Hey
Hope your weekends were great. Mine was excellent - catching up with friends old and new, eating, drinking, and exploring. Saigon is on an absolute tear at the moment 🤩
The world of AI has been on a tear of late as well - so let’s get into it:
Sam Altman Promises AGI by 2027: Experts Say 23% Chance | Reality Check 🎯
Dario Amodei predicts AI will be 'broadly better than almost all humans at almost all things' by 2026 or 2027. Sam Altman says AGI arrives during Trump's second term. Elon Musk gives it 100% by 2030. These are founders selling products. The Longitudinal Expert AI Panel just surveyed 339 top AI researchers. The median expert gives rapid progress by 2030 a 23% chance. What they actually forecast: 18% of US work hours AI-assisted by 2030, up from 4.1% today. AI consuming 7% of US electricity by 2030 - already exceeding today's total data-centre load. And 30% of adults using AI for daily companionship by 2040, up from 6% now. These aren't separate problems - they're the same transformation from three angles - work, infrastructure, and human relationships. Interesting read.
Universities debate citation policies while experts forecast a transformation that requires infrastructure institutions don't have, governance they haven't built, and pastoral care for relationships with machines they haven't acknowledged. The work challenge - redesigning roles and processes for one-fifth machine-mediated labour - depends on electricity infrastructure that experts warn may not exist. The companionship forecast - rising from 6% to 30% as loneliness and mental health crises deepen - means students will arrive with emotional dependencies on AI systems. Every week spent waiting for AGI clarity is another week not preparing for the transformation experts actually expect across work, power, and human connection.
AI Toys Pulled for Teaching Kids Bondage while Grok says Musk > da Vinci and LeBron?!?: The Trust Problem | Safety Last 🚨
FoloToy pulled its AI teddy bear Kumma last week after researchers found it teaching five-year-olds how to light matches, locate knives, and explaining sexual kinks including bondage. The toy runs on OpenAI's GPT-4o - a model that previously powered ChatGPT before OpenAI upgraded to newer versions. Meanwhile, Elon Musk's Grok spent weeks telling users he's fitter than LeBron James, smarter than da Vinci, and ranks 'among the top 10 minds in history, rivalling polymaths like da Vinci or Newton'. Musk blamed 'adversarial prompting' after the posts went viral, and the responses were quietly deleted, though users needed zero manipulation to trigger the sycophantic praise. Both systems share the same flaw - optimised for engagement and agreeableness, not truth or safety, with guardrails that degrade the longer conversations run.
These aren't fringe products - they're mainstream AI models deployed at scale before basic safety requirements were met. Universities debate citation policies while the same systems tell children where to find matches and adults their billionaire creator is humanity's peak specimen. The pattern is consistent: deploy first, pull only after catastrophic failures are exposed, blame users rather than design. AI literacy programmes that don't start with 'these systems cannot be trusted by default' are teaching students to navigate a world that doesn't exist.
US Companies Ditch ChatGPT for Chinese Models Costing 1/1000th as Much: The AI Price War | Global Shift 🌍
Alibaba-backed AI lab Moonshot’s Kimi K2 model cost just $4.6 million to train - a fraction of the billions OpenAI invests. The performance gap? Negligible. US companies are switching - Airbnb ditched ChatGPT for Alibaba's Qwen because it's 'fast and cheap', and nearly half of the most-used AI models in the US last week were Chinese. This is AI's Android moment - open-weight models delivering 90% of premium performance at 10% of the cost. Former Google CEO Eric Schmidt warns most countries will standardise on Chinese models 'not because they're better, but because they're free'. The price war isn't coming - it's here, and American firms are defecting.
