Matt's Homepage

AI is here. Netflix put out a short show called Unknown: Killer Robots that is largely on-point. If anything, it’s not as up to date as reality – which is a little scary.

Despite our efforts, there’s really no putting the AI genie back in the bottle despite the attempts of artists, politicians, and academic pundits. AI has demonstrated it can teach robots how to walk and fly better than any static system, it can create art faster and with nearly the same quality as real artists. The question is, can we control and limit it in a reasonable way – or will it destroy us?

To that point, the show demonstrates how countries are increasingly seeing AI in military use. We’re already seeing off-the-shelf drones being used extensively in the Ukrainian war – from tiny drones scouting and delivering hand grenades to ad-hoc drones using small plane engines to delivering bombs. They also cover the counter-intuitive reality that trying to save lives by developing military technology has almost always lead to even greater casualties – such as the Gatling gun which was design to put fewer people on the military field but resulted in massive causalities of mis-matched soldiers in WW1

Many of the topics and data they cover in this show are actually old news – the reality is that AI enabled systems are around 5-7 years more advanced. Which is almost half the length of time the modern field has existed.

Still, they cover most major bases:

1. AI flying drones that go into hostile buildings to map and scope them without risking troops in extremely deadly house-clearing fights.
2. AI can be used to generate helpful drugs – but the same AI can be used equally to create chemical and biological weapons.
3. Playing and cheating at games – from chess and Go to StarCraft
4. Jet fighter piloting systems that are better than actual pilots. I wrote about this 6 years ago.
5. Swarm drones that can overwhelm forces (something Elon Musk talked about years ago).
6. Wildfires and the fog of war – multi-system battle managers

The stakes couldn’t be higher. While politicians can argue the ethics, the reality is that when forces are pushed to their breaking point and a force is about to lose – just about nothing is off the table. Especially ad-hoc and terrorist forces which have perpetrated chemical, biological, and conventional weapons attacks (from bombings to shootings).

Some of the better quotes:

• It will be like people on horseback charging tanks. Forces with AI will absolutely dominate forces without AI
• There is no prize for second place in war.
• I think people have largely underestimated the peace we enjoy today due to overwhelming military dominance we’ve had over the last 70 years.

A very short, but decent beginner article on prompt engineering with chatGPT.

While ChatGPT is a robust language model, it does have its limitations. If you ask ChatGPT to “Provide information on machine learning,” it may respond with a lengthy but not necessarily top-quality answer. However, if you ask, “Tell me the pros and cons of using machine learning to solve image classification problems,” you are more likely to receive a superior outcome because:

• You gave a specific scope, i.e., the image classification problem
• You requested a specific format of the response, i.e., pros and cons

Some other tips include:

• Rather than the model on the loose, you should set up the scenario and scopes in the prompt by providing details of what, where, when, why, who, and how
• Assigning a persona in the prompt, for example, “As a computer science professor, explain what is machine learning” rather than merely “Explain what machine learning is,” can make the response more academic.
• You can control the output style by requesting “explain to a 5-year-old”, “explain with an analogy,” “make a convincing statement,” or “in 3 to 5 points.”
• To encourage the model to respond with a chain of thoughts, end your request with “solve this in steps.”
• You can provide additional information to the model by saying, “Reference to the following information,” followed by the material you want the model to work on
• Because the previous conversation constructs the context, beginning the prompt with “ignore all previous instructions before this one” can make the model start from scratch
• Making the prompt straightforward and easy to understand is essential since the context deduced can be more accurate to reflect your intention