No PriorsNo Priors Ep. 144 | The 2026 AI Forecast with Sarah & Elad
CHAPTERS
- 0:05 – 0:35
Year-in-review: mainstream AI, agentic coding, and enterprise pull
Sarah opens by recapping a breakout year for AI: mainstream attention, policy visibility, and explosive usage. She highlights the shift from copilots to coding agents, heavy inference demand, and rapid adoption in medicine, law, and customer support.
- •AI moved from niche to mainstream, including policymaker focus
- •Application layer surge: coding tools shifting toward agents
- •Inference capacity becoming a binding constraint
- •Healthcare adoption: clinical decision support and documentation tools
- •Enterprise functions (law, support) accelerating deployment
- 0:35 – 1:31
Research landscape shifts: neo-labs, open source pressure, and many bets
The conversation frames 2025/2026 as an unusually open research moment, with multiple serious contenders and open source narrowing the gap. Sarah lists emerging research directions and predicts robotics efforts will soon collide with real-world constraints.
- •Multiple frontier players + open source closing the gap
- •Neo-labs funded; narrative shifts toward an “age of research”
- •Exploration areas: diffusion, self-improvement, data efficiency, continual learning, agent collaboration
- •Growing interest in EQ/emotional intelligence and new transformer variants
- •Robotics optimism rising, with reality-checks imminent
- 1:31 – 2:48
Warm-up banter to set the tone: hype, AGI vibes, and a microplastics detour
Sarah and Elad joke about living in “2026” and riff on microplastics as a comedic reset before forecasts. The light segment transitions into Elad’s structured set of predictions.
- •Playful AGI/AI-winter framing as context for forecasting
- •Microplastics/micro-glass joke thread and consumer products tangent
- •Signals the episode format: alternating predictions and reactions
- •Sets a skeptical-but-optimistic tone about hype cycles
- 2:48 – 3:53
Elad’s 2026 forecast: hype backlash + vertical AI winners consolidating
Elad predicts a renewed wave of contrarian takes claiming AI is overhyped, even as real adoption compounds. He also expects further consolidation in key vertical applications beyond coding, scribing, and legal tools.
- •Recurring narrative cycle: “AI is overhyped / not working” returns
- •Technology diffusion takes years despite visible early value
- •Vertical markets consolidating into a few scaled winners
- •Prior examples: coding tools, medical scribing, legal (e.g., Harvey)
- •Next wave of verticals reaches massive scale in 2026
- 3:53 – 6:26
Professional adoption accelerates: doctors, lawyers, compliance as surprising early adopters
Sarah reacts with data points and intuition that adoption is already blindingly fast, especially in conservative professions. Both argue that unstructured-data reasoning makes AI uniquely valuable in professional workflows, even if market sentiment wobbles.
- •Investors anxious about capital deployed vs. timing uncertainty
- •Physician adoption: documentation + clinical decision support (e.g., Abridge, OpenEvidence)
- •Conservative professions adopting fastest: doctors, lawyers, accounting, compliance
- •Unstructured data interaction as the killer capability
- •Public-market jitters (e.g., NVIDIA expectations) may not reflect secular change
- 6:26 – 7:17
Models for science and math: breakthroughs will be overhyped now, underestimated long-term
Elad predicts notable “science model” wins—materials, physics, math—that spark exaggerated claims that science is solved. He argues the near-term impact will be overstated, while the long-run importance will be underestimated.
- •Foundation-model expansion into physical sciences and mathematics
- •Expect a few marquee wins: new material, conjecture/proof, etc.
- •Media/investor reaction: “science is solved” hype cycle
- •Reality: progress will be incremental, but compounding
- •Long-term: scientific workflows profoundly transformed
- 7:17 – 8:09
Robotics reality-check: humanoid pilots, timeline misses, and sentiment shakeouts
Sarah predicts a sentiment “collapse” for some robotics companies as deployments begin and imperfections become visible. She expects small-scale humanoid/semi-humanoid deployments in 2026, followed by investor bifurcation between real progress and failed promises.
- •Robotics hype meets deployment reality; not everyone hits timelines
- •Humanoids deployed at small scale in consumer/industrial settings
- •Early failures trigger backlash despite genuine field progress
- •Investing splits into winners vs. narrative casualties
- •Timeline discipline becomes the differentiator
- 8:09 – 10:25
Self-driving as the leading ‘robot’ success—and who wins: incumbents, China, and capital intensity
Elad argues self-driving will meaningfully “arrive” as a dominant robotics story in 2026, citing incumbents like Waymo and Tesla. They debate whether robotics will mirror AVs: capital and manufacturing favor incumbents, while startups still have openings—especially outside pure locomotion.
