Eleven AI Developments in July 2023: Robotics, Models and Safety

In July 2023, I collect eleven AI developments from a busy week: robotics, model scaling, generated media, alignment and safety research. The demonstrations and forecasts belong to that period and should not be read as present-day product claims.

Google’s RT-2 combines vision-language pretraining with robotics data to output actions. In one demonstration it selects a toy dinosaur when asked for an extinct animal; in another it selects an object useful for a task. The presenter sees this as evidence of useful generalization in the demonstrated settings.

The presenter says an intermediate plan improved performance on some tasks and notes that the paper discusses unfamiliar objects and settings. He speculates about domestic robotics, but the examples are research demonstrations rather than proof of a general-purpose household robot.

Inflection AI’s Mustafa Suleyman predicted that models much larger than GPT-4 would be trained within roughly 18 months. The presenter emphasizes the scale of that forecast; the comparison and schedule are a 2023 prediction, not a verified outcome in this video.

The video briefly shows an AI-generated video example and then discusses an Atlantic article about OpenAI. The presenter summarizes the article’s account of model scaling, possible agent development and risks seen in a base model.

He reports that the article described potentially harmful responses from an unaligned base model and discusses concerns about misuse. These accounts illustrate why safety testing matters; the video does not establish that every deployed system behaves the same way.

The article also presents Ilya Sutskever’s concern that advanced systems may not reliably follow human intentions. The presenter describes proposals to use AI to assist alignment research and notes the difficulty of directing future systems toward human values.

In another interview, alignment researcher Jan Leike describes using AI to help address alignment as an ambitious plan that he nevertheless considers promising. The presenter also reports Leike’s view that a rapid transition from human-level to much more capable AI is possible; this is a forecast.

The presenter highlights a wearable captioning demonstration for deaf users and advances in synthetic speech. He sees accessibility benefits while recognizing that increasingly convincing generated media also creates challenges for identifying authentic material.

He notes that OpenAI withdrew an AI-text classifier because of low accuracy. He predicts that distinguishing generated media will become harder; that forecast should not be confused with proof that every detection method is impossible.

Stable Beluga 2, based on Llama 2, is discussed as an instruction-tuned model with strong scores on some 2023 benchmarks. The presenter compares reported results with GPT-3.5-era baselines; the results describe those tests and that time, not all tasks or current models.

Researchers also reported automated adversarial prompts that could transfer across models. The presenter uses this to question broad claims that a particular system’s safety defenses are solved. The general lesson is that safety evaluations need to consider adaptive attacks and their limits.

The video turns to a 2023 US Senate hearing. Anthropic CEO Dario Amodei describes concerns that more capable systems might lower barriers to biological misuse and forecasts growing risks. The presenter also cites Sam Altman’s acknowledgment of concern. These are risk assessments and predictions, not evidence that a specific attack occurred.

Amodei also advocates securing the AI supply chain and advanced systems. The presenter closes by connecting those proposals to the week’s broader question: how to gain benefits from advancing capabilities while improving safeguards.