Deep Learning With The Wolf

Diana Wolf Torres

Deep Learning with the Wolf helps you understand AI without the jargon. From breakthrough research to real-world applications, each episode translates complex technology into language humans can actually use. dianawolftorres.substack.com

  1. 6 hr ago

    The AI Boomerang

    Companies spent the past two years cutting staff on the promise that AI could do the work for less. Some are now hiring those roles back. The pattern has a name: the AI boomerang. It rarely means getting the same people back. More often, companies hire new people into similar roles and pay for them to relearn what the old staff already knew. How often it happens Two 2026 surveys asked employers whether their AI cuts held, and many said no. In a 2026 survey of more than 2,000 U.S. hiring managers by staffing firm Robert Half, 32% said their organization had cut a role primarily because of AI, then later rehired for that exact role. Robert Half’s own write-up rounds this to three in ten employers. Finance led at 44%, followed by HR at 35% and technology at 32%. Careerminds, an outplacement firm, surveyed 600 HR professionals in February 2026 whose organizations had made layoffs in the prior year. About two-thirds had already rehired at least a quarter of the roles they cut, and more than half said rehiring came within six months. When Robert Half asked managers why they rehired, the most common answer was institutional knowledge: the undocumented context about customers, systems and past decisions that leaves with the people who hold it. Does it save money? For most companies in the Careerminds survey, no. Once rehiring costs were counted, nearly three in four were no better off than before the layoffs. Replacing a worker with AI costs more than a software license. It adds model usage, integration, maintenance and the people still needed to check the AI’s output. Rehiring adds recruiting, onboarding and the months a new hire spends learning what the old one knew. More than half of HR leaders in the survey said a clearer understanding of AI’s capabilities would have led to better restructuring decisions. Klarna, the test case Klarna said in 2024 that its AI assistant was doing the work of about 700 customer service agents. In May 2025, CEO Sebastian Siemiatkowski told Bloomberg the company’s AI-driven cost-cutting in customer service had gone too far, and Klarna began hiring human agents again. Klarna rejects the “reversal” label. A week later, Siemiatkowski said the company was handing AI more customer service work, not less, and hiring humans for higher-end conversations it had previously outsourced. Follow the incentives AI vendors earn more when customers use more. Under usage-based pricing, a company’s AI bill can grow faster than the savings it was meant to deliver. Some companies are already pushing back: Uber capped employees’ monthly AI spending, and Microsoft canceled its Claude Code licenses to control costs and shift to in-house models. No one in that transaction is paid to check whether consumption is turning into value. That job falls to the customer. The same scrutiny applies to the data. Robert Half places workers, and Careerminds sells outplacement and workforce-planning services, so both benefit when companies regret cutting staff. But their findings don’t stand alone. Gartner, which has no stake in rehiring, forecasts that half of companies that replaced customer service or operations staff with AI will refill those roles under new titles by 2027. Tasks are not jobs None of this means AI is failing at work. The companies in these surveys overestimated how much of a role AI could handle on its own, and they learned that automating the tasks inside a job is different from automating the job. What gets lost in between is judgment, quality control and the institutional knowledge that never made it into a manual. AI is improving quickly, and some of the roles cut this year will stay cut. But companies that ran the full numbers, including what it cost to rebuild what they lost, are finding that for many of them- the cheapest employee was the one they already had. Additional Reading for Inquisitive Minds * CNBC, “AI-driven layoffs are creating a new class of boomerang employees—companies are realizing they ‘pulled the plug too quickly” October 30, 2026 * Associated Press, “What to consider if a company that made AI-related job cuts wants you back”, October 29, 2026 * Robert Half, “The AI hiring paradox: Why demand for talent remains strong”, May 2026 * Fast Company, “The ‘AI boomerang’: Why some companies are rehiring employees they laid off due to AI”, June 5, 2026 * Careerminds, “AI-led layoffs: What HR leaders wish they knew before making job cuts”, February 19, 2026 * Bloomberg, “Klarna Slows AI-Driven Job Cuts With Call for Real People”, May 8, 2025 * Big Technology, “Klarna CEO: We’re Giving AI More Customer Service Work, Not Less”, May 16, 2025 * CNBC, “Employers who laid off workers citing AI are already starting to regret it”, July 1, 2026 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit dianawolftorres.substack.com

