Last week a guy named Jacob Coxon quit his job at Anthropic. He was a pretraining researcher, which means his whole job was making the models more capable. He spent three years doing it, at OpenAI and then at Anthropic. And when he resigned he did it in public, and he said neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.
Now normally the company says he is one disgruntled guy and moves on. That is not what happened. Anthropic’s own alignment science lead, Evan Hubinger, replied to him and agreed. His words. We really do earnestly believe AI could kill all humans. I personally think it is greater than ten percent within the next decade.
The man building the engine and the man building the brakes, same week, same company, both saying the car is not safe. That post passed a hundred and seventy-two million views. Three days later the CEO of that company published an essay asking the entire industry to slow down. That is where this week starts.
Hey guys, this is Olga. Welcome to Practical AI episode fifty-seven. This is the place to be for your weekly dose of practical AI. Every Friday we go through what actually shipped this week, what it means for your business, and what you can do with it on Monday. And worst case scenario, you’ll be the most interesting person at dinner this weekend.
Chris, hold that resignation for a few minutes, because I want to start somewhere useful. Everybody watching has a CRM they hate opening.
Chris, when is the last time a rep on your team did call prep properly? Not skimmed the account. Actually prepped.
Answers. The honest answer is never on a busy Tuesday.
Right, because it takes twenty minutes and you have a call in four. So Monday Anthropic shipped Salesforce inside Claude. Thirty-seven prebuilt sales skills. Account research, call prep, pipeline review, CRM updates. You are not in Salesforce. You are in Claude, and Salesforce is the thing feeding it.
It is an open beta on all paid Claude plans, your admin requests access through AgentExchange, and then you log in with your own Salesforce credentials, so your data comes in under the permissions you already have. Anthropic says writes need your approval by default. I am saying it that way on purpose. That is their blog post, not an audit, and by default means somebody can switch it off.
And Chris, this is the one you called on episode thirty-nine. You said the future is actually no UI, and I spent that whole episode proving it with an API. This is that, except now Salesforce shipped it themselves.
Takes the callback.
Here is what you do with it. Pick the two things your team skips because they take too long. For most people that is call prep and pipeline review. Those are the two this does first, and it does them with the actual account in front of it instead of a blank page.
Chris, how many times this week did you get a good answer out of an AI and then spend ten minutes copying it into a doc to make it look like something?
Answers.
That step is gone. Tuesday, Claude Cowork and Claude chat merged into one product. One prompt box, no mode toggle. The system decides on its own whether what you asked is a quick answer or a multi-step job that runs in the background.
And they launched Claude Docs and Claude Slides. Both in beta, both on paid plans. So the conversation becomes the deck. Not exports to a deck. Becomes one. Pro and Max get it first over the coming weeks, Team and Free later, and if you are on Enterprise your admin gets at least thirty days’ notice. Nothing you already have moves. Chats, projects, artifacts, connectors, skills, all stay where they are.
What you do with it. The next time you would have said “write me the outline and I will build the deck,” stop at the first half. There is no second tool anymore.
Chris, what happens to your website when nobody clicks through to it anymore?
Answers. This is his lane.
So two things happened this week. On the tenth, Amazon started letting its advertisers run campaigns inside ChatGPT. A pilot, select US advertisers, Delta Vacations first. Which is funny, because Amazon spent months blocking outside AI bots from crawling its store and then opened this door anyway.
Then on the sixteenth OpenAI launched Sponsored Agents. Instead of clicking out to a landing page, you talk to the brand’s own agent right there in the thread. Advertisers feed in their whole product catalog and hook up tools so the agent can take your lead or finish the sale without you ever leaving.
And it is reported, this is Sensor Tower, not OpenAI, that ads per user per hour on ChatGPT’s US mobile app went up a hundred and sixty-three percent in August against April. That is four months, and it is how many ads you see, not how much anybody spent. Same report has financial services going from two percent of ad spend to thirteen.
Chris, this is your episode thirty-four call. You said agents talk to agents and humans navigating websites becomes legacy. That now has a price list.
What you do with it. If you sell anything, your landing page is not the destination anymore. Your product information has to answer a question somebody asked in a sentence, not match a keyword. One site for humans, one for the agents.
