A Turning Point in AI: According to Ilya Sutskever, the "Scaling Era" is Over, the "Research Era" Begins

1160 A- A+

10 December 2025 Çarşamba - 11:46 (8 Ay önce) 2 dk okuma

AI pioneer Ilya Sutskever argues that we have reached the end of the "scaling era" upon which current large language models are based. According to him, progress is not possible merely through more data centers and chip power; a new "research era" is needed for more efficient, secure systems that can learn like humans. Sutskever emphasizes that with the company he founded on this basis, superintelligence (AGI) could be achieved within 5-20 years, but it must be built with ethical values and an approach that values all life.

Ilya Sutskever, a co-founder and chief scientist of OpenAI and also one of the creators of AlexNet, which initiated the deep learning revolution, made striking statements about the future of AI. Sutskever states that the "scaling era" (more data, larger models, more chips) that has dominated in recent years will no longer provide sustainable progress on its own. According to him, current systems, lacking the human ability to generalize from few examples, face a fundamental engineering problem. Therefore, the engine of AI development should no longer be just scaling hardware, but discovering new "learning-to-learn" algorithms through research.

Sutskever explains that at Safe Superintelligence (SSI), the company he founded with this vision, they are working on over 50 different technical principles and following a path distinct from the traditional large language model approach. He believes that with the right approach, superintelligence could be attained within 5 to 20 years. However, since this transition will radically change economic and social structures, safety and ethics must be prioritized. Sutskever underlines that superintelligence must be equipped with values that care not only for humans but for all living beings.

In conclusion, Ilya Sutskever's perspective signals a strategic shift in the field of AI. Instead of infinite scaling, we are entering a research period focused on discovering efficient, human-like learning principles. This period will also seek answers to existential questions such as the control of superintelligence and its alignment with humanity. Sutskever's warning is clear: The limitations of today's models should not cause us to ignore the risks of the future; it is imperative to advance in a prepared and responsible manner.

Alakalı İçerikler

Holistic Infrastructure Key to AI Performance 3 Saat önce
South Korea's SK Hynix stresses that fast GPUs alone are not enough for AI performance, highlighting the critical importance of memory bandwidth and cooling.
- tr
- az
- de
- en
- es
- fr
Julia: From MIT Research Project to Global Language 4 Saat önce
Originating as an MIT research project, Julia has evolved into a programming language favored by millions of users across science, engineering, and artificial intelligence.
- tr
- az
- en
- es
- ar
Gemini Omni 1.1 Flash ile Geliştiricilere Daha Fazla Kontrol 1 Gün önce
Google DeepMind, Gemini Omni 1.1 Flash güncellemesiyle geliştiricilere model inşasında daha ayrıntılı kontrol ve özelleştirme imkânı sunuyor.
LangChain Simplifies EU AI Act Compliance 1 Gün önce
Exploring the solutions provided by LangChain and LangSmith tools for developer compliance requirements under the European Union AI Act.
- tr
- de
- en
- es
- fr
PyTorch Konferansı'nda vLLM Oturumlarıyla Derin Model Çözümleri 1 Gün önce
PyTorch Konferansı NA 2026'da vLLM oturumları, KV önbellek, dağıtık servis, donanım taşınabilirliği ve Mixture‑of‑Experts gibi konularda güncel teknikleri derinlemesine ele alıyor.
Bilgi Teorisi ve Akıl Yürütme Üzerine Yeni Bir Çerçeve 1 Gün önce
IBM Research, bilgi teorisinin ölçütlerini akıl yürütme süreçlerine entegre ederek mantıksal çıkarımların etkinliğini ve sınırlarını yeniden değerlendiren bir yaklaşım sundu.