My Quick Update on Burnout Recovery and What 2026 Actually Has in Store for AI
Hey everyone, David Shapiro here. First, a personal note — my burnout recovery is going better than expected. Doubling down on gut health turned out to be the missing piece. I had a rough couple of days with the usual die-off reaction, but the numbers are finally moving in the right direction, and I genuinely feel better than I have in years. That alone has given me the energy to finish my book, *The Great Decoupling*, and to start talking to a lot more people inside (and adjacent to) the industry.
All of that context matters because what I’m about to share isn’t the speculative futurism I used to do as an outsider. This is what I’m hearing directly from the people building the future — what is actually coming down the pipeline in 2026.
Technical: Multimodal Becomes the Default, Video Changes Everything
The single biggest technical shift next year is that multimodal is no longer a feature — it becomes the baseline. We already have strong image understanding and decent voice, but native video understanding and generation is the leap that’s landing in 2026. Grok 5 and several other frontier models have this essentially locked in.
Once models can watch and generate video in real time, the game changes. Language was the perfect bootstrap because text is easy to tokenize, but now we’re learning how to tokenize everything. A friend pointed this out to me over a year ago, and it’s proving prophetic: there really does seem to be magic in just throwing more modalities at the transformer and letting it figure the rest out.
Anyone still claiming LLMs can’t scale to AGI because “language is a dead end” is missing the point. We haven’t even finished discovering what else we can tokenize, let alone exhausted the algorithmic and scaling improvements still ahead.
Benchmarks will keep saturating, of course, but we’re reaching the familiar tipping point every transformative technology hits — the moment it stops being a toy and becomes undeniably production-ready across a widening set of tasks.
You’ll also see the first serious world models and real-time interactive environments — essentially the early versions of AI-native game engines. Google DeepMind has already dropped strong hints. I don’t expect a full replacement for Unreal Engine 5 or Frostbite in 2026, but we will see playable proofs-of-concept and the first SDKs that make developers realize the creative destruction potential.
Business: 2026 Is the Year Enterprises Finally Say “Agents Are Actually Good Now”
Combine the technical leaps above with maturing software architectures, design patterns, and risk-management frameworks, and 2026 becomes the year a critical mass of companies look up and go, “Wait — agents are legitimately useful.”
Many enterprises are already quietly running pilots today — especially in HR and legal (the two departments every company has) and in whatever niche paperwork or research is specific to their industry. Right now these are controlled, heavily logged, human-in-the-loop systems, but the capabilities are improving fast.
I’ve lived through enough technology-adoption curves to recognize the pattern. Think of where VMware was with GSX (useful for enthusiasts) versus when ESX made server virtualization enterprise-ready for everyone. Or the difference between a BlackBerry and the first iPhone. Agents in 2025 feel like late-stage GSX. 2026 feels like the iPhone moment.
Jobs: The Next Recession Will Be the Catalyst for Mass Agent Adoption
Top-line unemployment still looks great on paper, but dig one layer deeper and the picture darkens — gigification, precarious employment, discouraged workers, and youth unemployment are all rising. A recession (which many smart people still expect in the next 12–24 months) would normally trigger mass layoffs followed by rehiring during the recovery.
Except this time, when the recovery comes, a lot of companies will ask, “Why rehire humans when agents can handle the workload?” We saw the same dynamic with remote work during the pandemic — necessity forced adoption, and the new standard stuck. The next downturn will do the same for AI agents. Jobless recoveries will become the new normal.
Humanoid Robots: 2026 Is When We Hit Product-Market Fit
I called 2026 the “year of the robot” almost exactly a year ago, and everything I’m seeing reinforces that. We went from POC [proof-of-concept] (Optimus walking on stage, Atlas doing parkour) to MVP (minimum viable product) in 2025. In 2026 at least one — and probably several — companies (Figure, Tesla Optimus, Unitree, Boston Dynamics Atlas, others) will cross into genuine product-market fit (PMF).
That doesn’t mean the robots will be perfect or cheap overnight. It means the capabilities will finally justify the price tag well enough that demand explodes and the race to scale manufacturing begins in earnest.
Bonus Reality Check: The “Shortage” Conversation
I asked around (industry contacts, AIs, Twitter) whether anyone is genuinely worried we’ll run out of chips or power. The consensus: there will be a short-term squeeze — electricity prices near data centers may jump 20% or so, and leading-edge GPUs will stay constrained — but it’s a squeeze, not a hard wall.
Power grids adapt slowly, but co-location with new generation (gas peakers, solar farms, restarting nuclear plants) is already happening. Chip fabs take years to build, yet existing fabs keep churning, and the price signal is so strong that new entrants and expansions are pouring in. Where massive money flows, bottlenecks get solved surprisingly fast.
So yes, 2026–2028 might feel tight in places, but the market is already routing around the constraints. Progress won’t stall.
That’s the view from where I’m sitting today. Thanks for reading — I’ll talk to you all again soon. Take care.
