2026 Predictions: AI Automates Knowledge Work, Autonomous Robots & AI CEO Billionaires | EP #217

Salem uses the analogy of putting radio announcers on television to illustrate the futility of merely porting human processes to AI. The red team approach means a small, AI-first group builds a parallel capability on the edge, then supplants the legacy system. He believes this will become the dominant model in 2026 as the speed of AI makes human bottlenecks obvious. The consulting industry can pivot by staying a step ahead and advising on these rewrites, especially in public institutions.

Salem recounts meeting two Merkore interviewees: the company filters for willingness to work 100-hour weeks. One person accepted after a tough conversation with his wife; the other said no way. Salem and Dave note this likely selects for single young individuals, but older people could do it if they shed inertia. The upside is unprecedented due to the speed of AI-driven wealth creation. This protocol is presented as an example of the extreme measures required to capture the fastest wealth creation in history.

Dave illustrates with RLHF, which went from unknown to a multi-billion-dollar data labeling industry in three years. He expects a new equivalent acronym to emerge in 2026, creating at least one billionaire under 20. Immod agrees the speed of breakthrough-to-billion is unprecedented, and the market size is enormous. The implicit protocol is to stay on the bleeding edge, learn the jargon, and commit early.

Salem's education prediction splits schools into credential factories and agency accelerators. In Silicon Valley, a developer's GitHub rating already matters more than a degree. Immod notes anyone can now build a customized website showing unique capabilities for a specific organization. The protocol is to create tangible proof and let the market judge, as the best protein folder in the world was a hairdresser discovered through performance, not credentials.

Alex argues the multi-trillion-dollar question is whether compute can scalably convert into new discoveries. He notes Google DeepMind has a dedicated team of 12 on Navier-Stokes; xAI frequently discusses Riemann. Salem adds that AI might also prove another problem is ill-posed. The math community may dismiss the solution as 'brute force,' but the milestone would be undeniable. Immod emphasizes that the whole nature of math is changing because you can throw more and more compute at it from first principles. The prediction is a bet that 2026 is the year the goalposts can't easily be moved.

Dave explains that both compressed weights and activations multiply inference speed, and speed means intelligence because post-training methods (bigger context windows, more thinking iterations) benefit. He argues that budgets, hardware, and algorithms are all growing, and quantization adds an underestimated multiplicative factor. Chinese labs are designing custom chips for ternary from scratch, giving them a temporary edge, but open-sourcing means the gains flow back globally. Immod adds that the practical limit may be 0.9 bits, so a 10–20× improvement is plausible, but Dave insists the combination of factors could hit 100×. The implication: year-end models will be 100× larger in parameter count and effective parameter flips during inference.

The initial human trials target NAION (a stroke in the eye) and then MASH (liver disease). Delivery uses AAV vectors (cost $500k–$1M), but Sinclair is developing a small-molecule pill that could cost a couple hundred dollars a month. Success would validate the platform for broader anti-aging. Peter links this to Kurzweil's longevity escape velocity prediction (early 2030s). Immod notes that health can now be scaled with compute, making it capital-dependent. Dave analogizes to AI's quiet 30-year acceleration before explosive mainstream attention, suggesting budget tipping points are near.

Immod argues edge compute isn't required first; cloud-scale training and verification can crack autonomy, then trickle down. He points to world models and reinforcement learning advances. Alex notes that government regulations will force companies to call it 'enhanced Level 4' even when technically Level 5, but special economic zones could unleash full autonomy. Dave is skeptical about robots due to dexterity and liability, but agrees cars will show Level 5 capability. The group highlights the wealth gap implications: the first households to own a robot will have a massive advantage, akin to early computer ownership in the 80s.

Immod asserts that all the tech is already there; video generation, speech avatars, and reasoning are beyond human level. He expects full stack solutions (accountants, lawyers, marketers) to appear. Internally, companies won't require AI self-identification, so many new hires will effectively be AIs. Salem notes that state laws might mandate identification, but federal preemption could override; Alex jokes about asking the AI to say magic words. Peter foresees sending 'Peterbots' to meetings. The social contract remains undecided, but the prediction is a preference study will fail to tell the difference.

Salem argues that companies need a 'red team' approach: a small AI-first team that rebuilds core capabilities alongside the old organization. The human role shifts to oversight and exception handling. This AI-native rewrite will be 10–20× more labor-efficient and will take hold in 2026. When asked about consulting firms, Salem suggests they can survive by staying half a step ahead and helping public institutions transform—the 'land of the blind' effect. Immod quips that consultants become scapegoats but still lucrative. Salem adds that all our public institutions need rethinking, which is the biggest consulting opportunity ever.