LLMs faking it, Big Brother and Photoshop AI obsolete
Notes From the Desk: No. 9 - 2023.09.29
Notes From the Desk are periodic posts that summarize recent topics of interest or other brief notable commentary that might otherwise be a tweet or note.
LLM’s are faking it
AI Explained reviewed a new paper The Reversal Curse in which it is revealed some significant flaws in LLM reasoning:
We expose a surprising failure of generalization in auto-regressive large language models (LLMs). If a model is trained on a sentence of the form “A is B”, it will not automatically generalize to the reverse direction “B is A”. This is the Reversal Curse. For instance, if a model is trained on “Olaf Scholz was the ninth Chancellor of Germany”, it will not automatically be able to answer the question, “Who was the ninth Chancellor of Germany?”
The takeaway expressed was that LLMs exist somewhere in between just memorization and intelligence. It can perform some rather complex tasks while simultaneously failing at what should be extremely simple tasks.
The concern I would emphasize is that this sets up the scenario where LLMs or something like them become overly trusted due to their impressive capabilities and then fail in spectacular unexpected ways.
This was something I mentioned at the beginning of the year.
So who needs intelligence when you can still produce meaningful results? The lack of true intelligence will likely reveal itself in unexpected ways. Which brings back the issues of trust again. AI can be deceitfully good at appearing competent when it actually has no understanding of the material. This might lead us to put too much responsibility on AI systems that then fail in some surprisingly catastrophic manner.
You probably don’t want LLMs automating important components of your business unsupervised.
More from the spy files
One thing AI seems to be really good at and never disappoints is its ubiquitous applicability for spying.
The rule for all new technology adoption is, if it can be used for spying it will be used for spying.
CIA Builds Its Own Artificial Intelligence Tool in Rivalry With China
Meta announces next wave of AI integrations
Photoshop content aware fill already obsolete?
Meta demoed the upcoming enhancements to their apps. Among those are significant AI image generation and manipulation capabilities which will now be built into Instagram.
Yes, in only a few months, Photoshop generative fill is no longer needed for a lot of current use cases. Social media is going to build these capabilities directly into the apps. The capabilities Meta demonstrated are very good and have the convenience of simply being embedded in the app. No need for that Adobe subscription any longer.
This is the nature of AI competition, in one moment you are on top of the world providing new capabilities unimaginable and in the next moment you are already obsolete.
Meta announces new AI glasses - We are all Big Brother
These new glasses are now approaching the design level of where they appear like normal glasses. Which means they will likely have much higher adoption than any other such previous attempts.
However, what is interesting will be the side effects of such adoption. Such as expressed below.
“What I would kill for is always-on video recording glasses that let me hit a button and save the last 10 seconds of what I've seen.
Imagine your kid doing something cute and just being able to save it forever, without having to get a camera out in advance.”
Essentially, we will become a willing part of the spy apparatus to spy on each other. Social media has already been a large part of everyone giving up their personal information voluntarily. Now there will be a camera recording everyone, everywhere, at all times so you can capture that moment in hopes of going viral. Big Brother is now all of us.
Meta also introduced new AI chat bots that will emulate celebrities or you
One of the more disturbing features Meta is rolling out is an emulation of people. There are several celebrity chatbots that have realistic avatars with which you can have conversations. They are also going to offer the ability to train new chatbots and avatars with your own likeness and personality.
We are rapidly building the civilization of illusion. I’m wondering how many people actually want this. I can perceive it as an interesting novelty for a moment, but who wants this to become ubiquitous in which nearly all interactions are fake or undiscernible as real?
The myopic AI art debates continue
The ethics, copyright and ownership arguments are always front and center in the debates of AI art. However, I feel these arguments are myopic. They focus on current capability in the present and current methods.
What happens when the AI can produce art, better or equal to humans, without the need for consuming the whole world of data? When it can replicate your style without training and produce infinite variations. It is inevitably where this goes unless AI fails to become the AI of intention by the builders. Furthermore, AI will dominate all mediums of expression including personhood itself. It is a replacement for humans, that is the target. How do you adapt to that outcome?
I previously wrote in more detail about AI art and the implications in AI art challenges meaning in a world of illusion.
No compass through the dark exists without hope of reaching the other side and the belief that it matters …
The biggest issue is that AI is indeed a replacement for personhood without being a person in any way. I do think that artists will remain unique in being human, and thus having actual embodiment, feelings, etc. That said, humans are easily deceived by superstimuli so even though the product is genuinely less rich and meaningful, it may swamp out human art.
I should add that you and Ztiv are my favorite AI commentators. Thank you for keeping up.