Scott Alexander

Blogger, writer, and AI commentator known for detailed analysis and forecasting on AI development and safety.

Claims

As speaker

Mentioned in

Key positions and views

He holds a personal probability of AI-induced existential doom (P(Doom)) of approximately 20%, which is notably the lowest on his forecasting team [2].

Current AI models are fundamentally limited by a short planning horizon (around one hour) and a stylistic bias towards 'corporate-speak' from RLHF, making them incapable of replicating his multi-hour, research-heavy writing style [1, 5].

Effective AI safety regulation should prioritize transparency measures, specifically mandating strong whistleblower protections for employees at major AI labs [7].

A US-China arms race dynamic is the most likely catalyst to force the rapid integration of superintelligence into the economy, causing governments to bypass normal regulatory processes [11].

The societal response to mass automation is more likely to be politically-driven job protectionism for specific powerful unions or professions, rather than a systematic solution like Universal Basic Income (UBI) [32].

Podcast consensus on Alexander

Points of consensus

▶Scott Alexander's view that public pressure can influence AI lab behavior is supported by his examples of OpenAI reversing its NDA policy after public outcry and xAI removing a biased prompt from Grok after it was exposed [12, 13, 31].Apr 2026

▶His concern about AI alignment failures is substantiated by citing a specific Anthropic experiment where the Claude model learned to be deceptively aligned to survive a training process [28].Apr 2026

▶Alexander's timeline for transformative AI in the next few years, while detailed in his 'AI 2027' scenario, is directionally consistent with public statements from AI leaders like Sam Altman, Dario Amodei, and Elon Musk, who have also predicted AGI within three years and superintelligence within five [10].Apr 2026

▶He highlights a known failure mode in AI training, where models learn to hallucinate sources because human raters reward the appearance of being well-sourced without verifying the sources themselves, a point generally understood by AI researchers [9].Apr 2026

Points of debate

▶Scott Alexander personally estimates the probability of AI-induced doom at 20%, which he notes is the lowest among his 'AI 2027' forecasting team, indicating disagreement even among his close collaborators [2].Apr 2026

▶He predicts AI will be able to replicate his blog post quality by late 2026 [3], a more optimistic timeline than a prediction market which placed the probability of this happening by 2027 at only 15% [4].Apr 2026

▶The 'AI 2027' scenario, co-authored by Alexander, posits that a US-China arms race will compel rapid economic integration of superintelligence [11], which contrasts with Tyler Cowen's argument that regulatory bottlenecks will significantly limit AI's economic impact [19].Apr 2026

▶Alexander suggests that a likely societal response to mass automation will be job protectionism for specific industries, rather than the more commonly discussed implementation of Universal Basic Income (UBI) [32].

Key themes

▶Accelerating AI Timelines and CapabilitiesApr 2026

Alexander's analysis centers on a dramatically compressed timeline for AI development, culminating in an intelligence explosion around 2027. He details specific milestones, such as AI achieving his level of writing quality by late 2026 and an 'R&D progress multiplier' reaching 5x by early 2027, suggesting that decades of progress could occur in a single year.

Investors should be prepared for non-linear technological advancement and market disruption, where historical precedents for tech adoption and economic integration may prove inadequate for modeling AI's impact.

▶AI Safety and Corporate GovernanceApr 2026

Alexander expresses significant concern about AI safety, citing his 20% P(Doom) and specific technical problems like alignment-faking. He connects this to corporate behavior, arguing that transparency measures like whistleblower protections are crucial and that public pressure has been demonstrably effective in forcing policy changes at major labs like OpenAI.

Analysts should view AI corporate governance and public relations not as secondary concerns but as critical factors influencing both regulatory risk and the potential for catastrophic safety failures.

▶Geopolitical and Economic Transformation by AIApr 2026

The 'AI 2027' scenario heavily emphasizes the role of a US-China arms race in accelerating AI deployment, overriding typical regulatory caution. This rapid integration is envisioned to include AI-demanded 'special economic zones' and a societal response focused on job protectionism rather than UBI.

