No Priors Ep. 144 | The 2026 AI Forecast with Sarah & Elad

Summary

• AI adoption is outrunning the backlash, but public markets may still treat any Nvidia stumble as a referendum on the whole cycle. Elad expects recycled “AI isn’t working” and bubble claims even though technology waves take “like 10 years to propagate”; he says adoption already looks “blinding,” especially among doctors, lawyers, accountants, and compliance teams. Sarah locates the anxiety in investor positioning and says a weak Nvidia quarter could trigger panic despite having little to do with the underlying secular change. Elad’s answer to skeptics: “So fast I don’t even know what people are talking about.”
• The next application value should concentrate in vertical winners, while Sarah makes the episode’s most explicit trading forecast: somebody will earn “hundreds of millions of dollars” using LLMs in markets in 2026. Elad points to coding, medical scribing, and legal software—including Harvey—as categories already consolidating into a handful of players, and expects the next verticals to reach massive scale. Aaron Levie later argues that agents tied to workflows, organizational data, context engineering, and change management have the clearest enterprise path.
• Robotics is set up for a sentiment break before a business-model break: small-scale humanoid and semi-humanoid deployments will arrive, some will fail, and investors will overreact. Sarah expects bifurcation as projected timelines slip; Elad counters that self-driving took roughly 15–17 years and is only now truly working. Their sharper dispute is structural: Elad points to Waymo and Tesla as self-driving’s current winners and sees a likely Tesla-robot winner, while Sarah argues that capital, hardware, manufacturing, and supply-chain needs can favor incumbents but that locomotion does not solve manipulation, leaving real room for startups and possibly Chinese companies.
• A major AI-lab IPO could become a reflexive benchmark trade rather than a clean fundamental decision. Elad’s hedge-fund example is stark: retail wants a pure-play AI winner, managers fear “missing Nvidia,” and annual benchmarking can force them to buy “regardless” of their company view. Elad expects one successful deal—if not priced too aggressively—to unlock many followers and raise huge sums for labs, even as investors worry whether AI demand supports capex, how credit and pay-on-delivery obligations distribute risk, and how concentrated the market is in Nvidia.
• Elad had not expected many distinct AI consumer products beyond ChatGPT, but now sees “magical” agent experiences he wants to use; Sarah agrees, while Elad still expects most new hardware to fail. Sarah’s durability test is “escape velocity”: without a network effect or another defense, a lab or Google can copy the feature two or three years later and catch up through distribution. The opportunity remains open because incumbents scare founders and too many teams merely rebuild last-generation products with current-generation models.
• Foundation-model competition is reopening around science, alternative architectures, and self-improving systems, yet scale still pulls capital toward a handful of winners. Sarah frames Ilya’s “age of research” as a chance to test compute-efficient ideas, diffusion models, and SSMs rather than fight only a resource war. Elad predicts one or two physics, materials, or math wins will prompt premature claims that “science is solved,” while the long-run impact is understated; his speculative fast-liftoff path is “code plus self-evolution.”
• Capital will keep spilling into defense autonomy and peptide medicine, though defense budgets and the boundary between fringe and mainstream adoption remain unresolved. Sarah expects drone-based defense and startup activity to accelerate; Elad agrees on the need but warns that spending must actually move from prime contractors to new companies at scale. Elad calls GLP-1 adoption “inexorable” and still underrated, while Sarah points to second-order effects and greater interest in engineered peptides and hormone therapies. They discuss fringe biohacking as a possible early indicator, not settled proof.
• The closing forecasts converge on proactive agents and “intelligence per watt” as 2026’s product and infrastructure constraints. Contributors expect AI to gather context, use screens, stop waiting for prompts, and bring a “Claude Code experience” to broader knowledge work; enterprise gains require agent harnesses and economically useful evals. Ben and Ash Spectre argue power is the near-term data-center constraint, but chips matter more over time because they depreciate faster and, at roughly 10 cents per kilowatt-hour, cost about an order of magnitude more than power over five years.

Deep dive

1. Adoption compounds faster than the backlash narrative

• Elad expects another round of pundits proclaiming AI overhyped or ineffective. His counter is temporal: technology waves take “like 10 years to propagate,” while users already receive “enormous value.” The annual backlash will mostly be “a bunch of waste of time.”

• Sarah locates the anxiety in investor positioning: large amounts of capital are deployed against technical and adoption assumptions that managers lack first-principles confidence in. A weak Nvidia quarter could trigger panic, but she sees that as largely disconnected from the underlying secular change.

• An Offcall report on doctor adoption showed amazing uptake across documentation and clinical decision support, including products such as Abridge and OpenEvidence alongside general models. Elad’s broader surprise is that people who tended to be the slowest adopters of technology are now adopting this stuff fast, including physicians, lawyers, certain accounting types, and compliance teams.

• Elad sees coding, medical scribing, and legal software such as Harvey already consolidating around a handful of players, with the next vertical cohort reaching massive scale in 2026. Sarah’s sharper monetization call: somebody will make “hundreds of millions of dollars trading markets with LLMs.”

2. Robotics will disappoint before it scales

• Sarah predicts “some collapse of sentiment” around robotics—not because the field stops progressing, but because companies are promising timelines they cannot all meet. Humanoid and semi-humanoid robots should reach small-scale consumer or industrial deployment; imperfect performance will provoke overreaction and investor bifurcation.

