 |
|
|
|||||
Yann LeCun, Metas outgoing chief AI scientist, says his employer tested its latest Llama model in a way that may have made the model look better than it really was. In a recent Financial Times interview, LeCun says Meta researchers fudged a little bit by using different versions of Llama 4 Maverick and Llama 4 Scout models on different benchmarks to improve test results. Normally researchers use a single version of a new model for all benchmarks, instead of choosing a variant that will score best on a given benchmark. Prior to the launch of the Llama 4 models, Meta had begun to fall behind rivals Anthropic, OpenAI, and Google in pushing the envelope. The company was under pressure to reassert Llamas prowess, especially in an environment where stock prices can turn on the latest model benchmarks. After Meta released the Llama 4 models, third-party researchers and independent testers tried to verify the companys benchmark claims by running their own evaluations. But many found that their results didnt align with Metas. Some doubted that the models it used in the benchmark testing were the same as the models released to the public. Ahmad Al-Dahle, Metas vice president of generative AI, denied that charge, and attributed the discrepancies in model performance to differences in the models cloud implementations. The benchmark-fudging, LeCun said, contributed to internal frustration about the progress of the Llama models and led to a loss of confidence among Meta leadership, including CEO Mark Zuckerberg. In June, Zuckerberg announced an overhaul of Metas AI organization, which included the establishment of a division called Meta Superintelligence Labs (MSL). Meta also paid between $14.3 billion and $15 billion to buy 49% of AI training data company Scale AI, and tapped Scales CEO, Alexandr Wang, to lead MSL. On paper, at least, LeCun, who won the coveted Turing Award for his pioneering work on neural networks, reported to the 28-year-old Wang. LeCun told FTs Melissa Heikkilä that while Wang is a quick learner and is aware of what he doesnt know, hes also young and inexperienced. Theres no experience with research or how you practice research, how you do it. Or what would be attractive or repulsive to a researcher, LeCun said. The division LeCun ran at Meta for a decade, FAIR (Fundamental Artificial Intelligence Research), was a pure research organization that picked its own areas of inquiry. An adjacent applied AI group worked closely with the lab to find ways to use the research in Metas own products. But the organizational changes werent the only reason LeCun wanted to leave Meta. He has long expressed doubts that the current thrust of Metas AI researchlarge language modelswill lead to human-level intelligence because such models cant learn fast and continuously. LLMs can learn a certain amount about the world through words and images, but the models of the future will also have an understanding of the real world through physics. And it’s those world models that LeCun hopes to invent at his new company, Advanced Machine Intelligence. LeCun will act as executive chair, which will allow him to spend much of his time doing research. Alex LeBrun, CEO of French healthcare AI startup Nabla, will become CEO of AMI. Im a scientist, a visionary. . . . I can inspire people to work on interesting things, LeCun told Heikkilä. Im pretty good at guessing what type of technology will work or not.
Category:
E-Commerce
When ChatGPT burst onto the scene, much of academia reacted not with curiosity but with fear. Not fear of what artificial intelligence might enable students to learn, but fear of losing control over how learning has traditionally been policed. Almost immediately, professors declared generative AI poison, warned that it would destroy critical thinking, and demanded outright bans across campuses, a reaction widely documented by Inside Higher Ed. Others rushed to revive oral exams and handwritten assessments, as if rewinding the clock might make the problem disappear. This was never really about pedagogy. It was about authority. The integrity narrative masks a control problem The response has been so chaotic that researchers have already documented the resulting mess: contradictory policies, vague guidelines, and enforcement mechanisms that even faculty struggle to understand, as outlined in a widely cited paper on institutional responses to ChatGPT. Universities talk endlessly about academic integrity while quietly admitting they have no shared definition of what integrity means in an AI-augmented world. Meanwhile, everything that actually matters for learning, from motivation to autonomy, pacing, and the ability to try or fail without public humiliation, barely enters the conversation. Instead of asking how AI could improve education, institutions have obsessed over how to preserve surveillance. The evidence points in the opposite direction And yet the evidence points in a very different direction. Intelligent tutoring systems are already capable of adapting content, generating contextualized practice, and providing immediate feedback in ways that large classrooms simply cannot, as summarized in recent educational research. That disconnect reveals something uncomfortable. AI doesnt threaten the essence of education: it threatens the bureaucracy built around it. Students themselves are not rejecting AI: Surveys consistently show they view responsible AI use as a core professional skill and want guidance, not punishment, for using it well. The disconnect is glaring: Learners are moving forward, while academic institutions are digging in. What an ‘all-in’ approach actually looks like For more than 35 years, Ive been teaching at IE University, an institution that has consistently taken the opposite stance. Long before generative AI entered the public conversation, IE was experimenting with online education, hybrid models, and technology-enhanced learning. When ChatGPT arrived, the university didnt panic. Instead, it published a very clear Institutional Statement on Artificial Intelligence framing AI as a historic technological shift, comparable to the steam engine or the internet, and committing to integrating it ethically and intentionally across teaching, learning, and assessment. That all-in position wasnt about novelty or branding. It was grounded in a simple idea: technology should adapt to the learner, not the other way around. AI should amplify human teaching, not replace it. Students should be able to learn at their own pace, receive feedback without constant judgment, and experiment without fear. Data should belong to the learner, not the institution. And educators should spend less time policing outputs and more time doing what only humans can doguide, inspire, contextualize, and exercise judgment. IEs decision to integrate OpenAI tools across its academic ecosystem reflects that philosophy in practice. Uniformity was never rigor This approach stands in sharp contrast to universities that treat AI primarily as a cheating problem. Those institutions are defending a model built on uniformity, anxiety, memorization, and evaluation, rather than understanding. AI exposes the limits of that model precisely because it makes a better one possible: adaptive, student-centered learning at scale, an idea supported by decades of educational research. But embracing that possibility is hard. It requires letting go of the comforting fiction that teaching the same content to everyone, at the same time, judged by the same exams, is the pinnacle of rigor. AI reveals that this system was never about learning efficiency, it was about administrative convenience. Its not rigor . . . its rigor mortis. Alpha Schools and the illusion of disruption There are, of course, experiments that claim to point toward the future. Alpha Schools, a small network of AI-first private schools in the U.S., has drawn attention for radically restructuring the school day around AI tutors. Their pitch is appealing: Students complete core academics in a few hours with AI support, freeing the rest of the day for projects, collaboration, and social development. But Alpha Schools also illustrate how easy it is to get AI in education wrong: What they deploy today is not a sophisticated learning ecosystem, but a thin layer of AI-driven content delivery optimized for speed and test performance. The AI model, simplistic and weak, prioritizes acceleration over comprehension, efficiency over depth. Students may move faster through standardized material, but they do so along rigid, predefined paths with simplistic feedback loops. The result feels less like augmented learning, and more like automation masquerading as innovation. When AI becomes a conveyor belt This is the core risk facing AI in education: mistaking personalization for optimization, autonomy for isolation, and innovation for automation. When AI is treated as a conveyor belt rather than a companion, it reproduces the same structural flaws as traditional systems, just faster and cheaper. The limitation here isnt technological: its conceptual. Real AI-driven education is not about replacing teachers with chatbots or compressing curricula into shorter time slots. Its about creating environments where students can plan, manage, and reflect on complex learning processes; where effort and consistency become visible; where mistakes are safe; and where feedback is constant but respectful. AI should support experimentation, not enforce compliance. The real threat is not AI This is why the backlash against AI in universities is so misguided. By focusing on prohibition, institutions miss the opportunity to redefine learning around human growth rather than institutional control. They cling to exams because exams are easy to administer, not because they are effective. They fear AI because it makes obvious what students have long known: that much of higher education measures outputs while neglecting understanding. The universities that will thrive are not the ones banning tools or resurrecting 19th-century assessment rituals. They will be the ones that treat AI as core educational infrastructuresomething to be shaped, governed, and improved, not feared. They will recognize that the goal is not to automate teaching, but to reduce educational inequality, expand access to knowledge, and free time and attention for the deeply human aspects of learning. AI does not threaten education: it threatens the systems that forgot who education is for. If universities continue responding defensively, it wont be because AI displaced them. It will be because, when faced with the first technology capable of enabling genuinely student-centered learning at scale, they chose to protect their rituals instead of their students.
