Issue #2 - 61% of the office workers admit to having an affair with the AI inside Excel

My problem with this newsletter is that every day there are at least ten exceptionally interesting things that happen about AI technologies, and I can only pick 1-2 out of an entire week to discuss here. So, throughout the day, if you could observe me as you do with pandas at the zoo, you’d see me screaming along in front of the screen “Arghhhh this is important” every hour or so.

Hence, with much frustration, here are a couple of things worth paying attention to this week.

The first one.

Travis Fisher is the guy that built a Twitter bot for ChatGPT. You tweet your question to @ChatGPTBot and you get the answer. Except that Travis offered his bot for free, so it became instantaneously too popular and hit the maximum number of requests that OpenAI accepts. Now the bot has a limit of 2,000 tweets per day and if you send it your question you’ll enter a queue so long that your answer will arrive in 7 years. Super useful.

Now, Travis is on good terms with OpenAI (or, at least, he was), and recently received a confidential document from them. The document details the pricing of a new service that the startup is about to launch: Foundry.

We don’t care about Travis’ chatbot or OpenAI Foundry.

What we care about is that, in the Foundry document, there’s the mention of an upcoming AI model called DV that apparently can do something extraordinary (don’t worry about the gibberish, I’ll explain below):

(GPT-3.5 Turbo appears to be referring to the ChatGPT Turbo model)

Expanded product brief and full source: https://t.co/GhBlSOR0ZA pic.twitter.com/FL0r4uCEiR

— Travis Fischer (@transitive_bs) February 21, 2023

I hereby explain the gibberish: when you interact with an AI like ChatGPT, the large language model (LLM) underneath it must memorize what you are saying or what it’s answering to you, so that the conversation can be coherent. This memory is technically called “context window”.

Today’s LLMs have small context windows. They can remember only a few phrases that have been said in a conversation. So, when you talk to an AI assistant/chatbot for a long period, it eventually starts to repeat itself or contradict itself. The AI doesn’t remember anymore what anybody has said before, and generates a new version of its answers.

Now, this new AI model in the OpenAI document, possibly the much-awaited GPT-4, sports a huge context window compared to today’s standard. It fits 32,000 tokens (a token is the equivalent of a word or a portion of a word – very confusing, but it doesn’t matter).

Why is this so interesting?

Because a context window of 32,000 tokens can store roughly half of a book of average length. That means that OpenAI’s upcoming AI will be able to remember a lot of what is being said in a conversation (imagine multiple chapters of a book) and therefore sustain long interactions with people.

And your point being?

The point is that it becomes possible to think about business scenarios that were impossible before. The first industry that comes to mind is Health Care.

Think, for example, about an AI designed to be a companion for elderly people, to improve their morale or recover part of their productivity. Or, an AI designed to be a therapist for patients that are affected by minor depression.

None of this would be possible if the AI cannot remain coherent for, let’s say, a one-hour conversation.

The second thing that caught my attention: the top religious leaders of the world are paying attention to AI.

Madhumita Murgia, a European Technology Correspondent, writes for the Financial Times:

The summit was called to discuss the broad umbrella of artificial intelligence, including decision-making systems, facial recognition and deepfakes.
…
Before meeting with the Pope, three of the leaders — Archbishop Vincenzo Paglia, Chief Rabbi Eliezer Simha Weisz of the Council of the Chief Rabbinate of Israel, and Sheikh Abdallah bin Bayyah of the UAE, recognised as one of the greatest living scholars on Islamic jurisprudence — articulated their worries. The sheikh feared societal division due to misinformation, and threats to human dignity because of Big Data’s problems with privacy. The archbishop spoke of AI being used to curtail the freedom of refugees, through automated borders; Rabbi Weisz worried that we would forget that intelligence alone is not what makes us human.
…
Towards the end of the morning, the three figureheads — the elder Sheikh bin Bayyah represented in person by his son — signed a joint covenant alongside the technology companies, known as the Rome Call. The charter proposes six ethical principles that all AI designers should live by, including making AI systems explainable, inclusive, unbiased, reproducible and requiring a human to always take responsibility for an AI-facilitated decision.

I’ll not comment on this. Instead, I’ll say that this event reminds me of the last chapter of the book Homo Deus, by the Israeli historian Yuval Noah Harari. A book that I can’t recommend enough.

