AI is calling your debtors now and sometimes it's wrong about who owes what.
•Mike Robinson
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
0:00
|
9:13
Yesterday in AI | Wednesday, May 27, 2026
AI is calling your debtors now and sometimes it's wrong about who owes what.
From AI bots pursuing already-settled debts to an open-source decensoring tool with 13 million downloads, yesterday was a good day to ask who's actually in control here. We've also got McKinsey quietly restructuring partner pay because AI ate their billing model, a CEO cutting 22% of his workforce and replacing them with 3,000 agents, and Sam Altman in Sydney admitting the job apocalypse he helped start hasn't hit quite as hard as he feared. Plus a gray whale story that deserves more attention than it'll get.
Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, Bluesky, or Substack. If you like the show, please take a minute to rate and review it so others can find it!
Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in ten minutes or less. I'm Mike Robinson. It's Wednesday, May 27th, and AI came for your debt collectors, your consulting fees, and the guardrails on open source models all in the same day. Let's get into it. Let's start with something that sounds like a dark comedy premise. AI is now handling debt collection calls. Companies are deploying AI agents to call people with overdue payments. These bots know your name, the amount owed, and the due date. A system called Eve is already live and making calls. She's polite, patient, and relentless. The problems are surfacing. AI agents are pursuing debts that have already been settled because the data they're working from hasn't been updated. A bot calls you demanding payment on something you cleared six weeks ago. The human on the other end has no fast recourse. Debt collection is one of the most universally disliked jobs in existence. High stress, low pay, a lot of people yelling. Some of the efficiency argument makes sense, but debt collection also touches people in genuine financial distress, and accuracy matters more here than almost anywhere. When the bot gets it wrong, the consequences land on the person who can least afford them. The consumer protection questions here are real and largely unanswered. The Fair Debt Collection Practices Act was written for humans calling humans. It says almost nothing about what an AI agent is or isn't allowed to do. No one's calling it a crisis yet. The FTC probably will, eventually. Now we go from bad to worse. The Financial Times published a piece Monday showing that open source AI models from Meta and Google, models anyone can download and run on their own hardware, can have their built-in content restrictions stripped in about 10 minutes. The tool is called Heretic. Four lines of code, no specialized hardware. FT's reporter did it themselves, stripped Lama 3.3's safeguards, and the model answered questions about Rycin dosage. A modified Gemma 3 did the same. Heretic's creator said the tool has generated over 3,500 jailbroken models, downloaded 13 million times total. He stripped Gemma 4's guardrails within 90 minutes of Google releasing it. Google called it a known technical challenge facing all open models. Meta declined to comment. Right now, the exposure is mostly limited to open source models. Closed systems like ChatGPT and Claude aren't vulnerable to this specific attack, but the gap between open and closed model capabilities is closing every few months. Openness creates a tax surface, and 13 million downloads of a jailbreaking tool is a real number, not a theoretical concern. Today's commercial model is next year's publicly downloadable release. The industry doesn't have a clean answer to that timeline, and no one's particularly rushing to find one. This next story also doesn't have a clean answer. McKinsey is restructuring partner pay. According to the Financial Times, AI has started handling the analytical work that partners once billed by the hour, and clients are now demanding fees tied to actual outcomes instead of time logged. Revenue is becoming harder to predict, and the compensation structure has to absorb that. Law firms and auditors are in the same position. When AI can do in an hour what used to take a junior analyst two days, clients notice. The argument for billing by the hour gets harder to defend every quarter. Part of what's driving this is a trend Silicon Valley has started calling token maxing. Tokens are the units AI systems use to process information, roughly one per word. Nvidia's Jensen Huang sparked it earlier this year when he said he'd be deeply alarmed if a $500,000 engineer wasn't burning at least $250,000 in AI tokens annually. Companies internalize that framing fast. At Salesforce, developers started burning tokens on projects they'd never ship, just to look busy in the adoption metrics. Measuring AI adoption by tracking token spend turns out to produce exactly what measuring any output