It would be ludicrous to assert that YouTube video viewing compresses the learning time, allowing one to become an Olympic coach in a sport without ever having played it.
Yet, it the article claims that humans can merely define intent and supervise the AI’s output - the same humans that used the AI to avoid the tedium and time commitment required to obtain mastery from actual work.
Perhaps I’m sorely mistaken. All those coworkers who went up the management chain after 2-3 years experience are the foremost technical experts in their field. They merely press a single button like George Jetson and the AI will figure out all the details!
This article is about current level AI, not future AGI. The common definition of AGI is intelligence that matches a human's, so there would be no need for a human to supervise every action of the AGI any more than a CEO supervises every action of their employees since by definition the human supervisor would do no better than an AGI performing the same role.
> we thank ChatGPT, Claude, Gemini, and Grok for tirelessly traversing the combinatorial space of this manuscript.
This is a premise so fanciful this might as well start with "assume a goose that lays golden eggs" that's then been extrapolated into 100+ pages of AI generated text.
Lead author is a crypto guy. Having written any paper on crypto, especially as an economist, is negatively correlated with my interest in reading your arguments.
> For companies, the core strategic insight is that verification is no longer a mere compliance function, but a primary production technology—and increasingly, their most defensible one.
Sure, but the real wins are when the “intent-execution-verification sandwich” is done to the verification layer.
Have the agent produce its own verifiers according to a human supplied spec, then let the agent install it in its own workflow so it can later use it to check its own work.
Software development is a highly repeatable process as long as the agent is building - and verifying! - its own verification suite as it goes. By producing its own tooling to let itself define and codify “working software” and to later use that ever-growing verification suite to ask and get answers to that question, no human is needed for working software to be (eventually, after potentially lots of self-iteration) produced.
Test-driven development works for this.
It’s massive leverage - humans define high-level what the verification suite should contain (ie. plan mode), and the suite is built by the agent then used by that same agent to check its own work. Work that fails the suite’s checks can be fixed by the agent and re-checked as many times as needed before it is allowed to declare itself done. It’s just linting on epic steroids.
For non software work it is possible to get a similar result if hypothetically you can build a human-curated library of examples of “good work” and “bad work”, then have the agent build a verification suite using those.
For example, let’s say you want a blogging agent that never gives away its agentness.
You could prompt for a blog post and get something you can bang into shape by yelling at your agent over and over, or you could automate the part where you yell at the agent over and over!
Here’s how:
Gather a set of shitty AI prose that you hate seeing, and have an agent write a script that receives prose as input and will error out and reject if said prose containing those AI-isms. The prose-checker script needs to spit out the problem line numbers & offsets. That’s the most important thing as it will help any agent using this script - if it has a line number or an offset, it knows where to make targeted edits to improve its first draft.
When you provide that script to your blog post generating agent and tell it to refuse to stop and continue making improvements to its work until the prose-checker script does NOT error out, it can iterate as many times as it needs until the checker is satisfied. Only then is it going to present you its work. Voila! No more AI-isms! (At least, not the ones your checker disallowed)
You can do the same in a positive direction too - for example you could automate making prose sound more like you wrote it yourself.
Have your agent build you a script that invokes an LLM which, given samples of your writing style, can “grade” the prose based on closeness to yours. The grader needs to give feedback to the invoking agent to help it edge its output closer to the goal.
It’s powerful to think of AI as a tool-building tool that can use tools!
Both scripts are like 10-20 lines of code; the first wraps a regex and the second wraps `claude -p`. Or they could be one script. The point is, you’re no longer going to tell the original agent what to do - you’re giving it a prompt and a suite of validations - tools that let it check if it’s reached the goal you have for it, and which if it hasn’t, give guidance on how to inch closer.
Your job as a software engineer DID get taken - by you, the software verification specification engineer! And the agents DO write over 90% of the code. Because why would you code by typing when you can have so much more leverage using coding agents to codify your goal.
Articles like this feel overly optimistic.
It would be ludicrous to assert that YouTube video viewing compresses the learning time, allowing one to become an Olympic coach in a sport without ever having played it.
Yet, it the article claims that humans can merely define intent and supervise the AI’s output - the same humans that used the AI to avoid the tedium and time commitment required to obtain mastery from actual work.
Perhaps I’m sorely mistaken. All those coworkers who went up the management chain after 2-3 years experience are the foremost technical experts in their field. They merely press a single button like George Jetson and the AI will figure out all the details!
This article is about current level AI, not future AGI. The common definition of AGI is intelligence that matches a human's, so there would be no need for a human to supervise every action of the AGI any more than a CEO supervises every action of their employees since by definition the human supervisor would do no better than an AGI performing the same role.
18 emdashes in just the first 5 paragraphs is crazy work. Might as well just ask claude myself.
> we thank ChatGPT, Claude, Gemini, and Grok for tirelessly traversing the combinatorial space of this manuscript.
This is a premise so fanciful this might as well start with "assume a goose that lays golden eggs" that's then been extrapolated into 100+ pages of AI generated text.
Lead author is a crypto guy. Having written any paper on crypto, especially as an economist, is negatively correlated with my interest in reading your arguments.
> For companies, the core strategic insight is that verification is no longer a mere compliance function, but a primary production technology—and increasingly, their most defensible one.
Sure, but the real wins are when the “intent-execution-verification sandwich” is done to the verification layer.
Have the agent produce its own verifiers according to a human supplied spec, then let the agent install it in its own workflow so it can later use it to check its own work.
Software development is a highly repeatable process as long as the agent is building - and verifying! - its own verification suite as it goes. By producing its own tooling to let itself define and codify “working software” and to later use that ever-growing verification suite to ask and get answers to that question, no human is needed for working software to be (eventually, after potentially lots of self-iteration) produced.
Test-driven development works for this.
It’s massive leverage - humans define high-level what the verification suite should contain (ie. plan mode), and the suite is built by the agent then used by that same agent to check its own work. Work that fails the suite’s checks can be fixed by the agent and re-checked as many times as needed before it is allowed to declare itself done. It’s just linting on epic steroids.
For non software work it is possible to get a similar result if hypothetically you can build a human-curated library of examples of “good work” and “bad work”, then have the agent build a verification suite using those.
For example, let’s say you want a blogging agent that never gives away its agentness.
You could prompt for a blog post and get something you can bang into shape by yelling at your agent over and over, or you could automate the part where you yell at the agent over and over!
Here’s how:
Gather a set of shitty AI prose that you hate seeing, and have an agent write a script that receives prose as input and will error out and reject if said prose containing those AI-isms. The prose-checker script needs to spit out the problem line numbers & offsets. That’s the most important thing as it will help any agent using this script - if it has a line number or an offset, it knows where to make targeted edits to improve its first draft.
When you provide that script to your blog post generating agent and tell it to refuse to stop and continue making improvements to its work until the prose-checker script does NOT error out, it can iterate as many times as it needs until the checker is satisfied. Only then is it going to present you its work. Voila! No more AI-isms! (At least, not the ones your checker disallowed)
You can do the same in a positive direction too - for example you could automate making prose sound more like you wrote it yourself.
Have your agent build you a script that invokes an LLM which, given samples of your writing style, can “grade” the prose based on closeness to yours. The grader needs to give feedback to the invoking agent to help it edge its output closer to the goal.
It’s powerful to think of AI as a tool-building tool that can use tools!
Both scripts are like 10-20 lines of code; the first wraps a regex and the second wraps `claude -p`. Or they could be one script. The point is, you’re no longer going to tell the original agent what to do - you’re giving it a prompt and a suite of validations - tools that let it check if it’s reached the goal you have for it, and which if it hasn’t, give guidance on how to inch closer.
Your job as a software engineer DID get taken - by you, the software verification specification engineer! And the agents DO write over 90% of the code. Because why would you code by typing when you can have so much more leverage using coding agents to codify your goal.