Prompting the LLM is no substitute for knowledge work

Artificial intelligence, especially large language models such as ChatGPT are all the rage now. ChatGPT has the distinction of being the fastest growing product in the history of technology - it reached the first one million users in just a few days. As such, it is the talk of the global village now. There is no shortage of overly excited people posting their takes on how it will change the world for the better and what you should be doing right now to benefit from that. On the other hand, some are worried about ending up jobless with no place to go in the dystopian reality of 2029. Who are right? We cannot tell without knowing the future, but I don’t think any of the extreme predictions are reasonable. In many cases they are not even intellectually honest (more on that later). I am not saying that LLMs are not a big deal. Building the computational machinery to make human language tractable by software has long been a Holy Grail of research fields such as computational linguistics and natural language processing. Making that work is a great technical achievement that should not be taken lightly. However I don’t believe in black-and-white thinking. I don’t even believe in grayscale thinking. There is way too much complexity in the current social, technical, political and cultural realities to be sufficiently captured by mere notion of “goodness”.

Can the LLMs make you jobless by text-generating your job position away? Such question was posed to sci-fi author Cory Doctorow in 2020, during the Remote Chaos Experience conference. Back then, ChatGPT was not released, but GPT-3 was already available. His response was that only people who have very boring, terrible jobs should be worried about GPT-3. That answer should not have been surprising to people who actually played a bit with GPT-3. It can certainly be guided into generating text that compares to what a human could write, but if left unattended it can easily go off the rails and start generating complete garbage.

ChatGPT (a.k.a. GPT-3.5-turbo) is not exactly immune to generating nonsense either. Let us do a quick experiment with aichat - one of the OpenAI API clients.

$ aichat
Welcome to aichat 0.8.0
Type ".help" for more information.
〉.model gpt-3.5-turbo

〉.prompt What online courses could be taken to learn JavaScript deobfuscation?

There are a few online courses that could help you learn JavaScript deobfuscation:

1. "Advanced JavaScript Concepts" by Udemy: This course covers advanced JavaScript topics, including deobfuscation techniques.

2. "JavaScript Deobfuscation and Malware Analysis" by Pluralsight: This course focuses specifically on JavaScript deobfuscation and how to analyze malware that uses obfuscated JavaScript.

3. "Practical Malware Analysis" by the Malware Unicorn: This course covers malware analysis techniques, including how to deobfuscate JavaScript used in malware.

4. "JavaScript Deobfuscation for Web Developers" by Packt: This course focuses on deobfuscation techniques for JavaScript used in web development.

5. "Reverse Engineering Malware" by OpenSecurityTraining: This course covers a variety of reverse engineering techniques, including deobfuscating JavaScript used in malware.

We can see a factually wrong output here. Courses 2 and 4 don’t exist. Courses 1, 3 and 5 do exist, but they are not very relevant to JS deobfuscation. So we got an answer that is factually wrong, but written in a very confident wording. We got a problem of trust here. If we rely on ChatGPT to produce a text that is factually correct we cannot accept the possibility of some percentage of output being simply wrong. Thus it cannot be trusted to do all the work for us - to benefit from ChatGPT and similar systems, we have to work with them to achieve a desired result. That includes checking the generated text for factual correctness.

But we’re talking fairly obscure, niche stuff here. In fact, I kind of did it on purpose to demonstrate a major failure mode of LLMs. But what if ChatGPT was used to help with something more relevant to general society, such as preparing a lawsuit against an airline? One Steven Schwartz, an attorney with some three decades of experience, was using ChatGPT for (scary quotes) “research” when preparing paperwork to sue a certain airline over an injury allegedly sustained by a customer. That went horribly, scandalously wrong. Like in our quick experiment above, ChatGPT hallucinated references to prior legal cases and court decisions that simply did not happen. Schwartz himself got in trouble for submitting legal paperwork with false data and the story made it’s rounds across the press outlets. You know you really fuck up your work when your oopsie makes it’s way to New York Times.

GPT-4 is more advanced than ChatGPT, but is susceptible to hallucinating stuff as well (just ask the same question about deobfuscation courses). But what if there was a sufficiently advanced LLM that did not hallucinate and only generated text that was factually correct (unless specifically instructed to write fiction)? Realistically that is not even possible, as it relies on perfectly correct information being sourced for training of such model. However, for the sake of argument, let us assume that there is such a system running. If we have a perfect knowledge machine, can we use it to replace a human work? Well, the thing is, we still need to tell it what to do. Even a perfect LLM is GIGO: garbage in, garbage out. When it comes to non-trivial things (such as developing large software systems), describing them in sufficient detail in a way that captures all nuances and edge cases is quite difficult type of work. This can be seen in real world software projects, especially when they involve non-technical stakeholders. It’s naive for manager or entrepreneur to assume that soon enough they will be able to get rid of those pesky developers, as they will just tell a chatbot what code it should write. Quite likely the activity of getting an LLM to do things (generate or edit blog post, develop code, write marketing copy, etc.) will be new kind of work - prompt design/engineering.