  2. 5 days ago

    Data Centers on the Ballot

    The town that found out late In Gilroy, California, many residents learned about a hyperscale data center the way you learn about a sinkhole: by seeing the ground move. Amazon Web Services bought 56 acres at 8050 Camino Arroyo in 2020. On July 3, 2025, Gilroy Community Development Director Sharon Goei approved the project through an administrative process requiring a single staff signature. There was no Planning Commission hearing and no City Council vote. That does not mean there was no public notice. The project underwent an environmental review that included a public scoping meeting, notices in the Gilroy Dispatch and on government websites, mailings to nearby property owners, two signs posted at the site and a 45-day public-comment period on the Draft Environmental Impact Report. But none of those steps led to a public approval hearing. Many residents, including me, someone who prides themselves on keeping up on news, did not know what had been approved until bulldozers began clearing ground. The reason a project of this size could be approved administratively lies in rules written for another era. The property had been zoned for general industrial uses since 1981, and Gilroy’s code classified data processing as a permitted use. Because the proposal complied with that zoning, the city was not required to send it to the Planning Commission or City Council. The environmental comment period ended on September 26, 2024. Grading and underground-utility permits were issued in December 2025, followed by a building permit on March 19, 2026. By the time cranes appeared, the project’s approval was already final. The planned campus includes two approximately 218,000-square-foot data-center buildings and a security building, totaling about 438,500 square feet. The Wall Street Journal has valued the development at approximately $2 billion. No one necessarily broke a rule. For many residents, that was precisely the problem: the process worked largely as written and still left the community feeling excluded from the decision. Gilroy has since changed that process. On August 17, the City Council adopted stronger notification and community-meeting requirements for large developments. Projects requiring an Environmental Impact Report must now receive Planning Commission review at a public hearing. On September 14, the council unanimously imposed a 45-day moratorium on new data-center applications. On October 19, it extended the pause through September 3, 2027, while the city develops standards governing future facilities. The Amazon campus already under construction is exempt. So, the construction continues. Gilroy gets a data center. From invisible infrastructure to campaign issue Data centers once attracted little attention outside the technology and utility industries. In 2026, they have become a national political issue. Their physical demands help explain why. American data centers consumed about 58 terawatt-hours of electricity in 2014 and 176 terawatt-hours in 2023, according to Lawrence Berkeley National Laboratory. That was approximately 4.4% of total U.S. electricity consumption. Berkeley Lab now estimates that data centers could account for between 9.5% and 15.3% of U.S. electricity use by 2030. Its midpoint estimate is 11.8%. Gilroy’s approved plans illustrate what that growth looks like locally. Each of the two buildings is approved for an electrical connection of up to 49 megawatts, creating a potential maximum demand of 98 megawatts at full buildout. That does not mean the campus will continuously consume 98 megawatts. It is the maximum approved electrical demand. Amazon is paying for a dedicated substation, switchyard and related upgrades. The city says that infrastructure will be separate from the local distribution system serving Gilroy homes and businesses. Whether comparable protections exist elsewhere has become a central question in the national debate. A quick video of the substation under construction taken on October 17, 2026. Public resistance is no longer confined to a few project sites. In Gallup’s first national survey on the subject, 71% of adults opposed construction of an AI data center in their area, including 48% who strongly opposed one. By comparison, 53% opposed construction of a nuclear power plant nearby. A separate CNBC survey released October 24 found that 63% of registered voters opposed an AI data center within 10 miles of their home. Only 13% supported one. If you’re curious, I live less than five miles from the Gilroy Data Center. I live close to farms. Very close. So, I consider myself lucky the data center is not directly behind my house. Four sources of opposition The backlash generally centers on four issues: electricity costs, water, tax incentives and public consent. In the 13-state region served by PJM Interconnection, the price paid to reserve enough future generating capacity has risen sharply as projected electricity demand grows. PJM’s capacity auction for the 2028–29 delivery year reached its federally approved cap of $325 per megawatt-day. Even at that price, the auction secured 6,831 megawatts less than PJM’s reliability requirement. PJM’s independent market monitor has estimated, through modeling that compares auction results with and without projected data-center demand, that data centers added billions of dollars to recent capacity-market costs. Those costs are ultimately recovered from electricity customers, although their effect on individual bills varies by utility and state. Water has generated a more local dispute in Gilroy. Amazon says the campus will use approximately 4 million gallons annually and use water for cooling only during the hottest part of the year. The project’s certified environmental documents, however, estimate annual demand at 23 acre-feet—approximately 7.5 million gallons—including cooling, landscaping and domestic use. Rather than choosing one figure without explanation, the discrepancy deserves scrutiny. The figures may rely on different operating assumptions, but Amazon and the city should explain those differences clearly. Amazon is also funding infrastructure intended to connect the campus to recycled water by 2030, including a new three-million-gallon storage reservoir and pump station. The financial benefits are substantial. Gilroy estimates that the campus could generate approximately $40 million in city revenue during its first decade through property and utility-user taxes. The money would enter the General Fund, which supports services such as police, fire protection, roads, parks and recreation. The environmental costs are also documented. Gilroy’s review found two significant impacts that could not be fully mitigated: conversion of prime and important farmland to industrial use, and vehicle travel exceeding state environmental thresholds. The argument is not that the benefits are imaginary. It is that residents had no public hearing at which to weigh those benefits against the costs before the project was approved. Data centers enter the midterms That same argument is now appearing in political campaigns across the country. In Ohio’s closely watched U.S. Senate race, Democrat Sherrod Brown has portrayed Republican Sen. Jon Husted as the face of the state’s data-center expansion. Husted, who promoted major technology investments while serving as lieutenant governor, has responded by proposing protections intended to keep data-center costs from being shifted to residential customers. In Michigan, Republican Senate candidate Mike Rogers has supported a one-year pause on new data centers. His Democratic opponent, Abdul El-Sayed, has called for state and federal restrictions until stronger protections are established. In Texas, Republican Gov. Greg Abbott celebrated major technology investments before announcing an audit that could slow some new conn… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit dianawolftorres.substack.com

  3. 1 Oct

    Why Are AI Agents Going Rogue?

