Which Side of the Singularity Are We On?The surge of optimism surrounding our approach to AGI might not be unwarranted. On January 4, Sam Altman (regardless of your opinion of him, he is currently the most influential figure in AI)
described his current status as: “near the singularity; unclear on which side.” Two days later, in an interview, he clarified that AGI would be created
during Trump’s presidency—by 2029. On the same day, he elaborated in his
“Reflections”: “We are now confident that we know how to build AGI in the traditional sense. We believe that by 2025, we could see the first AI agents integrated into the workforce, fundamentally transforming company outcomes. […] We are beginning to target […] true superintelligence.”
If Altman’s prediction proves accurate, the implications for humanity will be monumental. In comparison, any events we currently consider critical—such as tomorrow’s inauguration of Trump—will pale into insignificance against the backdrop of the world being drawn into the black hole of the singularity.
So, what happened that brought us so close to the singularity, with AGI predictions suddenly accelerating? What caused timelines to plummet so drastically
(as seen in the Exponential View chart referenced)?Only Sam himself knows for sure. However, a compelling and straightforward explanation was
offered by Gwern Branwen in a comment on Ryan Kidd’s
post about the new paradigm of scaling outputs:
“I think what’s missing in the discussion of the scaling paradigm is the self-play feedback loop: much of the value of a model like o1 isn’t in its deployment, but in generating training data for the next model. Every problem solved by o1 now serves as a training data point for o3 (e.g., any o1 session that eventually stumbles upon the correct answer can be refined to eliminate dead ends and produce a clean walkthrough to train more nuanced intuition).”
You can test this hypothesis yourself with a simple four-step experiment:
1. Take a transcript of your conversation with ChatGPT on a problem it failed to solve satisfactorily.
2. Submit this transcript to Claude, asking it to solve the problem.
3. Review Claude’s response, which, even if not a perfect solution, will likely be significantly better than ChatGPT’s initial attempt.
4. Now reverse the roles: for another problem, give Claude’s unsatisfactory response to ChatGPT. Most likely, the same pattern will repeat—the second model will outperform the first by leveraging its output as training data.
This aligns with the historical development of AlphaGo versions: AlphaGo Fan, AlphaGo Lee, AlphaGo Master, and AlphaGo Zero. The last version trained entirely on self-play without human data. In just two years, the Elo rating of new versions skyrocketed from 3300 (already above most world champions) to an unimaginable 5185. After 2018, DeepMind ceased developing AlphaGo or participating in official matches, as it had become both uninteresting and meaningless—humans couldn’t come close to competing.
If Gwern Branwen is correct, this reflects the powerful effect of a “data flywheel,” where each new model generates data to train an even better one.
And if that’s the case, then Sam Altman is also right. Tomorrow’s inauguration of Trump, an event considered globally significant today, will fade into a trivial footnote as Earth enters the era of the singularity.
#AGI #AI #science
