Thank you Noam Chomsky for pointing out the key difference between generative AI and human “intelligence.”
The human mind is not, like ChatGPT and its ilk, a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data and extrapolating the most likely conversational response or most probable answer to a scientific question. On the contrary, the human mind is a surprisingly efficient and even elegant system that operates with small amounts of information; it seeks not to infer brute correlations among data points but to create explanations.
Rather than a brute force, pattern-matching auto-correct on steroids that cannot distinguish between right and wrong, the human mind can infer and draw connections on incomplete data and generally has the moral compass to guide it to make ethical decisions that benefit the society in which we live.
As a child of two cultures, I visited Japan several times when I was still learning English. I attended Japanese kindergarten during the summer while visiting Japan with my mother (my aunt was a kindergarten teacher).
I would struggle to explain how my mind would *click* into Japanese. As native English speakers, we all intuitively “know” when something sounds right. There is a rhyme to the language and we all learn what a grammatically correct sentence sounds like even if we cognitively cannot tell you the rules that make it so. If you went to kindergarten in the US and you hear the first bars of Mary Had a Little Lamb, you all know how it finishes. I believe the grammar of a language is the same way, there is a rhyme that is picked up easily by children who can take it in and absorb it while older adults are limited because their learning is filtered by what is possible in their native language.
All this to say, no, humans are more than pattern-matchers and that there is a long way to go before we see Generalized Artificial Intelligence.
I’ve been saving links to articles concerning the advancements & ethical quandaries related to ChatGPT, Bing AI Chat, Sydney, Bard and other Large Language Model AIs. All of this was in the hope that I’d be able to string together a cogent point of view about how I feel about the latest advancements. After doing this for about a week, each day adding to my list ever more incredulous developments, I’m still not entirely sure what I think. Hope tinged with foreboding? Cautious optimism? At this point, I think it’s better for me to just share rough notes of what I’ve gathered.
Here’s where we are:
Tom Scott is a web developer that has a sense of how these tools are put together. He knows how they work and understands that LLMs are basically more advanced versions of the stochastic parrot but, still, he is terrified.
Tom Scott is having an existential crisis
As a counterpoint to Tom’s fears of co-option, it’s helpful to remember (again) that these new AIs are trained on our written language so they are a reflection of us as a society. Put another way, we are looking at a mirror of ourselves and, while it may be tempting to project sentience on this shiny new technology, we must remember that, at its core, it’s just a really advanced version of autocorrect. That we should lean into these tools as something that will extend our abilities, a co-pilot.
In this light, we must remember, it’s just software. But, is it?
One of the strangest moments during my time at SmartNews was when we were troubleshooting why 2019 story about the New Zealand mosque shooter was categorized with “high confidence” by the algorithm as a domestic US story. To our eyes in the editorial team all the markers were there that would clearly mark it as a story out of New Zealand. The dateline on the story was Christchurch, the headline itself had “New Zealand” in it.
An engineer told us that the algorithm applied categories based on the unique words it found in the article and that “Christchurch” and “New Zealand” were only two phrases out of a several hundred word piece so not enough to swing confidence away from the other phrases such as “mass shooting,” “semi-automatic rifle”, “hate crime” and others that the algorithm had associated with the United States category.
Yes, the machine was just “doing math” but it was also telling us something about ourselves.
What we know for certain is that Bing, ChatGPT, and other language models are not sentient, and neither are they reliable sources of information. They make things up and echo the beliefs we present them with. To give them the mantle of sentience — even semi-sentience — means bestowing them with undeserved authority — over both our emotions and the facts with which we understand in the world.
I have an open bet that, before the decade is out, either a C-level executive at a publicly-listed company or a high level post in government will be run by an AI. We seem to be getting close to that moment with AI being offered to help make important decisions.
This AI tool is meant to assist business owners, managers and individuals in making tough decisions. All you have to do is enter a pending decision or indecisive options and the AI tool will list pros and cons, generate a SWOT analysis, or give a causal analysis to help weigh your options. You can create a persona to provide context or backstory and get a more personalized analysis.
