A Question

Here’s a story I like to tell, with audience involvement. Sometimes people ask me about artificial intelligence,¹ and what I think. In response, I like to ask them the following question:

What layer is SCSI in the OSI Model?

In order to answer this question, you need to know two things:

1. What is SCSI?
2. What is the OSI Model?

This is a question that I have been asking a lot of people recently. Naturally, many of these people have never heard of SCSI or the OSI Model before because of these terms’ esoteric, dated, and domain-specific nature.

The explanation of the two terms that I usually give out loud is something along the lines of:

• SCSI (or out loud “Skuhzee”) is a physical interface for connecting computers to storage or other peripherals. It looks like a USB plug from the 1980’s.²

• The OSI Model (pronounced O.S.I.) has seven layers: (1) physical, (2) data link, (3) network, (4) transport, (5) session, (6) presentation, (7) application. While a general framework, the OSI Model might best be explained by applying it to the internet. For example: (1) copper wire, radio frequency, etc; (2) rules to communicate over physical layer, (3) internet packets (IP), (4) messages made up of IP packets (TCP/UDP protocols), (5) session is a sequence of messages, (6) encryption or compression, (7) website or app data sent over network. The importance of the model is the general idea, not remembering every layer.³

What layer is SCSI in the OSI Model?

At first glance, it might appear that this question is beyond reasonable expectations for someone to correctly answer, if they’re unfamiliar with SCSI and OSI.

However, I haven’t heard a wrong answer so far. The challenge of the question isn’t the rote knowledge of the Computer Science terms. The challenge that the people whom I’ve spoken with have solved is the ability to reason about new thoughts. They are able to make connections about new information (SCSI and the OSI Model) without any previous knowledge about the relationship between those two distinct concepts.

The punchline, after the listener gets the answer correct, is that all the AI models (that I have seen so far) get the answer wrong.⁴ This serves as a starting point to discuss the capabilities of AI. What it can do in practice, what it can do in theory, and why we shouldn’t be surprised that it can make simple logic errors that appear obvious to humans who have access to the same necessary information.

The Answer

A hypothesis for why ChatGPT and Bard are unable to answer this question is because there’s no data available⁵ that states the correct answer. Both of them are able to correctly define SCSI and all seven layers of the OSI Model before they go on to give the wrong answer.

There is no answer given here. By giving an answer, we might defeat the test. Alan Turing originated the idea of a Turing Test, which posed the question of if a test could be devised to tell humans and computers apart.⁶ Many websites today use Captchas,⁷ which are a form of Turing Test. We will give an MD5 hash of the answer so that it can be revealed in the future. The hash is:

17c69c1ec0963bafd905466d6cc9ae07

Knowing what SCSI and the OSI Model are is not enough for AI to pinpoint where on the model that SCSI belongs. This suggests⁴ that on even simple questions, artificial intelligence can draw incorrect and erroneous conclusions. On the very same questions, humans are able to create logical connections in the absence of prior information and to operate in the midst of uncertainty. Is it possible that every single logical connection that an artificial intelligence makes is only possible because someone else had made that connection for it before, somewhere online or in its private training set?

Imagine if the title of this piece was “SCSI is Layer X on the OSI Model,” where X is the correct answer. Would ChatGPT and Bard get the correct answer—after consuming this post? If so, then this would suggest a fragility of AI that would be prone to the influence of a single result in the absence of any other data. The absence of data on a particular topic on the internet is sometimes called an information void or a data void.⁸

Perhaps the best way to think of generative AI is as search engines, which also have the challenge of data voids. In The New Yorker, Ted Chiang explains generative AI by using the analogy of lossy compression and raises the idea that the technology has the capacity to replace “traditional search engines” in the future.⁹

Thank you to everyone who has let me ask them this question as well as for the thoughtful feedback from TGIF.

Notes