Abhay Venkatesh and Emiliano Garcia-Lopez
October 6, 2026. Also published on X.
Many shall run to and fro, and knowledge shall increase.
Daniel 12:4
When AI was first coming out, everyone thought that the first people to get automated would be all the “wordcels” [2]: the poets, the executive assistants, and the marketers. The opposite seems to have happened. Software engineering has become almost fully automated and most recently, AI has produced a proposed solution to a Millennium Prize Problem. [3] And yet it cannot function as your executive assistant. It can solve hard math puzzles but has no idea how to not totally screw up your calendar.
In the Valley we tend to overrate intelligence [4] and treat it as the be-all-and-end-all skill: if you can solve math puzzles, you can do everything else. But that’s not how intelligence actually works. In AI, strength along one axis of skill doesn’t reliably carry over to others. Many would argue that this is just a matter of time until RSI takes off and it learns everything else. We don’t expect this to happen. Our prediction is that AI will become extraordinarily good at a growing number of things and stay shockingly bad at others, often within the same field. We call the AI we actually observe “patchwork AGI”: an AI that grows patch by patch, trained by humans or by AI-assisted humans.
Patchwork AGI is not the “jagged frontier” others have described: an uneven edge around a filled-in region, in domains like math and coding. [57] [58] We actually make a stronger claim than this: because models learn through massive data coverage instead of general principles, most of the region isn’t filled in at all, even in the relatively stronger domains. This is why we see strange phenomena like this one: in Stanford’s 2026 AI Index, the same top model that beats human baselines on PhD-level science questions also fails to read a simple analog clock about half the time. [62] If models’ abilities were general, the first task would cover the second. Models only appear generally intelligent because benchmark suites average over many tasks at once. So scores rise together, and gaps only show up when you look at the individual tasks the training data doesn’t cover.
And the patchwork doesn’t stand still: as patches get filled, new patches get created, because new AI capabilities transform the nature of work people do (e.g. instead of writing code, you are now managing agents that write code, and so on).
We think that the current paradigm of AI progress will not become a strict superset of human or all useful intelligence. In particular, we do not expect what is known as “fast takeoff” [5]: once AI crosses a certain level of threshold, it automatically takes over becoming more intelligent than all humans in an accelerating period of time (often imagined as being driven by recursive self-improvement, or “RSI”). [6] [7]
We are not AI pessimists. We think that there is not an AI bubble, but there is an everything but AI bubble. The AI Economy [8] will eclipse the everything-but-AI economy. And we think this will be fast and transformative. But that is different from saying that we are birthing a machine god that is omnipotent, all powerful, and out-of-control and recursively improving in short duration.
Our mental model is that rapid progress will be persistently uneven. Existing “gaps” in AI capabilities can close quickly without actually being a superset of all useful intelligence. Furthermore, those capabilities will create further use-cases for AI that are more “meta” in nature, leading to a perpetually renewing “patchwork” of capabilities (e.g. with the proliferation of AI slop, the need for AI slop checking). If you ask us “what comes after AI?”, we will say “20 more years of AI”. [9]
Learning tasks for an AI is more than just about abundant data. For example, there are plenty of examples of “good writing” out there in the world, and the AI has been trained on it, but AI—as of Oct 6, 2026—is still unfathomably bad at writing. The AI can produce text that “looks” like good writing but it isn’t actually good writing, because good writing is often about having something new to say.
For people, writing is about having higher-level ideas: learning concepts and then abstracting them by writing them in words. AI learns the other way around, from the “words up” (by learning to predict the next word [10]), and as far as we can tell, never learning the actual concept, mostly what sounds plausibly like a concept. [11] For example, a model that writes an algorithm well in one programming language—let's say Python—can do much worse in another, even when given that language’s rules. [12] [13] It knows what the algorithm should “look like” in a language it has seen often more than it knows the algorithm itself.
