@WISPL is a very simple user interface that is designed primarily with scientific uses of large language models (LLMs) in mind.
Features
Formatted output. The user interface handles responses from the language model that may include simple markup (e.g., bullet points), software code, and equations.
Google searches. The LLM is instructed to issue dedicated search commands when it needs current information to answer a question. E.g., "Who is the current British PM?" will likely result in the LLM querying Google first before answering.
Computer algebra. The LLM is instructed to rely on computer algebra (the Maxima computer algebra system) when it needs its capabilities. For instance, "What is the integral of $(\sin x \cos x)/x^2$ with respect to $x$?" will likely result in a Maxima query before the LLM answers.
Plotting. The LLM is given the option to embed simple function plots. E.g., a request like "Describe the Bessel J0 function" may result in a response that includes a plot of that function.
Location-specific information. If you allowed the browser to share your location data with you, the LLM receives your geolocation data and may be able to answer questions like, "What will the weather be like at my location tomorrow?"
The LLMs are not deterministic. Their use of these tools is not always perfect. They may mess up the syntax when invoking the tool, attempt invoke the tool inappropriately (e.g., in the middle of a response) or forget about the tool altogether. Sometimes, it helps to remind the LLM. (E.g., "Use the SEARCH command to verify your answer." Or, "can you show me a plot of this function?")
When Google or Maxima were used by the LLM, this is indicated below the answer. Clicking the triangle icon can help expand that line, so you can see what Google or Maxima responses were received by the LLM.
Models
The @WISPL software can use different language models. Currently, it is set up to use GPT 3.5 and GPT 4 from OpenAI, and Claude from Anthopic.
As the use of the cloud-based models incurs a cost, by default users are granted access only to Eliza, a "dumb" algorithmic language model from the 1960s that is included for test/validation purposes. If you wish to access any of the "true" LLMs, we need to discuss it via e-mail.
The models to which you have access can be viewed by clicking or tapping the down arrow in the Ask button. This down arrow disappears once you initiate a conversation with a selected model.
To initiate a conversation, make sure the right model is selected, enter your query in the text field provided, and click Ask. Alternatively, Control-Enter also works when you are editing your query.
The software can also handle image generation by DALL-E, OpenAI's graphics engine. Access to DALL-E is restricted to select users. If you have access, you will see a separate button, "Generate image".
Conversations
The LLMs are "stateless" machines: they have no memory of earlier parts of a conversation. This illusion of continuity comes from the user interface, which preserves the entire conversation and sends it back to the LLM, along with the latest user entry, for "completion". This means that over time, the length of the conversation grows. If the subject is meandering, it can confuse the LLM and result in lower quality responses. So if you wish to switch topics, it may make sense to start a new conversation. You can do so by reloading the page or clicking/tapping the logo in the upper left.
Conversations can become too long for the LLM to handle. In this case, the user interface requests the LLM to offer a summary, which is then used internally in place of the earlier conversation. While some details will likely be lost as a result, it allows a conversation to continue effectively indefinitely.
Other features
The @WISPL user interface makes it easy to create a permanent record of a conversation: Just print it from the browser. The appearance is optimized for printing, either to hardcopy or in the form of PDF output.
When printing answers that include responses from Google or Maxima, you may wish to expand these items first, so that their full content appears in the printed page.
Additionally, conversations can be saved on your computer by clicking/tapping the Save button. The save file format is a "JSON" formatted data file that can only be read by the @WISPL user interface itself. To load an earlier conversation, use the Load button and select a previous savefile. This will allow you to continue a conversation where you left off.
Additional features include language selection (click/tap the language icon/national flag) and password change (click/tap the "hamburger" menu in the upper right and select Settings).
LLM @ WISPL