Apollo Research publishes a starter guide to testing AI models

In short

Apollo Research, a group that tests AI systems, has published a starter guide to model evaluations, often called "evals". The guide explains what evals are, which skills help, and how a newcomer might begin. It is a how-to guide, not a report of a specific incident, and the authors say evals alone do not make models safe.

What happened

Apollo Research published the original report on 8 January 2024. It was led by Marius Hobbhahn, with feedback from four colleagues. The authors say evals are a new field, so their advice is a "current best guess" that may change quickly.

Apollo describes evals as "the systematic measurement of properties in AI systems." An eval might ask whether a model can solve a coding problem, or whether it tends to seek power.

The guide separates several kinds of testing:

• Red-teaming means actively trying to find a specific ability or tendency in a model. It shows that something can happen, but not how likely it is in real use.
• Benchmarking estimates how often a model behaves a certain way on a set of test cases, with as little back-and-forth with the model as possible.
• Capability evaluations ask whether a model can do something. Alignment evaluations ask whether it tends to do it. Apollo says a model could be able to do something harmful yet be unlikely to do it in practice.

The guide then lists skills it considers helpful. These include writing good prompts, fine-tuning (further training a model on extra data), building software around a model so it can act as an "agent", and keeping a scientific mindset. It names Inspect, a tool maintained by the UK AI Security Institute, as its main recommended tool. It also suggests two ways to start: pick a measurement and iterate, or learn general skills first.

The authors note limits. They say behavioral tests cover only a small slice of possible inputs. They also warn that a benchmark can stop being useful once people treat it as the only measure of what it targets.

What was done about it

The source describes no incident, so there was no response to report. The guide itself is Apollo's contribution. It says the authors shared a draft with participants of an Apart Research evals hackathon on 24 November 2023 and thanked them for feedback.

Why it matters

Apollo says evals increase knowledge about what AI systems can do, how they tend to behave, and where they fail. It says this helps the public, AI companies and lawmakers make decisions. It notes that evals already underpin Responsible Scaling Policies, which are company rules for deciding when to release more powerful models. The authors stress that evals are not sufficient on their own. They compare them to testing during a pandemic: useful for decisions, but not a safety measure by themselves. They say high-confidence claims should not rely on evals alone.