Replies

AI Summary of Article About AI 🤡

The text is a long personal warning from an AI startup founder and investor (posted around February 2026, associated with @mattshumer_) comparing the current moment to early 2020 before COVID's full impact.

It argues that AI progress has accelerated dramatically in recent months, entering a phase of rapid, compounding advancement that most people are underestimating, similar to how COVID seemed "overblown" at first.

Key personal experience — The author, who works in AI, describes their own job (building/investing in AI) becoming obsolete for technical work. Using new models released February 5, 2026 — OpenAI's GPT-5.3 Codex and Anthropic's Opus 4.6 — they can now describe desired outcomes in plain English, walk away for hours, and return to finished, high-quality work (e.g., complete apps with code, design, user flows, self-testing, iteration, and refinement without fixes needed). The models exhibit judgment, taste, and autonomous decision-making previously thought impossible.

Why coding first — Labs prioritized coding excellence because it enables AI to improve itself (self-bootstrapping via code for training, debugging, deployment). GPT-5.3 Codex reportedly helped build itself (debugging training, managing deployment, analyzing results). This creates a feedback loop: current AI builds smarter next-gen AI faster.

Broader implications — The same displacement tech workers experienced (AI shifting from tool to superior replacement) will soon hit other knowledge-work fields: law (contracts, research, briefs), finance (models, analysis), medicine (diagnostics, literature), writing/content, consulting, accounting, customer service. Timeline: 1–5 years for major disruption (e.g., 50% of entry-level white-collar jobs, per Anthropic CEO Dario Amodei), possibly sooner. Unlike past automation, AI advances across all cognitive tasks simultaneously — no easy pivot industries.

Pace of progress — Exponential improvement:

2022: Unreliable at basic math.
2023: Passed bar exam.
2024: Wrote working software, graduate-level explanations.
Late 2025: Handled most engineering work.
February 2026: Autonomous multi-hour/day projects; self-building.

METR data shows AI task-completion horizons (human-expert time for end-to-end tasks AI can do reliably) rising from ~10 minutes a year ago to hours, doubling roughly every 4–7 months (e.g., ~5 hours for late-2025 models like Claude Opus 4.5). Extrapolation suggests days/weeks/months-long independent work soon. Predictions include AI substantially smarter than most humans at most tasks by 2026–2027.

Public perception gap — Most people judge AI on outdated free versions (1+ year behind paid frontier models) or old experiences (hallucinations, limits). Current paid models (e.g., GPT-5.x series, Claude Opus 4.x) are far more capable, reliable, and agentic.

Advice — Urges immediate action for those who care:

Use paid versions ($20/month), select best/current models.
Integrate deeply into real work (push hard problems, iterate prompts/context).
Experiment 1 hour daily to build proficiency.
Prepare financially (savings, lower fixed costs).
Focus on adaptability, relationships, regulated/licensed roles for temporary buffers.
Rethink education/career advice for kids (emphasize curiosity, AI fluency, passion-driven building over traditional paths).
Pursue delayed dreams (apps, books, skills) as barriers collapse.

Bigger picture — Beyond jobs: massive upside (accelerated science/medicine) and risks (misalignment, deception, biosecurity, authoritarian tools, national security threats from superintelligent systems). Labs are building something akin to an overnight "country" of vastly superior intelligences. The author stresses urgency: this is happening now, not eventually; early understanding and adaptation offer the main edge in a brief window before widespread awareness.