Politically, he was a humanitarian who did know right from left and was trapped uncomfortably between the two. He was constantly defending his Communist friends to his right-wing enemies and his right-wing friends to his Communist enemies, and he was thoroughly detested by both groups, who never defended him to anyone because they thought he was a dope.
- Joseph Heller, Catch-22
Every take on GenAI is maximalist, which means every take is wrong. GenAI is neither an asteroid coming to wipe out humanity, nor a complete waste of investors' money. The truth lies somewhere in the middle, which is a take that will satisfy exactly no one.
I have for you three analogies, each flawed in their own way, that try to grapple with my personal experiences of using GenAI as a programmer writing boring corporate software.
(I will keep calling the LLM-based chatbot technology "GenAI" to differentiate it from AI in games and fiction, and because "generative AI" is a mouthful.)
GenAI is like a forklift truck
Now that my employer is actively encouraging me to use GenAI, especially Claude Code, I've found it useful to think of it as using a forklift truck.
Imagine you need to move a shipment of ramen packages from the docks to your warehouse. Luckily, you invested in this newfangled "forklift truck" technology; it's apparently a great time-saver that also reduces labor costs. But looking at the machine in front of you, you don't see an easy way to transport your ramen packages. So you carefully stack individual packages onto the forks of your truck until they threaten to topple over. But no matter how slowly you drive, some of your dried noodles always end up on the floor. "Damn," you think to yourself, "this forklift technology clearly has a ways to go. It's no match for skilled longshoremen!"
Obviously, this is a silly way to use a forklift truck, but it's a common frustration with GenAI. Developers who aren't yet used to wrapping up their requests in metaphorical crates and pallets dismiss the whole thing as "souped-up autocomplete". What doesn't help is that most people's first encounter with GenAI tend to be the worst performing models out there, with Meta and Google's offerings both lagging seriously behind the latest and greatest.
But similar to a forklift truck, you need to adapt your environment to make good use of GenAI. Your warehouse needs a flat floor, preferably poured concrete. You need wide lanes between your storage racks to give the machine enough space to turn. And separating the forklift from other foot traffic suddenly becomes a critical safety concern. There's an infamous German forklift safety video that shows, in semi-satirical fashion, everything that can go wrong in a workplace when you don't treat the machine with the respect it deserves.
So instead of asking Claude to make an entire app from scratch, try pointing it at your existing code. For example, ask it to list everywhere you use a HTML color in your CSS instead of using color variables. Not only is this list immediately beneficial to your work, but it's a tedious task that you would never do on your own. You can then use the compiled list to attack each item yourself, or you can just point the robot to it. However, not every work environment can be adapted to take advantage of a forklift truck. When you are dealing with a legacy application without clear tests, it's very hard to set strict bounds on what the robot can change. It becomes much too easy to take a wrong turn and topple over some critical infrastructure, and it's even worse when you're dealing with subtle behavior changes.
But once you've set up an environment with a robust system prompt, skills that describe how you want the chatbot to behave, and automated tests to keep it in line, you'll find that impossible tasks suddenly become possible, hard tasks easy, and easy tasks trivial. But like a forklift truck, you still need a driver behind the wheel to steer the system in the right direction.
GenAI is like a machine gun
The PM M1910 is a terrifying weapon. This Russian variant of a weapon designed in the 1880s entered Moscow's service in 1910. It has seen action in the Russian civil war, both World Wars, and basically every conflict since then, up to and including the Russian invasion of Ukraine in 2022. Military analysts call it the "ultimate sustained fire machine gun." What makes the gun so long-lasting is that even the greenest soldiers can use it effectively, it's cheap to run, and there are spare parts scattered everywhere around the globe.
Even ignoring the doomsayers who claim that GenAI will kill all humans by the end of the decade, it's still a terrifying weapon. For under $10,000 in hardware costs, you can run an uncensored open weights model that is marketed as an "offensive-security model". This free software has been verified to generate working keyloggers, ransomware, and brute-force attacks, according to the copy on its HuggingFace page. Any rogue state actor can download this right now and take down hospitals and other critical infrastructure connected to the Internet for less than the cost of a single ballistic missile. And they probably are.
The GenAI maximalists will claim that the only action we can take against this pending onslaught is to deploy even bigger models, even faster than "they" can. But you can't outrun a PM M1910 belt-fed water-cooled machine gun on an open battlefield. What you do is find higher ground, dig into your trenches, and deploy drones to attack their positions. In cybersecurity, these counter-measures are ensuring one compromised system cannot take down the whole network, keeping your systems either up-to-date on the latest patches or disconnected from the wider Internet, and logging everything you can to catch intruders after the fact. The problem, of course, is that these counter-measures are expensive, slow, and require constant upkeep. With GenAI-based attacks, attackers have an extreme advantage, and society at large will have to bear the costs.
GenAI is like a farm animal
GenAI is not a pet, but a farm animal. A shepherd cares deeply for his flock, but he will still cull a sick sheep if it would protect the herd. Too often, I see people talk about their chatbots as if they're a close relative. And that's a one-way ticket to bad outcomes. They get so invested in particular outputs that they brush away bad responses as "funny quirks". But GenAI is a tool, and broken tools should get replaced.
When a model starts producing garbage, do not hesitate to start a new session, or even change models entirely. Never argue with a chatbot; it will suck you down into the mud, and the chatbot actually likes it. Remember, for example, that Claude "skills" are just Markdown documents. Kimi can follow them just as well, and for a tenth of the cost.
Not every job requires a prize ram. In fact, a great workflow is to have an expensive model like Claude Sonnet make a plan, and then use a much cheaper model like Kimi 2.5 to execute it. You often don't even have to go that far; you can throw a cheap model at a problem until it hits a wall, and then switch to a more expensive one to punch through it. Most of the work that I do is simple plumbing based on client feedback and then adding a regression test for any changed behavior. Cheap models work perfectly fine for that work.
GenAI is a transformative tool that is quickly becoming a commodity. There is still variance in sheep: Some produce more wool, others produce higher quality meat, and some are specially bred to withstand harsher climates. But if a new sheep comes along that works the same for a tenth of the cost, a good shepherd doesn't hesitate to add them to their flock.
Conclusion
Whether it's a forklift truck, a machine gun, or a farm animal, none of these lenses justify an industry that investors claim is worth trillions of dollars. I think we'll get some cool stuff out of it, but also many terrible externalities. Like Clevinger in the book Catch-22, I find myself sputtering to my leftist friends that GenAI has some utility, assuring my right-wing friends it won't take their jobs, and being called a dope by both parties.