Explore AI psychosis, workforce burnout, and emerging talent profiles in 2026. Learn why AI fluency alone isn't enough for success in the AI-native era.
AI Psychosis in 2026: Understanding Talent, Burnout, and the Future of AI-Native Work
Executive Summary
- AI adoption has reached critical mass: Claude Code and Codex are no longer specialized tools but baseline capabilities for all knowledge workers in 2026
- AI psychosis is becoming prevalent: Workers experience constant dopamine surges from AI task completion, leading to exhaustion and burnout despite increased productivity
- Two talent profiles are emerging: Domain experts leveraging AI within their field, and meta-optimizers who delegate even goal-setting to AI systems
- Infrastructure and control layers matter most: Companies competing with general AI tools must own specialized inference infrastructure, customer data, and vertical-specific tools
- The era of AI applications is beginning: Beyond AI tools themselves, the next wave involves domain-specific harnesses and vertical solutions built on top of open AI infrastructure
- Slow AI movements are emerging as counterbalance: Some thought leaders are questioning the sustainability of relentless AI acceleration and advocating for mindful adoption
The Acceleration of AI Adoption: From Niche to Mainstream
Since early 2026, the pace of AI integration into workflows has become almost unrecognizable compared to just six months prior. What began as discussions around Claude Code and OpenAI's Codex has evolved into something far more comprehensive—these tools are no longer optional enhancements but rather foundational capabilities expected across all knowledge work domains.
The timeline speaks for itself. In January 2026, organizations were still debating whether to invest in AI programming tools. By May 2026, the conversation has fundamentally shifted. Not everyone uses these tools, certainly, but everyone is aware of them and understands their transformative potential. The knowledge workers who have integrated Claude Code or Codex into their daily workflows represent a rapidly expanding cohort that continues to grow at an accelerating pace.
Parallel to this adoption curve, the underlying AI models themselves are improving at a breathtaking rate. According to industry insiders and emerging data, model performance continues its steep J-curve trajectory. GPT-5.5 is now the current frontier model, with GPT-5.6 rumored to arrive soon. Meanwhile, Google's competitive response through initiatives like Google AI Studio suggests a shift toward more implicit workflows—rather than requiring explicit communication back-and-forth with language, users can state intentions roughly and the system delivers precise outputs.
The computational resources required to maintain this acceleration are staggering. Companies like Google, OpenAI, and others are continuously expanding infrastructure, improving efficiency at both engineering and algorithmic levels, and investing heavily in inference optimization. DeepSeek's V4 model, for instance, introduced algorithmic innovations that challenged the assumption that only NVIDIA-based systems could power frontier AI. This ecosystem diversification—moving away from NVIDIA dependence—suggests a maturing AI infrastructure landscape.
Model release cycles have compressed dramatically. What was once a major annual event has become a bi-monthly occurrence, with each increment representing what would have been a major update by 2024-2025 standards. Rumors of a two-month release cycle for frontier models reflect the intensity of competition and the pace of advancement. This creates a unique challenge: human cognitive capacity struggles to keep pace with technological change, even as the same AI tools are supposed to amplify human capability.
The Emergence of AI Psychosis: Symptoms and Manifestations
Within this context of explosive productivity gains, a concerning phenomenon has begun to dominate industry conversations: AI psychosis. This term, popularized by figures like Andrej Karpathy and discussed extensively across technology platforms, describes a state of hyper-immersion in AI-assisted work that generates persistent dopamine surges while potentially harming long-term cognitive health and work-life balance.
The mechanics of AI psychosis are deceptively simple. When using tools like Claude Code or OpenAI's Codex in recursive loops (often called "Ralph loops"), users experience a rapid cycle of task delegation and result completion. Each successful task generation triggers a dopamine spike—what some describe as a "slot machine effect." The user sets a goal, the AI executes it, and the immediate positive feedback creates strong behavioral reinforcement. Unlike traditional slot machines, however, this cycle appears productive. Work is genuinely being completed. Outputs are tangible. The user feels they are accomplishing remarkable things in compressed timeframes.
