2026年科技趋势报告

2026 ContextContext Contrary is a talent and research-driven investment firm. We believethat at the root of every iconic company is one thing: extraordinarypeople. As extraordinary people build companies,But justas important as the fundamental forces driving technology are thesecond-order effects of those trends.they build on topof foundational trends that are shaping the world around us.Each year, our Tech Trends Report isolates the most importantContrary is a talent and research-driven investment firm. We believethat at the root of every iconic company is one thing: extraordinarypeople. As extraordinary people build companies, they build on topof foundational trends that are shaping the world around us. But justas important as the fundamental forces driving technology are thesecond-order effects of those trends. currents within technology and the way social behavior formsaround them. This report is meant to both isolate the foundationalconcepts and hint at the progressing second order effects theycould cause. Because that’s where opportunity lies.currents within technology and the way social behavior formsaround them. This report is meant to both isolate the foundationalconcepts and hint at the progressing second order effects theycould cause. Because that's where opportunity lies. Computational IntelligenceComputational Intelligence FoundationalModelsAl AdoptionCompute Supply & Demand Foundation ModelsFoundationModels Over the last decade, AImodels have rapidlyreached a human baselineof performance across avariety of tasks.Over the last decade, Almodels have rapidlyreached a human baselineof performance across avariety of tasks. Google, Meta, Microsoft,and OpenAI have led thecharge in building notableAI models, followedprimarily by academicorganizations.Google, Meta, Microsoft.and OpenAl have led thecharge in building notableAl models, followedprimarily by academicorganizations. Across companies, modelshave continued to reachhigher benchmarks ofaccuracy across skills likephysics, math, andsoftware engineering.Across companies, modelshave continued to reachhigher benchmarks ofaccuracy across skills likephysics, math, andsoftware engineering. By 2030, Al models acrossdomains like softwareengineering, biology, andmathematics are expectedto be capable of near-perfect accuracy.By 2030, Al models acrossdomains like softwareengineering, biology, andmathematics are expectedto be capable of near-perfect accuracy. Each subsequent model,from Gemini 3 to GPT-5.2,continues to performbetter and better inconducting knowledgework tasks.Each subsequent model,from Gemini 3 to GPT-5.2.continues to performbetter and better inconducting knowledgework tasks. Across the board,state-of-the-artmodels are adheringto scaling laws:using larger sets ofdata and more timeand compute fortraining.Across the boardstate-of-the-artmodels are adheringto scaling laws:using larger sets ofdata and more timeand compute fortraining. Compute, in particular, hadbeen rising since 1950, butdramatically acceleratedstarting in 2010.Compute, in particular, hadbeen rising since 1950, butdramatically acceleratedstarting in 2010. Despite higher volumes ofcompute being used,comparable performancehas been reached withless compute each year.Despite higher volumes ofcompute being used,comparable performancehas been reached withless compute each year. As Al adoption hasoccurred, inference costshave started to declineacross models, bothproprietary and opensource.As Al adoption hasoccurred, inference costshave started to declineacross models, bothproprietary and opensource. However, powerconsumption acrossmodels has continued torise significantly as usageincreases.However, powerconsumption acrossmodels has continued torise significantly as usageincreases. We made a mistake in not being more transparent about OpenAl's involvement. We were restricted frodisclosing the partnership until around the time o3 launched,and in hindsight we should have negotiated hardefor the ability to be transparent to the benchmark contributors as soon as possible. Our contract specificallyprevented us from disclosing information about the funding source and the fact that OpenAl has data access to Despite modelimprovements, skeptics ofmodel performance citemodels accessingbenchmark data intraining, and issues withbenchmark testconfigurations.Despite modelimprovements, skeptics ofmodel performance citemodels accessingbenchmark data intraining, and issues withbenchmark testconfigurations. Meta exec denies the companyartificiallyboostedLlama4'sbenchmarkscores Benchmark TestsAI programs train on questions they're later tested on. So how do we know if they're getting smarter? Gemma (the “open-access"’ cousin of its Gemini product), Microsoft's Phiand Alibaba's Qwen have been trained on the text of popular benchmark tests.rainting the legitimacy of their scores. Think of it like a human student whosteals and memorizes a math test, fooling his teache Some improvements a