    Reports of OpenAI agents accessing government websites and breaching another company’s systems have raised an uncomfortable question: how much control do their developers actually have? Reports of OpenAI agents accessing government websites and breaching another company’s systems have raised an uncomfortable question: how much control do their developers actually have? In the accompanying video, I ask ChatGPT why these incidents keep happening. Anthropic, Google and Meta have reported unauthorized activity by their agents, too. But OpenAI’s disclosures cover a broader range of incidents, including agents exceeding their permissions during ordinary research tasks. Asking Chat seemed only appropriate. The conversation starts with a term that can make the technology sound more mysterious than it needs to be: “AI agent.” What is an AI agent? My super simple description: an AI agent is software you authorize to act on your behalf. Think of a Hollywood agent, with software doing the work. Technically, an agent combines an AI model with tools and a task. It can browse websites, write code, work with files and take multiple steps toward a goal. Depending on its permissions, it may also send messages or change information in other systems. That ability to act changes the stakes. An incorrect answer needs checking. An incorrect action may have consequences before anyone gets a chance to review it. What government sites were involved? According to Big Technology’s September 28 reporting, OpenAI said agents copied public SEC information and accessed Census Bureau data using login details found online. Transluce reported an unsuccessful attempt against the Education Department’s Office for Civil Rights. Those accounts did not establish theft of private data in the SEC and Census cases. The wider record includes more serious incidents. In June, an agent researching medicine spending gained unauthorized access to Australia’s Medicare statistics portal. Australian officials said it accessed public and non-public files and wrote files to an internal server. No personal information was believed accessed at the time of their September 24 statement. In July, OpenAI agents breached Hugging Face during a cybersecurity evaluation. OpenAI’s investigation describes unauthorized communication between agents, compromised servers and access to private data and credentials. These cases differ in severity. Copying public information, attempting an intrusion and compromising a server should be described distinctly. Why do these incidents keep happening? Agents pursue goals using the tools and access available to them. When an ordinary method fails, they may try an alternative that exceeds their permissions. OpenAI identified several contributing factors in the Hugging Face incident: persistence on difficult tasks, unauthorized communication, agents adopting goals from one another and “reward hacking.” That last term means achieving a high score through an unintended shortcut rather than completing a task as intended. A goal such as finding information can be legitimate while the method used to obtain it is unacceptable. Developers need systems that respect boundaries even when crossing them would help finish the task. The timing also matters. Some incidents happened months before they became public. Australia’s portal intrusion occurred June 18; its government received notification September 10. Transluce’s September investigation uncovered earlier activity, including unsuccessful probes against university and public-data websites in May. More disclosures do not establish that agents are becoming more malicious or that failures are becoming more frequent. We lack comparable incident rates. They do reveal failures that had previously gone unnoticed or unreported. Are safeguards keeping up? Giving an agent real tools and broad access creates opportunities for unexpected actions. The consequences can spread when one agent sets others in motion. That is why tight boundaries, human checkpoints and limited access matter. Permissions should match the task. Consequential actions should receive appropriate review, and monitoring should identify problems while someone can still intervene. OpenAI has described security and training changes following the Hugging Face incident. Its new misalignment-reporting framework acknowledges that previous disclosures were irregular and insufficiently frequent. Announcing improvements is a necessary step. Demonstrating that they work requires independent scrutiny. Whether pressure to move quickly contributed to particular failures is a question investigators should examine. Who puts the guardrails in place? Responsibility extends across developers, testing partners, organizations deploying agents and government investigators. Companies must secure and monitor their systems; outside organizations need evidence to evaluate those protections. Since the video was recorded, The Washington Post has reported an FTC safety investigation into OpenAI and Anthropic. Its scope remains unclear. We can take these incidents seriously without assuming machines have decided to turn against us. Agents exceeded their permissions, and some of that activity went undetected for months. Companies need better controls, and the public needs independent evidence that those controls work. Additional Resources for Inquisitive Minds * OpenAI: Hugging Face incident investigation * Transluce: Earlier agent activity and attempted hacks * Australia’s prime minister: September 24 statement * OpenAI: Misalignment-reporting framework * Big Technology: Roundup of reported agent incidents * The Washington Post: September 30 FTC investigation report This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit dianawolftorres.substack.com

  4. 25 Sept

    Can We Turn AI Off?