This creates a psychological trap. Because the work feels meaningful and the results are undeniable, individuals rationalize increasingly extreme work patterns. Teams observe colleagues working continuously through meals, using SSH connections or iPads to check on AI agents mid-lunch, assigning tasks at all hours. The perception spreads that this is normal, even necessary, for success in the AI era. Missing this wave of productivity means falling behind. The FOMO (fear of missing out) compounds the pressure.
Several concerning behaviors emerge from this state:
Task postponement through illusion of completion: When someone writes a task in their to-do list but then immediately delegates it to an AI agent, they experience the psychological satisfaction of task completion without actually doing the work. This creates a false sense of progress and can lead to accumulated, unfinished work at deeper levels.
Inability to disengage: The constant availability of AI tools, combined with the speed at which results are returned, creates a compulsion to keep working. There's always another task, another optimization, another possibility. The friction that once naturally limited work hours—the requirement to context-switch, find resources, or wait for human collaboration—has been eliminated.
Loss of learning and depth: Paradoxically, despite accelerated task completion, individuals may not be deepening their domain knowledge. If AI is handling problem-solving, analysis, and even goal formulation, the person becomes dependent on the AI for cognitive work. This represents a fundamental shift from work-as-learning to work-as-delegation.
Burnout despite apparent productivity gains: The irony of AI psychosis is that individuals can appear extremely productive while experiencing severe exhaustion. They're completing massive amounts of work, yet the constant context-switching, the dopamine cycles, and the inability to disengage create a state of chronic activation that manifests as burnout.
Andrej Karpathy's public acknowledgment of being in this state carries particular weight because of his expertise and position. He has cautiously acknowledged using tools like Cursor (an AI-enhanced code editor) and agents in ways that approach maximum automation, experiencing continuous "dopamine surges" that keep him in a state of over-immersion. Notably, even Karpathy—with his deep technical foundation and understanding of AI systems—frames this as problematic, suggesting that burnout from AI acceleration is not a failure of individual willpower but rather a systemic issue with how AI tools are designed and how society is incentivizing their use.
Two Emerging Profiles of AI Talent
As organizations grapple with AI adoption, a clear distinction is emerging between two fundamentally different approaches to leveraging AI tools. Understanding these profiles is crucial for companies seeking to hire, and for individuals evaluating their own relationship with AI-accelerated work.
Profile One: Domain-Expert Leveragers
The first profile consists of highly skilled professionals with deep domain expertise who use AI as a force multiplier within their field. These individuals have clear problem-definition capabilities and specific knowledge about what needs to be solved. Examples include experienced lawyers, investment bankers, life scientists, and domain specialists who have spent years accumulating field knowledge.
For these professionals, AI tools like Claude Code or Codex become extensions of existing capability. They understand the vertical domain deeply enough to:
- Precisely define problems within their field
- Recognize when AI outputs are appropriate versus dangerous
- Validate AI-generated solutions against domain-specific requirements
- Integrate AI work into established professional workflows
The archetypal example is the "10x Lawyer" or "10x Banker"—professionals who have completely rewritten their business by integrating AI while maintaining expert judgment over the process. They still control decision-making, still understand the implications of what they're delegating, and still bear responsibility for outcomes. They've achieved what might be called AI-native workflows within their domain: problems that once required teams are now handled by the individual plus AI agents.
These professionals are attractive to companies because they bring not just AI fluency but also the judgment to know when and how to apply it. They're less likely to be burned out because they maintain domain mastery and can exercise decision-making autonomy. However, they're also selective about which problems to delegate, maintaining some cognitive engagement with their work.
Profile Two: Meta-Optimizers
The second profile represents a fundamentally different approach: individuals who are unclear about specific goals but instead use AI to optimize the goal-setting process itself. They delegate not just task execution but also task definition to AI systems. They run automation loops (like OpenCode with embedded LLM loops) that continuously search for what they should be doing next.
This approach offers remarkable advantages. A meta-optimizer with limited specific knowledge in a domain can:
- Query research databases, papers, and existing companies
- Use AI to synthesize unfamiliar information
- Have AI identify patterns and opportunities within that synthesis
- Identify the shortest path to understanding or solving something
- Acquire functional domain knowledge rapidly
The paradigmatic example is using OpenCode to study longevity science when you have no formal background. After a month of directed AI-assisted learning, it's possible to have contextual conversations with domain experts. The AI becomes an eternal tutor, breaking down concepts, identifying connections, and adapting to your learning pace.
However, this approach carries hidden risks. While task completion can be sustained, personal development may not be. If an individual hasn't spent years accumulating foundational knowledge, their ability to judge AI outputs remains suspect. They may complete many tasks and build products without truly internalizing the domain knowledge required for wisdom. This is workable for many B2C and B2B applications (like building a character chat service), but it becomes risky in domains where depth matters (drug development, financial strategy, etc.).
Notably, successful meta-optimizers often have foundational knowledge they've accumulated over years, even if it wasn't their primary focus. The common pattern is someone who has read extensively about a field for a decade, then uses AI to rapidly integrate and apply that knowledge. The AI accelerates, but it builds on an existing foundation.
The Infrastructure and Control Layer Thesis
One of the most important—and least discussed—aspects of the emerging AI economy is infrastructure. While everyone focuses on model improvements and application possibilities, the real competitive moat may lie in the orchestration layer beneath the user-facing application.
Consider the task of building a product like Codex or Claude Code at enterprise scale. The superficial requirement seems simple: take a good model (like GPT-5.5), add computer use capabilities, integrate tools, and manage context. This describes the component level, but it misses the orchestration entirely.
In practice, deploying Codex or a similar system at scale requires understanding inference architecture deeply. Some workloads generate very large "prefill" blocks (long input context that must be processed before generating tokens). Others have strong "decode dependencies" (requiring sequential token generation with tight latency constraints). The optimal hardware orchestration, batching strategy, and service architecture varies dramatically based on these workload characteristics.
This is not a problem that can be solved by installing vLLM or SGLang and declaring victory. It requires the kind of infrastructure expertise that current frontier labs (OpenAI, Google, Anthropic) have spent years developing. In Korea, companies like Lablup have begun building comparable orchestration expertise, suggesting that frontier-level inference infrastructure is beginning to be productized.
This is analogous to the early 2000s, when Google's MapReduce and BigTable made computing and storage distributed and efficient. For years, this infrastructure was unique to Google. Engineers who saw Borg running in the wild often found it almost unbelievable. By the time open-source systems like Hadoop caught up, significant time had passed. The companies that wanted to compete at Google's level required similar infrastructure.
Today, if a company dreams of becoming an OpenAI or Google-level enterprise, even just at vertical scale, they will discover that:
- Orchestration layer is mandatory: Control and optimization of inference hardware, batching, context management, and service routing cannot be skipped or outsourced
- Control layer is the actual moat: Above the orchestration layer sits the application, and the application's competitive advantage comes down to a control layer that only the company can build
- Control layer has two axes: The toolset available to the AI agent (what it can actually do) and the customer data the company holds (customer history, preferences, behavior, transaction records)
These two dimensions—toolset and customer data—define what problems the AI can solve and how well it can solve them for that specific customer. A company that can orchestrate both dimensions together creates competitive advantage that cannot be easily replicated by simply deploying Codex.
The Application Era is Beginning: Unbundling ChatGPT
After years of discussing AI infrastructure and models, the industry is now entering what might be called the Application Era. This is not about building better chatbots or AI assistants. Rather, it's about building vertical solutions that leverage AI's general capabilities while wrapping them in domain-specific control layers.
