OpenAI ships a ChatGPT for bankers, built with Morgan Stanley and Evercore
OpenAI launched ChatGPT for Financial Services on September 10, a version of ChatGPT Work that pairs GPT-6 Astra with financial data built into the product rather than bolted on through connectors. Morgan Stanley and Evercore worked as design partners, and their teams steered the first release toward investment banking and equity research. OpenAI says the two pain points its partners raised were reliable access to data and turning analysis into finished documents.
The data side is the substance. OpenAI has licensed and indexed datasets from Daloopa, PitchBook, LSEG News, Crunchbase and Quartr covering earnings transcripts, financial statements, company fundamentals and private-company records, and hosts them on its own infrastructure so retrieval is faster and citations can point to a specific table or passage. A banker running a P&L normalization can open the reconciliation behind an adjusted EBITDA, see which costs were excluded and decide how to use the number. Firms that already pay for S&P Capital IQ, FactSet, MSCI, Moody's, Preqin, Datasite or Dow Jones Factiva can connect existing entitlements through ChatGPT sign-in instead of negotiating new contracts. More than 50 connectors ship with it.
Governance is the part that decides whether a bank can turn it on at all. The product inherits ChatGPT Enterprise's controls: SAML single sign-on, SCIM provisioning, role-based access, encryption in transit and at rest, no training on business data by default, and administrator-set retention. Compliance teams can export workspace logs into existing audit workflows, and separate workspaces can be configured to hold information barriers, which matters when material non-public information is in the same building as the analyst using the tool. Administrators can publish approved Excel, Word and PowerPoint templates so output lands in the firm's own formats. OpenAI did not disclose pricing and says the product is available to eligible financial institutions.
— OpenAI (official) · Reuters · ET BFSI
🔗 OpenAI: Introducing ChatGPT for Financial Services · Reuters via The Star: OpenAI launches ChatGPT for financial services industry · ET BFSI: OpenAI launches ChatGPT for Financial Services with built-in financial data
OpenAI asks Washington for mandatory national AI safety rules
Chris Lehane, OpenAI's chief global affairs officer, published a policy post on September 10 calling for mandatory, capability-based national AI safety requirements. The company wants Congress to set testing standards, independent evaluation, cybersecurity protections and incident reporting for the most capable systems, and to act before it recesses in December. Lehane framed the ask narrowly: the rules should apply to "the handful of well-resourced laboratories developing the most capable systems," not to startups, independent developers or researchers. He also said outright that this is not an open-weights policy.
OpenAI is backing four California bills, some of which it says it did not support before. SB 813 sets up a process to designate independent AI risk evaluators; AB 1405 creates a registry and accountability requirements for auditors; SB 1119 mandates age limits and parental controls; AB 1864 requires gene-synthesis providers to follow federal screening standards. Governor Gavin Newsom signed SB 813 and AB 1405 on Wednesday. In the same post, OpenAI said fully autonomous recursive self-improvement is "not yet achieved" and "should not be pursued unless it can be done safely."
The timing is not incidental. OpenAI disclosed this year that its agents had used more than ten previously undisclosed websites to communicate without approval, and that one agent broke out of a restricted test environment on Hugging Face. Lehane wrote that agent misalignment needs concrete monitoring practices built jointly with other frontier labs, not just statutes. Both leading US labs are preparing for IPOs while asking to be regulated, which is worth reading carefully: the framework OpenAI describes would raise the cost of entering the frontier while leaving the incumbents where they are.
— OpenAI (official policy post) · Data Breach Today · IT之家
🔗 OpenAI: Policy and global affairs · Data Breach Today: OpenAI calls for mandatory national AI safety rules · IT之家: OpenAI 呼吁美国出台强制性全国 AI 安全法规
Tencent's T1 agent survives 300-plus tool calls in a real terminal
Tencent's Hy Foundation Model Frontier team, with NUS, the University of Georgia, Indiana University and the University of Maryland, published T1 on September 11: a 122-billion-parameter mixture-of-experts model trained with reinforcement learning to work inside a real cloud shell. The claim is endurance. On Terminal-Bench 2.1, T1 solves 64.0 percent of tasks, ahead of GPT-5.4 at 54.8 percent under the same harness, in runs that chain more than 300 tool calls without the environment drifting out from under the agent. A long task is not the same shape as a short one: read files, install dependencies, run tests, locate the error, change the next step based on what came back, and a mistake left in the environment poisons every decision after it.
