Using language models to write single-use apps
George Veletsianos describes how he used ChatGPT to create a piece of software to solve a problem he faced during a research project.
What are some of the implications? Teaching people pseudo code is much easier than teaching people the ins and outs of different programming languages. Could a good understanding of pseudo code be what is most that is necessary for people to be able to write their own code and programs with the support of AI? This opens up a lot of venues for internet research for graduate students. Lots to think about here.
I mentioned something similar a couple of weeks ago, where Dave Nicholls described how he used generative AI to create an app for a specific purpose:
I asked it to build an app for the ISIH conference in 2024 that showed people cafes and restaurants within five kilometres of where they were staying, ranked by consumer reviews in Google Maps. It did it in seconds, and it’s amazing.
This is incredibly powerful…the ability to create single-use apps in a couple of minutes that solve the unique problems that each of us face every day. Imagine being able to tell your personal AI assistant what your objective is, and have it create the app you need to move forward?
Academic articles as data, not information (and definitely not as story)
Related to the idea of adding your own documents to a language model’s underlying dataset, is the idea that we might change how we interact with those documents. I’d not thought of this before, but we interact with almost all text in the same way. But what if the interface to the document was adaptive, responding to the purpose of our reading and providing the information in different ways accordingly?
The Semantic Reader Project aims to augmenting scholarly documents through AI-powered, interactive reading interfaces.
Scholarly publications are key to the transfer of knowledge from scholars to others. However, research papers are information-dense, and as the volume of the scientific literature grows, the need for new technology to support the reading process grows… . We describe the Semantic Reader Project, a collaborative effort across multiple institutions to explore automatic creation of dynamic reading interfaces for research papers… . We structure this paper around challenges scholars and the public face when reading research papers; Discovery, Efficiency, Comprehension, Synthesis, and Accessibility.
I was fascinated to come across this project, which aims to generate different reading interfaces for academic papers, depending on what your goal is.
Since coming across Explainpaper last year, I’ve been thinking of academic papers as data, rather than something containing data. From this perspective, it makes sense to interrogate those articles – through natural language – rather than reading them. It feels like this project is a step towards an environment where academic papers aren’t read at all. Instead, they’re queried collectively through prompts. Instead of engaging with single papers in series, you might engage with a topic, concept, or question, where the underlying dataset includes 1000 papers running in parallel.
I love the idea of an interface to data that adapts to your needs. Combined with a library of your personal information, this could completely change how knowledge work happens.
However, there are at least 2 main concerns with this, which I’ll expand on below:
- Context: existing generative AI systems lack personal context
- Trust: existing generative AI systems aren’t trustworthy
‘Embeddings’ that add context to LLM outputs
Embeddings allow you to supplement the underlying language model that ChatGPT uses, with your own resource library. When you prompt the model, it includes the content of your personal archive in its response.
Say you have a website that has thousands of pages with rich content on financial topics and you want to create a chatbot based on the ChatGPT API that can help users navigate this content. You need a systematic approach to match users’ prompts with the right pages and use the LLM to provide context-aware responses. This is where document embeddings can help.
The implications of this are significant, and we’re going to see a ton of apps and services popping up that help you to personalise the outputs of language models. Not only are we going to see more apps and services using this approach from commercial interests (see, for example, the recent announcement of Google’s Project Tailwind), but also for individuals interested in personal knowledge management (see, Ton Zylstra’s post about a personal ChatPKM).
Combined with system messaging that help you to steer the model in certain directions – possibly by defining different ‘personas’ – this is going to have a huge impact on the ecosystem.
Building trustworthy AI systems
Bruce Schneier explains that one of the biggest problems we face with existing generative AI systems isn’t that they’re incompetent, but that they’re not under our control and therefore inherently untrustworthy. He suggests that trustworthy AI would need to be:
- Under your direct control.
- Which means you need to be able to run the model on your own device (or on a cloud service that you control).
- And that the technology needs to be transparent, showing you what it’s doing, when, and how.
- You also need to know how it works (at least at a basic level).
- And when you don’t know how it works, the model should be able to explain its reasoning to you.
- While also providing sources that support its choices.
Obviously, we’re never going to get trustworthy AI from OpenAI or Google, because that objective isn’t aligned with their profit motive.
So it was great to this announcement from the Allen Institute, about an open language model made by scientists, for scientists:
We will be making all elements of the OLMo project accessible; not only will our data be available, but so will the code used to create the data. We will open-source the model, the training code, the training curves, and evaluation benchmarks. We will also openly share and discuss the ethical and educational considerations around the creation of this model to help guide the understanding and responsible development of language modeling technology… . The artifacts released as part of the OLMo project will include training data, code, model weights, intermediate checkpoints, and ablations. A release strategy for the model and its artifacts is in development as part of the project.
It’s not just the model that’s open source, but everything about it. And the financial support of the Allen Institute means that this kind of model will be competitive in scale with products created by OpenAI, Google, and Facebook.
Summary
We’re moving towards a generative AI ecosystem with trustworthy models running on your devices, with access to own data, customised with personas you define, creating personalised apps that solve unique problems for you.
Cheating on essays seems like a rather quaint notion at this point, don’t you think?