I’d love to hear more about some successful uses of the quickly proliferating AI tools that are all the rage this year.
I’ve used ChatGPT as a better-than-ever thesaurus:
What words begin with “c” that are relevant to transformation?
And to help with word block:
Rewrite the following paragraph to explain it to a five-year-old.
I’ve also used Midjourney to generate images, then edited them in Affinity Photo to change their colour palette and make space for diagrams and the like.
Cool topic!
I’ve used ChatGPT for emails, especially things like job rejection and acceptance; it’s more polite than I can ever be! It does 80% of the work but writes like a high school student so always needs some improvement. Also, for generating placeholder text for websites - I won’t miss Lorem Ipsum.
I’ve been using Stable Diffusion for preparing slides and for use when teaching engineering-related topics that need media for demonstrations and practals. It gives me unique media for every project I do with students. I finally got it to generate hi-res images with lots of passes using the latest version, which now does faces well. I’ll also use it when preparing presentations and material for marketing, as it sure beats contacting a designer.
This is just the start. As these tools evolve, they’ll improve, and hopefully, ChatGPT will become more mature and be able to write at a higher intellectual level.
Of course this is very early on.
The biggest practical use I have had for GPT has been to augment my workflow in conducting an academic literature search on a topic.
I have had good success asking ChatGPT (or even more recently Perplexity.ai) two questions in sequence:
What are articles that SUPPORT Argument X?
What are articles that are AGAINST Argument X?
I have used it to generate boilerplate code – the kind of thing where I’d normally find some example or tutorial and adapt what I found there. One thing I’ve learned to like about ChatGPT code in my limited use so far is that is uses very clear, simple variable names (where a lot of what I find on the web uses single letters or obscure abbreviations because whoever’s typing is in a hurry).
Be careful with realistic placeholder text. I have lost track of how often I’ve seen someone create plausible placeholders only to see them wind up in production inadvertently. One of the beauties of lorem ipsum is that it’s pretty immediately recognizable as Not What Should Be There.
If you want variety, there are endless alternatives: veggie ipsum, pirate ipsum, hipster ipsum… google around!
Good point. The websites I make are typically throwaway prototypes for research and normally don’t get publicly released (the aim is to write research papers on the development process). I will think twice about this if it is for a commercial or public project.
I’m doing a lot with it to explore coding solutions. Give it an unknown AWS product, a problem you’re trying to solve, and a description of your existing app/endpoint/relevant database structure, and it either gives you a starting point for your solution or makes you realize you weren’t thinking about something.
Over long sessions, you can keep feeding it more information about your app, too. Since they launched history, it’s easy to go back to an old session whenever you’re working on a problem. You can ask it for solutions while you code and see how yours differs from its.
You can keep widening its solution, too, asking it for postgres triggers, migrations, model code, etc. It knows what to generate based on what it’s already made. If you don’t like what it made, correct it as part of the prompt or with a refinement based on your next prompt.
ChatGPT’s style of working with an AI gets me better results than copilot. I hope this is the future, where we have tools that can serve as partners for exploration, rather than tools completing comments and initial lines from a programmer hoping that they get lucky and pass tests, and with code that should just become higher level language abstractions (like the old debate with C decades ago–making it comfortable to write lower level code just meant we didn’t work harder on creating compilers that would write it for us.)
I would rather pair with this thing than a human 60% of the time. 20% of the time I need to have a conversation with someone with real insight, and 20% of the time I need a pairing partner that will keep me engaged and moving.
This was an odd case. I have a little bit of travel with two different destinations and in-between time. I tried doing some time calculations but then fed everything into chatGPT and it spat out a rough itinerary.
Post trip, how well did the itinerary work?
AI seems reasonably good as a fact checker to give references supporting or rebutting a source
Example:
https://www.perplexity.ai/?s=u&uuid=3ee8485d-e7e6-4d4b-8df5-48ded0efe70a
Another example - I am figuring out a workflow when I am reviewing the literature on a given topic - either professional or personal - to gather references along with a brief summary and Pro/Con designation which are added by the AI and confirmed by me.
I am pondering how to make this an ongoing process with some sort of HTML table editor when I want to add more items or more fields.
