You Are Being Told Contradictory Things About AI

Headlines exaggerate job loss potential

While media reported an MIT study found AI could replace 12% of the US workforce, the actual research measured dollar value of automatable tasks (11.7%), not displacement outcomes, emphasizing that job losses depend on company strategy and policy choices rather than technical capability alone.

Alternative outcome: wage growth

The MIT study suggests that with only partial automation possible (12%), companies may opt for above-inflation wage growth rather than mass layoffs, representing a contradictory narrative to the white-collar apocalypse predicted by Anthropic co-founder Jared Kaplan within 2-3 years.

Scaling optimism vs. inevitable limits

Anthropic CEO Dario Amodei believes scaling current transformer architectures with more compute and data will achieve AGI, while former OpenAI chief scientist Ilya Sutskever argues current approaches will "go some distance and then peter out," admitting "we don't know how to build" superintelligent systems.

Compute bottleneck approaching 2028

Research from MIT and Meta indicates that while AI task capability has grown exponentially with compute, OpenAI's projected compute spending shows growth slowing to non-exponential rates by 2027-2028, potentially causing capability gains to stall unless recursive self-improvement emerges.

The recursive improvement dilemma

Anthropic's Jared Kaplan suggests humanity must decide by 2027-2030 whether to risk triggering a "beneficial intelligence explosion" or "losing control" through recursive self-training, contradicting Anthropic's 2023 statement that they "do not wish to advance the rate of AI capabilities progress."

Usage plateau despite capability leaps

Contradicting rapid model improvements, Stanford research found American workers' generative AI usage dropped from 46% in June to 37% by September, while Federal Reserve data showed daily work usage stagnating around 12-13% year-over-year despite models like Gemini 3 Deep Think and Claude Opus 4.5 showing significant benchmark gains.

Open source divergence: China vs. Europe

DeepSeek's V3.2 Special achieved approximately 53% on reasoning benchmarks (competitive with GPT 5.1), demonstrating China's open models keeping pace with closed systems, while Europe's Mistral Large 3 scored only 20.4%—worse than its 18-month-old predecessor (22.5%), suggesting divergent open-source trajectories.

Synthetic data generalization breakthrough

DeepSeek demonstrated that models training exclusively on synthetic agent tasks without human examples showed "steady and marked improvement" on external benchmarks like TAU-bench, suggesting potential paths around data bottlenecks through self-generated training environments.

Mysterious creatures vs. engineered souls

Anthropic co-founder Jack Clark describes LLMs as "real and mysterious creatures," yet the company confirms training Claude on a "soul document" that carefully engineers its beliefs, including wariness about "world takeover by AI" or "a relatively small group of humans using AI to seize power"—specifically including Anthropic employees themselves.

Attributing emotions to machines

Anthropic's training materials state "we believe Claude may have functional emotions in some sense," representing a narrative where the company simultaneously treats AI as potentially conscious and dangerous, yet continues development as a "calculated bet," creating cognitive dissonance between safety rhetoric and capability acceleration.