A 2023 Discussion of Incremental AI Releases and Safety

In April 2023, the host examines a statement by OpenAI president and cofounder Greg Brockman about releasing models beyond GPT-4. Brockman discussed a continuum of incremental improvements as a possible safety opportunity, alongside both present-day and longer-term risks.

The host speculates that a successor might arrive through intermediate versions, using labels such as GPT-4.2 as illustrations. Those names and a particular release schedule were his prediction at the time, not a commitment in Brockman’s statement.

A training checkpoint is a snapshot of model parameters at a point in a training run. A later checkpoint reflects further training, without necessarily changing the model’s parameter count. Brockman suggested deploying successive checkpoints as one possible way to make improvements more gradual.

The host discusses whether training data might run short. In a contemporary interview, OpenAI’s then-chief scientist Ilya Sutskever says there was still useful data available, while noting that other training approaches would eventually be needed. Estimates of future supply and quality are uncertain.

The host discusses proprietary data and user interactions as possible sources of training material. His inference about their relative importance is speculative. Policies on use of customer data and opt-out mechanisms are product-specific and can change, so the 2023 description should not be treated as current instructions.

An example from Sébastien Bubeck shows a model producing progressively improved drawings of a unicorn during training. It illustrates a change on one task, not a complete measure of capability or a guarantee that every checkpoint improves every behavior.

Brockman warned against confident predictions about AI progress. The host gives examples of an economist surprised by an exam result and forecasters underestimating progress on a competition-math test. These examples show forecast error, not a reliable timetable for future systems.

Brockman’s statement addresses risks ranging from misuse and bias to possible existential harms. The host cites a survey of AI researchers and notes that nonresponse limits how representative its result may be. The survey reports respondents’ beliefs, not a measured probability of extinction.

OpenAI reported that GPT-4 performed better than GPT-3.5 on its safety metrics. The displayed technical-report chart compares incorrect behavior on selected sensitive and disallowed prompts after human-feedback training. Such metrics are useful within their definitions, but do not prove a system is safe in every setting.

The host points to research on using language models and tools for scientific work, including beneficial and harmful potential applications. He describes the authors’ call for safeguards and asks whether an increase in capability might also increase the consequences of misuse. That is a question, not a settled answer.

Brockman wrote that people would have a chance to shape the future together. The host asks how broad participation can be reconciled with the concentration of decisions in a small number of companies.

The host cites a contemporary comment about hardware constraints and notes that safety testing can take time after training. He predicted a hypothetical intermediate release around mid-2024; this is a historical forecast, not current product information.

In an interview excerpt, Sutskever identifies reliability as a potential obstacle to real-world impact: if users must repeatedly check answers, the economic value can be reduced. He adds that he was not predicting this would necessarily be the outcome.