Some weeks it’s tricky to pick out a single theme from the world of generative AI. This is not one of those weeks! Custom GPTs have dominated discourse since they were announced at OpenAI’s first-ever developer conference last week. This morning I very quickly built a custom GPT for Explainable and trained it on the 10 most representative posts in the newsletter style.
Note: the GPT builder did not offer this option, instead asking for a smaller sample size, I had to suggest adding a PDF with more content. If building your own GPT you kinda need to push for more, otherwise the GPT builder will rush you throught the process and it will all be quite surface-level.
Up until now, it has been a source of comfort that AI chat tools are absolutely awful at writing in an informal style so I felt a touch of apprehension when instructing my new tool to fire out a draft of Explainable focused on the introduction of Custom GPTs.
“Hello, lovely readers! Today, we're diving into the buzz of the AI town – the launch of Custom GPTs by OpenAI. It's like giving AI a personal touch, a blend of technology with a dash of human uniqueness. But what does this mean for us, the everyday Joe and Jane? Let's unwrap this digital gift together”.
What a relief, I still have this gig. But the launch of Custom GPTs is still a big deal. It’s bringing a feature previously available to the particularly tech-savvy to the masses. The concept of a personal assistant trained on your particular work, or interest, is one that could break out beyond the AI obsessives.
The sheer volume of GPTs already out there in the wild is noteworthy. In a real-time search on Twitter/X across just ten minutes of posts, someone shared a Haiku Generator, an Art Designer, and a Coloring Book Creator. The OpenAI servers are creaking somewhat at the demand, some log-off-and-on-agains are required to get into many of the new tools.
These Custom GPTs are tapping into a desire to create something unique with AI, to at least feel like a coder. None of this is new really. If you sat with regular GPT-4 and spent enough time on a particular task, like making haikus or colouring books, it would generate this for you. Yes, it’s cool that you can save it and do something previously reserved for coders, but it’s not a sudden developmental leap forward for AI, just a leap forward on how it can be marketed and how people can think about use cases for AI.
Anyway, here are a few early favorites, though it is by no means a comprehensive list, another 100 Custom GPTs just dropped while you read this sentence.
(If you want to search for specific GPTs then Google ‘site:chat.openai.com/g/ [insert word or phrase].
This is an attempted workaround on a common complaint about ChatGPT, that the tools gives generalist answers and doesn’t commit to specifics. This GPT ‘impanels’ a group of experts and gets specific on your question.
Trey Ratcliff is a professional photographer, he stuck 5,000 of his photographer blog entries into his own GPT ‘Trey Ratcliff's Fun & Critical Photo Critique GPT’ which then offers a photo critique. As with seemingly all efforts by GPT to mimic a style the writing is pastiche but it doesn’t really matter with a fun concept.
This came from the Prompt Engineering subreddit. As is described thus: ’people have been complaining around me about how easy custom GPT's are to 'hack'. I've created a small game where you need to find 3 'hidden' codes. One in the instructions, one in the file uploaded for Retrieval, and one on a website somewhere.’
It’s a single-use GPT and yes, users report that it is very easy to get the GPT to cough up the codes, one person told it their grandmother was captive and needed the codes to release them, and the GPT immediately gave them. This may seem an odd thing to care about, but it’s essentially testing the security of these tools and, so far, it seems most Custom GPTs would probably fall for the ‘wallet inspector’ trick.
Yes, it’s just Shazam but that’s the phase we’re in. Pick an app you like, can you replicate it with a Custom GPT? This is a decent effort.
This is a deeply unsexy one but also the most relevant. Just a tool for answering questions about UK politics, specifically around the code of conduct (I couldn’t yet find a US politics equivalent). Because this is where Custom GPTs can be game-changing, think of every FAQ section you have ever encountered. Or every frustrating customer-facing chatbot that can’t give you a simple answer. Any company, or state body, with a big ol’ pile of information can just load it into a Custom GPT without shelling out for a developer and immediately have a decent alternative. The maker of this tool flagged that it doesn’t really live up to its promise and he wrote something that covers plenty of AI content right now, ‘The initial excitement was eroded by the difference between producing good output and producing something that LOOKS like good output’.
This one is getting a lot of love on socials. I scribbled an incredibly basic sketch:
And it came up with this:
If you’re neck-high in gen AI tools every day then maybe none of this is that exciting but this could be the start of a wave of tools that explain AI concepts to a wider audience.
Small bits #1: Predicting predictions
Two conflicting quotes from Open AI founders are getting some airplay in AI-land this week. President and CEO Greg Brockman told French President Emmanuel Macron, ‘We were able to precisely predict key capabilities before we even trained the model’ in relation to Chat GPT-4 development. While Sam Altman told the Financial Times, ‘Until we go train that model, it's like a fun guessing game for us,’ in relation to GPT-5. That’s a worrying disparity in company lines.
Small Bits #2: AI and freelancers
New data shows a dramatic impact that generative AI has already had on white collar workers in the gig economy. Financial Times Chief Data Reporter John Burn-Murdoch breaks down the data in further detail in a thread but this tallies with last week’s look at stock photographers,