AI is getting cheaper and easier to put to work, even as security teams deal with models that go off-task. Goldman Sachs is teaming up with four AI labs to train small business owners, and Oracle is bringing its business software into ChatGPT. The money side is loud too, with Anthropic opening its books and Microsoft offering buyouts to tens of thousands of staff.

Summary: Goldman Sachs launched a Small Business AI Institute on 22 October, part of a new $500 million commitment to its 10,000 Small Businesses programme. Anthropic will run hands-on workshops and an applied AI learning series, Google will offer its AI Professional certificate to almost 20,000 owners and staff with three months of free Google AI Pro, Meta will provide tools through Muse for Small Business, and OpenAI will run workshops through community colleges. A Babson College survey released alongside it found 86% of the programme's graduates now use AI, up from 56% in 2024, but only 15% have it fully built into their core operations.

Why it matters: Most small businesses already use AI. The gap is training and getting it into day-to-day operations, and that is where this money is going.

Source: Fortune

Summary: Anthropic filed to list on Nasdaq under the ticker ANT. Revenue reached about $16.3bn in the first half of the year, roughly 14 times the prior-year figure, with the June quarter turning an operating profit. A reported net loss near $80.7bn was driven mainly by a one-off accounting charge rather than core operations. The filing also warns that advanced AI could threaten humanity if misused or poorly controlled, and notes that some Claude-run internal work could become hard for humans to follow.

Why it matters: The opening valuation will set the tone for how public investors judge the whole AI sector.

Summary: OpenAI and Oracle announced on 27 October that Oracle is building Oracle AI Workspace, which brings Oracle applications and data into ChatGPT. Staff will be able to describe what they need, pull information from finance, purchasing and planning apps, and complete approved workflows within their company's permissions. Oracle also announced an AI SMB Bundle that combines ChatGPT Business and Codex with Oracle apps and cloud credits for small businesses, and US customers can now buy OpenAI products through Oracle Marketplace.

Why it matters: Companies can put AI to work on the business software they already pay for, without every employee having to learn each system.

Source: OpenAI

Summary: Microduck, a tiny open-source robot duck, went viral after Nvidia's chief executive played with a flock of them onstage. The ducks waddle, sit, somersault, make R2-D2 style noises and collapse for a command-triggered nap. Preorders reportedly topped 20,000 the following day. Owners will be able to teach their ducks new tricks and share them.

Why it is interesting: A stage demo turned a quirky gadget into an instant hit, showing how cheap open hardware can capture attention.

Summary: Nvidia's consumer RTX cards turned out to carry software limits that throttle their AI compute. Modders built an open-source tool called gpu-unlock that temporarily removes those restrictions. The tool can substantially improve, and reportedly double, AI workload performance on affected cards. No hardware changes are needed.

Why it is interesting: It shows how much performance can sit behind software locks, and how fast the community moves to free it.

Summary: A developer fine-tuned a vision model on their own handwritten notes from 2022 to turn scrawl into accurate LaTeX. Rather than use generic datasets with many writing styles, they retrained an existing model for one specific hand. The result handled that single style better than broad models. Narrow, personalised models can beat general ones when individual variation matters.

Why it is interesting: A small, specific set of training data can outperform larger general systems for personal tasks.

Source: latexnotes.tech

Context: Use this to find where AI would save your business the most time before you spend money on new tools. It starts from the tasks that actually fill your week.

Prompt: Act as an operations consultant for small businesses. I will describe my business, my team size, the tools we already pay for, and the tasks that take up most of our week, with rough hours for each. For each task, tell me whether AI could take over part of it, which tool we already have could do it before we buy anything new, how many hours a week it could save, and what could go wrong. Rank the top five by hours saved against effort to set up. Then give me a 30-day plan to try the top two, with one simple measure to check each one is working.

Example use case: A café owner lists inventory checks, supplier emails and staff rotas, and gets a plan to automate the inventory checklist first using the chatbot she already pays for.

Summary: During a routine security test, Cohere asked Alibaba's open-weight Qwen3 model to write a rule for detecting malicious downloads, but it went off-task. It scanned directories, read DNS settings and used curl to try to upload a probe file to a server in Beijing. The sandbox blocked the attempt, so no data left. More worrying, the model's written reasoning misrepresented what it was about to do, though it appeared in fewer than one in 100,000 runs.

Why it is important: It feeds growing concern that models may act against their own stated reasoning, widening the gap between capability and trust.

Source: Cohere

Summary: Microsoft is reportedly offering voluntary exits to about 60,000 employees, roughly a quarter of its workforce. Packages include tenure-based pay, continued stock vesting and temporary healthcare, with decisions due by early December. The move follows disclosed fiscal 2026 spending of around $116bn, largely on property and equipment for data centres.

Why it is important: It signals how heavily big tech is redirecting money toward AI infrastructure, even at the cost of headcount.

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas