Hello World of AI 🤖
Progress in computer technology has been steady but incremental since the early 2010s, yes our cameras got more pixels, our networks got more Gs, and our phones are folding now; but it’s been a lot more of the same. The last time I was mind blown by a piece of technology was the Google Duplex demo in. Barring that, it’s been quite a lull.
That lull has recently been shattered. Since December 2022, there’s been a feverish excitement in the air. ChatGPT took over the airwaves and I (like many others) am genuinely impressed by its capabilities. Generative AI feels like a truly transformative technology to me that will change a lot of workflows. It’s already making an impact! People are using GPT based tools for everything from composing documents and emails, summarising information, to converting code from one language to another.
And it has had such a great impact in barely a year, can you imagine where we’ll be in 10 years? No wonder ChatGPT was one of the fastest growing consumer apps and a feverishly fast-moving ecosystem has emerged around it.
So let’s unpack this AI thing and look beyond the hype. What is it? What does it mean for us?
What is AI?
Simply put, ChatGPT is autocomplete on steroids.
Every time you type a query into Google search or type something on your phone keyboard you get a prediction of what the next word will be. ChatGPT is exactly this, however instead of predicting the next word, ChatGPT is built to predict hundreds and thousands of words more. Which means, it can take your query and spew out paragraphs and pages of responses. Because it is trained on a broad body of human knowledge which includes Wikipedia, Reddit, etc. It is able to provide realistic sounding responses. The responses are nothing more than a probabilistic guess and are often wrong, but they are right so many times that people get fooled into thinking that ChatGPT is alive.
For a more detailed intro, see Stephen Wolfram’s awesome post: What Is ChatGPT Doing … and Why Does It Work? — Stephen Wolfram Writings
How to think about AI — Meta commentary
Before we get to the implications of AI, here is some meta commentary to keep in mind. These are thoughts that should influence your thoughts when you’re thinking about AI. Understood? Cool!
#1 AI impact is not a single factor conversation
We tend to look at any potentially disruptive technology in a very single factor or mono causal fashion. However, there are always multiple trends occurring simultaneously.
An example from manufacturing: While automation threatened to take away manufacturing jobs in Europe and America in the 70s, that fear didn’t pan out at all. European and American companies didn’t automate their existing factories, they instead set up altogether new factories in Asia to exploit cheap labour. First to Japan, then Taiwan, then China, now Vietnam and India. Next probably Africa. So even though automation as a technology could have made an impact, economic factors which we can broadly lump under globalisation were the deciding factor in shaping the manufacturing landscape.
We always have multiple hands moving pieces on the chessboard. Politics, economics, social norms, cultural norms, etc. It’s incorrect to think of AI as a big tsunami washing over the entire world, though that makes for a wonderful investor pitch!
#2 Shove then love
How humans handle technological change has a funny but consistent pattern. First, we shove the technology into existing workflows. Later, as we get it working reliably, grasp the full spectrum of the possibilities and overcome inertia — we create new workflows around a technology. For example, the first TV shows were just motionless adaptations of radio shows until people adapted to the new medium.
The pattern is playing out again. Google and Microsoft are putting AI powered summaries in their search engines. The workflow of opening up the search engine and entering the query is still the same. Do you think someone in 2040 will still be typing into a text box? Or will an ambient AI agent on your phone intelligently interject your conversations with answers at just the right moment?
Relatedy, technologies roughly follow the hype cycle. Generative AI, today is at the peak of inflated expectations. So, very soon (<2 years), the over-hype-ing is going to lead to a crash.
#3 It’s all feelings, no facts today
You can easily divide people into camps based on their feelings.
- Fear — Fear of automation, job loss, change, inability to adapt, uncertainty that this will unleash
- Excitement — Excitement about getting doing work faster, getting rid of drudgery, applications/possibilities of generative AI, economic opportunities of all sorts
A chaos tunnel is coming and no one can see around it. All obvious successes 10 years from now will just be ex post facto rationalisations.
As a techno-optimist, I stand with both feet planted in the land of fields of excitement. My mind flies into 2030 when I have a Jarvis like assistant answering my messages, writing initial drafts of my documents, scheduling my meetings while I focus on creative/important work like sleeping well after reading a good book. Shudder! Will there still be books?!
We’ve known since long that the fabled knowledge work involves a fair amount of drudgery. No wonder Wikipedia and StackOverflow have been the most visited sites since forever. People think in terms of automation, not in terms of value creation. People are thinking about speeding up the old, not about creating the new. We as a society, need time to absord, play, and invent things with this new technology.