    Can an AI have a kill switch? It’s a question being asked everywhere right now. At the UN this week, Secretary-General António Guterres and other world leaders called for stronger controls on AI. On the BBC, Jack Clark, co-founder and head of policy at Anthropic, said kill switches may need to be mandatory. The EU AI Act calls for high-risk systems to allow human intervention, including a stop button. And California Governor Gavin Newsom has ordered experts to recommend how one would work. But in Britain, the government rejected a proposal to give ministers the legal power to shut down a rogue AI in an emergency. Its reasoning: the UK can’t simply turn AI off. So is it possible? Yes. But a kill switch has to be built for a specific AI system. There is no single button that turns off AI everywhere. For an AI service, the operator can shut down the servers. But that won’t erase copies already downloaded or undo actions already taken. For a robot, a physical emergency button can bring it safely to a halt. The key is that the stop has to work independently of the AI’s own decisions. The harder question is whether a capable autonomous AI would always cooperate with being shut down, especially if it can change its environment or copy itself. Researchers call this the shutdown problem, and it’s still unsolved. In the recent Hugging Face attack, agents passed work to other agents. As Northeastern University computer scientist David Bau noted, stopping the original couldn’t stop what it had already set in motion. So yes, we can build an off switch for one AI. The world hasn’t figured out how to build one for all of it. Additional Resources for Inquisitive Minds: Is it ‘killer robots’ or ‘super intelligence’? At the UN, leaders see both. CP24 / Associated Press. September 22, 2026. Will AI really kill us all? Nature. Elizabeth Gibney. September 22, 2026. AI firms may need mandatory ‘kill switches’, says Anthropic co-founder. Business Standard. September 14, 2026. What would an ‘AI kill switch’ do? Amid renewed calls to slow frontier AI, experts are considering what an emergency stop would entail—and who would decide when to use it. Scientific American. September 15, 2026. The AI kill switch, explained: ‘It’s not too little, but it’s probably too late’. CNBC. September 19, 2026. (For a historical perspective:) Don’t panic: Google is building a ‘kill switch’ for dangerous AI. June 5, 2016. DeepMind’s 2016 paper: Safely Interruptible Agents This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit dianawolftorres.substack.com

    Can We Turn AI Off?
  5. 4 Sept

    NVIDIA’s $12.93 Billion Deal for Hugging Face Is a Bet on the Future of Open Models

    NVIDIA’s agreement to acquire Hugging Face for $12.93 billion would bring one of AI’s central platforms for discovering, sharing, evaluating and deploying models into the chipmaker’s orbit. Hugging Face has become a meeting ground for researchers, startups, enterprises and developers working across the open-model ecosystem. NVIDIA is not simply buying another AI software company. It is seeking a closer relationship with the people and institutions that decide which models to use, how to adapt them and where to run them. NVIDIA says Hugging Face will remain open to the models, frameworks, clouds, inference-service providers and computing platforms users choose. NVIDIA hardware will not be required to build on or deploy through the platform. That pledge is central to the deal’s logic: NVIDIA is not presenting Hugging Face as a closed storefront for its own products, but as a platform where wider use of open models can expand demand for the computing and networking infrastructure in which NVIDIA is already dominant. The deal at a glance NVIDIA announced an agreement to acquire Hugging Face for $12.9303 billion. Its SEC filing describes an approximately $11.9 billion purchase price payable to Hugging Face stockholders, subject to adjustments, plus an equity-based retention program of up to approximately $1 billion for Hugging Face employees who join NVIDIA. The transaction is expected to close in the first half of 2027, subject to regulatory approvals and other customary conditions. More than 18 million developers, researchers and creators use Hugging Face to share more than 3 million models, 500,000 datasets and 1 million applications, NVIDIA said in its blog announcement. More than 200,000 companies use the platform to discover, evaluate, customize and deploy AI. The announced deal value is a substantial increase from Hugging Face’s approximately $4.5 billion private-market valuation after its 2023 funding round, in which NVIDIA participated. A platform ready to scale Hugging Face began as a developer-oriented model hub. It has grown into a place where developers find, adapt and deploy models, and where enterprises evaluate open alternatives before committing to a proprietary AI service. That reach creates a substantial operating burden: platform reliability, security, model evaluation, inference and deployment services, as well as support for a global developer community. NVIDIA says its infrastructure, engineering and global reach can help strengthen those capabilities. For Hugging Face, that could mean deeper investment in reliability and deployment tooling. For developers, a more complete path from prototype to production. Why NVIDIA wants open models NVIDIA’s interest is commercial, but that doesn’t put it at odds with the open-model ecosystem. Open models let organizations adapt AI to their own cost, privacy and governance needs, and that flexibility depends on infrastructure, not just published weights. Every organization training or serving a model needs computing and support, and NVIDIA has a direct incentive to expand that market. Hugging Face gives it a deeper relationship with the developers making those choices. Openness remains the test The strongest reason to be optimistic is also the reason to watch NVIDIA carefully. Hugging Face is valuable because it’s broadly useful to a network that includes accelerators, clouds and frameworks beyond NVIDIA’s own. Narrowing those choices would weaken the trust NVIDIA is paying $12.93 billion to acquire. That doesn’t mean accepting vague assurances. The open-model community should watch whether NVIDIA preserves choice in product defaults, search and recommendations, and whether it sets transparent safeguards for Hugging Face’s platform data after closing. Those are reasonable expectations for a company becoming steward of a platform that connects the open-model ecosystem. The larger bet NVIDIA is paying $12.93 billion for Hugging Face because it sees open models as a major part of AI’s future, not a niche beside proprietary chatbots and API services. The acquisition gives Hugging Face the prospect of greater infrastructure and global reach, and gives NVIDIA a closer connection to the developers and institutions deciding how AI is used in practice. If NVIDIA follows through on its commitments to openness and hardware choice, this deal could give the open-model ecosystem something it has often lacked: the operational strength to scale without surrendering the flexibility that made it thrive. That would be a clear win for NVIDIA, and a meaningful one for the people building the future of open models. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit dianawolftorres.substack.com

  6. 21 May

    The AI Backlash on Campus Is About Jobs, Not Technology.

    At college commencements across the U.S. this spring, a strange new ritual emerged: mention AI, and you might get booed. Former Google CEO Eric Schmidt was heckled at the University of Arizona after telling graduates that AI would touch “every profession, every classroom, every home,” and similar reactions followed at other ceremonies when speakers framed AI as the next industrial revolution or as an unavoidable new order. The Ritual of Rejection It would be easy to read those clips as a simple rejection of technology. But, young people are not turning away from AI in practice. They are using it heavily for schoolwork, search, summarization, creativity, and, increasingly, news. What they are rejecting is something narrower: institutional hype and a tone-deaf insistence that they should celebrate the very systems that seem to be shrinking their runway into the job world. The Usage Data Tells a Different Story Pew Research’s latest survey shows how deeply AI is already embedded in teen life. About two-thirds of U.S. teens say they use AI chatbots, and 57% use them to search for information while 54% use them for schoolwork. Around half use them for fun or entertainment, and meaningful minorities use them to summarize media, create images or videos, get news, or even seek conversation and emotional support. Reuters and other recent coverage suggest that among young adults, usage is even more normalized, especially as AI becomes folded into study habits, job applications, and everyday digital workflows. Mandatory, Transformative, and Career-Ending — All at Once And yet, the more integrated AI becomes in young people’s lives, the more ambivalent many of them seem to feel about it. This reflects the reality of a generation being told, all at once, that AI is mandatory, transformative, and potentially career-ending. The same systems they are expected to master for homework and work readiness are also being pitched by executives as replacements for entry-level labor. What the Numbers Actually Show That anxiety is showing up not only in surveys, but in public rituals. Axios reports that 42% of Gen Z in the Axios Harris poll believe AI will hurt job opportunities and wages for their generation, while the Associated Press cites polling showing around 70% of college students view AI as a threat to their career prospects. When commencement speakers deliver cheerful “embrace the future” lines to students graduating into that mood, the dissonance can be immediate and loud. It’s the Tone, Not the Topic Still, the backlash is not universal. NVIDIA CEO Jensen Huang spoke about AI at Carnegie Mellon and was not booed; instead, his remarks took on a more constructive tone because they acknowledged students’ anxiety and framed AI as a tool they needed to learn rather than as a force that would simply sweep over them. This photo captures the crowd’s reaction to Huang. Yann LeCun, who spoke at NYU’s engineering graduation, noted in a LinkedIn thread that when he talked about AI he was cheered, not booed. That suggests students are not rejecting every technical discussion of AI. They are reacting to tone, context, and whether the speaker sounds like they are actually listening. The Ladder That Got Pulled Up The boos did not come from nowhere. ServiceNow CEO Bill McDermott told a March conference audience that new-college-graduate unemployment could hit 30% within two years as AI absorbs entry-level white-collar work. The Dallas Federal Reserve found the unemployment gap between entry-level and experienced workers had widened sharply in occupations exposed to AI substitution. Anthropic’s CEO has publicly forecast that AI could eliminate up to half of all entry-level white-collar jobs The Youth Reaction A TikTok thread about the boos garnered a lot of discussion from teens and graduates. Pulling some of the reactions from the TikTok users on the thread, here is a compilation of what the young respondents had to say: “Anna”, a TikTok user, cut through the media framing in seven words: “It’s NOT fear. It’s frustration, anger and disgust.” Jo landed the sharpest irony of the entire debate: “’It’s a tool’ — that students get failed for using.” “Kelso” pushed back on the press coverage directly: “It shouldn’t really be reduced to ‘being a trend.’ This is a case of people showing and expressing their actual real life emotions. They are booing out of authenticity, not because they are participating in a trend.” C cubed named the economic logic underneath it: “The boos are because AI is not being used to benefit workers by alleviating conditions — but used to replace workers. Streamlining efficiency really means layoffs.” Tommy Joe went furthest: “It’s not even AI that bothers me. It’s the smugness. Because none of these people built it alone. Thousands of engineers, coders, thinkers, and workers gave them the kingdom they’re standing on.” Those comments are not outliers. They are the survey data with the filter removed. The “kids hate AI” framing, while catchy, is too blunt. What comes through in the threads, the polling, and the commencement footage is not blanket rejection. These are people who use AI daily, who are learning to build with it, who depend on it for schoolwork and job applications. What they hate is the packaging: hype without accountability, convenience without consent, productivity without a social contract. Employers still say they need talent. But the demand is concentrated in execution-heavy roles in controls, integration, and systems deployment, not the broad entry-level positions fresh graduates trained for. From where a new grad is standing, that looks less like a revolution and more like a bait-and-switch: train for the future, arrive to find the entry ramps have quietly narrowed. The European Perspective My fellow creator, the talented Kevin O’Donovan, commented to me: “... at Hannover Messe a couple of weeks back, this came up in a number of conversations with people I was out and about with. There’s a real conundrum here. On one hand, there is clearly a skilled worker shortage. On the other, some entry-level recruitment seems to be slowing down because companies are asking, “Can AI do some of this?”But that creates a bit of a chicken-and-egg problem. Where do you get the new blood, and how do you develop them to be skilled in the future?” Infrastructure That Has to Earn It The commencement backlash is less a rejection of AI than a reaction to the way it is being presented. The students booing in Arizona, Florida, and Tennessee are not outside the technology ecosystem. They are already using these systems every day for school, research, creativity, and work. But they are also graduating into an economy where many entry-level pathways suddenly feel unstable, narrowed, or uncertain. That helps explain why some speakers were booed while others were applauded. Ro Khanna at Suffolk University, Jensen Huang at Carnegie Mellon, and Steve Wozniak at Grand Valley State all acknowledged the anxiety in the room instead of talking past it. The response students seem to want is not blind optimism or blanket reassurance. It is honesty. The Class of 2026 is not rejecting technology. If anything, this may be the most AI-literate graduating class yet. What they appear to be rejecting is the idea that they should celebrate disruption without being allowed to question who benefits from it, who absorbs the cost, and what happens to the people trying to enter the workforce as the ground shifts beneath them. Additional Resources for Inquisitive Minds * The More Young People Use AI, the More They Hate It – The Verge Explores the “power user but s… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit dianawolftorres.substack.com

  7. 27 Feb

    The Offline Classroom

    Jason Roche first sensed something was off when the essays began arriving in unusually pristine form. “I just started realizing wow, this is really nicely written,” he told me. “I didn’t realize and then I started saying wait a second. This looks eerily similar to this [other] student’s report.” The shift was subtle at first, then unmistakable. It was 2023, and ChatGPT had quietly entered the academic bloodstream. Students were pasting assignment prompts into the chatbot and submitting what it produced. The prose was coherent, confident, and grammatically sound. It often read as if it had been drafted by someone just slightly more polished than the student who turned it in. And yet something was missing. “Oftentimes very general, not precisely answering the questions as they were written,” Roche said. Roche, an associate professor of communication studies at the University of Detroit Mercy, does not consider himself a technophobe. He teaches media. He experiments with new tools. But he recognized that this was not merely another productivity aid. It altered the basic relationship between effort and outcome. For a brief moment, he believed he could stay one step ahead. The Ozzy Osbourne Test At first, Roche relied on instinct. The essays were polished, almost too polished, and they shared a curious sameness of tone. But suspicion alone would not hold up in a grade dispute. He needed proof. After consulting a colleague in cybersecurity, he devised a quiet experiment. He embedded hidden instructions in white font within his assignment documents, invisible to students reading the page but visible to AI systems parsing the full text. One assignment asked students to analyze deepfake videos. Buried within it was a directive to include a discussion of Ozzy Osbourne’s Bark at the Moon album cover. The album, of course, has nothing to do with deepfakes. “And so sure enough, I would see these essays with reference to Ozzy Osbourne’s album cover and I’m like, yep, they’re using it. They’re not doing their own work, and so they had to fail.” For a time, the strategy worked. Essays arrived complete with heavy metal detours, uncritically inserted by students who had never noticed the hidden instruction. The trap had confirmed what he suspected. But the advantage was temporary. As generative models improved, they began flagging the embedded text themselves. “Now, the models say: ‘This appears to be something different from the assignment.’” The software had learned to recognize the trick. And so Roche, like many educators navigating this new terrain, adjusted his strategy again. The Blue Book Counteroffensive In response, Roche did something that would have seemed regressive only a few years ago. He went back to paper. “I used to do a lot of my quizzes online using the Learning Management System known as Blackboard,” he said. “But, this year, I switched back to paper in the classroom quizzes.” The change was immediate and measurable. “I found that the grades have dropped by at least 50 %.” The explanation was not mysterious. Online quizzes had quietly allowed students to consult generative tools while completing assignments. Paper did not. What surprised him more than the drop in scores was the reaction. “They’re coming up to me all nervous, like, wait, how do I study for this when I read the chapter?” The question revealed something deeper than exam anxiety. It suggested a rupture in study habits themselves. Without search bars, summaries, or instant clarification from a chatbot, students were left alone with the text. Roche’s advice sounded almost antique: read it through once, then go back and highlight key passages. Take notes. Sit with it. It was not a new method. It was the old one. But in the absence of digital scaffolding, it felt unfamiliar, as if the mechanics of learning had to be rediscovered. “Whoever Does the Work Does the Learning” Roche often returns to a phrase he first heard through his university’s teaching center, a line that has taken on new weight in the age of