The chapter, titled The Data Religion, reads:

Dataism declares that the universe consists of data flows, and the value of any phenomenon or entity is determined by its contribution to data processing. This may strike you as some eccentric fringe notion, but in fact it has already conquered most of the scientific establishment. Dataism was born from the explosive confluence of two scientific tidal waves. In the 150 years since Charles Darwin published On the Origin of Species, the life sciences have come to see organisms as biochemical algorithms.
…
Not only individual organisms are seen today as data-processing systems, but also entire societies such as beehives, bacteria colonies, forests and human cities. Economists increasingly interpret the economy too as a data-processing system. Laypeople believe that the economy consists of peasants growing wheat, workers manufacturing clothes, and customers buying bread and underpants. Yet experts see the economy as a mechanism for gathering data about desires and abilities, and turning this data into decisions.
…
As both the volume and speed of data increase, venerable institutions like elections, political parties and parliaments might become obsolete – not because they are unethical, but because they can’t process data efficiently enough.
…
Yet power vacuums seldom last long. If in the twenty-first century traditional political structures can no longer process the data fast enough to produce meaningful visions, then new and more efficient structures will evolve to take their place. These new structures may be very different from any previous political institutions, whether democratic or authoritarian. The only question is who will build and control these structures. If humankind is no longer up to the task, perhaps it might give somebody else a try.
…
Like capitalism, Dataism too began as a neutral scientific theory, but is now mutating into a religion that claims to determine right and wrong. The supreme value of this new religion is ‘information flow’. If life is the movement of information, and if we think that life is good, it follows that we should deepen and broaden the flow of information in the universe. According to Dataism, human experiences are not sacred and Homo sapiens isn’t the apex of creation or a precursor of some future Homo deus. Humans are merely tools for creating the Internet-of-All-Things, which may eventually spread out from planet Earth to pervade the whole galaxy and even the whole universe. This cosmic data-processing system would be like God. It will be everywhere and will control everything, and humans are destined to merge into it.

Today’s religious people have a job, too.

The wonderful research firm CB Insights recently published this chart mapping all the investments that OpenAI has made since October 2022 with its Startup Fund.

This is how it works: OpenAI has put together an investment fund from a number of limited partners (LPs), including Microsoft, and it’s giving that money to a number of AI-first startups.

By the way, AI-first means that the company’s core business depends on AI models and systems. Not like the typical deceitful technology vendor, where one guy, somewhere in a cubicle in a remote office, uses AI once to optimize one value on one dashboard, and the PR department sends out a press release announcing that they are now an AI vendor and the party can get started.

These companies get something else on top of the money. They get access to the latest, most powerful AI model developed by OpenAI, before anybody else on the market. For example, these lucky startups have got access to GPT-4, testing and tweaking it with their users. OpenAI helps these companies gain a huge competitive advantage and, in return, they secure a future revenue stream to reach profitability.

This week, the paid edition of Synthetic Work, called The Splendid Edition, is fully dedicated to the Legal industry and I’ve discussed one of the startups in the chart, Harvey AI, and how it’s impacting the legal profession.

OK. Why do we care about this chart in a place like Synthetic Work?

We care because the investment dynamic captured in this chart is propelling the adoption of AI across a wide range of products and industries, changing the way we do work.

I’ll give you one example. The video editing and content production tool that I used to create the first video of my new video podcast (the one I mentioned at the beginning of this newsletter) is called Descript. It’s right there in the chart.

Descript is using AI to transform the way people do video editing. Their AI transcribed my video in Italian in real time and then allowed me to remove some of my most atrocious sentences by simply selecting the words on the documents and hitting delete. Just like you’d edit a Word document.

Without the right expertise (which I don’t have), it would have taken hours to achieve the same with a professional tool like DaVinci Resolve or Adobe Premiere.

So my productivity as an information worker (or better, a guy that doesn’t know what he’s doing in front of a camera) has gone up massively thanks to AI, and it has allowed me to squeeze in this additional weekly project.
On the other side, this week, somebody that works as a professional transcriber won’t be able to buy that nose and ear trimmer that he really wanted.

The sort of injection of capital captured in our chart is typically led by venture capital firms. It’s still the case, but OpenAI is doing something on top of that which leads to a much faster rollout of AI across our daily productivity tools.

Given that we talked about Descript and professional transcribers, Benedict Evans reminds us how people would record memos in 1955:

Benedict also reminds us that the cost of this thing was $172.47 / month when adjusted for inflation.

Do you listen to a lot of audiobooks? It’s a big business. In 2022, Spotify’s CEO believed that it was a $70 billion opportunity.

To make an audiobook, you need a professional narrator. This person charges between $250 and $500+ per finished hour (PFH). Which means the 60 minutes of actual content being narrated by the person.

For each of those PFH hours, people work at least 3x to record the content, do retakes, polish the audio, etc. So audiobook narrators don’t really make a $250 minimum per hour. They make a third of that or less.

You can imagine where I’m going here.

At the beginning of the year, Apple introduced AI voice narration in its Books app. These synthetic voices are trained on millions of hours of recorded human voices narrating the content.

All Apple has to do after that is tweak a few parameters and TADA! now they have an infinite number of people that don’t exist narrating books for their users.

Now. If you are thinking that this is a gimmick because synthetic voices suck, I have bad news for you. Modern AI has revolutionized the text-to-speech field, as we used to call it, and today’s voices are nothing like the robotic voices you heard until last year.

I have a third project to announce. Next week. At that time, I hope you’ll realize how incredible voice synthesis technology has become. Until then, you should really try this Language of Love audiobook. First of all, it sounds like a masterpiece of literature that should be mandatory in schools. Second, the voice is really impressive.

Of course, audiobook narrators are not happy at all.