Metaphors for AI
#1 AI is the steam engine of knowledge work
One way to understand AI is to map to the previous technology that took away power from humans: the steam engine. The steam engine produced movement that competed with the labour of the human, horse, donkey, etc. However, the steam engine wasn’t just more horses. The steam engine rewrote the structure of the world through subsequent inventions of railways, mass manufacturing, etc.So it is with AI. Computers today are still human directed to a large extent. AI and LLMs are the steam engine of cognitive labour, bringing with them the ability to automate existing work. The second order effects will again be the more interesting ones as we push the technology into new and newer domains; and rewrite large parts of society and industry. Mass manufacturing is viewed as net positive because it was massively deflationary. But it did create lots of its own problems from urbanisation, epidemics and health problems, wealth inequality, etc. as society transitioned from agrarian to industrial. Computational mass manufacturing is unlikely to be without its own problem. You can expect lots of displacement during the transition
#2 AI is encoded knowledge
Another way to understand AI is that we are now programming computers with data instead of instructions.
I have personal experience with this phenomena in the field of computer vision. When I was a Bachelor’s student, one of my projects was about detecting security threats from camera feeds. An initial goal was to detect humans; we used hand-tuned parameters in shape detection algorithms to identify humans from other shapes. This is also how initial OCR programs worked. Later, I discovered how models trained on thousands of labelled images were providing a more robust way to detect humans. Instead of tuning manually, you just fed the models lots of data and got them to learn
LLMs work the same way. Instead of providing a long list of instructions, we provide a long list of examples. During the training process, the computer then stores its knowledge in the form of weights — which is nothing but a list of numbers. This is more deeply explained in Andrej Karpathy’s blog post: Software 2.0. I sometimes see people refer to neural… | by Andrej Karpathy | Medium
The encoded knowledge in an LLM has the same problems as a student. The books might be outdated, there can be questions from outside the syllabus, and the student may just make up answers to things it doesn’t know. This is exactly what happens with LLMs too.
Implications of AI
Consumer goods & experiences
The most visible impact of AI is going to be in consumer experience. All the fun and not so fun stuff from movies and science fiction books is coming soon.
Every device and software will have AI agents helping you along. Agents will be everywhere — Your operating system will have one, so will most of the software you interact with. In fact, some software will just be agents. Eventually these agents will figure out how to talk to each other; just like how we figured out how to share things across apps; whether it is through programatically through APIs or manually through the share button.
Today, Agents exist in the form of lamely scripted and occasionally human chat bubbles. In the future, they’ll be like Jarvis from Ironman, taking orders and doing a lot of the work for you. Just like Jarvis, they’ll have the ability to act on your behalf. A sort of power of attorney. They’ll have access to all your data — your documents, schedules, contacts, messages, etc. Like a personal secretary, they’ll lay down your schedule for you, taking away the trouble of planning your day manually. This will free you up for more important work. Of course, you’ll always have the power to override them. AI assistants of the future won’t be like the robotic automatons of today. They’ll be endowed with emotion and personality to make your interaction with them more acceptable and enjoyable. A common annoying situation is when someone leaves behind their phone in a room and it starts ringing. In the future, you’ll just be able to angrily scream Stop ringing! and the phone will mute itself. It will have awareness.
AI is only as useful as the data it has access to. While today our data is scattered across our phone, laptops, and various cloud services; in the future your agent will manage all your data for you. While it may still will be physically be stored in different places; your AI assistant will have access to it and will index it; so, it’s always ready when you need it. This will take the quantification of self movement to a whole new level.
OK some bad news: There will be shitty programmers out there. And your bank’s AI agent will probably be unreliable and sucky.
Social, political and economic impact
(No) Job loss — One of the biggest fears around AI is job loss. However, we’ve seen time and time again that automation creates new jobs. Why? Because humans don’t just automate things and sit there. We invent new things together, some on top of the previously automated things. Sure in the short run, a lot of data analyst jobs go away, but soon they’ll get replaced with prompt engineers, model QA and more. Additionally, if being a data analyst becomes easier thanks to AI; there’ll be a lot more data analyst jobs. They will however be but lower paid compared to existing roles.
This phenomenon is called the lump of labour fallacy. E.g. When automobiles and trains replaced horses, we didn’t just zip between towns faster. We squished multiple towns together and got dense cities, high rises and a whole new urban paradigm.
And remember you’re not the horse, you’re the driver! We shouldn’t forget that we do have agency to react to this change. Companies and governments should be pushed to reskill existing workers into newer skills.
Also, change is not one-sided. It’s actually more cost effective to replace the CEO with AI. How many layers of “management” “leadership” do you need? Cost-benefit wise it more impactful to replace highly paid execs with AIs than to replace cheaper rank and file employees.
Knowledge work: Barbell impact.
Just like knowledge work today, a 100 years ago manufacturing work was a lot of bespoke processes involving highly knowledgeable workers. Then came scientific management and assembly lines, and we got highly specialised workers. It’s time for Taylorization of knowledge work. Even today, lot of knowledge work is ridiculously inefficient with poorly organised information, overload of meetings and an overall lack of focus on output. Today every person juggles lots of tasks and has lots of tacit knowledge; like the workers in pre-assembly line car factories There might be an opportunity to create good knowledge work factories.
