Human in the Loop, but Where?

 I recently had some imaging done. The good news is that everything is fine. Thanks for your concern. The experience made me think about a process being touted to make AI more palatable to the masses called human in the loop.

The AI Model is Fundamentally Different

See, one of the big issues with AI is that it makes mistakes, but we assume it doesn't. That's a terrible assumption, though understandable. Until very recently, the dominant, if not only, model of computing was deterministic. Programs follow specific paths as designed by a human. If the program is correct, the program doesn't make mistakes. Mistakes are human. An error in software output would be the result of a human programming something incorrectly, a human putting in the data incorrectly, or a human interpreting the results wrong. They could only do what it was told to do with the data given to it, but it always did it the same. Enter the same data, you always get the same result.

Not so with AI. AI is probabilistic meaning it is based on models governed by probabilities. You can ask the same question with the same data and get different results. Using AI means the computer can make mistakes. In the terminology of probability, you can have false positives and false negatives. A problem emerges from this social-computing system - we assume that the machine is always correct even when it isn't. Think of it as a presumption of correctness driven by our previous interactions with computers. Simply put, we are used to computers delivering correct results all the time, but AI doesn't work that way. Applying the previous presumption of correctness to AI results in many bad decisions.

Human in the Loop AI

Human in the Loop helps to alleviate some of this problem by having a human being, presumably an expert, supervise the AI and make the actual decision based on recommendations from the machine. There is a problem, however. Human in the Loop will work differently depending on when humans are looped in. This is where the imaging idea came to mind.

AI and Radiology

Say you need some imaging - a simple Xray, a CAT scan (which is a complex Xray), or MRI. A technician takes the images but does not interpret them. Interpretation is the job of a highly trained and experienced doctor called a radiologist. In current practice, the radiologist, using all their learning and clinical experience, looks at the image and decides if something is abnormal and categorizes what that might be abnormal about it in highly clinical language. This language is understandable to other doctors. Radiologists are. like all humans, prone to human mistakes. They might miss something or overinterpret something, yielding false negatives and false positive. Those mistakes, however, are rendered less common by more experience and training. Their results are also checked by other specialists tasked with delivering care. They rely on the radiologist's report but also their own clinical judgement.

Let's assume that the radiologists decide to use AI to augment their work, perhaps even to reduce their own chance of mistake. They first make their determination then ask the AI to point out what it sees. If the AI sees something that radiologist didn't, then the doctor takes a look at it to determine if it's a false positive. If it doesn't see what they saw, they assume it's a false negative. In this scenario, clinical judgement comes before machine judgment. Too many mistakes on the part of the AI and the radiologist stops using it because they see it is a distraction not a help. The mistakes against clinical judgement destroy the presumption of correctness and the doctor no longer relies on the now flawed machine.

Now, let's switch the position of the doctor and the machine in the process such that the machine looks first. The AI points out anomalies for the doctor to check. This is the point where confirmation bias takes effect.  Having been presented with something to look at, something that seems right, the doctor stops looking and moves and misses something else. If the AI shows nothing at all but it's plain to the doctor something is there, the radiologist specialist will ignore the AI. If, on the other hand, the AI shows something plausible - not right, not complete but plausible - might the doctor feel like enough has been done, have the AI right the report and move on? Yes, certainly. If the machine which has a presumption of correctness, tells the doctor is found something, the doctor may well not look seriously for other problems. 

Who Comes First Matters

In the first scenario, clinical judgement comes first with the AI only acting to potentially catch mistakes. The second scenario has the AI coming first with the expert trying to catch its mistakes. Human psychology being what it is, with the added force of the presumption of correctness, the potential for false negatives goes up in the second scenario compared to the first scenario or the scenario without AI. 

This will be true in business processes as well. Humans conditioned to believe that the transactions they observe in software are generally correct, aside from human error, will set aside their own judgment for flawed machine judgement. The machine should be a check on human mistakes not the other way around. Otherwise, expect mistakes to balloon. 

So, Human in the Loop AI sounds great but where humans are supervising and, especially making decisions, matters if the goal is to reduce errors not increase them.

Looking into the Digital Graveyard

It's amazing how much cloud software is heading to the graveyard. Yahoo Notepad, which have been around since the 2000s is being retired. I have notes in Notepad that predate 2010 so it's at least that early. For context that's at least four years before Microsoft OneNote and three before Google Keep. I think it could be older but I'm having a hard time finding a record of Notepad's launch date.

Salesforce's Quip is also on the chopping block. I covered Quip in the 2010s, long before it was part of Salesforce. I liked it a lot and even used it for a real project. It could have become the next Notion, but Salesforce didn't really know what to do with it. They were just into "collaboration" at the time and grabbed something.  This was at roughly the same time that Zoho was expanding, and Google, Dropbox, and Box were adding document-based collaboration. I guess they thought they could get in on that too. It made just about as much sense then as now.

Microsoft Whiteboard just recently died. It's hard to believe that it launched 12 years ago. The Google version, Jamboard, bit the dust a few years back. The whole digital whiteboard thing seems to be done like dinner. At the time Whiteboard was launched, it was tied to the giant screen conference room table project, Surface Hub, that Microsoft was pushing back then. Those were happier times before the pandemic showed us that no one really wants to collaborate in person, especially around a giant embedded screen in a table. 

Frankly, I surprised as many of these have survived as long as they have. The 2010s were the era of collaboration, driven by the notion that if you give people collaboration tools, productivity will be higher.  This was also the age of the intranet, intercompany information portals that facilitated information sharing. Very few of them are still around, mostly replaced by other software. Chat apps first appeared at this time too, though many of them have not survived. Slack mostly won and Teams killed off any of the laggards. There are still some left, but they don't have the market share of Teams or Slack. The same is true of video conferencing which is now dominated by Zoom and Teams.

