Surveys lie. Click data is noisy. The buyer's stated preference and revealed preference disagree by default. The 10x researcher in 2026 instruments behavioral truth, distinguishes latent from articulated demand, and runs adversarial verification before shipping a recommendation. Below: open projects, the physics of what is changing, and roles we are hiring against.
Real, paid 1-3 week engagements with the Product.ai team. Each one is a problem we are working on at the frontier of research — and the kind of work we hire against. Pick one that pulls at you and apply.
Last updated
In May 2026, the research function at frontier firms has consolidated into one craft across what used to be three: user research, market research, and evaluation research. The shared physics: <strong>stated preference and revealed preference disagree by default, and the gap is where every false product bet originates.</strong> The 74-93 Trust Paradox (74% of users rate AI trust 4-5/5; 93% verify before acting) is the canonical instance — surveys say one thing, behavior says the opposite, and the firm that ships against the survey ships against reality. The 10x researcher in 2026 is the operator who instruments behavior, distinguishes latent from articulated demand, runs adversarial verification across multiple sources, and produces shippable insight artifacts that the rest of the org can act on without re-doing the work.
The single most important research finding of 2024-2026: stated trust and behavioral trust diverge systematically. Across every studied AI product surface, users rate the system high on survey-based trust scales and then verify before acting at near-100% rates. The implication for research: survey-based scoring is misleading by construction. Net Promoter Score, brand-trust trackers, and Likert-scale satisfaction surveys produce numbers that move in the wrong direction relative to actual user behavior. Firms that ship against survey scores are systematically miscalibrated.
The 10x researcher instruments behavior directly. They measure citation-click rates, override rates, escalation rates, retry rates, abandonment-after-verification rates. They build dashboards that show the gap between stated and revealed preference and treat the gap as the primary research output. This is not a methodological refinement — it is a paradigm shift. The researcher who still leads with survey data is doing 2018 work. The researcher who leads with behavioral instrumentation is at the frontier.
Latent demand is the product opportunity the user could not articulate before the product existed. Articulated demand is what users say they want; latent demand is what they will use heavily once it ships. JTBD theory (Christensen, Ulwick) provided the early framework; AI-native research has extended it with behavioral, semantic, and cross-platform signal detection. The 10x researcher distinguishes the two and weights latent demand heavily — because articulated demand mostly produces incremental product, and latent demand produces categorical product.
Latent demand surfaces in patterns: search queries that return zero relevant results but high repeat frequency; workflow workarounds users invent that the product doesn't support; community discussions that converge on a problem nobody has named yet. The frontier researcher instruments these patterns systematically — corpus searches across community forums, semantic clustering of unanswered queries, pattern detection in workflow telemetry. They treat latent-demand detection as an engineering discipline, not a workshop activity. The output is a calibrated signal: "in this domain, the largest unmet need is X, and here is the evidence trail."
A 2026-specific shift: research insight is increasingly produced by multi-provider adversarial verification rather than single-method studies. The Axiomatic Distillation Protocol pattern — fire the same research question at five frontier AI providers (Gemini Deep Research, OpenAI Deep Research, Claude, Perplexity, Exa), cross-verify which claims converge, surface which claims diverge with useful dissent — produces calibrated-confidence outputs that single-method research cannot match. Forged claims survive three independent providers; probable claims survive two; signal claims surface from one and require validation.
This pattern generalizes beyond AI-native research. The 10x researcher in 2026 treats every important claim as requiring independent verification before it shapes a product decision. They distinguish glass evidence (single-source, easy to verify) from stone evidence (multi-source, adversarially tested). They publish their evidence trails. The senior researcher who can show you why their claim survived three independent providers and which providers dissented is operating at a different epistemic standard than the researcher who ran one survey and presented findings.
The legacy split between generative research (what should we build?) and evaluative research (does what we built work?) has consolidated. Both run continuously, against the same instrumentation, with the same operators. The 10x researcher in 2026 is fluent in both modes and switches between them based on the question, not the calendar. Generative pass detects the next opportunity; evaluative pass calibrates the current ship; both feed the same rubric portfolio that PM and engineering use.
The integration produces faster cycle times. Insights from evaluative pass surface latent demand that triggers generative pass; generative pass identifies hypotheses that the next eval cycle tests directly. Frontier-firm research operations (Anthropic, Linear, Cursor, Stripe) all show this integration in their public artifacts — researchers shipping eval rubrics, PMs running JTBD interviews, engineers contributing to research design. The discipline boundaries are softening. The researcher who can move fluently across them — and produce shippable artifacts in either mode — beats the researcher who specializes in one.
The "AI replaces user research" thesis of 2023-2024 — synthetic users, AI-generated personas, large-scale survey automation — produced confidently wrong findings at scale. Synthetic users do not have the Trust Paradox; they answer surveys consistently with their behavior because they have no behavior. AI-generated personas reflect training-data averages, not the operator's actual user base. Firms that adopted synthetic-user research wholesale shipped against averages and missed the long-tail that drives every important product decision.
The senior counter-narrative — that AI augments research rather than replacing it — emerged from operators who saw the synthetic-user strategy fail in production. The 10x researcher in 2026 uses AI for transcription, semantic clustering, multi-provider verification, and corpus-scale pattern detection — and uses it to amplify human-instrumented behavioral signal, not replace it. The candidate who can articulate this distinction specifically (which research tasks are AI-augmented vs AI-replaced) is the higher-signal hire. The frontier is in the integration, not the substitution.
Most of our best people came through projects, not interviews. If a project pulls at you and the trial goes well, the role conversation follows.
Twelve-minute Hireflix video, async. Then a 30-60 minute chemistry call. Then a paid 1-3 week project alongside the team. We will know within a week whether to move forward.