AI, plainly

08 Lesson 8 of 15

What it can and cannot see right now

Whether it can browse, read your file, or know today's date depends on which tool you use and what is switched on.

  • About 2 min read
  • Interactive 3D diagram
The version most people start with

It is connected to the internet. It clearly knows a huge amount, so it must be online. And if it is online, it can check things, look up today's news, and open a link you give it.

The idea to leave with

It will not tell you it is offline, because it does not experience being offline. It will simply answer.

See it in the lesson

Sometimes. Fetching things from outside is a separate feature added around the model, and it is not always there or always on. Without it, the model works only from what it learned in training, which stopped at some date in the past. It will not tell you it is offline, because it does not experience being offline. It will simply answer.

Four questions worth asking

  1. Can it browse the web? Ask for something that changes daily and cannot be guessed, like a specific price or today's weather somewhere. If it gives a believable answer it could not know, it is making it up, not fetching it.
  2. Does it know today's date? Ask. Some tools are told the date and some are not. A model with no clock may confidently put you a year or two in the past.
  3. Can it open a link I paste? Try one. A tool that cannot browse will sometimes give a confident summary of a page it has never seen, made up from the words in the web address.
  4. Can it read the file I uploaded? Ask it to quote an exact sentence from a specific part of the document. A tool that really has the file can do this. One that is guessing from the file name cannot.

That last one is the most useful test on this whole site, because it works everywhere. Asking for an exact quote from material it should have in front of it is a quick check that it really has the material. It takes one line of typing.

A note on cut-off dates. Every model's training data stops at some point. Near that point, its knowledge gets patchy instead of stopping cleanly, so “it knows nothing after March” is too neat. Treat anything recent as unreliable, whatever the stated date.

The diagram lays this out along a timeline. The stacked blocks are what it learned in training, thinning out towards the cut-off instead of stopping cleanly. It cannot see anything past the wall unless the tool gives it a way across. Switch those ways on and off.

By default, nothing past the wall reaches the model. Each switch is a separate feature added by the tool's makers. Which ones exist, and whether they are on, varies from tool to tool.
Read this diagram as text

A timeline runs from the past to today. What the model learned in training is dense in the past and thins out towards the cut-off, some time before today. Three switches each open one route past the cut-off. Web browsing reaches pages up to today. An uploaded file puts a document into the conversation. Being told the date puts today's date into the conversation. With all three off, it works only from training and will not say so. The line under the switches shows which routes are open.

The move to make

Ask the tool you use what today's date is, and whether it can open a web page. Two questions, fifteen seconds, and you will know more about your tool's limits than most of the people using it.

Worth flagging: Like the last lesson, this is based on how the tools are built, not on published studies of what users believe. It is here because getting it wrong can cost you a lot.