Sorry, I write really long things here :)
My goal is to help inform future conversations. I don’t claim to be the definitive source of knowledge, rather, I want help inform and hopefully assist others in understanding better so that our knowledge as a group can lead to better outcomes.
So, to start, can we pretty please stop calling it AI. Right now, AI is nothing more than a marketing term driven by business people who know nothing about what is actually being created here and just are hoping that calling it AI will generate more investment. Its “ML” — Machine Learning — Nothing more, nothing less. It has no “Intelligence” and likely won’t for a very very very long time if ever.
/rant complete
But onto the real point of this diary…
AI / ML is here, it is changing the world, but its not going to take it over and destroy everything we know and love. Also, it needs regulation.
CONTEXT
I have worked on Machine Learning technology since roughly May of 2010 when I started working on what would become Bing at Microsoft. At the time, real world use cases of ML were few and far between. It was known to be obscenely expensive to do correctly, neural nets were the goto solution, and all the successful potential use cases for ML were better and more cost effectively solved using complex algorithms like Google Search and Microsoft Handwriting Recognition were utilizing.
I now work in software consulting and have helped lead a few recent AI / ML research projects with a variety of companies.
IN THE BEGINNING
Modern machine learning technology really started to hit mass market with a little remembered product known as Kinect for Xbox360. For all intents and purposes, Kinect was the first mass market product utilizing machine learning and it showed. It showed both the promise and the failure points of ML. On one hand, it was mind blowing how well Kinect could track our bodies. On the other hand, it was mind blowing how much it got wrong from tracking pets to having extremely high latency as a control device.
Regardless, Kinect woke up the word of computer engineering to the potential of ML. Internally at Microsoft, everyone was starting to ask “what can we use ML to do”. Out of this, many projects emerged including Bing itself. Microsoft had always wanted to compete with Google in search, but the issue was one of time and resourcing. It would take a decade to hire up 5,000+ engineers with the skillsets needed and likely at least a decade to make a competitive algorithmic search engine. With ML, in theory, they could build a competitive search engine in maybe 2-3 years with just 500 or so engineers. No guarantees, but why not try, and they did.
It was so successful, roughly 2 years after Bing launched, we could tell the moment Google themselves switched to using an ML-driven search engine to remain competitive.
THE SILENT DECADE OF INNOVATION
As with a lot of technology, things take off, then they crash back to reality, then they slowly build back to be something actually useful for consumers.
ML did just that. For the better part of a decade from 2013 to 2023, nothing really happened in public view for ML. Behind the scenes things like GPT and MidJourney were continuing to innovate. Few outside of the ML community were paying attention as GPT developed slowedly, but it was moving forward. Then suddenly, right around GPT3, people started to notice, though still mostly in the tech community. Shortly after GPT3.5 launched, a bunch of non-technical people got involved in trying it out and laughing about the results, but now it became real, and alongside it becoming real, the panic set it.
DESTROYING THE WORLD, OR NOT
For better or worse (in my view much worse recently), anyone and everyone in the modern day can say or publish anything on a topic. This leads to most new announcements or innovations getting hit with hundreds of “news” articles and YouTube videos that misrepresent or otherwise abuse issues with the technology to create fear around it or to mock it. Afterall, that generates clicks and views.
Around the time Microsoft launched the “new Bing Chat”, the “AI Wars” started. If you went on YouTube and searched to learn more, video after video fit one of a limited set of topics.
- AI is going to destroy the world
- AI sucks and is useless, but will still destroy the world
- AI caused someone to commit suicide, let's all be scared and shut it all down
Now, nothing. Almost no one on YouTube or in the news in general is talking about AI / ML at all anymore. Why? Because nothing anyone predicted or fearmongered around happened or likely will happen.
What has AI / ML actually done this year?
- A new more conversational way of getting search results
- Gradually improving image generation technology that almost no one is willing to use in a professional setting due to copyright concerns (which is a good thing)
- A bunch of research projects that are testing waters of the future of the space.
- A ton of people flooding Amazon Kindle with books no one wants to read.
The reality is, we just had the next huge peak of interest in the space like what happened after the Kinect launched. Then almost immediately, the hype died down and people in the industry got back to real work.
It will likely be at least 3 to 5 years before most of the world cares again.
WHAT’S NEXT FROM A TECHNICAL POINT OF VIEW?
So what is next? For starters, not an AI revolution, once again, there is no intelligence here. This isn’t the next industrial revolution and this technology won’t be what causes the next work revolution. It just another tool in our collective toolbelts. Increasingly people will do things like the following...
