Generative AI applications for generic business enterprise use cases — i.e. threats to the professional-managerial class — are well-trodden, so I put down a few slightly oblique possibilities1 outside that, prompted by ChatGPT-4’s release.
A few of these require always-listening omni-directional mic with speech-to-text transcription and API upload to cloud. Yes, this means minimum privacy.2
In most cases you could probably have ChatGPT-4 write the code for you and run it yourself. Say, on your MacBook using something like Sveinbjorn Thordarson’s hear. Always-listening components and most API-enabled integrations are pretty straightforward.
The “bonus” options in most cases couldn’t be built this week, but everything is on the way. The possibilities will also expand further with image support, which isn’t yet public.
1. The Divorce-Enjoyer: Always-on marriage counselor
A dystopian option to kick it off.
An app for couples to track conversational marital patterns. How can you work on being more supportive? Who's always right? (Caution: may swiftly cause divorce).
Correlate with heart rate data from Watch/Oura, HRV, sleep data via API on back-end.
Enriched insights
More stress? Eating habits, sleep quality?
Does conversational health improve after, before exercise? Are there correlations with HRV and other metrics?
Are there certain times of day you have a healthier marriage? On vacation? With or without the kids?
An AI warning in your ear telling you not to be too flippant this morning because past conversations with this stat profile have led to degraded marital health.
There are technical implementation challenges with including integrated data sets, but the likeliest solution is to store the “dumb” (i.e. non-generative) collected data in the application storage and select what data to upload to GPT for correlation with the right crafted prompt.
As with other use cases, this will become more trivial and technologically accessible when competitor models3 emerge with different training sets and capabilities.
Bonus: Remove the "stats," tracking, etc. from user view and place recommendations in your ear, make assessments of your marriage — opaquely and because the AI says so. The Steve Jobs approach of making the technology invisible to the user.
Extra dystopian bonus: Incorporate it into a humanoid physical entity with normalized human-like language, e.g. Siri, and keep it around the house to watch and judge you. $899 Marriage Counselor Bot.
Privacy score4: 1 of 5 (1 because it can work with differential privacy methods)
2. The Television Show Enjoyer: Help me understand what I’m watching
For those watching the Tencent version adapting Liu Cixin’s Three-Body Problem (YouTube playlist: Three-Body).5
An app that listens to your show or movie while you watch and explains what’s going on or improves on translation if you’re watching in a foreign language with subtitles.
Pause, ask “what just happened?” and receive an explanation.
Pause, trade theories as to what might happen next.
This may suffer from inability to identify characters based on audio alone, but you may be able to provide the .srt (subtitle) file — or wait for processing and training power to support video with some clever API solutions.6
This can be done with audiobooks as well.
Summarize where you left off in The Confusion and what the hell Jack is doing in Ahmedabad.7
Bonus: Learn your habits, interests, and levels of understanding across genres to prepare you with advice such as whether to watch before bed, whether it’s hard to follow for you, or preemptive background the AI knows you’ll need.
Spooky bonus: Identify government stake in content produced by heavily stated-intertwined production houses like Tencent in the People’s Republic of China and generate national interest assessments based on content, funding, and distribution.
Privacy score: 2 of 5
3. The Content-Creation Enjoyer: Stream-of-consciousness onto paper
Personal scribe to listen to your insightful conversations with friends/colleagues or your partner, then translate them into coherent articles and tweets.
Never lose a great idea again.
No need to stare at your screen four hours trying to come up with the opening sentence.
You'll need to edit, but it'll be 1/20th of the work.
Bonus: Speak a cue term, such as "that was really deep, bro/sis! I think I'm a genius!" to trigger the AI to analyze the last five minutes of conversation only.
Dystopian bonus: The AI generates these without asking, sorts them and analyzes them by likelihood of going viral or likelihood of being accepted into specified publications, and presents them to you every Monday morning for selection.
Extra dystopian bonus: Integrate with a Bluetooth sobriety test device. Sell the module for $199.
Super-extra nationstate dystopian bonus: The same, but connect it to government censors.
Privacy score: 1 of 5 (1 — differential privacy — until authoritarian governments operationalize it, then 0)
Better privacy possible with on-demand recording instead of always-listening, but you still upload to someone else’s cloud (OpenAI in this case).
(indeed, this is all as if the world needs more “content…”)
4. The Presentation-Giving Enjoyer: Generate draft decks
Google Slides will soon support generative AI image and audio creation, but what we really need is a PowerPoint/Slides/Canva generator based on dictation.
Bonus: Let me critique and edit slides inline in a conversation with the AI.
Extra bonus: The AI learns from my tone, preferences, and rehearsals; then edits appropriately without prompting.
