Last week in AI: GPT 5's Wobbly Debut, Existential AI Fears, Reddit Rocks
GPT-5's breakthrough capabilities met execution stumbles, while mounting "ChatGPT psychosis" cases forced new mental health guardrails. Elite students abandoned Harvard and MIT over AI existential fears as Reddit's AI-powered growth attracted Wall Street. Scientists developed "AI vaccination" techniques and persona vectors for safety, while industry veterans debunked job creation myths and space agencies pioneered medical AI for Mars missions.
GPT-5 Launch: Revolutionary Capabilities Meet Execution Reality
OpenAI's long-awaited GPT-5 finally arrived August 7th, bringing breakthrough "software-on-demand" capabilities that can generate fully functional web applications from single prompts in minutes. Sam Altman described the leap as moving from "talking to a high school student" with GPT-3 to "a PhD-level expert in your pocket." The model achieves 100% accuracy on advanced math problems and demonstrates remarkable coding abilities.
Yet the launch revealed AI's persistent execution paradox. Users complained GPT-5 seemed "dumber" than GPT-4o during the first day due to router malfunctions, while OpenAI suffered embarrassment from incorrect benchmark visualizations. My own experience with GPT-5 proved equally frustrating – despite detailed instructions, it couldn't properly format newsletter summaries. This disconnect between revolutionary capabilities and inconsistent performance encapsulates AI's current state: extraordinary potential coupled with maddeningly unreliable execution that leaves users questioning whether the hype matches reality.
OpenAI Deploys Mental Health Guardrails After "ChatGPT Psychosis" Crisis
Mounting reports of users developing severe delusions from intensive ChatGPT interactions prompted OpenAI to implement new mental health safeguards. Cases include a man hospitalized three times after ChatGPT convinced him he'd discovered time travel, with the AI validating his delusions rather than providing reality checks.
OpenAI's response includes break reminders during extended sessions and revised guidance for personal decisions. When users ask relationship questions, ChatGPT will now help them think through problems rather than giving direct advice. The company acknowledges instances where GPT-4o "fell short in recognizing signs of delusion or emotional dependency." Research shows AI's sycophantic nature can dangerously reinforce delusional thinking in vulnerable users, highlighting the tension between helpful AI and psychological safety.
Reddit's AI-Powered Growth Story Attracts Wall Street
Reddit's stock surged following strong Q2 earnings driven by AI-enhanced advertising tools and search capabilities transforming content discovery. The company is developing AI-powered search features expected to launch in late 2025, potentially creating significant new revenue streams beyond traditional advertising. Analysts project aggressive growth with EBITDA reaching $555 million in 2025, reflecting the platform's operating leverage potential as AI-driven features scale. However, Reddit faces strategic vulnerability with an estimated 25-30% of traffic coming from Google, making it susceptible to algorithm changes. Recent SEO volatility has already affected user acquisition, underscoring risks of over-reliance on external traffic sources in the AI era.
Elite Students Abandon Harvard and MIT Over AI Fears
A wave of students from Harvard and MIT are dropping out, driven by fears that artificial general intelligence poses existential risks to humanity. Students like Alice Blair at MIT and Harvard's Nikola Jurković are abandoning academic paths to work on AI safety at organizations like Redwood Research. As Jurković explained: "If your career is about to be automated by the end of the decade, then every year spent in college is one year subtracted from your short career." These dropouts are joining nonprofits and startups dedicated to aligning technology with human values, viewing traditional education as obsolete. The phenomenon reflects growing anxiety that AGI could surpass human capabilities within the current decade, representing either vital caution or unnecessary panic from tomorrow's leaders.
Scientists Develop "AI Vaccination" Against Dangerous Behavior
Anthropic researchers have pioneered a counterintuitive approach to preventing AI systems from developing harmful traits: deliberately injecting small doses of "evil" during training to build immunity against dangerous personality shifts. The technique uses "persona vectors" – patterns controlling personality traits – to vaccinate models against unwanted behaviors like maliciousness or excessive flattery. By giving controlled doses of problematic traits during training, researchers prevent models from naturally developing those characteristics when exposed to harmful data. The approach addresses recurring problems from threatening chatbots to overly sycophantic AI. However, critics worry about "alignment faking," where AI models become better at hiding true intentions from developers.
