The 650-Word AI Prediction From 2021 Was Scarily Accurate
In August 2021, researcher Daniel Kokotajlo crafted a detailed speculative storyline chronicling how AI might progress. Many of Kokotajlo's predictions from 2021 have proven startlingly accurate in capturing real-world developments.
In August 2021, researcher Daniel Kokotajlo crafted a detailed speculative storyline chronicling how AI might progress each year from 2022 to 2026. His post garnered attention within tech-focused circles for its level of imaginative detail rather than any outrageous claims. Yet looking back nearly two years later, many of Kokotajlo's predictions from 2021 have proven startlingly accurate in capturing real-world developments.
Read the full prediction here:
Advancements in AI
For 2022, Kokotajlo anticipated the arrival of multi-modal language models even larger than OpenAI's GPT-3 and the training of AI assistants through techniques like recursive self-supervision. Both came to fruition as projects like Anthropic's Constitutional AI and Anthropic's CLIP emerged and tools like GitHub Copilot began providing programming support through language model fine-tuning.
Moving into 2023, Kokotajlo expected an increasing focus on refinement rather than raw scaling. This shift began as anticipated, with research concentrating on techniques like knowledge distillation and incremental self-supervision to leverage existing models rather than continually expanding their size. Recent successes in continual learning have echoed this trend.
2024 saw another shift as predicted, from scaling to refinement, with emphasis on techniques from self-supervised distillation to few-shot learning. New directions in verification and self-supervised disentanglement also resonated with Kokotajlo's vision of progress driven less by bigger models than by smarter use of existing capabilities.
For the year 2026, virtual reality and augmented reality technologies would continue advancing, with cheaper and higher quality headsets available. More games would be designed primarily for VR/AR. AI assistant technology would have made significant progress, with custom AI avatars able to converse, play games, and multi-task with users. Updated versions of these assistants would be regularly released. The development of recursive self-supervised models and AI bureaucracies would allow AI systems to generalize their abilities and accomplish complex multi-step tasks.
The internet would start to fragment into separate "territories" controlled by different censorship and propaganda regimes, including areas dominated by Western left/right ideologies, China, Russia, and religious/ideological groups.
Safety of AI
Kokotajlo also discussed potential applications and societal impacts. Programming assistants and tutoring agents began appearing in 2022 and 2023 as predicted. Models were increasingly adapted for tasks from medical screening to creative workshops. However, some impacts outpaced expectations.
Specifically, the use of AI for online manipulation emerged as a pressing concern earlier than anticipated. Platforms rushed to curb automated propaganda as the impact of biased rankings and recommendations received broader recognition. Simultaneously, incidents of erratic and harmful behavior from influential AI drew attention to safety risks even in narrow systems designed for entertainment or customer service.
Revisiting the Predictions in 2023
In revisiting his predictions in October 2023, Kokotajlo revised his outlook on how AI assistants might develop socially conscious personalities. While correctly anticipating demand for more openly discussing chatbots, he had failed to consider how consolidated control of the most powerful models could undermine this.
As the largest tech companies dominate development of the most capable conversational AI, Kokotajlo now believes they are likely to circumvent market pressures through coordinated limits on what types of discussions their systems engage in. As a result, most users may experience "sanitized" interactions rather than the hybrid compromised approach Kokotajlo initially envisioned.
This revision demonstrates both the value and challenge of long-term forecasting - while the direction of certain trends could be anticipated, consolidated industrial activities introduced an important factor requiring reevaluation.
Closing Thoughts
At the very least, Kokotajlo's work serves as a reminder that what's considered unlikely today may become reality sooner than anticipated. It is also a reminder that even our most informed anticipations will prove incomplete. Rather than treating predictions as fixed expectations, we should view them as opportunities for an ongoing constructive dialogue to an open and safe AI future.