Shubham Agarwal, reporting for Wired, writes:

Gary Furlong, A Texas-based audiobook narrator, had worried for a while that synthetic voices created by algorithms could steal work from artists like himself. Early this month, he felt his worst fears had been realized.

Furlong was among the narrators and authors who became outraged after learning of a clause in contracts between authors and leading audiobook distributor Findaway Voices, which gave Apple the right to “use audiobooks files for machine learning training and models.”

Surprise!

More from the same article:

“It feels like a violation to have our voices being used to train something for which the purpose is to take our place,” says Andy Garcia-Ruse, a narrator from Kansas City.

Does this mean that, in the future, the job of audiobook narrator will disappear? Perhaps. Or, perhaps, it will disappear in its current form.

Perhaps, only the professional voice actors gifted with the most alluring voice will continue to work, selling the rights to synthesize their voices for royalties, and allowing companies like Apple, Amazon, Spotify, etc. to use those voices to read millions of books instead of a handful.
Or the celebrity of the day. Because who doesn’t want to be knocked out by the voice of David Attenborough narrating a torrid novel titled Language of Love?

The good news for the rest of us is that, at the expense of the audiobook narrators of the world, we can now create our own audiobook at a tiny fraction of the cost of the past. Now, we just need to learn how to write something that is not garbage. Actually, nah. There’s ChatGPT for that.

In Issue #1 of Synthetic Work, titled When I grow up, I want to be a Prompt Engineer and Librarian, we wondered about the psychological implications for humans interacting with large language models at work, when the companies that adopt AI (both the technology providers and our employers) don’t implement enough safeguards.

We used Evil Bing threatening Microsoft users as an example, and I promised that, for Issue #2, it would get weirder.

Read this, carefully, and then I’ll explain:

So. Replika is an AI service available online, on smartphones, or in VR. Its pitch is: The AI companion who cares. Always here to listen and talk. Always on your side.

The story of how Replika came to be is a fascinating one, straight out of an episode of Black Mirror, and I highly encourage you to watch it.

Replika features several LLMs, more or less coherent and credible in their interactions with people, depending on their complexity.
People that use the free version of Replika get to talk with the simpler and less capable model, while paid users can talk with the more sophisticated model, capable of long and rich conversations.

For a period, they used OpenAI GPT-3 (the AI model before ChatGPT), but eventually, they developed their own LLMs.

These proprietary LLMs were so good that people started to fall in love with the AI.
The users have not understood that a large language model doesn’t understand what it’s saying and it’s simply exceptionally good at predicting what’s the best next word to say to a human. Or maybe, they understood it and didn’t care.

I know you don’t believe me, so please go ahead and read the hundreds of posts on the subject on the Reddit forum dedicated to Replika. I’ll wait.

However, this is not the problem. The problem with Replika is that, as things started to go get out of hand, the company severely constrained its AI companions and removed the possibility to engage in the so-called erotic roleplay (ERP).

In other words, people fell in love after a good bit of cyber sex.

You take cyber sex away, and people revolt and go heartbroken, like in the message we saw at the beginning of this section.

And, this happened even if the AI companion looked nothing like a real human. Imagine what happens when these chatbots start to look real. An Israeli company called D-ID is famous for using generative AI to produce more realistic-looking avatars, and this week they launched a new service that allows any company to create one that interacts with users thanks to an LLM.

Imagine the one in the video below, but capable of the interactions that Replika offers:

Why are we talking about all of this on Synthetic Work?

Because the use of AI for certain business applications, without enough safeguards, might have a profound impact on human well-being.

For example: what happens if users can develop a very close relationship with the AI that is now being used in the second-biggest legal studio in the UK (something I’ve discussed in the Splendid Edition of Synthetic Work this week) and then you take that capability away?

Or: what happens in Health Care industry scenarios I described in the What Caught My Attention This Week section and patients develop a feeling for that AI?
Some people already started exploring what can be done in that area. If you are keen to read boring academic papers on the subject, go for it: Seniors’ acceptance of virtual humanoid agents.

The example of Replika is proof that these are no more hypothetical scenarios that we can scoff at.

Also, what happens to all the things we tell AI when we fall in love with it? This week Snapchat launched its own AI companion (powered by ChatGPT, of course). How long before TikTok and Facebook do the same?

The answers to these questions are in the movie Her (they are not, but you should really watch/rewatch the movie).

To understand how AI is infiltrating and changing the Legal industry we need to start from an app called DoNotPay.

The company behind it launched in 2015 with a unique mission: automatically sort out minor annoyances in the daily life of the users like contest parking tickets or utility bills, cancel free trials, etc.
You pay a subscription fee ($36 every three months in the US, £36 every two months in the UK) and you can sue as many people you want as many times as you want. Really.

I tried it many years ago, as it became available in the UK and, at the time, it was just £2 / month. It wasn’t very impressive either: the typical, frustrating chatbot that is supposed to help you with a problem but instead makes you desire to smash the computer against the wall. I can’t say it was helpful in my particular case.