32 2. KAIST reports up to 69.6 percent less fine-tuning time 02:29 3. ADNOC signs an AI development deal with OpenAI and G42 03:14 4. Grok API text watermarks are scheduled for November 7 04:04 5. Seven OPEC+ countries keep December oil targets unchanged 04:49 Visit Hot Tea Disclosure Narration uses an AI-generated voice. Transcript Welcome to Hot Tea for Monday, November second, twenty twenty-six. This briefing uses an A I-generated voice. Today, researchers select less repetitive data to shorten model training. An oil company signs a development deal with OpenAI and G forty-two. Grok documents a future change to generated text. And seven oil producers agree to keep their December targets unchanged. We'll keep the reported results separate from the plans. SemiAnalysis's new A I work test tops out at 12 percent. Claude Opus five point five leads SemiAnalysis's November second leaderboard. GDPx has two hundred and six tasks from fourteen specialist data partners, across eleven fields. It puts agents into document collections, engineering software, and simulated laboratories, where they have to find information, use tools, and deliver the work. SemiAnalysis reports single-attempt performance across three attempts, with each field weighted equally. Most tasks allow eight hours, though some allow less. Its twelve percent overall score is weighted across these selected fields. It does not measure the share of all jobs that A I can do. SemiAnalysis and its partners designed and scored the benchmark, and no independent replication has been established. The publisher also says safety filters lower some scores in cybersecurity, biology, and alignment. For buyers, the useful comparison is the work closest to their own. On a new workload, count total selection and training costs, then include review and correction. KAIST reports up to 69.6 percent less fine-tuning time. The team grouped training data by its effect on learning, chose a smaller set, and spread it across graphics processors. KAIST announced the work on November second. The preprint dates to September, and the team presented the findings at an October conference. Its experiments used models with about seven billion and thirteen billion parameters, three coding and medical datasets, and eight performance benchmarks. KAIST reports nearly the same performance with about thirty-one percent of the training data. The maximum time reduction comes from selected experiments, not every model. Medical question-answering benchmarks do not establish a benefit for patients. On a new workload, an independent rerun needs to count selection and training time together, plus review and correction costs. A shorter run matters only if the model still does the work it is meant to do. ADNOC signs an A I development deal with OpenAI and G42. The companies announced their multi-year agreement on November second. They plan to link oilfield development planning with commercial decisions, using OpenAI Frontier on G forty-two infrastructure in the United Arab Emirates. ADNOC says human experts will oversee consequential decisions. The company wants to prepare development plans in hours instead of months. That is ADNOC's goal, not a measured result from this new agreement. The deal does not show a better investment return either. Completed deployment, checked planning results, and records of human review would give us more to assess. Until then, the agreement shows planned work, not a delivered result. Grok API text watermarks are scheduled for November 7. Documentation updated on October thirty-first describes the planned change for supported text models across the company's A P I endpoints. The watermark changes token choices without adding visible characters. Team administrators can request an opt-out for individual models under the documented limits, though some models can remain locked on. Detection is statistical, so a result can be uncertain. Access to the detection service also requires approval. These are the company's specifications for a scheduled rollout. They do not show that the new defaults are active now. After the start date, developers can compare response headers with the documented settings. Independent tests would still need to check whether people can edit the text and leave the mark detectable. Seven OPEC+ countries keep December oil targets unchanged. OPEC says Saudi Arabia, Russia, Iraq, Kuwait, Kazakhstan, Algeria, and Oman made the decision at their November first meeting. Reuters reported the decision as well. December required production will stay at September levels, and the seven countries plan to meet again on November twenty-ninth. That is a policy target, not a record of what producers pump, load, or deliver to buyers. An unchanged target cannot tell us where fuel prices or shipping costs will go. Before the next meeting, reported production and delivered volumes are the useful comparison. Those figures show whether the barrels arrive, rather than only what the quota permits. 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.