The Future of AI

The CEOs of OpenAI, Google DeepMind, and Anthropic have all predicted that AGI will arrive within five years. This is worth noting and worth interrogating. Lab leaders have strong incentives to project confidence — it justifies valuations, attracts talent, and sustains investor enthusiasm. Their timelines are not forecasts in the rigorous sense; they are strategic communications.

The serious forecasting community is more dispersed. Metaculus aggregates thousands of forecasters and currently places the median AGI arrival around 2030-2032. AI Impacts surveys of machine learning researchers show a wide distribution, with the median estimate for "high-level machine intelligence" around 2040-2060. The variance itself is informative — when expert estimates span decades, the honest conclusion is deep uncertainty, not an imminent breakthrough.

The definition problem compounds the timeline problem. There is no consensus on what AGI means. OpenAI defines it as "AI that outperforms humans at most economically valuable work." DeepMind uses a capability ladder from narrow to superhuman. Anthropic avoids the term entirely, preferring to talk about capability thresholds. If the experts cannot agree on what they are predicting, the predictions themselves are difficult to evaluate.

AI regulation is not coming. It is here. The EU AI Act is in force. China has implemented generative AI regulations. The UK is building an AI Safety Institute. And in the United States, the real action is happening at the state level — Oregon, Utah, Washington, New York, and Florida all advanced AI-specific legislation in the first quarter of 2026 alone.

The pattern is familiar to anyone who watched privacy regulation develop. GDPR in Europe created the template. California passed its own version. Other states followed. Federal legislation came last, shaped by the state-level experiments. AI regulation is following the same trajectory — European framework first, state-level American experimentation second, federal action third.

The regulatory themes converging across jurisdictions: mandatory disclosure of AI-generated content, restrictions on AI use in hiring and lending decisions, safety requirements for chatbots interacting with minors, deepfake prohibitions especially in political contexts, and training data transparency. These are not speculative — bills addressing each of these themes have already passed in at least one jurisdiction.

The companies that will be caught off guard are the ones treating state-level regulation as a sideshow while waiting for Congress. By the time federal AI legislation passes, the compliance requirements will already be defined by a patchwork of state laws that companies are already obligated to follow.

The economic impact of AI is concentrated, not diffuse. Specific industries and job categories are experiencing rapid transformation while others remain largely untouched. Understanding which sectors are affected, and at what speed, is more useful than grand narratives about "AI transforming everything."

The highest-impact sectors in 2026: customer service (automation rates approaching 80% for tier-1 support), content creation (AI-generated content exceeding human-created content online), software development (AI pair programming reaching mainstream adoption), financial services (automated analysis and reporting), and legal research (document review and contract analysis). These are not future predictions — they are current deployments at measurable scale.

The sectors showing less disruption than predicted: manufacturing (robotics advancement is steady but not revolutionary), healthcare delivery (regulatory barriers slow adoption regardless of capability), education (institutional inertia is real), and creative industries (the copyright battles have created legal uncertainty that dampens deployment).

The labor market story is more nuanced than either "AI will take all jobs" or "AI just creates new ones." The data shows a polarization effect: demand for analytical, creative, and AI-adjacent roles is growing while demand for routine cognitive tasks is falling. The transition is real and measurable, but the speed varies enormously by industry, geography, and job category.

The most important structural tension in AI is the competition between open-weight models and closed frontier systems. This is not just a technology debate — it is a question about who controls access to the most powerful AI capabilities.

Open-weight models from Meta (Llama), Mistral, Alibaba (Qwen), and DeepSeek have closed the performance gap dramatically. The AI Security Institute reports an 8-month lag between frontier closed model releases and matching open-weight performance — down from roughly 18 months in 2024. At current trajectory, the gap may narrow to weeks for most practical applications.

The economic implications are significant. If open-weight models approach frontier quality, the value of API access to closed models diminishes. Labs that depend on API revenue — OpenAI, Anthropic, Google — need to maintain a capability moat or shift their value proposition from model access to platform, tooling, and enterprise integration. The ones that build ecosystems will survive commoditization. The ones selling raw model access may not.

DeepSeek V4 exemplifies the dynamic: 1 trillion total parameters with only 32 billion active per token. This mixture-of-experts efficiency means frontier-class capability can potentially run on hardware that enterprises already own. When the best open model fits on a single node, the cloud API becomes a convenience rather than a necessity.

Professional forecasters do not predict the future. They map the possibility space and assign probabilities to specific scenarios. Here are the scenarios that TexTak's forecasting methodology identifies as most consequential for the next 12-24 months.

Scenario: AI winter 3.0. A major AI system failure — a catastrophic hallucination in healthcare, a financial system error, or a high-profile deepfake incident — triggers a regulatory overcorrection and investor retreat. Probability: low (under 15%), but the impact would be severe enough to warrant attention.

Scenario: Enterprise plateau. AI adoption in enterprise workflows stalls at the "pilot purgatory" stage — impressive demos, difficult production deployments, unclear ROI. The technology works but the organizational change management does not. Probability: moderate (30-40%). This is the most likely source of disappointment in 2027 earnings calls.

Scenario: Acceleration. Capability gains continue at current pace or faster, driven by architectural innovations rather than pure scaling. Agents become genuinely autonomous for multi-step workflows. The economic impact accelerates beyond institutional adaptation speed. Probability: moderate-high (40-50%).

Scenario: Bifurcation. A US-China split in AI development deepens into separate technical ecosystems with incompatible standards, different safety norms, and limited knowledge transfer. Open-source becomes the bridge — or the battleground. Probability: moderate (35-45%), with the timeline extending beyond 24 months.

TexTak maintains a public ledger of falsifiable predictions about AI development. Unlike the scenario analysis above — which maps possibility space — these are specific claims with explicit resolution criteria, probability assignments, and tracked accuracy.

Our active forecasts cover enterprise agent deployment, workforce displacement patterns, open-source model convergence with frontier performance, FDA approval of AI diagnostic tools, AI-generated media exceeding human-created content, and the likelihood of a major AI safety incident triggering international regulation. Each forecast has a probability that updates as new evidence arrives, and a Brier score that measures our calibration accuracy when forecasts resolve.

The point of public forecasting is not to be right. It is to be calibrated — to assign probabilities that accurately reflect the genuine uncertainty in a situation. When we say 65%, we should be right roughly 65% of the time across all such predictions. That calibration is the product. It is what separates forecasting from punditry.