+ "markdown": "---\ntitle: Knowledge-Creating LLMs\nauthor: Tom Cunningham\nengine: knitr\nbibliography: ai.bib\ndraft: true\n---\n\n<!-- https://tecunningham.github.io/posts/2026-01-29-knowledge-creating-llms.html -->\n\nKnowledge-creating LLMs have distinct economic implications.\n: There has recently been a burst of excitement about LLMs which are advancing the frontiers of human knowledge. I think there are some distinct economic implications of knowledge-creation LLMs that I haven't seen described elsewhere:\n: 1. **Knowledge-sharing LLMs will be used by *non-experts*,** and will be widely available.\n 1. **Knowledge-creating LLMs will will be used by *experts*,** and will be closely held by the labs, or sold with exclusive licenses.\n\n Below I give a general argument, then a very simple formal model.\n\nSarah Friar, Jan 2026, [\"A business that scales with the value of intelligence\"](https://openai.com/index/a-business-that-scales-with-the-value-of-intelligence/)\n\n: > *\"As intelligence moves into scientific research, drug discovery, energy systems, and financial modeling, new economic models will emerge. Licensing, IP-based agreements, and outcome-based pricing will share in the value created.\"*\n\n\n\n## Model of LLMs for Discovery\n\nIt's useful to distinguish between two types of LLMs:\n: 1. **LLMs that share existing knowledge (old LLMs)**-- they are trained on human-produced and human-judged data.\n 1. **LLMs that discover new knowledge (new LLMs)** -- they are trained against new data directly from the real world, e.g. math, verifiable problems, computer use, actions in the world.\n\nOld LLMs share existing knowledge.\n: Traditionally LLMs are trained with human judgment as the ground truth, using labels from paid raters, or from LLM users. As a consequence they can answer questions and solve problems up to the limits of human expertise but rarely beyond.\n\n An implication: they will be used by people outside their areas of expertise, and by firms that are followers, to catch up to the frontier.\n \n As a consequence they decrease knowledge rents -- people and firms whose value is from their existing knowledge.\n \n They increase home production (you can solve problems yourself instead of paying for it), and so decrease GDP.\n \n They decrease the returns to innovation (and news-gathering), insofar as they cause new knowledge to diffuse more quickly.\n\n This business has high fixed costs -- collecting all the knowledge to train the model -- and relatively low marginal costs in sharing that knowledge.\n\n \n<!-- (more speculative) they decrease firm size, because you don't need an in-house specialist anymore. -->\n\nNew LLMs discover new knowldge.\n: Over the past year there have been various announcements of LLMs used to advance the state-of-the-art on various specific problems, i.e. creating new knowledge. \n\n They will be mostly used by *leader* firms instead of followers, and will be used *inside* their area of expertise instead of at the fringers.\n \n We should expect much higher variable costs, i.e. expenditure on advance the state of konwledge on a single problem.\n \n The demand for new knowledge is much less elastic than the demand for existing knowledge. Selling knowledge to one person is much more valuable than selling to two people. Thus labs will prefer to restrict output, e.g. by selling the knowledge to just one firm, instead of selling the ability to generate knowledge.\n \n Our benchmarks for new LLMs will be qualitatively different. Instead of seeing if they can answer questions which we already know the answer to, we want them to answer *new* questions, e.g. Erdos problems, or setting records on optimization benchmarks.\n\n Examples of knowledge-creating LLM applications:\n\n - Predict stock prices\n - Optimize algorithms\n - Optimize technology\n - Solve scientific problems\n - Create a movie\n\n\nA simple model with recipes:\n: 1. **Baseline: everyone buys from the person who knows the best recipe.** Everyone has a unit of labor. There's one consumption good, but various recipes for producing it, $r\\in R$, which determine the labor-cost of producing the good, $c(r)$. In equilibrium the person who knows the lowest-cost recipe ($c_1$) will sell the good in return for others' labor. Their margins are equal to the difference to the next-lowest-cost recipe, $c_2-c_1$ (assume Bertrand competition).\n 1. **Knowledge-sharing LLMs eliminate rents.** Now you invent a knowledge-sharing LLM, which can reveal the lowest-cost known recipe, $c_1$. You cannot make substantial profits from this knowledge: once two producers have the same cost then margins will be driven to zero. Assuming the recipe does diffuse, total output remains the same but the surplus is now distributed equally. If we additionally assumed some trade cost $\\delta$ then the knowledge will have value equal to $\\delta$, but notably there's no value to *exclusively* license your LLM. Also notably the returns to innovation fall: there's much less incentive to discover a new low-cost recipe if that knowledge will be immediately shared.\n 2. **Knowledge-creating LLMs generate additional surplus.** Next we introduce a knowledge-creating LLM, which generates a new recipe $c_0<c_1$. The inventor can monetize this either by producing the good themselves or licensing the recipe to a single producer. Now exclusivity is important: if they sold the recipe to *two* producers then profits will be driven to zero, and the value of the recipe will be zero. In equilibrium total output increases, the extra surplus is split between consumers and the owner of the new recipe.\n\n\nThe model can be extended to multiple goods. \n: If people have Cobb-Douglas preferences across goods then they will spend a fixed fraction of labor on each good, and so each goods market can be treated as independent.\n\n You can visualize the distribution of costs as follows:\n\n - Old LLMs are the minimum cost among existing humans.\n - New LLMs *lower* the cost.\n\n\n\n::: {.cell}\n::: {.cell-output-display}\n{width=672}\n:::\n:::\n\n\n\n<!-- \n| | LLM that shares knowledge | LLM creates knowledge |\n| ----------------------- | ------------------------- | ---------------------- |\n| | | |\n| Superhuman performance? | Only in special cases | Often |\n| Who uses it? | Followers in an industry | Leaders in an industry |\n| What types of use? | Specializations | Non-specializations |\n| | | | \n-->\n\n# More to Do\n\nThere are obvious implications for intellectual property.\n\n: A specific worry: if we maintain the same intellectual property law then there will be a land-grab, firms will rush to be the first to discover new technologies, and will then get an exclusive license, but that exclusivity will be inefficient (i.e. it wasn't necessary to motivate the research, the technology would've been discovered anyway).\n\n\nIt would be more satisfying to have a generative model.\n: I'd really like to sketch out a very simple model in which both humans and LLMs learn recipes from experimenting against the real world.\n\n\n\n\n\n# Related Notes\n\nSarah Friar, Jan 2026, [\"A business that scales with the value of intelligence\"](https://openai.com/index/a-business-that-scales-with-the-value-of-intelligence/)\n\n: > \"As intelligence moves into scientific research, drug discovery, energy systems, and financial modeling, new economic models will emerge. Licensing, IP-based agreements, and outcome-based pricing will share in the value created. That is how the internet evolved. Intelligence will follow the same path.\"\n\n@yuksekgonul2026learning, \"Learning to Discover at Test Time\"\n: > \"We report results for every problem we attempted, across mathematics, GPU kernel engineering, algorithm design, and biology. TTT-Discover sets the new state of the art in almost all of them: (i) Erdős’ minimum overlap problem and an autocorrelation inequality; (ii) a GPUMode kernel competition (up to 2×faster than prior art); (iii) past AtCoder algorithm competitions; and (iv) denoising problem in single-cell analysis. Our solutions are reviewed by experts or the organizers.\"\n",
0 commit comments