AI is amazing and chaotic at the same time. I use it every day to write code and research questions about product design and marketing. It can help me work 100x faster to complete certain programming tasks, and it can help me work 2x slower as it thrashes on a task that I eventually have to solve myself the old fashion way. This is the current state-of-the-art for the most compelling use of AI by a domain expert-coding. Coding agents are powerful and advancing quickly, but cries that human programmers will soon be replaced by robot programmers are laughably premature. There is simply no repeatable evidence to support this claim.
Exponential, naught. The AI maximalists (boosters and doomers alike) would have you think that artificial intelligence is improving exponentially and that that growth is leading to exponential growth in productivity. You hear things like “I am no longer needed for the actual technical work of my job.” (@mattshumer) But the people who say such things don’t actually provide specific examples of what that technical work is that is no longer needed, or the failure rate when nothing of value is produced.
And let’s be clear, the only metric of AI that is arguably growing exponentially is the cost of training LLMs. With each release of a frontier model there is an impressive jump in their efficacy, but these are incremental not exponential improvements. The comparisons to viruses are silly. Viruses with an R0 (basic reproductive number, R naught) of, let’s say, 2 (in the COVID range) means that for every unit of progress, two units of progress will occur in the same amount of time, and then again in the same equal period after that, and so on until a point of exhaustion is reached. As impressive as the growth of generative AI has been, it has not been exponential by any measure.
Tool shaped objects everywhere. What we are seeing is a explosion of tool shaped objects (@WillManidis) that appear to do work, but are simply a thing to work with without doing useful work. It can be both true that a tool can be used to create objects (wood carved, chairs built, etc) that are shaped like solutions to real problems (shaping, sitting, etc), and to create things that are shaped like the solution to a problem that doesn’t exist: “make me an AI to write emails from bullet points, and also make me an AI to summarize emails as bullet points”. (Something @benedictevans likes to joke about).
All technological revolutions so far followed the Gartner Hype Cycle. The present one is unlikely to be an exception.
We are clearly still on the way up to the Peak of Inflated Expectations. This doesn’t mean that the technology isn’t already fabulously useful, just as trains and fiber optics were in the early days of their respective revolutions. But two things have to happen to reach the Plateau of Productivity: production of solutions to real problems, and wide dissemination of those solutions. Both forces are external to the advancement of the core LLM technology. Both move a more normal speed of social, commercial and political change.
Railways needed customers and supply chains that needed distribution of goods and people. The internet needed to solve actual problems of commerce and information transmission that demanded lots more bandwidth provided by the laid fiber. AI needs deterministic software tools that is able to leverage probabilistic inference to provide real value to their users.
The Peak of Inflated Expectations is reached when:
- Users realize that general purpose agents cost more than the value they provide. There are and will continue to be compelling demos of OpenClaw doing ridiculously impressive feats, but those won’t generalize to tools that are useful and safe for a mass market
- Venture funding dries up for foundation labs which means they can’t pay their data center bills because they lack of large scale adoption and price competition. Incumbents will be fine because they have the cash flow to cover their expenses and they have use cases for AI that lead to more cash flow
- Negative public sentiment grows because of the seemingly chaotic application of AI to every aspect of life and commerce causing a backlash that dampens demand for consumer AI. Again the incumbents who have built-in use cases (ad targeting, search, content moderation / personalization) will largely avoid these headwinds. Also, providers who focus on professional domain experts (Claude Code) will probably be fine as they provide real value to a market that can pay for those tools.
We’ll know we are in the Trough of Disillusionment when:
- The build out of data centers and LLMs transitions from being funded by financial capital to being funded by production capital (Technological Revolutions and Financial Capital). “Financial capital [enables] early-stage innovation and production capital [ensures] sustained, widespread economic and potential social progress in the later stages of a techno-economic paradigm.”[5]
- Broad applications of AI emerge in highly valued domains such as medicine, law and finance. This will allow for the shift from speculative funding by financial capital to production capital.
- Development of new form factors for AI powered products. At first we built horseless carriages but then we built Model Ts. The emergence of new form factors that are AI native will be a good sign that we are starting up the Slope of Enlightenment
So what something big is happening?
The deployment of an extremely powerful technology based on long understood statistical computation. Its unlock at this moment is due to a change in scale, not in kind. Sure, there were breakthroughs in how the models are configured (most famously Attention Is All You Need from Google) but even those were not enough without the shear power of the GPUs and scale of the training data used to compress the worlds information into a form that supports context-dependent text (and image) generation.