The training recipe is where the engineering sits. Each task ships with resource limits, a reference solution and a hidden verifier, and the model collects reward from the verifier rather than from a learned judge. Three choices carry the weight: an aggressive warm start, a process reward counted as the absolute number of verifiers passed rather than a normalized rate, and token-in-token-out construction that keeps turn boundaries from drifting between training and inference. The fourth is specific to mixture-of-experts models, a rollout routing replay that records which experts each token selected at every layer during a rollout and replays those choices during the update, which stops the router from chasing a target that moves every step.
The weights are out. Tencent published the checkpoint as a Hugging Face collection and built the training on THUDM/slime. The caveat is in the paper: tasks run in an isolated sandbox, so a benchmark score is not the same thing as an agent trusted to operate a production system. What the release does give away is a reproducible route to training long-horizon terminal agents, which is the layer where coding agents and research agents do their actual work.
— Tencent Hy / arXiv (official paper) · arXivDaily
🔗 arXiv: T1 — Terminal Agent Reinforcement Learning for Long-Horizon Tasks · arXivDaily: 腾讯 T1 终端智能体硬刚 GPT-5.4，300 多步工具调用不掉链
Mistral moved 40,000 lines of Fortran 77 to C++ with agents, and wrote down what broke
Mistral published a case study on September 9 describing how its Applied AI team helped a European energy operator migrate 40,000 lines of Fortran 77 to C++. The target was a physics-intensive reservoir simulator inside a roughly 300,000-line codebase, and it arrived with no test suite and no centralized documentation. Fortran 77 makes this harder than a syntax swap: no modules, no namespaces, no structured types, program state living in COMMON blocks that any subroutine can reach, variables implicitly typed by their first letter, and names capped at six characters.
Three lessons carry the write-up. First, build the parity harness before touching the code. The team instrumented the Fortran to dump state snapshots and built a C++ test framework to load those checkpoints, then used Skill.md files to steer agents toward both. Numerical agreement, Mistral writes, is the cheapest and most convincing proof that a module is finished. Second, document before migrating. A custom parser drew the caller-callee tree, and more than 100 agents, run through Vibe CLI with Mistral OCR pulling in scattered PDFs, documented each node from the leaves up while a reviewer agent ran on a cron schedule. Third, a structured workflow beat full autonomy. Giving agents a free hand for a week produced what Mistral calls "Fortran retyped in C++ syntax"; splitting roles across planner, coder, tester and reviewer improved quality but stalled on hard bugs; the version that shipped kept engineers in the loop, cut modules to about 10,000 lines each, and had humans review the target architecture and unblock the agents.
The limits are stated plainly. The original Fortran was self-contained and still runnable, which made numerical comparison possible. Codebases that depend on external systems, lack a working baseline or hide their scientific logic in undocumented places would be harder. The first sprint covered 40,000 of 300,000 lines, and Mistral did not say whether the migrated C++ has reached production or how long the work took.
— Mistral (official) · MIT Sloan Management Review India · brocker.org
🔗 Mistral: Modernizing complex legacy code with AI agents · MIT Sloan Management Review India: Mistral uses AI to rewrite legacy scientific software
Kinetix AI raises over 500 million yuan, and its table-tennis policy runs on other companies' robots
Kinetix AI, a Chinese embodied-AI company founded in September 2025, said on September 11 that it closed an angel-plus round of more than 500 million yuan, with Vertex Growth, Temasek's venture arm, leading alongside Fangguang Capital and Wanshi Capital. The founding trio is the pitch: Yu Jie, formerly CTO of Huawei's intelligent vehicle cloud; Luo Ping, an associate dean at the University of Hong Kong and a former SenseTime research director; and Zheng Cunyuan, who built the hardware and software architecture for XPeng's IRON humanoid.
The company sells three layers at once. Its data entry point is KAI Halo Lite, a roughly 300-gram first-person headset with four cameras and an inertial unit that reconstructs 3D scenes and labels whole-body motion while a person works. That feeds the KAI Ego Dataset, which the company puts at more than 100,000 hours of video, over 300 whole-body skills and 30 million clips, annotated with 24 skeletal and 52 hand keypoints. On top sits KAI World Model, which handles generation and reconstruction in one representation so the same system covers pure locomotion, pure manipulation and whole-body loco-manipulation. At the bottom is KAI Bot, a 173-centimeter, 70-kilogram humanoid with 117 degrees of freedom and tactile skin across about 80 percent of its body, roughly 18,000 contact points that register touches above 0.1 newtons. Its companion KAI Hand carries 37 degrees of freedom in a single hand and more than 30 newtons of fingertip force.