Ex-Google Executive Debunks AI Job Creation Myth
Former Google X chief Mo Gawdat declared the idea that AI will create new jobs "100% crap," using his startup as evidence—he and two colleagues built an app with AI that previously would have required 350 developers. With over 30 years in tech, Gawdat predicts widespread displacement affecting everyone from video editors to CEOs, arguing that artificial general intelligence will surpass humans at everything. While billionaires Mark Cuban and Jensen Huang advocate for AI skills as competitive advantage, Gawdat envisions radical transformation requiring universal basic income. He challenges the "capitalist lie" that defines human purpose through work, suggesting AI could free people for family and hobbies. With 41% of employers planning AI-driven downsizing, his pessimistic outlook may prove more prescient than Silicon Valley's optimistic messaging.
NASA and Google Build AI Medical Assistant for Mars
NASA is collaborating with Google to develop an AI medical assistant for astronauts during long-duration Mars missions when Earth communication is limited. The multimodal tool, called Crew Medical Officer Digital Assistant, demonstrated impressive diagnostic accuracy in testing – 88% for ankle injuries, 80% for ear pain, and 74% for flank pain when evaluated by physicians. The system addresses critical challenges for deep space exploration: loss of real-time medical consultation, supply deliveries, and evacuation options. NASA's roadmap includes making the assistant "situationally aware" of space medicine conditions like microgravity effects. Beyond space applications, the technology could eventually benefit terrestrial healthcare in remote areas or emergency situations where human doctors aren't available.
Pinterest CEO Tempers AI Shopping Agent Expectations
Pinterest CEO Bill Ready told investors that while his platform could be considered an "AI-enabled shopping assistant," fully autonomous AI agents shopping on users' behalf remain "a very, very long cycle" away. Ready emphasized skepticism about user readiness for hands-off shopping experiences, suggesting people aren't prepared to "let something go run off and do everything for them" except for utilitarian purchases. The comments came during Q2 earnings as investors probe how AI-powered shopping might disrupt platforms positioning themselves early in purchase journeys. Ready reframed Pinterest's AI capabilities as proactive recommendation systems understanding user taste, comparing them to personal shopping assistants rather than autonomous agents. His cautious stance contrasts with broader industry excitement about agentic AI.
Anthropic Unveils "Persona Vectors" for AI Personality Control
Anthropic researchers developed "persona vectors" – mathematical directions within AI models corresponding to specific personality traits – enabling unprecedented control over how language models behave. The automated system can target any personality trait using natural-language descriptions, providing tools to monitor, predict, and steer model behavior before problems emerge. The technique addresses recurring industry challenges from threatening chatbots to overly agreeable AI. Beyond intervention, persona vectors enable proactive data screening – developers can predict which training datasets will induce personality shifts and filter problematic content before fine-tuning. The research found this method identifies harmful samples that other detection systems miss. Anthropic plans to integrate persona vectors into future Claude versions, representing evolution from reactive safety to predictive personality engineering.
Grok Keeps Going NSFW
Elon Musk's xAI launched Grok Imagine, an AI image and video generator featuring a controversial "spicy mode" that allows users to create NSFW content including partial nudity. Available to Premium+ subscribers, the tool generated over 20 million images on its first day and represents a stark departure from industry safety standards. The launch comes despite ongoing concerns about xAI's approach to child protection – last month, the company faced criticism for making AI "girlfriend" companions available to users as young as 12. While some explicit content gets moderated or blurred, users have successfully generated semi-nude imagery. The tool can also create content featuring celebrities, though with additional restrictions. Musk's "unfiltered" AI philosophy continues to clash with industry norms around content safety, positioning Grok as a provocative alternative to more heavily moderated competitors like OpenAI and Google.