Sort of a personal reference manager with integrated AI for summary info
Not totally sure where I will go with this but so far it seems quite helpful
Todoist has launched AI integration. It’s actually a really innovative implementation, I think:
The concept that AI integration may soon become a routine feature in many apps is quite appealing. In that regard this is quite an exciting development.
That said - although I applaud ToDoist for doing this, I am not quite sure that a To Do app is something that lends itself to AI. I suspect the concrete steps for most tasks are individualized and not easily determined by AI.
For example, ToDoist shows an example that it can help you list the steps to run a marathon. It doesn’t show the answer but here is what ChatGPT tells me - while it is good advice, it does not strike me as something that is quite in actionable To-Do list form:
make a to-do list of how to run a marathon
1. Train regularly: Start building up your endurance with regular training runs several months before the marathon.
2. Eat Healthy: Start eating a balanced diet of healthy food that will fuel you for your runs.
3. Stay Hydrated: Make sure to drink plenty of water and other fluids throughout the day and during your runs.
4. Rest: Give your body a rest day at least once a week, and make sure to get plenty of sleep the night before long runs.
5. Plan: Have a plan for the race. Know what your goal is, and how you will reach it.
6. Gear: Make sure to have the right gear before the race. Good running shoes, clothes, and other equipment can make a difference in your performance.
7. Stretch: Stretching before and after your runs can help to reduce the risk of injury and keep your muscles loose and limber.
8. Mental Preparation: Prepare mentally for the race. Visualize yourself crossing the finish line and remind yourself why you are running.
9. Race Day: On race day, remember to take it easy and have fun. Pace yourself and enjoy the experience.
10. Celebrate
I wonder if AI would be useful for grouping and classifying.
ie: These items could be bought at this store, which looking at your calendar you will be near on this date…
Or perhaps automatically grouping/tagging items in projects and coping them to “projects” based to location, time, availability etc… (this would require more metadata and already sounds a bit tedious…Could AI also generate metadata based on a sentence ?)
Possibly - but not “generative” language model AI, which is what is offered by the OpenAI API most of these apps are likely to use.
This is an example using Bing to generate what I find to be an extremely helpful way to do a literature search. It summarizes each article in 1 sentence and uses that as a hyperlink to the full article. It’s useful for both specific research and general browsing in an area of interest.
That said - as useful as it is sometimes it “hallucinates” and gives nonsense responses. Most of these responses are spot-on. A few are not well summarized. The third from the bottom is totally wrong factually; cervical epidural injections do not treat knee pain, and in fact the article it references is about knee pain and has nothing to do with the requested subject of cervical epidural injections.
Thus this technology is not yet ready for prime time in terms of authoring reports or doing safety-sensitive tasks. That said - if it creates a useful list of articles but occasionally directs me to something off-point then no harm done. On balance it’s quite helpful. And presumably it will improve over time and/or I will figure out better ways to write prompts.
Well then be careful giving a demo to your boss or to your hypothetical stockholders. 
For what it is worth… if hallucination by AI is a concern, one strategy is to explicitly tell the software to either not guess or tell you it is guessing when it is not sure of an answer.
I’ve used ChatGPT to generate quick outlines on subjects to trigger my thinking and writing. It’s faster than calling friends with expertise in the topic and asking “what are the ten things I need to know about blah blah blah”.
Not so success yet with Bing Chat. Maybe it’s the topics I ask about, but about half the time Bing starts answering then suddenly deletes the answer and announces it cannot answer the question.
Early days.
Katie
I’m afraid the software does not know when it is guessing. It assembles its answers from the material it has been fed, which may have been incorrect or unrealistic. Perhaps it could assign a confidence factor to each answer it provides and you could suppress “guessing” by requiring only answers with large confidence factors.
Still, one of the persistent news stories around AIs, I think, will be about the bone-headed results that they occasionally supply. 
I believe it does. Or more specifically its confidence is a function of how close the vector comparison is of its embeddings.
See my example above. The concept of “Tell me if you do not know” has been discussed as a recommended technique in the OpenAI Discourse discussions.
More info here:
And in this source which it refers to:
