Hey
Happy New Year! Hope the silly season treated you well and you had a chance to recharge - however that looked for you, wherever you are. Mine was excellent - spent it chilling on the beach in Phu Quoc with extended family. Excellent food, great weather, and fantastic company. Came back happy, tanned, and with a bit more wobble than I went away - goodtimes.
Quite a bit's happened while we've been away. If you missed it, I pulled together the five themes that defined 2025 as HE's AI crisis year (link here) - worth a read to set the scene. But let's not dwell on last year. It's 2026 now, so let's get into the latest headlines:
Gemini 3 Flash is a Beast, But: Google Still Can’t Build the Puzzle | Platform Fragmentation 🧩
Google released Gemini 3 Flash this week - and it is a beast, running at at 3x the speed and a fraction of the cost of previous Pro models while maintaining edge performance (release report complete with nerdy benchmarks here), now rolling out as the default in Gemini app and Search. The technical achievement is real. But then Google do themselves no favours with implementation across the apps in their vast family - every one uses it a bit differently - and the inimitable Ethan Mollick points out that this is actively preventing them from creating the single unified interface for agentic work that would let this capability actually matter (great review of the last 3 years from him here as well btw) They've built all the pieces but haven't assembled them yet.
So what? for HE: we’re already living this fragmentation. Claude for analysis and writing, Gemini for… pretty much anything now, Copilot for Microsoft integration. A recent Google DeepMind paper suggests this might be AGI's actual form - collective intelligence distributed across specialised networks rather than a single superintelligent system - interesting parallels with the human brain. If true, orchestration capability trumps vendor selection. The 2026 winners won't be institutions that picked the ‘best’ model - they'll be the ones who developed capacity to integrate and deploy platforms, tools, and approaches across contexts. Are we building that connective tissue, or waiting for a unified platform that may never arrive?
PS. Long-term, if Google gets their house in order, I feel they take this AI arms race once and for all. They have enormous asymmetrical advantages - Gemini is now a highly competitive frontier model, Google DeepMind keeps pushing things forward (think Alphafold), they have near-infinite multimodal data in YouTube, 3.8bn users on Chrome, even more on GSuite, their own TPU AI chips, and a market cap roughly equal to Germany or Japan. It’s theirs to lose at this point.
The "Federalist Papers" for the AI Age: When Top Economists Start Proposing Robot Taxes, It’s Time to Listen | Governance Framework 🏛️
While universities debate vendor whitepapers, top economists - David Autor, Erik Brynjolfsson, Joseph Stiglitz - just released The Digitalist Papers Vol. 2, their attempt to do for AI governance what Hamilton and Madison did for the US Constitution. All killer, no filler but one standout essay is Autor's "Beyond Job Displacement", introducing the "expertise framework": AI doesn't just automate tasks, it reshapes the value of expertise itself. When AI automates simple tasks, remaining work becomes specialised and wages rise. When AI automates expert tasks, work becomes accessible and wages fall. Same technology, opposite outcomes. Autor explicitly discusses taxing AI services to preserve economic stability - when leading economists propose "robot taxes," disruption isn't hypothetical.
So what? for HE: If AI allows a Year 1 grad to perform like a Year 5 senior, the economic premium on experience - and the degrees that signal it - collapses. I’ve been banging this drum for two years, but if you won’t take it from me, take it from MIT’s Professor of Economics: universities built their value proposition on expertise scarcity. That scarcity is ending. We're moving from an economy of "expertise scarcity" to one of "adaptation speed." Are we preparing students for adaptation speed, or still selling credentials based on knowledge accumulation that AI makes obsolete?
Goldman Sachs Asks "AI Bubble?" While Universities Sign 3-Year Vendor Contracts | Financial Risk 💸
Late last year, Goldman Sachs published "AI: In a Bubble?" and their answer is a nervous "not yet, but...". Then Oracle released earnings on December 11th: shares plunged 13%, wiping $90 billion in market value as the company revealed it burned through $10 billion in cash in six months on AI infrastructure. NYU's Gary Marcus, interviewed for the Goldman report, is blunter: "Generative AI is still autocomplete on steroids" and "we're in a financial bubble even if we're not in a tech bubble”. Writing on his Substack in December, Marcus went further: "Without world models, you cannot achieve reliability. And without reliability, profits are limited. The economics don't make sense, and never will”. A Deutsche Bank survey this month found 57% of institutional investors now cite "AI valuation plunge" as the #1 risk for 2026. The shift is stark: in 2025, companies were rewarded for announcing AI plans - in January 2026, they're being punished for the cost if immediate profit isn't visible.
So what? for HE: Universities are signing 3-5 year contracts with vendors burning VC to keep prices low. If the bubble bursts: vendors collapse, prices 10x when subsidies end, and the "infinite progress" narrative stalls. Are our pilots resilient to vendor bankruptcy? Are our budgets ready for actual compute costs? Or are we building strategy on VC subsidies that were never meant to last?
Evidence Over Speed: ACSES Did the Messy Work Many Unis Skipped | The Patient Build 📋
When it comes to AI, choosing evidence over speed feels radical in 2026 because not everybody does it. Australian universities deployed AI tools just like everyone else - panic-banning ChatGPT, then unbanning it, launching workshops without efficacy data. But the Australian Centre for Student Equity and Success spent three years doing what institutions felt they couldn't always afford to do: building governance on learning science, not panic. The Australian Framework for AI in HE, released in December by researchers including Lodge, Southgate, Gulson, Henderson, Slade, and Bower, deliberately waited for assessment reform work to complete and for research on how students actually learn with AI. The result prioritises what evidence shows: human connection over efficiency, equity-first design, durable capabilities over skills that expire. Lodge frames it as engaging with "the messy, human work of higher education" - the work most institutions skipped.
So what? for HE: Choosing evidence over speed costs money universities don't have and time markets won't wait for. If AI is collapsing the ROI on degrees by making expertise accessible, rushing tools without efficacy data accelerates the crisis. Evidence-based design could preserve value - but requires resources institutions often lack. Many chose the cheaper path - deploy fast, measure later, hope for the best. The Framework proves an alternative exists. The question isn't whether evidence-based governance is right - it's whether the capability-governance gap is now so wide that doing the right thing has become unaffordable.
2026 Predictions: Where Experts Agree (and Where They Don’t) | The Year Ahead 🔮
2026 is when AI meets economic reality - at least according to experts from MIT, Stanford, Wharton, and industry. They converge on three points: measurement replaces hype, Chinese models disrupt everything, applications matter more than infrastructure. But they split hard on fundamentals. Mollick (Wharton) sees no slowdown - C-suite confidence surging, shift to agentic work. Christin (Stanford HAI) predicts bubble deflation as studies show moderate gains and environmental costs. Galloway (NYU and Prof G) calls "massive rerating down" as China effectively weaponises AI dumping. The convergence? Brynjolfsson (Stanford HAI)'s real-time dashboards replace arguments with data, 80% of A16Z startups use Chinese models, and Swisher (Pivot)'s robotics+AI shows value shifting from infrastructure to deployment.
So what? for HE: The divergence matters more than consensus. If Mollick's right about momentum, universities can't wait for perfect governance. If Galloway's right about Chinese disruption, expensive Western tools become unaffordable fast. If Brynjolfsson's dashboards show graduate employment cratering in real-time, the ROI crisis accelerates. These aren't abstract market forces - they determine what's pedagogically possible in September. Vendor collapse means scrambling for alternatives. Chinese dominance means every student has frontier AI regardless of policy. Visible graduate outcome collapse makes unavoidable the curriculum reforms that seemed too hard. The only agreement: arguing about AI's impact ends in 2026. Measurement begins.
So where are we at the dawn of 2026?
Google has frontier models with no unified interface. Autor warns expertise's premium is collapsing. Oracle burned $10 billion proving the financial engine might stall. ACSES released evidence-based governance most can't afford to implement. And experts split on whether 2026 brings momentum or collapse - they only agree measurement replaces arguments.
The pattern across every story: the capability-governance gap widens daily. Students already experience AI-mediated education designed by vendors optimising for engagement, not learning outcomes. The frameworks exist. The research is done.
The question isn't whether AI transforms higher education - it's whether institutions can close the gap between knowing what to do and having resources to do it before the ROI crisis these stories describe becomes irreversible.
2026 is when we find out.