The analogy, articulated by analyst Benedict Evans, is instructive: The past 20 years of computing were about "unbundling Oracle." Oracle (the database) could theoretically do everything, but instead, the industry fragmented into countless B2B SaaS and B2C applications, each focusing on a specific domain and building around it.
The next 20 years will likely be about "unbundling ChatGPT." ChatGPT and Codex can theoretically do everything—order Samdasoo water on Coupang, plan a birthday party, book a restaurant, plan a weekend trip. Yet the market will not consist of everyone simply using ChatGPT for everything. Instead, specialized service providers will emerge in each domain.
Why? Consider Coupang's situation. If Coupang unbundled its entire service into a clean API that ChatGPT could call, Coupang would lose everything that makes it valuable: the recommendation algorithms, the cross-sell opportunities, the media influence, the customer data advantages. By centralizing everything in ChatGPT, Coupang would be giving away its assets.
Therefore, Coupang will instead:
- Build its own specialized AI agent, potentially powered by OpenAI's infrastructure
- Layer domain-specific tools on top (Coupang's catalog, recommendations, pricing, logistics)
- Create context that incorporates customer history and preferences
- Deliver a solution that is better than "just use ChatGPT" for Coupang-specific tasks
This is not competition with ChatGPT as a general tool. Rather, it's competition within a specific domain by providing superior control and integration. The same pattern will repeat across every vertical domain: e-commerce, logistics, healthcare, finance, legal services, bioengineering, materials science.
The scale of these opportunities is enormous. If there are dozens of major vertical domains, and each domain fragments into 5-50 distinct competitors, then thousands of new companies will emerge. Each building not a traditional application but rather a specialized harness: a combination of general AI capability, customized toolsets, and domain-specific data integration.
This is why several observers have noted that current startup activity looks chaotic—because it is. Companies are pursuing diverse directions, exploring different verticals, and testing different models. This is not yet the era of consolidation. Rather, it's the exploratory phase where capital is being scattered across many potential winners to identify which combinations of domain + toolset + data actually create value.
AI-Native versus AI-Assisted: A Critical Distinction
As enterprises grapple with AI integration, a fundamental distinction has emerged that deserves serious attention: AI-native versus AI-assisted.
AI-native workflows are those where AI completes nearly all work with minimal human intervention. When the desired outcome is achieved, human involvement ends. The workflow is fundamentally reimagined around AI capabilities rather than augmented around human capabilities.
AI-assisted workflows, by contrast, leave human processes largely unchanged while inserting AI to make specific steps faster or better. The overall structure, decision-making, and responsibility remain with humans.
In practice, most enterprises claim to be pursuing AI-native transformation while actually implementing AI-assisted workflows. Consider a common case: a company identifies that 20% of an employee's time is spent on manual data entry. They implement an AI tool to automate data entry, saving that employee 20% of their time. They declare victory, call it AI transformation, and move on.
But here's the problem: that employee's core role is not data entry. It's analysis, judgment, and decision-making. By automating only data entry, the company hasn't rewritten the workflow; they've simply made the human slightly more efficient at a task that was never the core value. This is AI-assisted, and it rarely generates the productivity gains companies expect.
True AI-native transformation would require:
- Analyzing what the actual goal is (not what tasks the human does, but what outcome is needed)
- Identifying which components of that goal can be delegated to AI entirely
- Removing the human entirely from certain sub-processes where AI can operate independently
- Restructuring incentives and workflows around this new reality
- Accepting that some traditional roles may disappear while new roles emerge
Many companies fail at this transition because it requires uncomfortable choices. If a significant portion of a team's work can be automated, the organization must either eliminate that position, redeploy the person, or restructure compensation. Many companies lack the appetite for these difficult decisions and instead implement AI-assisted solutions that provide modest gains without fundamentally disrupting the organization.
The catch is that companies implementing true AI-native transformation may outcompete those with AI-assisted approaches by 5-10x in specific domains. This creates pressure to move toward AI-native, yet organizational inertia and human resistance slow the transition.
