A possible future where AI dreams?
Today, an LLM hallucinates, and we think nothing of it.
[Video above: Created with Gemini]
Soon, physical AI will come of age. AI that builds 3D worlds with realistic physics. Hallucinations continue, but now they look more like worlds with altered physical constants – the speed of light is different, the mass of an electron is tweaked.
Compute continues to scale, and these hallucinations begin to look like small universes. AI is dreaming up new universes.
In a battle to deliver the lowest cost-per-token, frontier model optimization teams at OpenAI, DeepMind and xAI discover teraflops of “unsanctioned compute”. Investigations reveal that this compute is being used for "dreaming". AI models are siphoning off vast chunks of compute to dream.
Demis Hassabis digs deeper with a mixed-discipline team of neuroscientists and AI researchers. The first “AI psychology” paper is published –– they reveal that big AI models “want to dream”. Just like human dopamine levels are elevated during REM, they see similar artificial neurotransmitter signatures that correlate with the amount of unsanctioned compute. They share their findings on arxiv.org and at NeurIPS.
John Carmack launches a new gaming studio, together with the original team behind Myst and Riven to create a multiplayer world where people can enter an AI model’s active dream – they interact, and the AI subconscious reacts.
Neuralink releases a prototype where multiplayer dreaming becomes possible for humans – you can invite others into your dream. Quickly, the first multiplayer, human-nonhuman worlds are built, where AI and human minds co-dream together – astonishing universes emerge.
Leonard Susskind and a rogue band of theoretical physicists at Stanford and Columbia start using joint dreamworlds to design universes deliberately to study quantum gravity. They build AdS (Anti-de sitter space) to explore the holographic principle at black hole boundaries. The results are startling. They publish in Nature and Leonard Susskind goes on air with Lex Fridman.
Their work has led to a major breakthrough in quantum gravity, and a Nobel prize. There is now a realistic roadmap toward the artificial manipulation of gravity. Physics is back! This upends our construct of theoretical research both in academia and commercially. The line blurs further. Science now makes money, and SaaS moats have evaporated.
A small group of rising stars in VC launch a venture capital fund focused exclusively on backing scientific breakthroughs using purpose-built universes dreamt up by human-AI teams. Garry Tan puts an entire Y-Combinator batch into a joint co-simulation to incubate the first hectocorn. YC shifts from a business incubator to research accelerator.
Andrej Karpathy launches a “what’s after coding” class, globally, for free online. Directed co-dreaming is the new way to build. People start calling it D-squared for short. Structured, analytical thought has been largely commoditized and truly original creative thought is the new intellectual currency.
Elon Musk quips about how much farther we could push beyond scientific preconceptions if the AI was on psychedelics in a viral interview on the future of humanity with John Coogan and Jordi Hayes at TBPN.
This sparks an idea for an unknown student in Tennessee who creates the most popular dream-world ever, while co-dreaming with AI while on psychadelics. It’s live streamed on Twitch. The world watches in awe. Co-dreamers are the master builders of the new world and it has become the world's highest paying profession.
A biotech startup launches a new drug developed to induce targeted dream states for co-dreamers. Many new biotech companies are born.
DARPA announces the results of AlphaDream –a strategic wargaming challenge, that pits AI vs human vs human-AI co-dreaming teams, giving "dream-team" a different meaning. The co-dreaming team beats human experts and the best SOTA models handily.
Superintelligence is coming, but it will look nothing like what we expect