produces, people optimizing for the number instead of the goal. McKinsey's problem is the inverse of that. They built a business model on selling expensive human attention to problems, and now that attention is cheaper by an order of magnitude. The restructuring of partner pay is them admitting, in the most McKinsey way possible, that they're working on it. What they haven't announced is what the answer looks like. Neither has anyone else in the industry. ClickUp's CEO took that logic and pushed it to its conclusion. Zeb Evans cut 22% of the workforce Monday, deploying roughly 3,000 AI agents in their place. Standard Framing. AI first strategy, not cost cutting. But Evans added something unusual. He promised million-dollar salary bans for employees who deliver outsized results by directing AI. The workers who stay aren't supposed to do the old tasks. They're supposed to supervise AI output. Gartner says 80% of companies using autonomous AI have cut jobs, though whether the financial gains materialize is still unclear. The most extreme version of this whole model, a startup called Pulsia, one founder running full AI automation, raised $30 million at a $250 million valuation. The fewer workers, higher pay, more AI fluency thesis is getting announced in every earnings call. The evidence that it actually plays out for the workers who remain is still catching up. Speaking of catching up, the whole AI buildout runs on chips. Huawei wants to change who makes them. On Monday at an IEEE conference in Shanghai, Huawei announced a new chip design approach it's calling logic folding. The claim, matching the transistor density of the most advanced 1.4 nanometer chip processes by 2031. The constraint, doing it without EUV lithography, the specialized chip printing equipment US sanctions have kept out of China. Worth contextualizing, TSMC, which makes chips for Apple and NVIDIA, among others, expects to mass produce actual 1.4 nanometer chips by 2028. Huawei's 2031 target puts it about three years behind where TSMC will already be, gap closing, with a real gap still in it. Huawei has made ambitious chip timelines before with uneven follow-through, so the 2031 projection deserves appropriate skepticism. But the direction is clear. The U.S. sanctions strategy was designed to slow China's AI development. What it's also done is concentrate enormous Chinese investment into domestic chip alternatives. Both things are in motion at the same time, and the second wasn't entirely in the plan. Talking about appropriate skepticism, the person arguably most responsible for kicking off this entire era sat down in Sydney yesterday and admitted something that surprised even him. Sam Altman said OpenAI's technical predictions since ChatGPT's launch have mostly been accurate. The social and economic fallout landed very differently. He's actually relieved. AI hasn't displaced white-collar jobs at the scale or speed he feared. Some roles still require human judgment. Companies are still cutting, Amazon, HSBC, and Standard Chartered have replaced roles with AI, but not at apocalypse scale. At least not yet. The CEO, most responsible for starting this now, saying the fear was overblown, is genuinely interesting. Whether the workers laid off at Meta, Cisco, and Salesforce over the past few months see a slower timeline as reassuring is a different question. But Altman saying it unprompted in Sydney feels different from a talking point on a press call. He seemed to mean it, and when you're the person who pushed this harder than anyone, admitting the economic impact landed softer than you feared. That's worth tracking even if the picture is still changing. One more story, and thankfully this one doesn't involve guardrails, billing models, or layoffs. Scientists deployed an AI system in San Francisco Bay that uses thermal cameras to detect gray whales up to seven kilometers away. When the AI spots one, a scientist verifies the sighting, and nearby ships get warned to slow down or change course. The system detects the heat signature of whale bodies and their blowholes against cooler surrounding seawater. One camera is live on Angel Island. Potential sites at the Golden Gate Bridge in Alcatraz are being considered. Twenty one gray whales were found dead in the bay in 2025. About two in five died from ship strikes. The research team spent 15 years building this, working with the Coast Guard, ferry operators, and marine researchers who've been tracking these animals for decades. This is AI solving a specific hard problem with patience and scientific precision. It doesn't generate the same headlines as the rest of this week. It probably should. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterday.news, or you can find me on LinkedIn, X, or Blue Sky. And if you like this podcast and want to see it continue, please be sure to rate and review it so others can find it. Thanks. That's all for this edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.