Forty-seven percent of US job seekers picked up AI skills in the last six months. A year ago that was forty-one. And the number that actually matters to me, the share teaching themselves went from twenty-two percent to thirty.
Employer-provided training? Flat. About one in six. It did not move at all.
And then Pew came out on Wednesday. In thirty-four out of thirty-seven countries, people think AI means fewer jobs, not more. Over seventy percent here, in Australia, in South Korea.
So almost everybody expects the jobs to go, and five out of six employers are not training a single person. Nobody is coming to teach you this. That gap is the whole opportunity, and it is the reason this show exists.
Back to Chris for a reaction, then move.
Chris, back to your cold open. Three days after that resignation, Dario Amodei published an essay called We Must Pace the Frontier. And the word is pace, not pause. His words, pacing does not mean halting model training or technical progress. He wants to slow the rate of capability gains to buy one or two years. Three steps. Outside evaluators with employee-level access, shared safety standards among democratic labs, and eventually international coordination including China where you can verify it. Anthropic committed to step one on its own, immediately.
And for about a day it looked like it might happen. Altman agreed publicly and said OpenAI would bring in independent evaluators too. Musk said Dario is right. Hassabis was softer, he said it points toward the right path. And Nadella came in a day later welcoming deliberate pacing and then pivoted to talking about ecosystems.
Then the real world answered. Zuckerberg broke ranks and said every lab has its own responsibility to move at the pace it can train safely, and pointed at Meta’s own hold on Muse. His line was, we didn’t call for everyone else to do this before we would, we just did it. Trump posted nine times on Truth Social calling the whole thing a hoax in capital letters. China’s foreign ministry called it fearmongering.
Chris, why can’t this work, even when the people who build it are the ones asking?
His beat. The shape of the answer: you cannot unilaterally disarm in an arms race. Losing today is a certainty, the risk is a probability, and the market always optimizes for surviving today.
So here is what you do with it, and it is not comfortable. The people building this asked for a brake. The governments and the biggest distributors said no. Nothing is slowing down. Stop waiting for a pause that is not coming and get fluent now.
Chris, Apple finally shipped it. Siri that reads your screen. Do you turn it on?
Answers.
Sunday, iOS twenty-seven and everything else twenty-seven, with Siri AI. It reads what is on your screen, pulls context out of Messages, Mail and Photos, and acts inside apps like WhatsApp and Audible. It has its own app now and the chats sync across your devices. English only in beta, more languages in October, and there is a waitlist. Not available in the EU or China at launch.
And the part everybody gets wrong. Apple’s models were trained with Google and distilled from Gemini under a deal from January. Gemini is not running on your phone. Apple’s own models run on the device and on their private cloud.
Now the honest part. Simplified AI polled their readers on whether they would let a screen-reading Siri drive their Mac. A hundred percent said keep it locked out. Zero handed over the keys. The capability shipped. The trust did not.
And Chris, this is the one that goes against you. Episode forty-one, you said tying yourself to one model is not the answer and everyone is leaning toward bring your own keys. Apple went and married Gemini.
Takes it. He is allowed to still be right long-term, but call it against him this week.
One more, and this is the one I actually want you to remember. A woman’s ten-year-old dog, Mopsy, had a dental abscess that was swelling into her eye. The eye specialist said the pressure was already damaging it and had partly torn the retina. Emergency hospitals do not do dental surgery. The specialty dentists were booked. Her own vet wanted a consult, then bloodwork, then a separate surgery appointment.
She made a dozen calls, got nowhere, and gave up. Then she asked ChatGPT. It found every practice within an hour’s drive, wrote individual urgent emails and sent them from her own Gmail, contacted more than twenty of them, and built her a live tracker of who had been asked. Half an hour later one vet wrote back. Bring her tomorrow morning, bloodwork on site, straight into surgery if she clears.
That is not a technical person. That is not even a business problem. But look at the shape of it. Twenty outreaches in parallel, one tracker, one reply that unblocks the whole thing. That is every hard week you have ever had at work.
If you got something out of today, subscribe. We do this every Friday at eleven Central. See you next week.