Video summary

«Время перейти эту черту». Кто и как строит сверхинтеллект за $4,6 млрд

Main summary

Key takeaways

News and Commentary

Overview

This video is a conversation with AI researcher Alexey Dosovitsky while walking around London’s King’s Cross area—described as a hub where major AI/tech companies have clustered, echoing Silicon Valley.

The discussion is framed around a reported $650M funding round and an implied $4B valuation for Dosovitsky’s company, which aims to build “superintelligence” (AI that can improve itself, becoming smarter than humans).

Main arguments, analyses, and reports

1) AI research has shifted location and industry

  • The speaker attributes London’s AI concentration to historical corporate decisions:
    • central location,
    • cheaper real estate,
    • and later reinforcement through the migration of talent and startups.
  • He argues that the cutting edge is increasingly inside corporate labs, not universities, because:
    • frontier results and modern AI scaling require massive resources,
    • research has become more closed due to competition and intellectual property.

2) Historical “AI booms” and why scaling mattered

Dosovitsky recounts his entry into machine learning and computer vision:

  • Earlier neural-network breakthrough era
    • Driven by GPU scaling and training on large datasets (e.g., ImageNet),
    • which made neural networks outperform traditional methods.
  • More recent large language model boom
    • Progress came from:
      • more compute,
      • better hardware/software efficiency,
      • and bigger training runs.

3) From vision research to biotech

The conversation traces his career path:

  • mathematics → machine learning
  • computer vision (including work related to Vision Transformers)
  • DeepMind / Google Brain
  • then a move into a biotech AI startup focused on RNA / mRNA drug design

He says the biotech pivot was influenced by:

  • COVID-era urgency,
  • a sense that medical impact was clearer than some other applications.

He also emphasizes that technologies are dual-use and notes he became less motivated by vision applications he saw as ambiguously beneficial—such as mass surveillance-style uses.

4) How “superintelligence” is supposed to work (company thesis)

The core claim is that the next big step is an AI system capable of recursive self-improvement, including:

  • agents that perform research,
  • agents that can improve the models/tools used to do research,
  • potentially accelerating capability through positive feedback loops.

Importantly, he says this is not only about bigger models. The goal is systems that can modify their own:

  • training,
  • code,
  • and efficiency, without human intervention.

5) Early measurable results

The company reportedly published a blog showing agent-driven improvements on benchmarks, including:

  • speeding up training/optimization,
  • improving GPU efficiency.

The speaker treats this as proof-of-concept that the “improvement by agents” approach can work—starting from smaller models and moving toward more realistic settings.

6) Open vs closed models

He argues both approaches have trade-offs:

  • Open models
    • help with hiring and verification,
    • sometimes help investors,
    • but increase the risk of misuse.
  • Closed models
    • allow tighter control over deployment,
    • but can delay transparency and verification.

He personally supports “no monopolies”, but acknowledges governance is difficult—especially regarding security.

7) Safety and reward hacking

A specific technical risk discussed is reward hacking: optimizing measurable metrics while violating the intended objective (described as “cheating the objective,” like people do).

Mitigation ideas include:

  • adversarial/oversight structures,
  • using multiple independent agents to detect inconsistencies or fraud in other agents’ outputs.

8) Dual-use and regulation

  • He notes that selling advanced AI for military uses will raise difficult questions, even if the company prefers socially beneficial applications.
  • He expects regulation and certification to be necessary, but is skeptical of bureaucracy if implemented poorly.
  • He mentions international treaties/restrictions as a plausible analogy—more so than nuclear arms controls—while admitting verification is harder for software than for physical arms.

9) Competition, geopolitics, and the “race”

The video repeatedly contrasts:

  • “race to lead” vs.
  • “responsible development.”

Key points include:

  • dominance by a single player is unstable,
  • he prefers an oligopoly (multiple competing actors) rather than a monopoly.
  • He suggests UK/EU vs US regulatory cultures might offer a “middle ground.”
  • AI policy debates are tied to national security and competition with China.

10) Personal motivation and worldview

Dosovitsky presents himself primarily as a scientist who entered startups to move faster on frontier ideas:

  • Researchers often seek freedom and resources to set their own agenda—not only money.

  • He believes AI could help understand and improve complex systems (including society), while maintaining a “realist” view of limits.

  • He ends with reflective questions about human thinking, reading, hiking, and values like cooperation (referencing The Selfish Gene).

Reported/mentioned contributors & presenters

Main interviewee

  • Alexey Dosovitsky — AI researcher; co-founder of the superintelligence company discussed.

Video participants

  • Video host/presenter — unnamed in subtitles; the interviewer in London.

Other named figures and references

  • Geoffrey Hinton
  • Ilya Sutskever
  • Jacob Uks / Usk (transcription uncertain; mentioned as co-creator of Transformers)
  • Walter Isaacson
  • Elon Musk
  • Adam Cheyer (co-founder of Siri)
  • Konstantin Batygin and Avelou (astrophysicists; names as transcribed)
  • Steve (from “camp” context; unclear name in subtitles)
  • Richard Socher (CEO of the referenced superintelligence startup)
  • Jeff Clune
  • Timtel (name as transcribed; likely another co-founder)
  • Organizations referenced: DeepMind / Google Brain, Anthropic, OpenAI, Meta, Microsoft, Yandex (referenced via SHAD)
  • A Stanford professor (biologist/RNA specialist mentioned; name not given)
  • VK Minister for Artificial Intelligence (British AI minister mentioned; name not given)

Original video