noticias ia para empresas agentes de ia automatizacion de procesos

Three things to start the week. First: the world's largest AI provider is not going public this year —Sam Altman said it on Saturday— and that means that until 2027 you will continue to contract blindly with a company that does not publish its accounts. Second: there are now four heads of the major labs calling for a slowdown, but each proposes a different control and none of them stop anything today. And third: the answer to the distillation scandal is not to ban it, but to legalize it in the United States. Below, also, the seven named agents that Salesforce has just put on sale, one of them capable of working for weeks without logging off. Yesterday's radar was about organization; today is about the contract.

Your largest AI provider will not go public this year: OpenAI postpones its IPO

Altman said Saturday that right now would be an ill-advised time to go public and, when asked directly about the year, replied not in 2026. He is pointing to 2027 and justifies it by the safety risks of the models, not by the market. OpenAI filed the confidential draft of its prospectus in June and its latest private valuation is $852 billion, following the $122 billion round it closed in March. For your company: if your automations rely on OpenAI, another year without quarterly accounts means you won't see margins, cash, or customer concentration before renewing. It's not a reason to run away, it's a reason for two specific things you can ask for this week: a ninety-day notice clause for any price or plan changes, and a quarterly portability test where you export prompts, configurations, and logs to a format you can read outside their platform. If you haven't done that exercise, your provider is not a provider: it's a dependency. Source

Now four are calling for a slowdown, and each wants a different control

What yesterday was Dario Amodei's essay with generic backing from Altman and Elon Musk, today there are four positions on the table, and therein lies the new fact: Demis Hassabis, from Google DeepMind, proposes a standards body funded by the industry itself —he compares it to FINRA, the U.S. financial sector supervisor— that examines frontier models thirty days before launching them, looking for cyberattack and biological risks. Amodei wants external evaluators embedded in the company during training; Hassabis, a prior examination from the outside; Musk limited himself to saying that Dario is right. For your company: if this prior examination is imposed, launches stop arriving every few weeks and your roadmap flattens out. You plan for that, you don't suffer it: assume that the model you have today is the one you will have for six months, demand by contract a thirty-day notice before withdrawing or changing the version you use in production, and have a second provider already tested with your own cases, not with a demo. Source

The answer to distillation is not to ban it: it is to legalize it in the United States

Yesterday we reported that Anthropic pointed to five Chinese labs for milking its model. The response came from within American venture capital: Garry Tan, CEO of Y Combinator, said he would do nothing about it and proposed an American distillation regime, so that U.S. open-weight labs can distill frontier models via API transparently and legally. His argument is that the big players didn't ask for permission to train on protected material either —Anthropic closed a $1.5 billion copyright deal— and that the dangerous scenario is for a single company to dominate the frontier. Tan does not endorse the credential fraud that Anthropic reported. For your company: the origin of the model you use has gone from being a technical curiosity to a contract risk, because no one knows yet which side the law will fall on. Ask each provider for two things in writing: indemnification for intellectual property claims on what its model generates, and a declaration of origin for the weights you are using. Source

Salesforce names seven agents and one of them works for weeks without logging off

Casey for support, Paige for IT and HR, Carter and Piper for commerce and lead gen, Fin for customer service, and Marshall for supply chain are now available; Hunter, for outbound sales, is in pilot with general availability in November and is the first to run on a long-horizon engine, capable of pursuing a goal for days and weeks instead of resolving a single conversation. What changes compared to the control plane the same company announced on Thursday: then it was selling the dashboard to govern third-party agents; now it sells its own agents that work on their own for weeks. For your company: an agent that works fifteen days in a row is not supervised by reading its chat. Before turning one on, write three things on a sheet of paper: who has the off button and how quickly they can press it, what its spending and volume limit is per week, and who reviews its work every Monday. Without those three lines, what you've hired is not a digital employee, it's a process without an owner. Source

What to watch this week?

Anthropic's IPO prospectus, expected in late September, because that's where the real computing commitments of a frontier lab will be seen for the first time. And if any of the four labs turns the talk of slowing down into a signed commitment with a date: that would change what you can demand in your next contract.