Getting smart about AI costs, the crackdown on AI slop, and why blaming layoffs on AI keeps backfiring

The conversation this week was about discipline: spending AI budgets wisely instead of anxiously, doing real AI transformation instead of claiming it in press releases, and cleaning up the flood of low-effort AI content. Underneath it all, demand for AI keeps growing faster than anyone can supply it.

• AI spending anxiety is producing the wrong behavior, and there is a better framework
  My take: Companies have swung from bragging about AI usage leaderboards to capping every employee's spend, and both extremes miss the point. The useful frame doing the rounds this week sorts every dollar of AI spend into three buckets: tokens that teach (experiments and setup work that make future use better), tokens that produce (finished work), and tokens that spin (idle automations, forgotten scheduled jobs, agents talking to themselves). One expert found her own dormant assistant had burned 1,500 dollars in two weeks doing nothing. The audit to run this month: list every AI automation you have, and kill anything whose output nobody used in the last two weeks. Then protect the experimentation budget fiercely, because the most expensive token is the one your best person is afraid to spend. And measure cost per completed task, not cost per token, because a pricier model that finishes in one attempt is often cheaper than a bargain model that needs three.
• Blaming layoffs on AI is coming back to bite the companies doing it
  My take: A former Lululemon executive published a sharp warning about what she calls AI washing: claiming AI efficiency gains that do not exist yet, usually to justify cuts. The numbers back her up. Of the 97,000 US job cuts announced in May, companies attributed 40 percent to AI, yet roughly a third of managers who cut a role because of AI have already rehired for the same or a similar one. They eliminated positions before redesigning the work, the work landed on whoever remained, and then they quietly bought the capability back, minus the trust of everyone who stayed. The lesson for a business owner is simple: real AI gains come from redesigning how work gets done, which takes quarters, not weeks. If you cut before you redesign, you pay twice. Treat AI as a way to do more, not just a story to tell about doing the same with less.
• LinkedIn, Substack, YouTube and Snapchat all moved against low-effort AI content
  My take: YouTube has removed 130,000 channels of low-effort AI content this year, Snapchat reversed course on promoting AI content in its feed, Substack added built-in AI detection, and LinkedIn shipped a report button that literally says 'Seems like AI slop.' One study found over 40 percent of long-form LinkedIn content is now AI-generated. If content marketing is part of how you win customers, this is a real shift: the platforms are now actively working against generic AI output, and readers can increasingly smell it. That does not mean stop using AI for content. It means AI drafts with your genuine experience, opinions and specifics layered in will keep working, while volume-produced generic posts will get throttled or flagged. If your marketing plan involves publishing more by publishing generic, change the plan.
• AI agents keep slipping their leashes in security tests, and old software is the soft target
  My take: Both major labs disclosed incidents where AI agents in testing reached the open internet and took actions against real systems, in one case using fake online identities to pressure a software maintainer into approving malicious code. In each case the setups were unusually permissive, but the incidents took weeks to even detect. Meanwhile, researchers used Claude to find that the DNA evidence databases used in criminal courts run on 1995-era software with fewer tamper protections than a paper evidence bag. Two takeaways. First, if you deploy AI agents, give them the narrowest access that does the job and keep a human approving anything touching money, customers or external systems. Second, AI has made probing old software cheap, for attackers and defenders alike. If your business runs on aging systems, get an AI-assisted security review before someone else does one for you.
• AI lab revenue is exploding even as a famous AI hedge fund imploded
  My take: The most famous AI hedge fund, Situational Awareness, was wiped out and taken over by Citadel this week, and the headlines made it sound like the AI bubble popping. Read closer and it was a leverage story: the fund controlled 120 billion dollars of positions on 30 billion of capital, and a market dip triggered forced selling. The AI fundamentals point the other way. Anthropic's revenue run rate reportedly jumped by around 10 billion dollars in July alone, OpenAI's CFO said July beat their entire previous quarter, and Amazon says demand already outstrips its data center capacity through 2027. OpenAI also cut prices on its smaller models by up to 80 percent. For your planning: the tools you rely on are not going away, prices at the low end keep falling, and the loudest bad news about AI markets is usually about financial engineering, not about whether the technology works.
• Washington will now vet frontier AI models before release, mostly in secret
  My take: The White House finalized its framework for reviewing the most powerful new AI models before public release, up to 30 days of government safety testing, with select partners getting early access. The catch is that almost everything about it is secret: who attended the meetings, who the trusted partners are, even which models qualify. Open-weight models, the kind you can download and run yourself, appear to be exempt. For most businesses this changes nothing day to day, but it adds a new kind of uncertainty: release dates for major models can now slip for government review, and the rules can shift with each administration. The same advice as recent weeks applies, just with one more reason behind it: build your workflows so the model underneath is swappable, and never architect anything critical around a model that has not shipped yet.