• Elad uses the 15–17-year self-driving curve as the patience benchmark: it is finally working, and 2026 should make autonomy matter in personal cars plus Waymo and Tesla cabs. He describes an autonomous vehicle as “one very complex, single-use-case robot”; Sarah likewise stresses that self-driving is only a single use case as she argues that general-purpose robotics requires more than locomotion.

• In their incumbent debate, Elad points to Waymo and Tesla as self-driving’s current winners, wonders whether Optimus and perhaps Waymo can carry over advantages, and says incumbents generally win as a numbers game. Sarah says capital, hardware, manufacturing, sensors, and supply-chain needs can favor incumbents, but identifies China, Tesla, and possibly a startup as likely robotics winners; she also says Tesla’s driving model may give Optimus locomotion without solving manipulation. She offers Google’s foundation-model comeback as an example of an incumbent slot whose eventual relevance she thought was predictable from its data, capital, hardware, and talent.

3. AI IPOs may become a reflexive benchmark trade

• Elad describes the feared market break: demand may not justify the capex cycle, counterparties may be “passing the ball” through credit and pay-on-delivery contracts for data centers and chips, and investors face concentration risk in Nvidia and a few adjacent names.

• His hedge-fund anecdote captures the reflexivity. A manager invested privately in several labs says he must buy a lab IPO “no matter what I think,” because retail could pop it and benchmarks punish abstention. Elad’s reaction: “This is not the investing job I know how to do.”

• Elad expects one major AI-company offering—provided it is not priced too aggressively—to perform extremely well and pull followers into the market. Retail wants an AI pure play beyond Nvidia, while an IPO lets a capital-hungry lab raise enormous sums.

• Elad says he had not expected so many unique consumer experiences beyond ChatGPT, but now sees “magical experiences” in consumer-agent software that he wants to use; he still expects a slate of consumer hardware to mostly fail. Sarah agrees on the opportunity. Her durability test is “escape velocity”: absent a network effect or another defense, a lab or Google can copy a winner two or three years later and catch up through distribution.

4. The “age of research” reopens the model stack

• Sarah’s neo-lab case starts with Ilya’s “age of research”: scaling still matters, but ideas can be tested above a finite compute floor, potentially producing faster or more efficient improvement. Capital can now test under-scaled architectures such as diffusion and SSMs. Elad’s constraint remains aggregation: scale and revenue eventually channel funding toward approaches already working.

• Elad predicts new physics, materials, and math models will produce one or two striking results—a new material or proved conjecture—followed by exaggerated declarations that “science is solved.” The one-off will be overstated; the long-run trend will be understated and “incredibly important.”

• A closing drug-discovery contributor supplies the concrete version: 2025 progressed from computer-designed small molecules to small antibodies and then full-length antibodies with drug-like properties, zero-shot. Their call is that “2025 has been the year of research” and 2026 becomes “the year of deployment,” including difficult targets traditional techniques struggle to reach.

• Elad’s honest uncertainty: “I always thought, perhaps incorrectly—I actually probably think it’s incorrect—that eventually you end up with evolutionary systems.” He analogizes to specialized brain modules and protein engineering’s move beyond analytical design, then offers “code plus self-evolution” as a possible fast-liftoff route, not a settled forecast.

5. Defense and peptides draw capital beyond the model layer

• Sarah expects drone-based systems and defense startups to drive a “massive reworking” of warfare, aided by the administration’s posture and growing startup density. Elad agrees the competitive need is obvious but warns that budgets must actually shift from prime contractors to new companies at scale. He also argues that booms attract capital and talent even when many companies die.

• Elad’s AI-side call is that GLP-1s remain underrated despite their popularity: adoption is “inexorable,” lower obesity creates broad second-order effects, and visible efficacy plus recognition that delivery matters could open a path for engineered peptide and hormone therapies. Sarah agrees that success will fuel more investment in similar opportunities.

• They discuss biohacking as a possible early-adoption cohort: Elad cites off-label GLP-1 use, peptides, and trips to Dubai for injections; Sarah adds ultrasound neuromodulation and stem-cell injections. His product question—“Is there a Hims of peptides?”—echoes the bodybuilding community’s earlier role in the broader use of creatine and other interventions.

6. Context and efficiency become the agent stack

• Across the closing forecasts, AI stops waiting for prompts: products infer intent from memory, screens, and relevant data sources, while faster inference unlocks new interfaces. One contributor expects a mass consumer agent with a ChatGPT-sized experience jump—“anyone’s ball game”—and another predicts a “Claude Code experience” for all knowledge work.

• Aaron’s enterprise mechanism is more demanding than model access: agents need organizational workflows, internal data, context engineering, and change management. He expects agent harnesses to produce “an order of magnitude improvement” through scaffolding, alongside economically useful evals that measure real knowledge-worker tasks.

• One contributor predicts the U.S. regains the largest-scale open-weight frontier after leadership shifted to China with DeepSeek at the end of 2024. Another expects AI to become a 2026 midterm issue and is “not sure which side’s going to win out”; a third says AI hostility rises even as perceptions of usefulness flip.

• Ben and Ash Spectre call 2026 the year of energy-efficient AI: grid connections and high-voltage equipment constrain data centers, while demand remains “insatiable.” At roughly 10 cents per kilowatt-hour, chips cost about an order of magnitude more than power over a five-year depreciation cycle—so “intelligence per watt” matters now, but chips matter more over the long term.