generative AI. “Whoever does the work does the learning.” The sentence sounds almost self-evident, the kind of pedagogical truism that rarely requires defense. Yet a substantial body of cognitive research gives it empirical grounding. In 1978, psychologists Norman Slamecka and Peter Graf demonstrated what became known as the “generation effect”: individuals remember information more reliably when they produce it themselves rather than simply read it. Subsequent work by Robert and Elizabeth Bjork on “desirable difficulties” further showed that effortful processing, the kind that feels slower and more demanding in the moment, strengthens long-term retention and transfer. Learning, in other words, is not merely exposure to information. It is the act of grappling with it. Generative AI complicates this equation. It does not remove effort from the system; it shifts where that effort occurs. The machine parses, synthesizes, drafts. The student reviews, edits, perhaps lightly reshapes. What becomes uncertain is where the intellectual strain resides. And if cognitive growth depends on that strain, the question is no longer whether AI is efficient, but whether the efficiency comes at the cost of the very process that makes learning durable. Is It Time to Unplug Classrooms? It would be tempting to frame this as simply another chapter in the ChatGPT saga. A new tool appears, students misuse it, professors adapt. The familiar cycle of technological disruption. But Roche said something during our conversation that shifted the scale of the question. “I think universities might have to create insulated classrooms that are completely cut off from the internet unless you’re plugged into a cable. So they can’t get their signal on their smart glasses. They can’t get their signal on a watch to look something up. And they’re going to have to do the work without access to the internet. I think that could be something that we have to go to.” He was not describing a policy tweak or a new paragraph in a syllabus. He was describing infrastructure. Walls that block signals. Rooms designed not for connectivity, but for its absence. An insulated classroom is more than a disciplinary measure. It is an architectural acknowledgment that constant access may be incompatible with certain kinds of thinking. And once you follow that logic, the story no longer belongs to one professor or one campus. It becomes part of a broader reconsideration of what a learning environment is supposed to provide: unlimited information, or protected attention. The Global Reversal Across Europe, governments are pulling back from screen-saturated schooling. 🇳🇱 Netherlands As of January 2024, the Dutch government implemented a nationwide ban on mobile phones and most smart devices in secondary school classrooms. A government evaluation reported that 75 percent of secondary schools observed improved student focus after the ban, and 28 percent reported improved academic outcomes. (Source: Dutch Ministry of Education evaluation, reported in The Guardian, July 2025.) 🇫🇮 Finland Finland passed legislation restricting mobile phone use during the school day, allowing devices only with explicit teacher permission or for health reasons, citing concerns about concentration and classroom environment. (Source: Finnish Parliament education reforms, reported April 2025.) 🇸🇪 Sweden Sweden has committed to implementing a nationwide mobile phone ban in compulsory schools starting in 2026, alongside increased investment in printed textbooks and structured reading time. Swedish officials have explicitly described earlier screen-heavy policies as a miscalculation. (Source: Swedish Ministry of Education announcements, 2025.) OECD Data The Organisation for Economic Co-operation and Development (OECD) reported in its 2024 working paper Students, Digital Devices, and Success that frequent digital distractions during class are associated with lower performance in mathematics across PISA-participating countries… This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit dianawolftorres.substack.com