Sure, you can use ChatGPT or GPT-4 to assist in software development work - generate code snippets, do some refactoring, get advice on improvements and so on. But at the end of the day LLM is merely a tool AI. It does not have it’s own judgement, values or will that human possesses. Even worse, it is debatable to what extent it knows things, as the fundamental way it operates is predicting what text should be generated next. It is the user that drives the LLM who is supposed to know what they are doing. For example, ChatGPT may generate a code snippet that is mostly correct in the light of provided requirements. It’s up to the user to make it completely correct. Now, what if the user was a non-technical entrepreneur? How well would it work out for them?

By amplifying, not replacing, human effort LLMs can be said to provide a leverage. As the leverage increases, the value of sound judgement increases as well. When it’s easier to turn your decision into a reality, the quality of these decisions starts to matter a whole lot more.

But the AI hype wave is in full swing now. There’s all kinds of social media influencers and content creators telling you to do this or that to benefit from the AI revolution to get rich. A lot of these people are digital snake oil merchants who are jumping from trend to trend. Some propose completely silly, unrealistic money making schemes just to get your attention. It’s all part of make-money-online content creation playbook. The truth is, these people mostly don’t make money from the thing they promote. Instead, they make money by promoting the thing they are promoting through ad revenue (esp. on Youtube), affiliate offers and their own info products. During the Bitcoin bull run and NFT craze there was no shortage of characters talking loudly about how you have to “invest” into this or that thingy on the blockchain. I suspect the vast majority of them could not tell the difference between Base64 and AES. Now you have great deal of people talking the same way about AI (mostly LLMs). How many of these newly minted “AI experts” even know what the fuck a linear regression is?

That is not to say that nobody ever gets rich during a gold rush. Some people swiftly grab a transient opportunity and milk it for what it’s worth. Others do some equivalent of selling shovels and pick axes during the Wild West period. Some simply gamble and happen to be winning. Speaking about LLM-based SaaS apps specifically, there’s a certain problem. If you have some SaaS app that has OpenAI API as a hard, critical dependency that means your SaaS business is essentially at the mercy of OpenAI. Right now, there are some important questions that pertain to generative AI. What exactly should and should not it be allowed to generate? If some data is scraped to train the LLM what happens to related copyrights? What will be done against gray hat and black hat usage of LLMs, such as fake product review generation, harassment, phishing, influence operations? Depending on how things play out, your entire SaaS business could die overnight by being cut off from API access as it may turn out not be kosher anymore in the light of new legislation or newly introduced corporate policy. Remember that Twitter API was free and open in the early years. There were many companion apps for Twitter users built on that API - unofficial clients, automation tools and the like. But over time Twitter introduced many restrictions into the official API and most of them had to die. This is something that indie hackers who rely on the goodwill and interest alignment from OpenAI (or Anthropic, or Google…) should keep in mind.

Furthermore, using an LLM that you don’t self-host can be problematic from data security and privacy perspective, as you are literally telling some other entity the specifics of your work. There was already a scandal of Samsung employees leaking trade secrets to OpenAI by using ChatGPT to help with their regularly scheduled programming. That is the reason why the usage of ChatGPT is explicitly banned in some corporations. A fundamental problem with information society is that many things are data driven, but we mostly don’t have any control how data about people and companies is processed, stored, transferred. Information functions by being always in motion and largely has a life of it’s own. If you tell your trade secrets to ChatGPT, can you trust ChatGPT not to tell them to your competitor?

Times are difficult now. We’re only 3.5 years into 2020s and so far we had a global pandemic, energy crisis, murderous Taliban psychos winning the Afghan war, a new major conflict in Ukraine, inflation wave causing cost of living crisis, tech downturn and banks failing. But even now I advice everyone to keep their heads cool. You probably cannot prompt your way into quick wealth, especially if you’re about average on coming up with LLM prompts. In economics there’s a concept of alpha. It’s a number quantifying how well an asset outperforms the overall market. Dumb luck aside, to achieve exceptional results that greatly outperform most of the competition one must posses or do something exceptional. The problem with doing easy things (e.g. generating basic ad copy) is that they don’t stay easy for very long in a competitive environment. It is unlikely that ChatGPT, GPT-4 and other AI systems will make everyone jobless due to considerations I have discussed. These systems are merely tools subservient to a human will and there will be a need for people working with them. Perhaps the daily reality of many occupations will change, perhaps some people will need to learn how to do something else, but I don’t expect the fully AI-driven world to happen anytime soon.

We cannot know what the future will be like. In a sense, we don’t even have a (predictable) future when things are so crazy, volatile and fucked up. All we can do is hedge our bets and roll the dice.

Keep calm and seek your alpha.

Trickster Dev

Code level discussion of web scraping, gray hat automation, growth hacking and bounty hunting

By rl1987, 2023-05-31