Privacy score: 1 of 5
5. Basic Cybersecurity Analyst: Generate event summaries
I’ve tested this one already. Take your CVEs, RSS feed alerts, and other content via API and run it through a prompt requiring formatting into an “actionable rapid alert product” as part of your security feed, your vendor’s email list, or whatever.
Since GPT-4 supports up to 25-32k words, you can give it a number of instructions on how to (appear to) write in the format you like.
Provide Words of Estimative Probability (WEP) as a baseline in the prompt. Explain WEP in your prompt.
Provide the core content of the US Office of the Director for National Intelligence’s Analytic Standards8. Optionally provide key content from Heuer.
Wayne Michael Hall is even better, but more suitable for larger-resourced custom training with competing models — or, minimally, larger token limits.
Insert the content to summarize (remember this is all via API).
Prompt for a title including bottom line up front and other key characteristics.
Bonus: Build a more custom script to update links and nodes in your threat intelligence platform of choice based on ingested data and construct your own automated front-line “analyst” team.
While hardly likely to yield the type of rigor required to avoid invading Iraq, it will produce better output than many commercial cybersecurity vendors.
Same use case can apply to generating priority intelligence requirements, rudimentary detection content & testing, and numerous other tasks.
Privacy score: 4 of 5. (The content is likely already public and you risk only the outputs, which probably carries licensing implications I haven’t explored anyway).
Bonus exercise: Go through the NICE Framework and identify which skills & abilities in which job series could be automated via API with a large language model. Do the same for US Federal Government job series.
Special US Government bonus: Do the above exercise, but under force of congressional mandate.
Dark Dystopia Bonus: The Life Choice Recommender
In a spin on #1, you can take the same concepts and apply them to making life choices. For that particularly malleable friend in your life. Or to take to the next level what social media has already wrought upon impressionable young adults.9
Always-on listening to you for one year to build a baseline, every moment.
Access to your metrics mentioned in #1, plus other data available such as social media via API.
Given any decision, such as where to move or who to date (available via API as well), the AI provides a recommendation.
As with #1, there are certain technical implementation challenges, but they’re achievable.
This could run into ethical alignment restrictions, although I think they are all evadable, but I don’t see why we shouldn’t expect it to be built. The key with applications like these is they don’t present themselves as all-encompassing intrusive devices — rather, you can expect it to be wrapped in a promise of improving mental health or life coaching. These applications may not be popular in every region, either, so merely because one culture is opposed to it does not mean there is no market globally.
Privacy score: 0 of 5. Zero point zero.
Spooky Bonus: The Propaganda & MISO Content Creator
You can use large language models to generate propaganda content and disseminate it via API with repeat less-detectable iterations.
As I no longer think this is novel in the discourse and I’m not inclined as a matter of intellectual honesty10 to have GPT-4 write the rest of my thoughts on this in the style of the blog above, I’ll link to my earlier thread on this.
Key takeaway is that creating propaganda is like other use cases for what is in some aspects a noise machine: it’s easy to create junk and measuring effectiveness remains dubious, but a trained hand could create content with the right prompt and strategic intent.
Afterword
My playing around with this doesn’t mean I’m a complete believer in these technologies in current form. They’re moving fast — so words are liable to get eaten later — and I think this has serious utility, but implementation across society won’t necessarily unravel the way prevailing narrative presents. Here’s a thread I find insightful on this topic worded better than I can offer:
(Plus a couple in the aforementioned category I wish I had since I regret to inform my readers I am, in fact, also still stuck in the PMC grind).
Shifting zeitgeist among an emerging privileged class notwithstanding, history proves “differential privacy” will apply and lead to products anyway if the value proposition is sufficiently compelling and otherwise painless. That means both for products where your data is sold and where it’s not.
There are a number of competitor models emerging such as Facebook’s now-leaked LLaMA and Google’s myriad projects which may partly explain why OpenAI no longer discloses significant technical details regarding its model’s training.
The privacy score is never going to be superb so long as you are uploading to massively pre-trained datasets like OpenAI’s ChatGPT. Training within your own storage is possible with other large language models. We have now re-entered the era of sub-enterprise and personal consumer interest in high-performance computing.
The Chinese (Tencent) adaptation of Liu Cixin’s Three-Body Problem recently released ahead of a Netflix adaptation coming later. This is worth watching for the compelling sci-fi ideas, but also as a look into one popular Chinese novelist’s perspective on existential conflict.
But read the novels first.
Copyright protections could make this difficult, but a future product could facilitate appropriate licensing agreements.
You should read, not listen to, The Baroque Cycle.
Freedom from political considerations will choke ChatGPT, but via API you can select an “alignment” and insert your own bias. This is one of numerous reasons this will generate summaries, not real human-like analytic output (where bias is omnipresent; but just harder to identify).
We should probably address this before it happens.
But what would stop me, or anyone?