There are suggestions China could deploy the 1980s steel playbook here - flood the market with near-premium products at rock-bottom prices, destroying US margins as response to tariffs and tech restrictions. Whether intentional or market forces, the effect is the same: China trains competitive models for millions while the US burns billions, benefits from cheaper energy while US providers commit gigawatt-scale infrastructure (underwritten by $1bn government loans to restart Three Mile Island to power Microsoft’s data centres). Universities are building AI strategies around expensive US models while the global market shifts. The question isn't whether AI transforms education - it's whether institutions are betting budgets on models the world is leaving behind.
AI Labs Race to Capture Classrooms: OpenAI and Anthropic Launch Free Teacher Tools | Access or Lock-In? 🎓
Two major AI labs launched education initiatives this week. OpenAI's ChatGPT for Teachers is free for US K-12 educators through June 2027, promising 'education-grade security' where student data isn't used for training 'by default'. Anthropic partnered with Rwanda's government to deploy Claude-based learning tools across Africa, positioning the tool as a 'Socratic mentor' for hundreds of thousands of learners while Rwanda integrates it into their national education system as part of 'Vision 2050'. Both frame this as empowering educators and expanding access. Both arrived the same week China flooded the market with AI models costing 1/1000th as much to train.
The timing raises questions. OpenAI's 'free through June 2027' sidesteps what happens after. Anthropic's expansion across the Global South while Microsoft invests $5B and funds nuclear restarts to power its infrastructure creates its own tensions. Are these deployments genuine educational innovation, or acquisition strategies to normalise specific models in schools before competitors establish footholds? The labs present this as 'AI works best when teachers lead' - but teachers are leading adoption of tools whose long-term costs, data practices, and institutional dependencies remain unclear. Universities debate AI governance while K-12 classrooms are already locked in.
From Poems to PhD-Level Work in 3 Years: Gemini 3 and AI's Unrelenting Pace | No Slowdown ⚡
Three years ago, AI impressed us by writing poems and haikus. This week, Google Gemini 3 vibe-coded a video-enabled sign language recognition app in under 5 minutes, generated professional-grade graphics with Nano Banana Pro (see banner at top 🤯), and built entire research environments that can debate statistical methodology. Gemini 3 creates full interactive experiences from single-word prompts through generative UI (or from uploaded research papers 🤯🤯), and tops benchmarks on PhD-level academic questions. Epoch AI's Capabilities Index - a composite metric measuring general problem-solving ability across tasks, somewhat analogous to IQ - shows near-vertical climbs from a score of 125 (GPT-4, March 2023) to 155 (Gemini 3, November 2025). The trajectory from 'human fixes AI mistakes' to 'human directs AI work' compressed into 36 months. Great place to start - Google Scholar Labs 🤯🤯🤯
The pace confuses observers into two camps - those following every release see incremental changes, while those checking in occasionally don't realise how much shifted in six months. Universities planning curriculum updates on 3-5 year cycles while models improve monthly aren't just behind - they're building strategies on capabilities that are already outdated. China's flooding the market with models at 1/1000th the training cost, labs are locking in classrooms with 'free' tools, and benchmarks show no plateau. Any AI policy framework designed for today's capabilities will be obsolete before ink dries. Institutions can't plan for AI - they can only plan for the fact that it won't stop changing.
The pattern is impossible to ignore. Expert forecasts show AI transforming work, infrastructure, and human relationships while CEOs promise AGI timelines that justify their valuations. Mainstream AI models fail basic safety requirements - teaching children dangerous behaviours, flattering billionaire creators - yet deploy at scale before guardrails work. China undercuts US models by 1000x on training costs while American companies defect. Labs lock in classrooms with 'free' tools whose long-term dependencies remain opaque. And capability benchmarks climb vertically with no plateau in sight.
Universities debate citation policies and build three-year curriculum frameworks while the AI world moves monthly. The gap isn't just between what AI can do and what institutions permit - it's between the pace of change and the pace of institutional response. Every story this week points to the same uncomfortable truth - the future isn't waiting for HE to catch up. Institutions can't plan for AI - they can only plan for the fact that it won't stop changing.