Ian Mulvany

August 10, 2023

Things I’ve been reading online, since May

I’ve been sharing some links around internally at work over the last few months, so I wanted to post them here. Clearly dominated by LLMs

Douglas Hofstadter has changed his mind about LLMs, and his perspective is quite sobering. I don’t give much time to folk who are very negative about the future prospects of these tools, but he is someone who is worth paying attention to, given his role at foundations of computing. I don’t fully buy into this view, but it’s probably an important one to pay attention to, and speaks to how disruptive people think these technologies are going to be. https://www.greaterwrong.com/posts/kAmgdEjq2eYQkB5PP/douglas-hofstadter-changes-his-mind-on-deep-learning-and-ai](https://www.greaterwrong.com/posts/kAmgdEjq2eYQkB5PP/douglas-hofstadter-changes-his-mind-on-deep-learning-and-ai)

“It’s not clear whether that will mean the end of humanity in the sense of the systems we’ve created destroying us. It’s not clear if that’s the case, but it’s certainly conceivable. If not, it also just renders humanity a very small phenomenon compared to something else that is far more intelligent and will become incomprehensible to us, as incomprehensible to us as we are to cockroaches.”

ThoughtWorks sharing how they built a tool to help brainstorming and ideation, that connects to OpenAI - https://martinfowler.com/articles/building-boba.html](https://martinfowler.com/articles/building-boba.html). It’s fascinating from a product and tech perspective. From a product perspective they have built a design thinking co-pilot. From a tech perspective they share a number of patterns on how to work with language models, in particular their use of LangChain and in-memory lightweight vector databases is worth paying attention to.

Use of ChatGPT comes with some uncertainty about how it uses the data, but not their API. Usage of API calls is clear, GPT does not use any of that information unless you opt in. WIth that in mind I’d like to point to Simon Willison’s suite of tools for working with GPT through the command line, Ive been having a lot of fun playing with these this week: https://til.simonwillison.net/llms/larger-context-openai-models-llm](https://til.simonwillison.net/llms/larger-context-openai-models-llm) https://simonwillison.net/2023/Jun/18/symbex/](https://simonwillison.net/2023/Jun/18/symbex/)

From the department of “no one know what is going on with all of this yet” - “We need to figure out under what conditions you need to tell somebody that generative AI is on the table,” he said. “If you change from a scribe to a computer, but the physician has the ability to look over the notes, do you need to tell them that’s ChatGPT?…Those types of questions I don’t think have been answered yet.” https://www.statnews.com/2023/06/13/ai-hospitals-patient-consent-health-care/](https://www.statnews.com/2023/06/13/ai-hospitals-patient-consent-health-care/)

The following blog post is very much a mirror of my own experience with GPT plugins. I got really excited for about a week, and then pretty much just stopped using them: https://matt-rickard.com/chatgpt-plugins-dont-have-pmf](https://matt-rickard.com/chatgpt-plugins-dont-have-pmf)

It’s important for us as an industry to get comfortable with these tools, and comfortable sharing what they are doing with these tools, because that’s the best way that we can figure out how to make the most use of them. This blog post gives a powerful argument in favour of that: https://www.oneusefulthing.org/p/detecting-the-secret-cyborgs](https://www.oneusefulthing.org/p/detecting-the-secret-cyborgs)

From the same author, some reflections on what it will look like to derive meaning from work, when most communications can be generated at the touch of a button https://www.oneusefulthing.org/p/setting-time-on-fire-and-the-temptation](https://www.oneusefulthing.org/p/setting-time-on-fire-and-the-temptation). The key question posed - will these tools take away meaning from work if they radically reduce our effort, and what then was the value of what we did in any case?

If you want to explore how tokenizers in LLMs work this is a great quick tutorial https://simonwillison.net/2023/Jun/8/gpt-tokenizers/](https://simonwillison.net/2023/Jun/8/gpt-tokenizers/)

You may have heard of “prompt engineering”, did you know that there are a few different types of prompt engineering? This blog post does a good job of covering these: https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/](https://lilianweng.github.io/posts/2023–03–15-prompt-engineering/)

Calling our UX and product colleagues - this is a nice meditation on how to think about the power that these models might have in reframing user experiences, beyond chatbots - https://maggieappleton.com/lm-sketchbook](https://maggieappleton.com/lm-sketchbook), which reminded me of what the https://www.kernelsearch.com/](https://www.kernelsearch.com/) team are trying to do.

Are you worried about your future career as a software developer? This is a good perspective on that - https://chenwang.org/posts/2023-05-ai-v-devs-p1/](https://chenwang.org/posts/2023–05-ai-v-devs-p1/) Newbies and expert-level devs look likely to benefit the most. For the rest, it could spell an existential crisis

If you want to spend 30 minutes seeing what Elsevier are doing with Machine Learning - Yvonne found this amazing presentation (you can watch at 2x speed!) https://www.youtube.com/watch?v=XGQBPtx_TmI](https://www.youtube.com/watch?v=XGQBPtx_TmI), though it’s probably rather out of date by now, as they have recently announced LLM integration into their product suite.

The video here is an amazing demo of what OpenAI’s Code Interpreter plugin can do in terms of automatically analysing data for you - https://dataliteracy.com/code-interpreter-for-chatgpt/](https://dataliteracy.com/code-interpreter-for-chatgpt/) (19 min, but if you watch it at 2X you can get the gist of it within about 5 minutes. I’ve had the chance of using code interpreter quite a lot over the last few weeks. It’s by no means a magic bullet, but it shows what a real conversational interface into a data set might begin to feel like in the future. Real vibes of talking to the Star Trek Computer.

LLMs are really not just a stochastic parrots - this blog post gives a reasoned example that claims to prove this - https://jbconsulting.substack.com/p/its-not-just-statistics-gpt-4-does](https://jbconsulting.substack.com/p/its-not-just-statistics-gpt-4-does)

Classifications from OpenAI: possible categories for the text:

  1. artificial intelligence (ai) and machine learning (ml)
  2. deep learning and ai advancements
  3. tools and technologies related to ai and ml
  4. perspectives and opinions on ai and ml
  5. user experiences and product perspectives
  6. career implications of ai and ml
  7. prompt engineering and tokenizers in llms
  8. presentations and demonstrations of ai and ml applications
  9. integration of llms in various industries (e.g., healthcare, publishing)
  10. analysis and capabilities of llms and gpt models.
the title was [[title]] The tags were: [[hashtags]]

About Ian Mulvany

Hi, I'm Ian - I work on academic publishing systems. You can find out more about me at mulvany.net. I'm always interested in engaging with folk on these topics, if you have made your way here don't hesitate to reach out if there is anything you want to share, discuss, or ask for help with!