Category:
E-Commerce
It’s the first week of January, and you’re already drowning in Slack messages. You told yourself this year would be different, that you’d set boundaries and stop overcommitting. But here you are, saying yes to another meeting you don’t have time for, staying late to fix something that could wait, feeling that familiar knot in your stomach every Sunday night. Across corporate America, 90% of employees are experiencing some level of burnout. For decades, weve been focusing on optimizing our physical health, tracking our sleep cycles, heart rate variability, while the part of us that actually drives our decisions at work, and quality of life, namely our beliefs and emotional patterns, remains almost entirely unmeasured. We blame our schedules, download another meditation app, and tell ourselves we’ll feel better once we find the right morning routine. But as companies prepare to spend $94.6 billion on wellness programs in 2026, it might be worth asking ourselves: What if we started to treat our minds as if they had capacity to improve instead of a crisis to manage? To change the pattern of anxiety and overworking, we need systems that support us on an ongoing basis. The same kind of structure that gets us results in the gym. That means specific targets as opposed to vague intentions, with consistent practice and a way to measure whether anything’s actually shifting. Awareness isnt the finish line Most resources available focus on self-awareness, particularly our ability to notice unhelpful thoughts and identify our triggers. This can help you spot the problem, but it doesnt build the muscle to change it. When were under pressure, we are most likely to default to the identity weve rehearsed the most. If we want different outcomes, we need to do different reps. The ‘mental fitness’ framing Physical training has three basics: assess, train, track. The inner version looks similar: 1. Assess the pattern, not the person Swap Im bad at strategy for Under time pressure, I rush to solutions and skip framing. That tiny pivot turns character judgments into coachable behaviors. 2. Train one thing at a time You wouldn’t walk into the gym and expect to have your desired physique by the end of the first session, so don’t try to reinvent yourself by Friday. Pick one thing that actually matters at work, whether its staying calm when nothing is clear or deploying deliverables when theyre 80% done instead of polishing until it’s perfect. Then do it for two to four weeks, just that one thing. 3. Track signals you can observe Pick leading indicators you can observe daily. Instead of asking “Am I better at communication?,” which measures the outcome, not the action, ask: “Did I pause for three seconds before responding in that tense Slack thread? Did I ask one clarifying question before jumping to solutions? Did I share context to help explain the reasoning behind my response? A simple four-week protocol any team can use In a culture obsessed with novelty, repetition can feel boring, but identity change is about repetition. The mind adapts through patterns, practicing a better version of yourself until it feels natural. Week 0: Baseline Write a short trigger map for the last two weeks at work. Note the situations that spark your worst habits (e.g., shifting scope, senior exec drop-ins, cross-team dependencies) Choose one thing to train, naming the opposite habit youre replacing. Weeks 12: Reps Create a 90-second routine that cues your new identity, such as reading a one-line intention (My opinion matters and I will speak up when needed), breathing for four counts, or previewing one clarifying question youll ask. Come up with three metrics to measure your progress with the new routine after encountering a trigger. For example, after facing a situation that would typically make you angry, ask: Did I pause before responding? Did I ask a clarifying question? Is there something I could have done better? Week 3: Progressive overload Add progressive overload. If you practiced in low-stakes meetings, maybe its time to bring the same behavior to a higher-visibility setting. If you trained with peers, try it with an exec. Week 4: Review and lock in Look back at your checkboxes. Where did the behavior hold under stress? Where did it collapse? Decide whether youd like to keep training this capacity for another block, or maintain it and choose a new one. What managers can do this quarter Leaders shouldnt be expected to fill the role of a coach to build mentally stronger teams. But they can make personal growth operational. This can look like: Normalizing capacity goals. Alongside objectives and key results, ask reports to name one thing theyre training for the quarter and the two behaviors that prove its working. Review those behaviors in 1:1s like you would a KPI. The key is framing it as skill-building, not fixing what’s broken to avoid direct reports feeling judged. Designing meetings for rehearsal so that, if someone is training concise communication, updates are time-boxed to 90 seconds. If another person is training direct feedback, they could be assigned devils advocate as a rotating role. Praise the rep, such as: You paused, reframed, and asked the right question, rather than the persona (Youre a natural). Teams are more likely to repeat what gets recognized. What this looks like in real life A product lead I worked with had a familiar pattern. Whenever requirements changed late in a project cycle, someone from sales would promise a client a custom feature, or leadership would pivot strategy two weeks before launch, she’d panic. She’d call emergency meetings to “align everyone.” Then, to prove she had everything under control, she’d build massive 40-slide decks covering every possible scenario and spend 20 minutes walking through each one while her team’s eyes glazed over. The meetings would drag on for an hour. People would leave more confused than when they arrived. Decisions took forever because there was too much information and no clear ask. She picked one capacity goal: “Create clarity with fewer words,” and to implement it, she did two things: Ask one framing question at the start, and end meetings with a single-sentence summary. Three weeks in, her team was making decisions faster because she changed the shape of conversations, starting with &8220;What decision are we trying to make today?” and ending with “So we’re moving forward with option B and revisiting the API integration next sprint.” Performance improved because she trained smarter. The quiet revolution In the 1970s, jogging was not a thing. Then exercise transitioned from medical advice into identity as people became runners, not because a brochure said so, but because practice made them that kind of person. Work is ready for a similar shift. We dont need more slogans about resilience. We need visible, repeatable ways to become the colleague, the manager, the builder we say we are. Treat your inner game like your training plan: pick the capacity, run the block, count the reps. Your calendar wont change for you. Your identity will, one powerful repetition at a time.
Category:
E-Commerce
All news |
||||||||||||||||||
|
||||||||||||||||||