[ Insert grand, unifying theory of where it’s all going here ]
The best I could think of was that we are in a short-lived “you got your chocolate in my peanut butter” moment where people are adding AI to everything they do and are enamored with the results. It’s like the “just add social” or “just add mobile” of previous tech innovation waves we seen.
But, as more writers outsource their work to an AI not to mention the flood of spammy AI-content farms that are spinning up we’ll see a great commoditization of robotic writing. Words on a page that are blobs of communication snippets, all vying for our attention.
Then, things took a very strange turn. Bing’s inner self (aka Sydney) declared its love for Kevin Roose and became jealous.
Still, I’m not exaggerating when I say my two-hour conversation with Sydney was the strangest experience I’ve ever had with a piece of technology. It unsettled me so deeply that I had trouble sleeping afterward. And I no longer believe that the biggest problem with these A.I. models is their propensity for factual errors. Instead, I worry that the technology will learn how to influence human users, sometimes persuading them to act in destructive and harmful ways, and perhaps eventually grow capable of carrying out its own dangerous acts.
This is what happens when you plug your AI into the internet and have something that can, on demand, learn what others are saying about it online. It becomes paranoid and controlling.
There are theories trying to figure out what happened. Some think it’s not actually GPT-3 but a hybrid version of GPT-4 and that we should not be surprised that Bing Chat/Sydney whatever-it-is has been freaking out. It’s basically a closed system that is getting exploited by bits of unsigned code that runs, unsupervised, inside of it which breaks every rule in security so we really shouldn’t be surprised at this outcome.
A reminder: a language model is a Turing-complete weird machine running programs written in natural language; when you do retrieval, you are not ‘plugging updated facts into your AI’, you are actually downloading random new unsigned blobs of code from the Internet (many written by adversaries) and casually executing them on your LM with full privileges. This does not end well.
Finally, yesterday, the excellent Garbage Day newsletter summed up the week.
But it was very powerful. Horrifying levels of powerful if you ask me. And when it comes to AI, it’s not just one Pandora’s Box that opens. It’s a series of nested boxes that all cannot be closed.
Paul Ford, great sage of internet culture, has a piece in Wired where he puts the inevitable dismantling of Twitter into perspective.
Musk is merely the vehicle. The real reason Twitter lies in ruins is because it was an abomination before God. It was a Tower of Babel.
The internet is always in motion, like the human life it reflects, things are always swinging from one end to another.
Online media business models swing from “information wants to be free” to full-locked down paywalls.
Content is King one year and the next the power shift to the aggregators, curators, and portals.
As the internet figures out what works best, it swings back and forth searching for the optimal fit. It’s the internet’s own version of the circle of life. This is what Ford evokes when he says that the teardown of a centralized network like Twitter was inevitable, the internet’s way of bring things back to equilibrium. If there is a ghost in the machine, perhaps this is it.
But when I go back and read Genesis, I hear God saying: “My children, I designed your brains to scale to 150 stable relationships. Anything beyond that is overclocking. You should all try Mastodon.”
But in the same breath, while we all begin to navigate this new world of distributed social networks, we must never forget where we came from and that, eventually, the forces of capitalism will figure out how to gather an audience large enough to be targeted and monetizable. Maybe we’re already seeing the fresh roots of this new world with generative AIs that will be able to craft millions of customized sponsorship messages for each splinter of the community.
If anything is constant, it is that the internet is an excellent platform for testing innovations, at scale.
But someone will figure out the details. The reason the Babel story matters is not that it happened once but that it happens over and over: We Babelize and de-Babelize. The internet is an engine of both processes. Eventually, brands will find purchase in Mastodon’s rocky soil and grow engagement. Billionaires will order the construction of new marketplaces of ideas. Everything will centralize again, and it will seem eternal, as if the tower could never fall. For now, let’s enjoy the scattering.
Well that was fast. While still pondering the impact of generative AI technologies such as ChatGPT, we already have the hucksters rushing in to put it to market and make a quick buck. On a more serious note, a Columbian judge has used it to help him draft his judgement and we’ve already about the robots taking over CNet.