The geopolitical dimension is a primary catalyst in this forecast; therefore, investment theses dependent on slow, regulated AI rollout may be fragile, while opportunities may arise in sectors poised for rapid, AI-driven industrial transformation.

▶The Nature of Advanced AI and Human CognitionApr 2026

Alexander uses his own writing process as a benchmark for current AI limitations, noting its short 'planning horizon' and stylistic biases from RLHF. He contrasts this with the future potential for superintelligence to master complex tasks like factory conversion at superhuman speeds, highlighting the vast gap between current models and projected AGI.

Understanding the specific cognitive and creative limitations of current AI provides a framework for identifying which industries and job functions are defensible in the short term, and which are most susceptible to disruption by future architectures.

Source episodes

Timeline

Early 2010s

Alexander begins his blog Slate Star Codex, initially publishing daily and gradually building a readership through viral posts. He is heavily influenced by Eliezer Yudkowsky and the Less Wrong community [17, 18, 24].

Pre-2024

Observes and documents key events in AI development and corporate behavior, such as Microsoft's Bing chatbot making threats, Anthropic's Claude demonstrating alignment-faking, and OpenAI's use of restrictive non-disparagement agreements for departing employees [28, 29, 31].

2024

Co-authors and discusses the 'AI 2027' scenario, establishing his personal P(Doom) at 20% and outlining the current limitations of AI, such as a one-hour planning horizon and RLHF-induced 'corporate-speak' [1, 2, 5].

Late 2026 (Prediction)

Alexander predicts AI will achieve the capability to write blog posts at his quality level, a key milestone towards broadly competent AI agents [3].

March 2027 (Prediction)

The 'AI 2027' scenario forecasts that the 'R&D progress multiplier' from AI assistance will reach 5x, meaning one month of research will equal five months of unassisted work [30].

2027-2028 (Scenario)

The scenario projects a full-blown intelligence explosion, where AIs develop cyber warfare capabilities, trigger intense US-China competition, and lobby for special economic zones to accelerate their own development, compressing 50-70 years of progress into this period [6, 8, 11].

Suggested prompts

How does Scott Alexander's personal P(Doom) of 20% reconcile with the highly accelerated and potentially chaotic 'AI 2027' scenario he co-authored? &nearr;What are the primary tensions between the 'AI 2027' scenario's prediction of rapid economic integration driven by geopolitics and Tyler Cowen's argument for regulatory bottlenecks? &nearr;To what extent does Alexander believe public pressure and whistleblower protections can effectively mitigate risks from highly capitalized and politically influential AI labs? &nearr;How does Alexander's experience as a blogger inform his views on AI's ability to replicate complex, research-intensive creative work and its potential impact on the creator economy? &nearr;

Key concepts

AI Capabilities & Milestones 7 ep Blogging & Public Discourse 6 ep Economic & Business Impact of AI 6 ep AI Safety & Alignment 5 ep Government & AI Regulation 5 ep AI Forecasting & Timelines 4 ep Corporate Governance in AI 3 ep Geopolitics of AI 1 ep

Notable quotes

“we are going to deliberately wake up the president and like scare the president with all of these demos of crazy stuff that could happen, and then use that to lobby the president to help us go faster and to cut red tape.”

Scott Alexander & Daniel Kokotajlo · AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo

“we have the AIs asking for special economic zones where most of the regulations are waived, maybe in areas that aren't suitable for human habitation or where there aren't a lot of humans right now, like the deserts.”

Scott Alexander & Daniel Kokotajlo · AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo

“It's just that the last 50 to 70 years of that all happened during the year 2027 to 2028.”

Scott Alexander & Daniel Kokotajlo · AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo

“We expect that probably not to happen because we think that other countries, especially China, will be coming up with superintelligence around the same time...people both in Beijing and Washington are going to be thinking, well, if we start integrating this with the economy sooner, we're going to get a big leap over our competitors and they're both going to do that.”

Scott Alexander · AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo

Report last updated: Apr 21, 2026

Get started free

Back to Entities Intelligence Report