To some extent, it’s already happening but AI will sharpen the divide further. We are going to get
- More wages but fewer jobs in cases where AI increases worker productivity. E.g Software developers. Paralegals, etc.
- New categories of lots of specialised low wage jobs where AI does most of the heavy lifting.
Augmentation = New knowledge work jobs
In many domains, the impact of AI will look closer to Google Maps than to self-driving cars. Google Maps allows a person with absolutely no knowledge of a city to drive perfectly well in it — anywhere in the world. Uber then takes that new skill and creates a market for it. Earlier these two things were bundled together like in the London Cabbie licence, that people would train multiple months for. Google Maps augments the navigation skills of an average person. This is where one of largest impacts of AI will be in many domains — Augementation.
This is good news because lowering the barrier to entry creates both a larger supply of workers. A larger supply of workers can create new categories. In existing domains, it reduces wages and increases demand.
In India, we have many problems in healthcare. One of them is that Doctors are not economically viable in rural areas. The cost of education for a doctor is such that the only places their career is viable is the top 30 or so urban centres where they command higher fees. The fees a rural patient can afford will not cover the training of the doctor. This opens an opportunity to create a medical assistant — this is someone with much less training but with AI + remote connectivity to city doctors when needed, we can train medical assistants for much cheaper and provide healthcare services to a lot more people. You can alternatively, think of this as upskilling or augmenting nurses into doing more of what a doctor does.
Hello data economy!
During the past decade the phrase “Data is the new oil” has been bandied about a million times. And indeed we’ve seen the magic of data powering everything from shopping recommendations to suspiciously targeted ads to our feeds.
Today there’s a quagmire of data brokers who trade information about you to improve advertising and marketing, but that world is inefficient and complex. The future will bring explicit data sale and targeting options. E.g. Netflix will have tiers like — Premium (HD, et), Basic (Broadcast ads), Free (Targeted ads). The Free with Targeted ads will only be available to those willing to link their data vaults to Netflix and also being targeting worthy.
We are likely to see people fall for the Internet trap all over again sacrificing privacy and data for convenience and/or careless adoption of new technology.
People aren’t footloose about what they post on the internet anymore. We are much more aware of the perils of carelessly shared information which means companies will have work harder or incentivize more to get people to share their data.
Since AI needs to be trained and connected to data to be relevant, it will take the data economy to a whole new level. Publishers today rely on ads next to human eyeballs reading their pages. However, those pages are perfect targets for training models on. From the New York Times to Reddit, all companies sitting on data hordes want to get paid for parting with their data. Just like with oil, we’ll likely have quality assessments, refineries and a whole long global supply chain to convert the latest babblings of humans into up to date generative models. Search engines and trending pages already do this to some extent.
Not all babblings are not created equal. My emails getting leaked might be a temporary embarrassment. The emails of a head of state getting leaked could be a national security issue. Consequently, privacy and data security will gain importance, companies in these fields will blossom.
Technology impact
Mainframe vs Terminal all over again
It’s going to be Mainframe vs Terminal all over again! Initial computers were mainframes that filled up entire rooms, this gave way to workstations that could fit in a person’s desk, which gave it to portable personal computers aka PCs , which that finally gave way to mobile phones. Today with huge models and giant training costs, we seems to be in the mainframe era. But innovation is rapid and models are getting more efficient as we speak.
Almost, in parallel to mobile we saw the emergence of cloud computing. This is likely to play out again too. Your phone will likely have small task specific models, while giant all encompassing models will run in the cloud. So it goes without saying that if the existing cloud providers don’t grok the innovation in AI, there’s room for AI focused cloud providers to emerge.
More decentralised that the web — A good thing so far is the models are crazily portable. The pool of information they draw from i.e., the datasets they train on are huge (Tbs) but they models are comparatively tiny. Llama from Facebook can be run on a laptop today. Can’t wait till we Moore’s Law this shit.
New paradigm, new concepts
When we went from local programs to using internet based services on our phones and laptops — we had to invent many new concepts for this new world. Logins, password managers, privacy settings, OTPs, etc.
If we succeed and take this AI thing all the through, we will new concepts here too:
- Data vaults: As the name says but with much holistic data about you. Imagine every click and action you take on your computer stored in data vault that is portable Or data from an always on camera that you wear on your body recording your entire environment. Data vaults will allow you to take your habits, patterns and data to any new computer.
- Data brokers: Agents enabling the transfer, access, and sale of data.
- Multi-modality: Future agents will be much more fluid — switching between programs, across image, video and text input just as humans switch context between mediums.
- Machine awareness: Today we have added a basic location sensor to our smartphones which enables so many applications. Imagine the future where machines are aware of your health, emotional state, etc. and can fine tune everything according to that. Sounds creepy but it can be something as simple as cancelling a bunch of appointments because you’ve fallen sick.
- Emotiveness: Consequently, we will not tolerate the shitty text based interfaces that we have today, we’re going to expect AI agents to be possess human-like emotiveness.