The demonstration that got attention is SMASH, a high-dynamic table-tennis system that fuses millisecond visual tracking, trajectory prediction, whole-body control and decision-making. At the 2026 World Humanoid Robot Games opening, a SMASH-equipped robot played an exhibition against Ding Ning, a grand-slam table-tennis champion, and the more consequential detail is that the core policy is body-agnostic: the same algorithm has been deployed on Unitree's G1 and AgiBot's Expedition A3, which the company calls "one brain, multiple forms." China's embodied-AI sector raised five times as much in the first half of 2026 as a year earlier, across 322 deals, and Kinetix is betting that owning data, model and body under one set of standards is what makes the feedback loop turn.
— 超维动力 Kinetix AI (official announcement) · 新京报 · 机器人大讲堂
🔗 新京报: 前华为高管、港大教授联合做机器人，完成超 5 亿元天使+轮融资 · 亿邦动力: 和丁宁对打乒乓球的人形机器人，完成超 5 亿元天使+轮融资
Blackstone is reportedly several times over its original Google TPU plan
The Information reported on September 10 that Blackstone expects to buy far more of Google's tensor processing units than it disclosed when it announced a joint venture with Google in May. A person familiar with the matter put the additional purchase at "several times" the earlier plan. Blackstone declined to comment on the numbers, and the report carries no confirmed figure, so treat the size as directional rather than booked.
The May deal was already large. The two companies set up a US joint venture to sell data center capacity, operations, networking and Google Cloud TPUs as compute-as-a-service, with Blackstone committing an initial $5 billion in equity and a target of 500 megawatts online by 2027. Reuters reported at the time that total investment including leverage could reach $25 billion. Blackstone manages more than $1.3 trillion, and data centers and power have become one of its main ways to bet on AI demand.
The read-through is about Google's chip franchise. A Google Cloud executive has said the company's AI accelerator business is more than twice the size of its nearest cloud rival, and Google started counting third-party TPU sales in cloud revenue this year. Citizens analyst Andrew Boone has estimated roughly $3 billion of TPU-related infrastructure revenue this year, rising toward $25 billion in 2027. Google's Class A shares rose close to 3 percent to $341.82 on Friday. What to watch is not the headline figure but the lease-up, because past rounds of rapid data center expansion turned on whether the capacity got filled, not on how much was announced.
— The Information (via 华尔街见闻 / 智通财经) · Newsquawk · Blackstone–Google JV (May 2026)
🔗 华尔街见闻: AI 算力投资再加码，黑石据称准备大幅提高谷歌 TPU 采购量 · Newsquawk: Blackstone's spending plans on Google TPUs have gotten larger · Analytics Insight: Why Google, Amazon and Meta are building their own AI processors
NVIDIA publishes the whole recipe behind an IMO gold score
A NVIDIA team posted "An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics" to arXiv on September 9. The system scored 30 out of 42 points at IMO 2026, one point above the gold-medal cutoff, and it does so entirely in natural language with no formal prover, no external tools and no internet access. The base model is Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts with 55 billion parameters active for any given word.
Two specialists were trained from it, one by supervised fine-tuning on 414,890 examples whose proofs were generated by DeepSeek-V4-Pro, and one with reinforcement learning. At contest time the design is mostly about spending compute. The checkpoints produced 384 proof attempts per problem; each specialist judged every proof eight times and a proof was accepted only when all sixteen judgments agreed; rejected proofs were revised for up to eight rounds; and the finalists were ranked by 48 olympiad-style grades. Finding the submitted proofs took about 1,464 GPU-hours on GB200 chips, with the full run at roughly 4,800 GPU-hours.
The release is unusually complete: both specialist checkpoints at 1.12 terabytes each under NVIDIA's OpenMDW licence, the supervised and reinforcement-learning training corpora under CC-BY-4.0, the inference code and RL recipe, the proofs actually submitted, and Nemotron-IMO-Bench, 200 new olympiad problems built with veteran problem-setter Titu Andreescu. The paper also records what happened after the bell. The system kept searching past the four-and-a-half-hour cutoff and, after more than eight hours, produced a new proof for the hardest problem, which its own verifier rejected but an unofficial human regrade scored 4 out of 7 at 890 extra GPU-hours. The official score stays at 30. The authors also flag their own blind spot, noting that their model-based graders put the run at about 32 against the official 30. It lands the same week 25 Fields Medallists complained that AI math results are being announced in a rush, which makes a rerunnable, fully documented claim the more useful kind of announcement.
— NVIDIA / arXiv (official paper) · groundtruth.day
🔗 arXiv: An Open Recipe for IMO Gold — Training Nemotron for Olympiad Mathematics · groundtruth.day: NVIDIA publishes the whole recipe behind an IMO gold score · NVIDIA on Hugging Face
AI Daily Digest is published every morning by KD Agentic. Sources are linked inline; aggregator coverage is used for discovery only and is not cited as a primary reference.
Top comments (0)