What the business school types didn't get was that there is a time and cognitive overhead to using many of these products. Checking your email, intranet, chat, and other apps for messages and acting on them absorbs both time and brain share while diminishing attention to any one task. As usual, the market also had just too many big folks chasing the same thing. Why use Quip in Salesforce when I have Microsoft Office 365 or Google Workspace to manage documents? You even have the choice of Zoho, Dropbox, or Box if you happen to hate Google or Microsoft. 

I admit that, at the time, I drank the Kool-Aid a bit. I saw a lot of promise in these new collaboration tools. There was promise, but it wasn't clear which would prove to be the way people want to work together. It got overhyped for sure, but there was something to the idea of being able to work on documents and communicate better. Once again, the pandemic taught us that chatting and sharing documents were a good thing precisely because they could operate offline. We just didn't need twenty choices and "live" collaboration really means desktop video conferencing. 

The next market to experience the shakeout will likely be in the space that Notion, Microsoft Loop, and Jira (kinda) inhabit. These are glorified Wikis whose whole purpose is to bring some organization to otherwise free form information. I'm betting that Microsoft Loop won't survive even though I like it. Jira is heavily tied to the Atlassian software development project environment and Notion is good for general purpose project information sharing. Loop has very good functionally and has the best integration with the Office ecosystem. It just doesn't seem to be making a splash. Much of what it already does is available in OneNote and Teams. It's a Notion copycat and Notion is just a fancy Wiki. As Microsoft moves more of its R&D focus to AI, I don't see it spending money on Loop.

This type of category death is common. I found a blog I wrote back in the 2000s. It talked about not trusting SaaS services because they can just go away suddenly. This is still the same in the 2020s only with more services to kill. 

Keep in mind, this type of product death is different than the one happening in AI right now. That's an example of too many expensive products flooding the marketing all at once. That kind of shake out always happens before the market solidifies. What's happening in the collaboration market is more about killing off the stragglers in a mostly successful market. 

Call it the Hyena Principle. This is the final thinning of the herd. Yahoo Notepad was at the beginning of the migration and products like Loop are towards the end. Now that the migration is almost over, there still a few products for the hyenas to eat before they turn on something else.

Great Test for Project Nomad

Lately, I have been implementing my own set of computing services. Some of you might recall my NextCloud project to implement my own set of cloud services. I did that primarily to reduce my reliance on the cloud services of a certain company in Redmon, WA. It's not that I think that are evil or, at least, no more evil than any other tech company. It's just that all of these cloud service companies have a history of dropping and changing services that I've come to rely on. It was also good geeky fun.

In that vein, I've just completed a new project to implement information services as self-hosted services. The idea behind self-hosted information services is to provide access to important information when the Internet is unavailable. I could potentially turn off my Internet service or go out into some place off-the-grid and still have essential information. 

Now there are three classes of people who do this:

1. Techie geeks me who do it because we can
2. Folks who live somewhere that internet service is unreliable or unavailable
3. Preppers

Many of these systems deliver internet-based information to schools in places that have no or poor connectivity. The Preppers are something else altogether. These are folks who want to be prepared (prepped) for disaster. There is a bit of overlap with survivalists but, as far as I can tell, preppers are more about readiness then wanting to hide from humanity or the government. For the record, I'm in the first category.

There were a number of options for software, but I chose to use Project Nomad. Project Nomad is an open source control plane for implementing and managing other services such as Kiwix. It runs on Linux (because Docker) and I was able to use an old laptop that I had already loaded Ubuntu onto. Project Nomad is one of the easiest bits of server software I've ever installed. It sets up Docker containers, installing Docker, if necessary, asks what you want to install, and then just does it. Once the core software is in place, you can choose what content to download which happen in the background.

There are five major components, and you can choose any or all of them to install. They are:

1. An information library, including Wikipedia. powered by Kiwix
2. Maps from OpenStreetMaps
3. A drug database
4. A learning platform with content from Khan Academy
5. An AI LLM

I went with the information library, maps, and the drug database. The AI component was not going to run well on the hardware I had, and I really didn't care if I had AI support. The learning platform was just not something I'd ever really use. It would be different if I was trying to run a school in a caven. Any service can be removed or added later as well as a few other services such as Jellyfin media server and Vaultwarden password manager. 

The next step is telling Project Nomad what content you wanted to download. I choose Wikipedia, Wiktionary, some Stack Exchange forums on technical subjects, a bunch of categories of books from Project Gutenberg, a collection of medical information including military medical books, the iFixIt website, and a bunch of prepper material. That last bit of information was mainly out of curiosity. There was an entire set of videos on how to put together a bug out bag. It's not what I normally think about.

A lot of the information is useful to have around, internet or not. The real test came when I experienced a substantial ISP outage. Project Nomad meant that I had a lot of useful information still available and I wasn't floundering. It worked perfectly with no glitches or hidden connections to internet sites. The only thing Project Nomad couldn't do was pull updates for the software and content. That's to be expected. 

That need to pull information down, at least initially, is the Achilles Heel of Project Nomad. You need a stable internet connection to install the software and content. If you have your Project Nomad server already set up before the zombie apocalypse, then you're fine. If you didn't think about it until the manure hit the rotating blades then you're out of luck.

At the moment, Project Nomad is mild assurance against sporadic internet outages. It could also be setup by someone and shipped to someone else off-the-grid. It's easy to install and pretty low on resource usage, if you skip the AI. That's fine because when the robot apocalypse happens, I'll be trying to kill AI not run it off soar energy in my cave.

1.