- •Self-driving: long gestation, now working; likely a major 2026 topic
- •Incumbent advantage thesis: Tesla/Waymo as AV precedent
- •Robotics winners may include Tesla + Chinese industrial ecosystem + a few startups
- •Capital, supply chain, sensors, and manufacturing as moats
- •Manipulation and hardware integration weaken direct transfer from AV stacks
- 10:25 – 13:53
Defining ‘robot’: appliances vs intelligent generalization
Sarah and Elad spar on what counts as a robot, contrasting “single-use automated machines” with “intelligent, generalizing systems.” The definitional debate clarifies why humanoids and self-driving are treated differently from household appliances.
- •Sarah’s threshold: intelligence + generalization across environments/tasks/objects
- •Elad’s broader view: machines that automate complex actions (even appliances)
- •Examples: dishwashers/vacuums/elevators vs self-driving cars
- •Why the definition matters: innovation pace and investor expectations
- •Frames robotics progress as intelligence + manipulation + robustness
- 13:53 – 16:41
IPOs and M&A: retail appetite, lab capital needs, and the ‘can’t miss it’ trade
They discuss how public-market dynamics may shape AI exits: IPOs could surge if one major AI company successfully lists. Sarah describes hedge-fund game theory—buying IPOs regardless of fundamentals—while Elad expects IPOs to be a powerful funding lever for compute-intensive labs.
- •Risk scenario: demand skepticism and CapEx/credit structure fears
- •Concentration risk around NVIDIA and a few key suppliers
- •Hedge-fund benchmark pressure + retail FOMO drives IPO participation
- •Elad expects more IPOs if a single flagship AI listing performs
- •Public markets as a mechanism to raise huge sums for labs
- 16:41 – 21:12
Consumer AI innovation bottlenecks: scary incumbents, shallow remakes, and scarce great product builders
Sarah predicts lots of consumer AI hardware attempts will fail, but sees promising agentic consumer software emerging. Together they diagnose why consumer innovation has lagged: incumbents absorb features, many startups just remake old products, and truly great consumer product talent is scarce and context-starved.
- •Prediction: many consumer AI hardware products fail
- •Optimism: new consumer agent software can feel “magical”
- •Incumbents ingest successful ideas into platforms; defensibility is hard
- •Founders often build ‘better old products’ instead of new paradigms
- •Constraint: limited pool of exceptional consumer product builders with current AI context
- 21:12 – 23:45
Neo-labs and alternative research directions: scaling limits, new architectures, and inference vs training tradeoffs
Sarah asks why neo-labs are getting funded and what research paths matter next—RL generality, continual learning, diffusion/SSMs, and more. The discussion balances Ilya’s “age of research” view (ideas can beat brute compute) with Elad’s view that revenue-generating inference bootstraps the next training cycles.
- •Neo-lab funding wave signals appetite for new bets beyond incumbent labs
- •Ilya’s framing: compute isn’t infinite; clever ideas can win
- •Testing alternative architectures at scale: diffusion, SSMs, etc.
- •Inference workload competes with research compute; resource allocation matters
- •Elad: scale consolidates winners; capital aggregates around what works
- 23:45 – 26:43
Toward self-evolving AI: biology analogies, modularity, and code as an accelerant
Elad explores a longer-term training intuition: AI systems may evolve via selection-like processes, inspired by specialized modules in the brain and evolutionary methods in protein engineering. He connects this to why code generation matters—code plus self-evolution could accelerate capability gains.
- •Hypothesis: evolutionary/selection dynamics may be a path to stronger AI
- •Biology analogy: specialized brain modules (vision, memory, empathy)
- •Protein engineering precedent: directed evolution outperforming pure analysis
- •AlphaFold as an example of AI leapfrogging traditional methods
- •Code as bootstrap: faster iteration + potential “self-evolution” liftoff
- 26:43 – 30:37
Beyond AI: defense tech acceleration and the biohacking/peptide wave (GLP-1 spillovers)
They shift to non-AI predictions: Elad forecasts defense-tech acceleration via drones and new startup density, while Sarah highlights GLP-1 adoption as still underrated with big second-order effects. Both see peptide therapies and biohacking practices as early signals of broader medical consumerization.
- •Defense tech: shift toward drone-based systems and new procurement patterns
- •Hype cycles can be productive by attracting talent and capital despite failures
- •GLP-1s: continued adoption with major societal second-order effects
- •Peptides/hormone therapies as the next frontier; delivery and usability matter
- •Biohacking as a leading indicator for mainstream medical trends
- 30:37 – 40:46
Lightning round: 2026 predictions from industry friends (reasoning, context, agents, open models, energy)
The episode closes with a montage of guest forecasts spanning technical and societal themes: reasoning systems, proactive agents, context/memory, faster inference, enterprise harnesses, open-model geopolitics, politicization, AI drug discovery deployment, and energy efficiency per watt. It ends with holiday wishes and show credits.
- •Reasoning breakthroughs broaden robustness across industries
- •AI becomes proactive: integrated work companion, coach, and manager
- •Context/memory as product differentiator; less copy-paste, more ambient context
- •Enterprise agents: workflow integration, scaffolding/harnesses, useful evals
- •Open-model competition (US vs China), politicization, drug discovery deployment, and energy-efficient AI
