A 2023 Look at ChatGPT Plugins, Browsing and Code Interpreter

In May 2023, Matthew Berman describes early access to ChatGPT plugins, browsing-enabled GPT-3.5 and GPT-4, and the alpha version of Code Interpreter. He wants to see how the tools connect a conversation to external services, web content and file analysis. The names, prices, availability and interface shown here are historical rather than current instructions.

Berman compares the plugin store to an app store and shows integrations including Kayak, OpenTable, Wolfram, Instacart, Expedia and Zapier. In the interface he demonstrates, up to three plugins can be enabled at a time. He selects OpenTable, Expedia and Instacart for the session, though the examples shown focus on Expedia and Instacart.

For a travel example, Berman starts by asking where he might go in Europe in July. ChatGPT suggests several destinations, and he chooses Amsterdam. Once he provides a three-day stay, the interface shows calls to Expedia and returns lodging options with ratings and nightly prices. The conversation narrows the search to luxury hotels.

Berman selects one suggested hotel and follows a link to its Expedia page, where a user could choose a room. The video shows a booking page, not a completed reservation. When he asks for a first-class nonstop flight from San Francisco, the plugin encounters errors; he retries and eventually sees flight suggestions, including an approximately $1,300 ticket. That historical output is not a verified current fare.

Next, he asks for a two-day healthy meal plan and grocery list. The assistant proposes meals such as yogurt, chicken and salmon, then asks whether to create an Instacart list. After Berman agrees, the plugin constructs a request containing the ingredients and returns a link to a shopping page. The demonstration reaches a list; it does not show a completed purchase.

Berman switches to browsing-enabled GPT-4. He explains that a model’s built-in knowledge has a cutoff and tests whether giving it a web URL can bring in newer context. Browsing can retrieve accessible material, but it does not give a model complete or guaranteed access to all information on the internet.

He supplies the URL for the paper [Generative Agents: Interactive Simulacra of Human Behavior](https://arxiv.org/abs/2304.03442) and asks for a summary. The interface displays steps such as reading, scrolling and following links, as well as several content-reading failures. It eventually produces a summary with references; the visible status messages alone do not establish that it read every part of the paper.

Berman asks how many agents inhabit the simulation. Further read failures appear, but the answer is 25, consistent with the paper’s abstract. He considers the result promising for summarizing, translating and retrieving information from web pages, while the test itself is one question rather than a broad reliability benchmark.

For Code Interpreter, Berman explains that a user can upload a file and ask the assistant to write and run Python for analysis, charts or other tasks. His example is a LEGO sets CSV downloaded from Kaggle. The assistant loads it with pandas and identifies fields for set number, name, year, theme ID and number of parts.

He asks for a plot of parts per set over time. Code Interpreter writes Python and returns a scatter plot with release year on the horizontal axis and parts on the vertical axis. Berman observes that larger sets appear more common later in the displayed data, though the plot also contains substantial variation.

Asked for the rate of increase, the assistant fits a linear regression and reports a slope of approximately 2.86 parts per set per year. This is a fitted average trend for the uploaded dataset under that model, not proof that every set gained exactly 2.86 parts each year or that the relationship remains linear outside the sampled period.

Berman next requests a line chart of the number of sets released per year. The generated graph shows an upward pattern in the data presented, covering years from around 1950 to just before 2020. He sees the ability to produce charts through conversation as potentially useful to data scientists and others, while results still depend on the data and generated analysis.

Finally, Berman uploads a Python file called memory.py from a larger project. The assistant describes code involving memory management and score normalization and notes that the file lacks the rest of the project context. When asked for method names, it writes Python to extract them. Berman closes by praising the alpha and mentions a $20 monthly subscription as his understanding at the time; this is historical pricing and not a claim about current access.