- Artists will use tools like MidJourney to help them prototype concepts in their heads before making something by hand
- Workers who can’t code will start to use tools like CoPilot to automate or accelerate getting visualizations of data without having to wait on someone else
- I will make better backgrounds for my YouTube thumbnails (or thumbnails at all)
- Rather than having to tweak your query 10 times on Google, you will have a conversation with the search engine and it will contextually understand better what you are trying to get to and from where
Looking out about 5 years, I see a few other innovations happening as some innovators and visionaries work to create the next generation of products driven by a more conversational interface
- Sites like SquareSpace and Wix will allow you to write out in natural language what you want out of a site and it will build it for you, including Database design in some cases
- Someone will make a basic game engine that lets you have a conversation with a system and it will code a simple platforming game, maybe even one with some combat elements
- The first “apps” will ship into iOS and Android marketplaces that are 100% coded at the direction of a human talking with one of these systems instead of hiring a consultancy like the one I work for. AKA — Build something for $500 instead of $500,000. Democratization of applications.
BUT YOU SAID REGULATION ABOVE, WHAT DO YOU WANT TO SEE?
Lots honestly, and some of it I can’t even think of right now. That being said, my ideas around regulation in the space is trying to fit less with protecting people as workers and more with protecting people from other people exploiting these tools to prevent them from working. Its kinda a weird line, but hear me out.
- Any work that is produced over 50% by an AI should not be copyrightable
I believe firmly that these tools will be widely utilized and should be going forward. In my view, this is no different than when people stopped using giant walls for spreadsheets. Sure, it changed the dynamics of the job market, but it also fundamentally improved productivity and resulted in many great advances.
That being said, if you want to copyright a work, you truly need a human involved at the majority level. This is going to be hard to define and enforce, but it will be important to do so.
The path to success I see here is requirements around watermarking concepts already making their ways around the art community. Some kind of fingerprinting or watermarking, but requiring it in all generated art and text works.
- Any work that leveraged AI during development, even if no AI created content remains in the final product, needs to credit the AI system(s) utilized
This is similar to how the laws require that any sponsored content. People who produce content online, that is sponsored, must declare the sponsorship. I want to see any content created leveraging these tools be required to declare which tools were involved in the generation of their content.
As an example, if I use MidJourney to help create my thumbnails for my videos, I should have to declare that somewhere visible and clearly identifiable in the image itself. If a music video uses any AI generated content, they need to clearly attribute that in the credits of the video. If a politician uses one of these tools to generate any audio for their propaganda, they have to properly identify such.
- No for-profit media can be created leveraging AI to replace a persons likeness, image, voice, etc… without the express permission of the person or the estate
This is around the idea that Hollywood studios or even independent content creators could easily steal and abuse the likeness or voices of well known public figures. As an example, I could imagine a Republican politician using ML tools to mimick Ronald Reagan’s voice and likeness and make it look and sound like he supports policies they are advocating for. Likewise a Hollywood Studio could try and produce a new Indiana Jones movie after Harrison Ford passes using entirely 3D generated images and AI generated voice.
If no estate exists for someone to grant permission, then we should default to “you cannot do it”. No one should have to worry about being misrepresented after they are gone.
- No government agency can utilize any non-government AI tech to do any analysis or creation of any work
This is primarily a national security issue. One thing I know all too well is how much data is collected to make ML systems work. In ethical hands, the data is properly collected, hashed, anonymized, etc… Unfortunately, Bing is the only team working with big data I have ever witnessed being ethical in this regard (I don’t even know if they are still ethical today with this new technology).
Going forward, there is too much risk that government employees could accidentally add classified or sensitive information to these services that may expose them to adversaries. As an example, I could envision a situation where a future President’s staff needs to take a bunch of documents regarding a security concern in a foreign country and turn it into a summarized document with charts and graphs for a President. While they all have appropriate clearances, Microsoft CoPilot or other services do not. If they use these tools to create the summaries and graphs, they could be exposing the information to 3rd parties that put national security or confidentiality at risk.
SO WHAT WOULD AI BE SINCE THIS IS ML?
Well, intelligent. It doesn’t just learn and spew back. It actually thinks and creates in a vacuum.
When I think of true AI, I don’t even think of what people call “General AI”. That would be a system that simulates everything a human can do.
First, we will get specialized AIs. AIs that can excel at one simple defined task or area of expertise. Imagine an AI that was specialized at housing. Said AI may be able to do the following.
- Look at an available lot
- Research and ingest all local, state, and federal laws regarding land use
- Determine what tests it needs to do in order to identify use options for the lot (soil, zoning, etc..)
- Research market needs for the area
- Create a bunch of sample houses, duplexes, etc… that could be built on the lot
- Create marketing/promotions to attract home buyers that may be interested in a house in the area.
- Negotiate with the purchasers and adjust options based on the discussions
- Identify contractors in the area that could be hired to build the project
- Negotiate rates with the contractor
- Make tradeoffs that don’t align with the original contract during construction that are mutually beneficial to the contractor, the future homeowner, and itself
- Do some form of relationship building with the goal of making the contractor or land owners want to work with them again in the future
Basically, an AI would be able to do more than just hard numbers, math based transactions. It would be able to independently work to mutual benefits on ambiguous objectives, make tradeoffs, some of which are not in its interest outside of long term relationships and other potential benefits.