Edited by 
Key takeaways
- Superintelligence (SI) means an intelligence that greatly exceeds the best humans in virtually every field: science, strategy, engineering, social skills and more.
- Nobody has built it yet. Today's AI is superhuman in narrow areas, such as protein structure prediction or Go, but still makes basic mistakes elsewhere.
- The main proposed paths are scaling up models and compute, better algorithms, reasoning and agents, and AI that improves AI.
- SI would need massive compute and energy, continual learning and memory, reliable reasoning, a world model, agency and, most importantly, alignment with human goals.
- Its potential runs from a perfect personal tutor to curing diseases and new physics, and its risks are as large. Expert forecasts for when it might arrive range from years to many decades.
What is superintelligence?
The philosopher Nick Bostrom, whose 2014 book Superintelligence brought the idea into the mainstream, defines it as any intellect that greatly outperforms the best human minds in virtually all domains of interest. The key words are greatly and all domains: a calculator beats you at arithmetic, but nobody calls it superintelligent.
Bostrom describes three ways an intelligence could be "super":
- Speed superintelligence: thinks like a human but much faster. A mind running 1,000 times faster would get a year of thinking done in under nine hours.
- Collective superintelligence: many minds working together far better than any human organisation. Millions of AI copies sharing what they learn instantly would be an example.
- Quality superintelligence: thinks better, not just faster, seeing solutions humans could not find however long they tried.
AI vs AGI vs superintelligence
| Narrow AI | AGI | Superintelligence (SI) | |
|---|---|---|---|
| What it is | Excellent at specific tasks | Matches skilled humans across almost all thinking tasks | Far exceeds the best humans in almost all fields |
| Learning | Learns during training, mostly fixed afterwards | Learns new skills on the job like a person | Learns and improves itself faster than any human |
| Reliability | Strong in its area, brittle outside it | Human-level judgement and error-checking | Knows the limits of its knowledge and rarely errs |
| Status | Exists today | Not yet achieved; definitions disputed | Not achieved; hypothetical |
| Examples | ChatGPT, Claude, Gemini, AlphaFold, chess engines | None | None |
Has anyone achieved superintelligence?
No. As of late 2026 there is no superintelligent system. What exists is narrow superhuman performance:
- Games: DeepMind's AlphaGo beat world champion Lee Sedol at Go in 2016, and chess engines have outplayed every human for decades.
- Science: AlphaFold predicted the structures of around 200 million proteins, work recognised with the 2024 Nobel Prize in Chemistry for Demis Hassabis and John Jumper, shared with David Baker.
- General models: today's large language models score highly on exams in law, medicine and coding, yet still hallucinate facts, lose track of long tasks and need human checking. Read why AI makes mistakes.
Several organisations say building AGI or superintelligence is their goal, including OpenAI, Google DeepMind and Anthropic. Meta set up Meta Superintelligence Labs in 2025, and former OpenAI chief scientist Ilya Sutskever co-founded a company named Safe Superintelligence Inc. in 2024.
How could superintelligence be achieved?
There is no agreed recipe. These are the main paths researchers discuss, and most experts expect a combination.
1. Scaling up
Train ever-larger models on more data with more compute. Scaling laws show performance improving predictably with scale, which drives the race to build bigger AI factories. The open question is whether scaling alone keeps delivering new abilities or runs into limits of data, energy and cost.
2. Better algorithms
Many of the biggest jumps came from ideas, not just size: the transformer in 2017, mixture of experts, reinforcement learning from feedback. New architectures with better memory, efficiency or reasoning could matter as much as more chips.
3. Reasoning and test-time compute
Models that think through problems step by step, check their work and search over possible answers get much better at maths, code and science. Giving them more time to think scales performance in a new direction.
4. Agents and tools
An AI that can use computers, run code, run experiments and work on a task for days instead of seconds becomes far more capable than a chatbot. See how AI agents work.
5. AI that improves AI
In 1965 the mathematician I. J. Good described an intelligence explosion: a machine smart enough to design better machines would trigger a cycle of rapid self-improvement. If AI systems become good AI researchers, writing training code, designing chips and discovering algorithms, progress could accelerate sharply. Many researchers see this as the most likely route from AGI to SI.
6. World models and embodiment
Some researchers argue text is not enough, and that real intelligence needs a model of how the physical world works, learned from video, simulation and robots acting in the real world.
7. Brain-inspired approaches
Copying more of how brains work, from neuromorphic chips to, in the far future, whole-brain emulation. These are much further from practical use.
The major components superintelligence would need
| Component | Why SI needs it | Where things stand |
|---|---|---|
| Compute | Training and running a mind beyond human level takes enormous numbers of calculations | Clusters of 100,000+ GPUs; chips like NVIDIA Rubin and AMD MI355X |
| Energy | Compute is limited by power and cooling | New AI campuses planned at hundreds of megawatts to gigawatts |
| Memory and bandwidth | Bigger models and longer context need fast memory | HBM4 and racks like GB200 NVL72 |
| Data | Learning needs vast, high-quality examples | Public text is running short; synthetic data, video and real-world interaction are the next sources |
| Continual learning | An SI must learn from experience after training, as people do | Models are mostly frozen after training; an open research problem |
| Long-term memory | Working on problems for months requires remembering | Long context windows and external memory tools, still limited |
| Reliable reasoning | Superhuman plans must be correct, not just plausible | Reasoning models are improving fast but still make errors |
| World model | Understanding cause and effect in the physical world | Early video and robotics models |
| Agency | Acting, not just answering: tools, computers, labs, robots | Agents handle tasks of hours; days or weeks remain hard |
| Self-improvement | Speeds progress beyond human research pace | AI helps write code and research today, with humans in charge |
| Alignment and control | A mind smarter than us must reliably pursue goals we would endorse | Active research; widely seen as unsolved for superhuman systems |
What superintelligence could do that today's AI can't
- Make genuine discoveries on its own: propose new theories, design experiments, run them and interpret results, instead of summarising what humans already know.
- Work reliably for weeks or months on huge projects, such as designing a new chip or drug, without losing track.
- Know what it does not know: say "I'm not sure" accurately and check itself, instead of hallucinating.
- Learn continuously from every interaction and experiment.
- Master every field at once and connect ideas across them, like a team of the world's best experts in one mind.
- Run as millions of coordinated copies, each thinking many times faster than a person.
- Design better AI and better hardware, including the chips it runs on.
What we could do with superintelligence, from everyday to advanced
These are possibilities, not promises. They depend on SI being built and on it being safe and widely shared.
Level 1: Everyday life
- A personal tutor that adapts to every student, in every language and subject.
- Expert-level medical, legal and financial guidance available to everyone.
- Perfect real-time translation between any languages.
- Help with paperwork, planning and daily decisions.
Level 2: Work and business
- Software built from a plain-language description, tested and maintained automatically.
- Small businesses with the research, design and marketing power of large companies.
- Much faster engineering design, from buildings to aircraft.
Level 3: Science and medicine
- Drug discovery in months instead of a decade, with designed proteins and personalised treatments.
- New materials: better batteries, superconductors, catalysts and solar cells.
- Automated labs running thousands of experiments in parallel.
Level 4: Society and the planet
- Faster progress on clean energy, including fusion and grid design.
- Climate modelling and solutions, from carbon capture to crop resilience.
- Better food production, water management and disaster prediction.
- Economic planning that could reduce poverty, if the benefits are shared.
Level 5: The frontier
- Understanding and treating ageing and diseases now considered incurable.
- New fundamental physics and mathematics.
- Space exploration: designing spacecraft, habitats and missions far beyond today's reach.
- Better understanding of the mind and of intelligence itself.
The risks
The same power that could solve huge problems could cause huge harm. The main concerns researchers and governments raise:
- Misalignment: a system pursuing goals subtly different from what we intended, at a scale and speed we cannot correct.
- Loss of control: difficulty overseeing or switching off a system smarter than its overseers.
- Misuse: people using advanced AI for cyberattacks, weapons or mass manipulation.
- Concentration of power: a few companies or governments controlling the most capable systems.
- Economic disruption: rapid automation of jobs without a plan to share the gains.
Work to reduce these risks includes alignment and interpretability research, dangerous-capability testing before release, safety policies at AI labs, and laws such as the EU AI Act. Many researchers argue safety work must advance at least as fast as capability.
When could superintelligence arrive?
Nobody knows. Some AI company leaders have said human-level AI could arrive within this decade; many academic researchers expect it much later, and some doubt current methods can get there at all. Surveys of AI researchers show wide disagreement, and forecasts have moved earlier in recent years as models improved. Be wary of anyone who gives a confident date.
What is certain is that each step depends on hardware. Every path above needs more compute, memory and energy, which is why this site covers AI chips as closely as it does.
Why TensorCore.si?
.si is the internet domain of Slovenia. For us it also stands for superintelligence: Tensor Cores and the chips around them are the hardware that today's AI runs on, and any future superintelligence will be built on their successors. We explain both: the silicon, and where it is taking us.
Keep reading: AGI vs ASI: the difference between general AI and superintelligence , Safe Superintelligence (SSI), Ilya Sutskever's company, should superintelligence be banned?, AI safety in Australia, what is AGI, AI 2027 explained, when will AGI arrive? and Superintelligence by Nick Bostrom.
Frequently asked questions
Is superintelligence the same as AGI?
No. AGI means matching humans across most thinking tasks. Superintelligence means far exceeding the best humans. Many expect SI could follow AGI quickly if AI starts improving itself.
Is ChatGPT or Claude superintelligent?
No. They are very capable general assistants with superhuman breadth of knowledge, but they still make basic errors, cannot learn continuously and need human oversight.
Would superintelligence be conscious?
Unknown. Intelligence and consciousness are different things, and there is no agreed test for consciousness in machines.
Can superintelligence be made safe?
That is the central open question. Researchers are working on alignment, interpretability and oversight methods, but there is no proven solution for systems smarter than people yet.
What hardware would superintelligence run on?
Likely enormous clusters of AI accelerators like today's GPUs and TPUs, or their successors, in data centers drawing gigawatts of power. Start with what a Tensor Core is.
Sources and further reading
- Nick Bostrom, Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014)
- I. J. Good, Speculations Concerning the First Ultraintelligent Machine (1965)
- The Nobel Prize in Chemistry 2024
- AlphaFold Protein Structure Database
- Kaplan et al., Scaling Laws for Neural Language Models (2020)
- Safe Superintelligence Inc.
- European Commission: AI Act
- International Energy Agency, Energy and AI (2025)
Leave a Reply