The Two Visions of the Future: Acceleration versus Deliberation
A significant tension is emerging within AI-engaged communities between two competing visions of how the future should unfold.
The acceleration thesis (held by most venture capital, startup founders, and frontier AI companies) argues that the pace of AI advancement must continue at maximum speed. Every bottleneck must be removed, every inefficiency must be automated, and every capability must be pushed to its limit. In this view, those who do not match this pace are left behind. The incentive structure of the market rewards those who can orchestrate AI tools most effectively. Therefore, individuals and companies must adapt to this pace or fail.
This thesis is reflected in hiring practices: companies like Anthropic reportedly seek people who can "enjoy chaos," who can make rapid decisions under uncertainty, and who can learn and execute at accelerated timescales. The assumption is that this capability is teachable and desirable.
The slow AI or deliberation thesis (articulated by figures like Mario Zechner and Mitchell Hashimoto) argues that relentless acceleration creates hidden costs that are not yet visible. When work is delegated to AI without comprehensive understanding, code can appear to work on the surface while harboring architectural flaws beneath. When learning is bypassed in favor of task completion, individuals may gain the ability to execute but lose the ability to innovate or judge. When decisions are made rapidly and continuously without reflection, systemic errors can accumulate.
From this perspective, "mind-sized bites" (a concept from Seymour Papert) represent an appropriate approach to learning and problem-solving. Rather than adapting human cognition to match AI speed, work should be paced to match human cognitive capacity. This may mean slower task completion but deeper understanding and more sustainable practices.
The tension between these visions is unresolved. Arguments for rapid adoption are undeniable—the market rewards speed, and there are genuine gains. But arguments for deliberation are also legitimate—unsustainable practices create eventual collapse, and depth enables resilience.
One possibility is that these represent a healthy natural dynamic: accelerators push the frontier, while deliberation advocates create feedback mechanisms that prevent excessive risk. But the current incentive structures strongly favor the acceleration narrative, potentially at cost to sustainability.
What AI Talent Looks Like in 2026
Given this complex landscape, what actually constitutes "AI talent" in 2026?
The answer has evolved significantly. Six months ago, the ability to use Claude Code or Codex fluently qualified someone as AI talent. But this baseline has shifted dramatically. Now, basic competence with these tools is table stakes—expected from most knowledge workers. It's necessary but no longer sufficient.
Beyond baseline competence, several characteristics distinguish genuine AI talent:
Domain mastery plus AI fluency: The individuals most effective with AI tools are those who combine deep knowledge in their field with fluency in AI. They can ask better questions, evaluate outputs more rigorously, and integrate results into a knowledge structure. This applies across domains: law, medicine, scientific research, engineering, product management.
Understanding of infrastructure and orchestration: As discussed, the next layer of differentiation involves understanding what lies beneath the application layer. How are models deployed? What are the inference constraints? How can tools be integrated? This knowledge is still relatively rare but increasingly valuable.
Judgment under uncertainty: With rapid change and incomplete information, the ability to make sound decisions, pivot when evidence demands, and operate effectively in ambiguous situations becomes critical. This is less about specific technical skills and more about decision-making capability.
Learning velocity: Not the ability to use existing tools, but the ability to learn new tools and paradigms as they emerge. Given the pace of change, the specific tools and techniques known today will be partially obsolete in 12 months. The ability to internalize new capabilities matters more than mastery of current ones.
Business acumen: Understanding customer problems, market dynamics, and how to create value is increasingly separate from pure technical capability. As AI tools become commodities, advantage accrues to those who understand where value actually exists.
Psychological resilience: The ability to sustain high performance and learning velocity without burning out is increasingly important and increasingly difficult. Those who can maintain perspective, set boundaries, and avoid psychosis-like over-immersion while still keeping pace represent a rare combination.
Notably absent from this list is "ability to click buttons in Claude Code faster." That's table stakes, not differentiation.
