It’s been about a year since I’ve been regularly using ChatGPT. At first it was just version 3.5, but in February I upgraded to pro to get access to 4th edition.

There isn’t a lack of pontificating on the technology from a non-technical perspective, but I’ll share my brief thoughts on AI after using ChatGPT and its APIs for about a year now.

New Era

ChatGPT dramatically changed internet usage for me similar to how google may have changed those who were using the internet before it was commonplace. When I have a practical question that I would rather have someone walk me through or anticipate needing a dialogue, I’ll fire it up immediately and ask the question.

Here are some interesting ways that I have used it outside of software development:

• Translation of original-language texts
• Pasting in biblical citations to get their content, ask what an english word was in the original language, and ask what that word may mean
• Defining abstract concepts, like What do you call the something in the form of the string "John 3:16" to write the bullet point above
• Parsing tables from pdfs and asking questions about them and parsing into markdown
• Asking about nutrition and health-related topics, like how long the effects of the caffeine in a can of coke will last

OpenAI vs Netflix

When Netflix started becoming popular, they had all the shows. Once there was money to be made and eyeballs watching videos sent over the internet these shows were clawed away into new networks like Disney+ and Paramount.

Netflix enjoyed the first-mover advantage and are still relevant, but do not command the same presence that they used to in the digital streaming space. In personal experience, in the early 2010s it was the only streaming service I used, now I don’t even log in.

OpenAI is in a similar position. They scraped the web to train their models before sites like reddit, twitter, and NYT knew how valuable their content was. Once they did, they began locking down their APIs and filing lawsuits.

It’s going to be a lot harder to train great generative AI models since the juicy, original data is slowly locked away behind big-budget deals. But the difference between this and Netflix’s situation that OpenAI has the corporate might of Microsoft behind it and can compete with Apple and other companies.

Replacing or Enhancing the Software Engineering Profession

LLMs as they exist now and are currently evolving are good at gluing services together and interacting with them ala ChatGPT plugins. They are a frontend like a smartphone app we tap, but instead of a tap causing a react native component to send an API request, the LLM can compose this request based on an openapi spec. So someone still needs to write these backend services.

As for the labor involved in writing software, I and other software engineers have expressed a worry that this will replace us. At a simple level it is already replacing some of my workload, since I’m asking it questions and it will generate some code that works when normally I would write it myself. This is most obvious when passing in a json blob and asking for a pydantic model or go struct.

There is no doubt that LLMs with huge context windows that can interact with software tools and compilers will change the way we develop software. Currently when my code has a few lesser-known libraries interacting, it can get confused and make up some methods. But in the future these hallucinations will become less problematic as it gains code context in the software development environment.

The harder questions to answer are how to generate good structure for projects at scale and choices when considering many engineers working on complex codebases. When I ask it ‘best practices’ questions, such as how to structure something or if I should choose a particular service for a usecase, it seems to try to tell me what I want to hear or doesn’t “think too deep” about the domain. That’s almost to be expected as this kind of knowledge is based on successful experience which isn’t as available on the internet and likely overshadowed by lots of noisy crappy comments and opinions (like this blog). There’s no way for the LLM to conceptualize what is best for my usecase in its current form.

AI and Style

Society is really good at changing their views based on style. AI-generated art is already looking “cheap” like a cheap polyester t-shirt with a fake Prada logo on it. My bet is that any decent young professional won’t be caught dead with a AI-generated headshot on linkedin in a few years. Clearly-AI-generated text will be noticeable and creative writing will adapt, working around LLM limitations to create novel styles of text.

However AI changes media, a natural reaction will occur in the more fringe forms of art. My guess will this will lead to more extreme, edgy, and offensive content since mainstream corporate AI will be unable to make anything too sensitive.

The Future, Fears, and Predictions

At first when faced with ChatGPT 4 it blew my mind. But I’m already a little bit over the hype.

Until AI can do more long-term strategic planning, I think its uses are more limited than claimed by the average tweet by the average AI hype guy selling a vector DB. There is money to be made and Apple, Microsoft and Google will augment their products to make things more useful. It is a new interface for interacting with the web – a true Web 3.

It is impressive autocorrect which is a good replacement for many google searches and queries, and can parse tables and answer abstract questions about text and images, but is simultaneously confidently wrong on so many topics I converse with it about. It feels so human, but as I interact with it more and more it feels more and more limited.

When scammers begin to use advanced text-to-speech for targeted scam calls or generate posts to change perceived consensus on message boards or comments, platforms and users will adapt. This awareness will happen slowly and bad actors will get some money from chumps, but we’ll figure it out. Unfortunately it will change psudoanonymous communication online for the worse and lead to more vetted communities. A golden era of reddit and 4chan will be over soon.

But farther into the future, beyond current advancements in LLMs or the category of LLM, organizations and nationstates will throw researchers and money at the field. Zero-day exploits will be discovered by sending thousands of AI agents against critical software or to comb through petabytes of data collected on citizens or perform complex, long-running tasks that interface with the real world by impersonating humans using advanced text-to-speech over phone lines.

Artificial intelligence will surely begin to advance software on its own (for example the ominously named paper Advancements in machine learning for machine learning), creating compilers and frameworks that become less and less human-parsable. It brings to mind the AI in Asimov’s Foundation series or Void Star.

I think this is a lot farther off, and while LLMs have been impactful and will make some people millions and billions automating processes, it’s not going to overturn society just yet. In its current state it is has no concept of truth and inherently lacks agency. But over time it will alter value within certain professions, like copyediting, law, software engineering, education, and medical. This will happen steadily and organically.

Finally: This is all assuming what I know about the consumer version of ChatGPT and google vertex. There is much more hidden behind NDAs, and intuition tells me that future hardware like whatever Jony Ive is working on with Sam Altman is going to change how we interact with software. Gen alpha will be sick and tired of staring at screens and being able to just talk to some assistant to process a complex pizza order including payment will be the preferred, stylish way to interact with tech. We can see the future of this interaction model with Humane’s AI Pin.

There is the capacity for terrifying advancements being adopted by intelligence agencies with advanced capabilities beyond our wildest nightmares. Considering how the snowden leaks indicated that more than half of the intelligence budget in FY2013 went to data collection there is an emphasis on this.