Blog & Research
Our research on reliable AI, and a dedicated track on superintelligence. Switch between them below.
A research track by SoReliable AI
Superintelligence
AI that greatly exceeds people in nearly every field is now the stated goal of several of the largest AI labs. We study what it means, how close it is, and how people can keep judging work that may surpass their own. Every article cites its sources at the end.
Six topics, in reading order. Choose one to filter.
Capabilities
How close are we? Measuring the length of work AI can do alone
One of the clearest ways to track progress is to ask how long a task an AI agent can finish on its own. The trend is steep. The details matter more than the headline.
Denis Ojua
/4 min read
Alignment
When the student outgrows the teacher: the alignment problem
Today's AI is shaped by people judging its answers. What happens when the answers are better than the judges? An explainer on the central problem of superintelligence.
Denis Ojua
/3 min read
Oversight
Scalable oversight: how people can check work they could not do
If AI becomes better than us at a task, how do we know its answer is right? Researchers have proposed debate, decomposition and assisted review. Here is what the evidence shows.
Denis Ojua
/3 min read
Interpretability
Reading the machine: what interpretability can show today
If we cannot always check an AI's answer, perhaps we can check how it got there. Interpretability research is starting to trace a model's steps, with real limits.
Denis Ojua
/3 min read
Governance
Superintelligence becomes policy: the US task force and the industry pledge
In one week, frontier AI leaders signed a voluntary safety accord at the White House and the US formed a Super Intelligence Force. What was agreed, and what it will take to make it real.
Denis Ojua
/4 min read
Open questions we are studying
We study superintelligence because the results will shape what we build. These are the questions we are working on next.
Research this with us- iHow can expert reviewers be measured on work that is harder than they could do alone?
- iiWhich signals show that AI is speeding up AI research, and how should they be checked?
- iiiWhen an agent's run counts as a success, how often does it survive expert review?
- ivWhat should an independent audit of a frontier AI system test, and who is qualified to run it?
- vHow should human judgment be recorded so that it can supervise systems that outgrow it?
Ready to try SoReliable?
Tell us what you are building. Any team, any project, any size. We find the specialists, train them on your standard, and evaluate the work against what your project actually needs.