There’s a long-simmering debate over whether LLMs will become a commodity market with thin margins on inference. My belief is that they will evolve to be something like a ladder market, an example of which is the DRAM market, where there is differentiation across products in terms of quality, but this is a moving target, with technology always improving. This gives the leaders pricing power, but they need to continue to invest in R&D to maintain their lead. In markets like this, the followers tend to benefit from incorporating the technological advances pioneered by the frontier producers at a reduced cost. I think this market shape, along with inference scarcity, already describes the current environment, where the frontier labs are able to charge the highest rates for their models, but there is enough competition to keep prices in check. Within a capability band, the models are largely interchangeable in terms of what they are able to do, but effectively using a smaller open-weight model as a substitute for a larger frontier model is not always possible. OpenAI and Anthropic are the two labs with the most capable models at the moment. They have the clearest moat in training data, compute resources, and technical know-how. Google has most of the same resources, so I expect them to catch back up at some point. When you look at the other labs, both in the US and China, they are clearly lacking in at least one, if not all three, of these factors. The other labs are innovating in their own way, but it’s clear that at least part of their ability to catch up is via distillation (the degree varies from model to model). I think, to date, distillation has been effective at helping bootstrap models, but it’s clear from using them that when you push them off-distribution or push them on long-running tasks, they degrade rapidly. The next frontier in training will push agents to run independently for longer, and I think training in conscientious behaviors will be key, but doing so will be hard and even harder to distill, which will give the frontier labs a durable moat for some time (12–18 months, perhaps).
Across the board, there are open-weight and frontier models that excel at one-off problem-solving, even for ambitious tasks. In software, what people have been noticing more and more is that these models produce code that incurs a lot of tech debt, and these models do not inherently clean it up effectively without a lot of external direction, either from strict linting, auto-refactor loops, or manual prompting. Long-running agents are the current focus of the frontier labs. This includes pushing models to the point where they can manage a codebase long-term. A recent paper introduced a helpful benchmark here that Dex at HumanLayer has called SlopCodeBench. This benchmark basically asks a model to build a piece of software, but the requirements are only slowly introduced as the model progresses, which mirrors how human engineers often have to work. None of the models do well on this benchmark, essentially because it requires building software that is well-designed to be extensible and carefully refactoring the code before making large changes when new requirements come in. The current models all just plow straight ahead and end up building a big ball of mud.
Thinking about the way the models behave right now, I think something they are lacking is conscientiousness, defined by my friend Google as:
Conscientious is an adjective that means wishing to do one’s work or duty well and thoroughly, guided by a strong sense of what is right. A conscientious person pays close attention to detail, takes obligations seriously, and puts care and effort into tasks.
Key meanings:
- Doing your job well: Working with great care, focus, and deep effort.
- Following a moral compass: Acting according to your inner sense of right and wrong.
Common traits:
- Organized: Keeping neat schedules, plans, and to-do lists.
- Responsible: Showing up on time and meeting goals.
- Thorough: Checking details and finishing tasks correctly.
Current models are intelligent, diligent, and hard-working, but do not exhibit the behaviors that a highly conscientious person does. In people, conscientiousness is the Big Five trait that correlates most consistently across many measures of success, both personal and professional, to the point that I personally feel like most social science research is just showing that things that correlate with conscientiousness also correlate with good outcomes.
I expect that in the near future, frontier models will start to show behaviors that resemble conscientiousness without being directly trained for them. RLVR (Reinforcement Learning with Verifiable Rewards) indirectly trains for traits that are broadly beneficial through evolutionary pressure. I expect that as the labs focus more and more on long-running agents, RLVR will begin to select for these types of behaviors; however, conscientiousness is one of those things that very rarely leads directly and immediately to success. I think this is something that distillation will fail to effectively pick up and that will be trained into models once very long-term and shifting tasks are part of their training set. Models that have had conscientious-like behaviors distilled into them might superficially look like they are exhibiting these characteristics, but this will be shallow and only appear on the happy path. When something goes wrong and they are off-distribution, they will fail to show these behaviors (see the DAgger paper), but this is exactly where these behaviors are the most valuable.
In the medium term, I think this will be the next wedge between the frontier and open-weight models. The frontier models will be able to power agents that can act more autonomously over long periods, even if they have similar outcomes on short-horizon tasks. Conscientiousness is something that is really hard to measure precisely in people, and I expect many people to fail to detect the differences across models well at first. This will be compounded by the open-weight models being able to do crazy one-off tasks, giving the appearance of strong ability, only to struggle to evolve a codebase over time. This feels like a durable moat that the frontier labs will be able to maintain for a time against the open-weight models. Learning to spot the subtle things that a conscientious agent does and knowing when this characteristic is important will be a key skill to develop in the coming months.