HomeFull speed ahead. Keep track of advancements in every frontier
SSI research programmeSafe Superintelligence—

SSI research programme

by Safe SuperintelligenceUSIL

ResearchStage 1 of 5

Pure research. No model, paper or product has been released.

Updated 27 Jul 2026Checked 25 Sep0 updates this week

Most important updates

  • 27 Jul 2026
  • 3 Jul 2025
  • Apr 2025
  • 4 Sep 2024
  • 19 Jun 2024

Current obstacles

  • No outside proofWith nothing released, outsiders can't judge if SSI's approach works. It wins chips and talent on reputation alone.

Physics limits

  • No way yet to read a model's goalsBehaviour is spread across billions of numbers with no labelled parts. Tests show what a model does in tested cases; they can't guarantee what it does in untested ones.
  • Models learn far less efficiently than peopleA person learns to drive in tens of hours; models see trillions of words and still fail on unfamiliar variants. Nobody yet knows how to get human-like generalisation.
  • Returns shrink as a power lawError falls only as a small power of training compute: each fixed step of improvement needs roughly 10x more compute, energy and money than the one before.

How it works

3 parts
Google's data centre in Eemshaven, Netherlands: SSI rents Google Cloud TPUs, which sit in campuses like this
Google's data centre in Eemshaven, Netherlands: SSI rents Google Cloud TPUs, which sit in campuses like thisPhoto: Wvdp · CC0 (opens commons.wikimedia.org)
Focus

One goal, no products

SSI says it will release nothing until it has safe superintelligence, so no compute or staff goes to products or customers.

Thesis

Beyond pure scaling

Sutskever argued in late 2025 that scaling pretraining is levelling off, and that the core problem is models generalising far worse than people.

Compute

TPUs and NVIDIA systems

SSI uses Google Cloud TPUs and, from a July 2026 deal, NVIDIA Vera Rubin systems expected to raise its compute about tenfold.

Update log

5 updates

Mon 27 Jul

  • Major: PressPartnership

Thu 3 Jul 2025

  • Minor: XPeople

Apr 2025

  • Major: PressFunding

Wed 4 Sep 2024

  • Major: PressFunding

Wed 19 Jun 2024

  • Major: BlogPeople

About Safe Superintelligence

The team behind SSI research programme

Safe Superintelligence

Superintelligence research, no products

Safe Superintelligence (SSI) was founded in June 2024 by Ilya Sutskever, OpenAI's former chief scientist, with Daniel Gross and Daniel Levy. It says its only product will be safe superintelligence and has released nothing yet.

  • Raised $2B at a $32B valuation in April 2025 with no public product.
  • Sutskever became CEO in July 2025 after co-founder Daniel Gross left for Meta.
  • Reported to use Google's TPU chips for its research.
Founded
20242 yrs
Headquarters
United StatesIsrael
Status
Private
Valuation
$32BprivateApr 2025
Raised
$3B2 rounds
Last round
Venture round · $2BApr 2025$32B post
Works in
AIFrontier models
Coverage
1 program · 5 updateslatest 27 Jul 2026checked 25 Sep
Lead investors
Greenoaks
People
Ilya SutskeverCo-founder and CEODaniel LevyCo-founder
Primer