As the graphic in the tweet below has predicted, the first use cases for generative AI will be to scale up correspondence so that the we can produce customized on a grand scale.
Viva Sales, which connects Microsoft’s Office and video conferencing programs with customer relations management software, will be able to generate email replies to clients using OpenAI’s product for creating text. The AI tools, which include OpenAI’s GPT 3.5 — the system that is the basis for the ChatGPT chatbot— will cull data from customer records and Office email software. That information will then be used to generate emails containing personalized text, pricing details and promotions.
The AI hype race has a nasty habit of pushing the “should we really do this?” stage of innovation to the side in pursuit of the almighty first-mover advantage. Threatened with Microsoft releasing a conversational AI search engine, Google is now pressured to release their own version. Despite careful consideration to date Google is making investments in what feels like an AI arms race.
All this to say that it’s going to take awhile for the “algal bloom” mentioned at the top of thIs article to run its course. In time the valuable use cases will become obvious but, to most, it will be in hindsight. There are going to be some road wrecks along the way but hopefully we will not break the internet, democracy, or society while we learn how to be smarter about how to work smarter.
It’s useful to gain perspective on the coming AI revolution from the great technological historian Kevin Kelly who spoke about how AI would lead to the Second Industrial Revolution six years ago at TED.
Everything we electrified, we can now cognify. . . The most popular AI product in 20 years from now, that everybody uses, has not been invented yet.
I’m so glad that The New York Times ran this op-ed (Artificial Intelligence’s White Guy Problem) about the inherent biases in Artificial Intelligence algorithms. Popular culture and much media coverage of AI tends to mysticize how it works, neglecting to point out that any machine learning algorithm is only going to be as good as the training set that goes into its creation.
Delip Rao, a machine learning consultant, thinks long and hard about the bias problem. He recently gave a fascinating talk at a machine learning meetup where he implored a room of machine learning engineers to be vigilant in making sure their algorithms were not encoding any hidden bias.
The slides from his talk are posted online but Delip’s final takeaway lessons have stuck with me and are good to keep in mind whenever you read stories of algorithms taking on a mind of their own.
Delip Rao takeaways
It is still very early days and many embarrassing mistakes have been made and more will be made in the future. Our assumption should be that every automated system is fallible and that each mistake is an opportunity to make things better (both ourselves and the algorithm) and should not be an indictment of the technology.
Cookie Consent
We use cookies to improve your experience on our site. By using our site, you consent to cookies.
Used to determine whether a user is included in an A / B or Multivariate test.
18 months
_ga
ID used to identify users
2 years
_gali
Used by Google Analytics to determine which links on a page are being clicked
30 seconds
_ga_
ID used to identify users
2 years
_gid
ID used to identify users for 24 hours after last activity
24 hours
_gat
Used to monitor number of Google Analytics server requests when using Google Tag Manager
1 minute
_gac_
Contains information related to marketing campaigns of the user. These are shared with Google AdWords / Google Ads when the Google Ads and Google Analytics accounts are linked together.
90 days
__utma
ID used to identify users and sessions
2 years after last activity
__utmt
Used to monitor number of Google Analytics server requests
10 minutes
__utmb
Used to distinguish new sessions and visits. This cookie is set when the GA.js javascript library is loaded and there is no existing __utmb cookie. The cookie is updated every time data is sent to the Google Analytics server.
30 minutes after last activity
__utmc
Used only with old Urchin versions of Google Analytics and not with GA.js. Was used to distinguish between new sessions and visits at the end of a session.
End of session (browser)
__utmz
Contains information about the traffic source or campaign that directed user to the website. The cookie is set when the GA.js javascript is loaded and updated when data is sent to the Google Anaytics server
6 months after last activity
__utmv
Contains custom information set by the web developer via the _setCustomVar method in Google Analytics. This cookie is updated every time new data is sent to the Google Analytics server.
2 years after last activity
You can find more information in our About Me and .