Sharp tools that fit the craftsman’s hand. CLI based coding agents in general and Claude Code in particular are taking the programming world by storm. Indeed, a perfect storm caused by access to huge amounts of code, novel LLM technology packaged into a neat little CLI wrapper with ready access to all of the Unix tools used by every experienced programmer. These tools, which effectively supercharge the developer experience by offloading multiple heavy burdens that range form remembering / learning esoteric Bash incantations to learning a new front-end framework, or God forbid getting the CSS right. Poof! All this is done by these magical programming genies. This is the 100x experience that many Claude Code users are reporting.
Craftsmen in the loop. Generative AIs don’t have craft. The lack of determinism, and more importantly taste, in the core engine means that a human expert needs to be the final arbiter of success for any task. The reason that coding agents are so effective is that expert programmers already have many tools for verification at their disposal, and the AIs are getting better at using those tools on their own. These tools include compilers, linters, test harnesses, which all enable syntactic and semantic checks. But even with these, a domain expert is needed to vet the work product to verify that it does what it is supposed to do.
The current state of AI technology simply cannot validate its own work. This is as true of Claude Code as it is of OpenClaw. That is not to say that such tools don’t sometimes produce problem shaped solutions. They most certainly do. And increasingly so. But we can’t trust them to always do so.
But the Claws! The use case for OpenClaw has been around since computers were a thing. At least from the early sixties there have been tools that enabled non-programmers to automate tasks for themselves. OpenClaw is this on steroids. These tools include: Logo, Smalltalk, Basic, Hypercard, VisiCalc, FileMaker, Flash, Dreamweaver, and modern no-code tools. There is a name for this: end-user programming. As the name suggests. End-user programming is when the developer and user are the same person or team. End users are by definition domain experts in the problem space they are developing in. The business owner who automates business processes specific to her business knows exactly what is the intended outcome. The difference of course is that the Claws do the thing by themselves. Amazing! Good luck with that.
There is no question that agents are a boon for end-user programming. I’d probably argue that the end-user programming revolution precipitated by OpenClaw will be more impactful than the more careful adoption of Claude Code by the professional programming class. This is because the security and privacy risks are arguably lower in the end-user programming use cases. But also the number of use cases is so much larger.
Human experts as essential guides. When I travel to Paris I can get around with Google Maps and a Rick Steves book, but if I really want to get to know a particular neighborhood or art collection, I hire a guide. AI can provide a map of an area, but it lacks human empathy and the intent that follows from that connection. You can’t one shot a legal defense or a SaaS app. There is intent all the way down. With each step of the technical, social, educational process, there is another decision to be made and an intention to guide it.
Long on Jevons paradox. Jevons paradox says that under certain conditions an increase in efficiency of production of a thing leads to an increase in demand for the thing and therefore more production.
The following conditions are necessary for a Jevons paradox to apply:[14]
- Technological change which increases efficiency or productivity
- The efficiency/productivity boost must result in a decreased consumer price for such goods or services
- That reduced price must drastically increase quantity demanded (demand curve must be highly elastic)
Let’s look at the conditions around the use of coding agents and demand for software.
- Code agents such as Claude Code are clearly increasing the productivity of individual developers and likely will do the same for teams with better tooling
- Greater efficiency and increased competition between software vendors will undoubtedly lead to decreases in price to consumer utility over time
- The software demand curve is virtually infinitely elastic and very sensitive to price. Most software has yet to be written because it is just too expensive to write
Bottom line. The number of software development jobs will increase year-to-year for the foreseeable future. This is not the say that the work that the human programmers do won’t change. Just as with high-level programming languages and more sophisticated tooling programming became more accessible to more people, the same will undoubtedly be true with generative AI. The reason is simple, demand for software is highly elastic and price sensitive so Jevons will rule the day.
Bottom bottom line. Agents will do what end-user programming tools have promised for decades, and do it well. This will be a huge win for businesses and other organizations that can replace human number crunchers and paper pushers by automating business processes. Given that demand for such work is relatively fixed, this will likely lead to a RIF in so called office work.
How then shall we work? As AI automates knowledge work of other kinds, the same will be true for those professions as for programmers. AI tools make knowledge work more efficient, the cost of doing related jobs goes down, competition forces savings to be passed to customers and, with elastic demand, providers who are experts get more work. Go to school. Get work experience. Become an expert. Use AI. Relax. You’ll be fine.