One specific cautionary note: very young people (teenagers, early 20s) may struggle with AI-accelerated work before developing their core thinking abilities. The risk is that delegation to AI prevents the development of independent judgment. Several thought leaders, including participants in the Dwarkesh Patel podcast ecosystem, have emphasized the importance of building personal capacity through challenging intellectual work, using tools like Anki for spaced repetition, and designing learning paths intentionally—not just doing whatever AI suggests.
The example of Dwarkesh himself is illustrative: he is sophisticated in AI tool use, but he also:
- Deliberately undertakes challenging intellectual projects (implementing AlphaGo from scratch)
- Uses structured learning tools (custom Anki decks)
- Learns from mentors and domain experts
- Designs his own learning path rather than following AI suggestions
This is not "just using AI to the maximum." Rather, it's using AI within a carefully designed personal development framework. This appears to be the emerging best practice for genuine talent development.
The Case for Skepticism and Alternative Paths
While the acceleration narrative dominates discussion, emerging counternarratives are gaining credibility. These don't reject AI but propose different relationships with it.
The slow AI movement, while still small, includes respected technologists who are explicitly questioning the sustainability of relentless acceleration. They argue that some problems genuinely require slower, more deliberate work. That test coverage and verification matter more than speed when delegating to AI. That architecture design cannot be entirely handed to agents.
Additionally, the emergence of subcultures and alternative value systems suggests that the singular "AI acceleration treadmill" is not the only viable path forward. Young people in particular are building communities around alternative values: environmental responsibility, creative expression, community building, rather than optimization for economic productivity.
Some people may rationally choose to opt out of maximum AI acceleration. They might:
- Use AI selectively while maintaining core cognitive engagement
- Prioritize personal development and relationships over task completion
- Build businesses in less AI-saturated domains
- Focus on creative or artistic domains where human effort has distinctive value
- Create technologies specifically for this slower, more deliberate approach
The market may support these alternatives if they can aggregate sufficient communities of interest. Someone building tools for "slow AI" or "thoughtful technology" could find audiences.
Conclusion: Navigating Uncertainty in the AI Era
In 2026, the relationship between humans and AI is at an inflection point. AI tools have moved from specialized capabilities to foundational infrastructure. Models continue to improve, and infrastructure complexity has become a source of competitive advantage. Talent that previously meant "someone who could use Claude Code" now means something far more complex and demanding.
Simultaneously, a concerning phenomenon—AI psychosis—has emerged, where the promise of productivity gains comes with real costs in terms of burnout, learning, and sustainability. This is not merely an individual problem but a systemic issue with how AI systems are designed and how markets incentivize their use.
Two talent archetypes are emerging: domain experts leveraging AI within their field, and meta-optimizers who delegate even goal-setting to AI. Each path is viable but carries different risks.
For organizations, the critical insight is that competitive advantage lies not in the AI models themselves (which are commoditizing) but in the orchestration layer, the control layer, and the integration of domain-specific tools and customer data. Building this layer is hard, requires specialized expertise, and is not easily outsourced.
For individuals, the choice is increasingly binary: develop the capability to work at the pace and intensity AI enables (risking psychosis) or deliberately choose alternative approaches that may be less economically rewarded but more sustainable. There is no universally "correct" choice, but the choice is increasingly unavoidable.
The next 12-24 months will see explosive activity as companies attempt to build vertical solutions on top of general AI foundations. Many will fail. Some will discover genuinely novel applications and value propositions. The industry will consolidate eventually, but we're in the exploratory phase.
For now, the most important capability is understanding this landscape—recognizing infrastructure as critical, understanding where value actually lies, maintaining judgment and depth alongside speed, and building intentionally rather than just riding the wave. Those capabilities will matter far more than any specific tool or technique that will be partially obsolete in six months.
원문출처: EP 97. AI Psychosis 시대의 사람들
powered by osmu.app