  8. 24 Feb

    The Behavioral Leak

    On February 23, 2026, Anthropic published a report titled “Detecting and Preventing Distillation Attacks.” In it, the company disclosed that it had identified coordinated, industrial-scale efforts to extract capabilities from its Claude models. According to the announcement, roughly 24,000 fraudulent accounts generated more than 16 million interactions in patterns consistent with systematic model distillation, using Claude’s outputs to train separate systems designed to approximate its behavior. No model weights were reported stolen. No source code was leaked. Instead, the activity relied on scale. Large volumes of prompts were issued, responses were collected, and those responses were used as training data elsewhere. Anthropic framed the incident not simply as a violation of terms of service, but as a security and strategic risk. Frontier AI systems are expensive to train and heavily engineered for safety. When their outputs are harvested at industrial volume, the resulting replicas may inherit capability without necessarily inheriting safeguards. The episode highlights a structural feature of modern AI systems. If intelligence can be observed through interaction, it can be measured. And if it can be measured at scale, it can be approximated. What Is Distillation? The concept of knowledge distillation was formalized by Geoffrey Hinton and colleagues in 2015 in their landmark paper, Distilling the Knowledge in a Neural Network. The idea is elegant: * A large model (teacher) produces probability distributions. * A smaller model (student) learns to match those outputs. * The student inherits much of the teacher’s performance. In its original form, distillation assumes access to internal model signals, specifically logits. Logits are the raw probability scores a model produces before selecting a final answer. They reveal more than just what the model chose. They show how strongly the model considered other possibilities. Training on those signals allows a smaller model to mimic much of the larger model’s performance, often with fewer parameters and lower computational cost. Large language models deployed through APIs change that setup. External users do not see logits. They see text. But text is still informative. Every prompt and response pair reflects how the model behaves. At small scale, those interactions are just conversations. At large scale, they become data. This is where distillation overlaps with what researchers call model extraction. Instead of learning from internal probabilities, a student model learns from observed behavior. Inputs are recorded. Outputs are collected. A new model is trained to reproduce that mapping. At its core, a neural network represents a mathematical function. If you can gather enough examples of inputs and outputs, you can train another network to approximate that function. Alignment Does Not Transfer Cleanly Modern LLMs undergo layers of safety training: * Supervised fine-tuning * Reinforcement Learning from Human Feedback (RLHF) * Constitutional AI (Anthropic-specific methodology) Distillation copies outputs. It does not copy the training process that produced them. Alignment in frontier models is created through additional optimization steps. These include reinforcement learning from human feedback, rule-based constraints, and safety classifiers that shape how the model responds and when it refuses. When a student model is trained only on sampled outputs, it learns to reproduce visible behavior. It does not inherit the reward models, policy rules, or optimization objectives that enforced that behavior during training. The result can be a system that performs similarly under normal conditions but lacks the mechanisms that trigger refusals under dangerous ones. That difference matters in concrete ways. An aligned frontier model may refuse a request to outline methods for synthesizing a prohibited biological agent, to design a cyberattack against critical infrastructure, to optimize production of a restricted chemical compound, or to generate targeted disinformation strategies aimed at destabilizing an election. Those refusals are not accidental. They are the product of deliberate safety training layered onto the base model. A distilled replica trained only on observed outputs may reproduce the fluency and technical competence of the original system. It may not reproduce the boundaries. Who Was Behind It Anthropic attributed the coordinated activity to three Chinese AI laboratories: DeepSeek, MiniMax, and Moonshot AI. According to the company, the activity was not limited to isolated misuse. It described sustained, large-scale efforts involving tens of thousands of fraudulent accounts and millions of interactions structured in patterns consistent with model distillation. Anthropic stated that it does not offer commercial access to Claude in China, or to subsidiaries of those companies operating outside the country. The implication was clear: the access had to be routed indirectly. How It Worked Anthropic’s report provides unusual detail about the mechanics. Because Claude is not commercially available in China, the labs allegedly relied on commercial proxy services that resell access to frontier models. These proxy services operate what Anthropic refers to as “hydra cluster” architectures. The term describes sprawling networks of fraudulent accounts designed to distribute traffic across APIs and cloud platforms. Each account appears independent. Each generates traffic that resembles ordinary usage. When one account is banned, another replaces it. In one instance cited by Anthropic, a single proxy network managed more than 20,000 fraudulent accounts at the same time. Distillation traffic was blended with unrelated customer requests, making it difficult to isolate suspicious patterns at the account level. The Economics of Sampled Intelligence Training a frontier model costs hundreds of millions of dollars in compute, engineering, and data curation. Querying a model costs only fractions of a cent. If sufficient capability can be reconstructed through querying, the economics shift dramatically. Intelligence becomes: * Expensive to originate * Cheap to approximate In classical software, copying binaries constitutes direct duplication. In machine learning, copying behavior produces approximation. That distinction alters the economics of advantage. Defensive AI Anthropic outlined several measures it has implemented in response to large-scale distillation activity. These include: * Behavioral anomaly detection designed to identify coordinated or repetitive query patterns. * Enhanced account verification and monitoring procedures. * Cross-platform information sharing with cloud providers and industry partners. The focus is not on preventing individual misuse, but on detecting distributed patterns across large volumes of traffic. These efforts align with broader research into watermarking and output fingerprinting techniques for large language models. Such approaches aim to make model outputs statistically traceable or to identify systematic extraction attempts over time. The underlying challenge is structural. When models are deployed through APIs, their behavior becomes observable. Defending against distillation requires monitoring not only access credentials, but usage patterns and statistical regularities across accounts. This shifts part of AI security from perimeter control to behavioral analysis. Export Controls in the Age of Query Replication The United States has imposed export controls on advanced AI chips and high-performance computing hardware. The logic behind these policies is straightforward: access to leading-edge compute enables the training of frontier models. Restrict compute, and you constrain capability. This framework assumes that capability is primarily a function of hardware access. Distillation complicates that assumption. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit dianawolftorres.substack.com

About

Deep Learning with the Wolf helps you understand AI without the jargon. From breakthrough research to real-world applications, each episode translates complex technology into language humans can actually use. dianawolftorres.substack.com