When AI gets good at some type of work, it changes the nature of the work around it. People take on work that once cost too much and expect more of their tools. [14] Consider the case of Cognition, the AI startup that makes the coding AI “Devin”, the so-called “first AI software engineer”. Companies use Devin but then they realize they need to govern its access, supply company context, and then actually get work done with Devin. Cognition worked with BlackRock for tools for that purpose. [15] Or take the case of Pangram AI, a company that determines whether something is AI generated slop or not. [16] AI has now made it so that it's really easy to generate text but then now we’re just inundated with slop. This is “recursive patchwork”: AI creates more problems for itself. It’s somewhat of a consequence of a paradox of progress or the paradox of technology more generally. We solve problems only to find new ones. [17] [18]
Let's try to see how exactly this patchwork manifests in practice. We think the patchwork really is a function of what it takes to train the AI. Favorable training has a few requirements. Can a program check the answer? Is a narrow procedure enough to pass? [19] [20] Does it need a model of physics, people, or causes? [21] How much experience beyond text does it need? [22] How much must it keep track of over time? [19]
Table 1. What a task requires
| Task | Can a program check it? | Is a narrow procedure enough? | World model needed | Experience beyond text | Tracked over time |
|---|---|---|---|---|---|
| Lean proof search | Yes | Often | Formal rules only | None | One proof |
| Coding puzzle | Yes | Often | Formal rules only | None | One problem |
| Maintaining a large codebase | Partly | Sometimes | How the system behaves; what the team wants | Some social | Months |
| Negotiating a deal | Partly | Sometimes | Other minds | Social | Days to years |
| Robot assembly in a new setting | Yes | Rarely | Intuitive physics | Perception, action | Minutes |
| Writing with something new to say | No | No | People and the world | Social | One piece |
| Running a wet-lab experiment | Partly, slowly | Rarely | Physical and biological causes | Perception, action | Days to weeks |
| Running an executive’s calendar | No | No | Other minds | Social | Weeks |
| Proposing a new theory | Only in hindsight | No | Causes, at depth | All of the above | Years |
Green: generally favorable for training. Yellow: mixed or conditional. Red: more demanding. The ordering is illustrative, not a measured ranking; difficulty depends on the task, tools, and setting. (see Appendix A.3)
Checkable tasks are the easiest to train. A test suite of proof checkers can grade millions of attempts with no person in the loop, so the strongest axes of patchwork AGI form where checking is available. At the same time, these proof checkers keep capability axes paradoxically narrow. They reward whatever passes, and the model just learns how to pass the checks rather than learning what the task is actually asking on a meta-level. [19]
For non-checkable tasks, such as those that require human judgment, the industry builds data by hand. Companies like Mercor pay humans to write tasks, compare answers, and build grading rubrics. Human data has far exceeded everyone’s expectations in its scale and durability. [23] [24] Everyone thought these companies would disappear as soon as RSI took over and yet it's entirely possible that Mercor may very well be on a path to a $100B+ IPO. We think data continues because we have patchwork AGI and not fast takeoff. (see Appendix A.1)
If capabilities transferred readily from one patch to the other, then perhaps our argument would collapse. Let’s study this dynamic from the literature. From what we can see, training on one skill sometimes does transfer, but only slightly and selectively when it does, and it can make other skills regress. [28] [29]
For example, in one controlled study by Chuxuan Hu, Daniel Kang, and colleagues at the University of Illinois Urbana-Champaign, math training raised math accuracy by 3.13 points while lowering average accuracy on other tasks by 1.81; math training helped coding but not more distant tasks (not surprising). [28]
But on the other hand, in another study, math training and coding training paradoxically lowered each other’s scores, while puzzle training raised all three (Table 2A). [29]
Table 2. Transfer varies by training domain
A. Math, coding, and puzzles [29]
| Training domain | Math | Coding | Puzzles |
|---|---|---|---|
| Math | +25.00 | −3.23 | +13.35 |
| Coding | −3.31 | +6.49 | +13.48 |
| Puzzles | +6.99 | +3.89 | +52.91 |
Qwen2.5-7B-Base; changes in domain-average accuracy, in percentage points. Positive values indicate improvement; negative values indicate decline.
What about non-technical skill learning? We see similar patterns. Training a small model on negotiation, for example, transferred best between two price negotiations, and weakly or negatively to scheduling and job interviews (Table 2B). [30] Some environments built from coding problems do raise customer-service accuracy, by 8.7 points on average, but the catch is that they train tool use directly. [31] Where gains do reach new skill areas or fields, it is usually due to training designed specifically for those areas (see Appendix A.6).
B. Negotiation and scheduling [30]
| Training domain | Craigslist | Marketplace | Job Interview | Calendar |
|---|---|---|---|---|
| Craigslist | +0.265 | +0.317 | −0.032 | −0.035 |
| Marketplace | +0.184 | +0.664 | −0.108 | −0.033 |
| Job Interview | +0.010 | −0.013 | +0.115 | +0.072 |
| Calendar | −0.151 | −0.002 | −0.168 | +0.239 |
Qwen3-4B-Instruct-2507; changes in mean utility (0–1 scale), calculated from SocialRL Table 3. Rows are training environments; columns are evaluation environments. Selected four-domain submatrix; price-negotiation models use supervised training before RL. Panel magnitudes are not comparable across models and measures.
If we think a bit deeper about how success is measured, then we realize that it may be measured in a more narrow way than would be optimal. For example, in SocialReasoning-Bench, built by Tyler Payne, Saleema Amershi, and colleagues at Microsoft Research AI Frontiers, frontier models usually booked the meeting or closed the deal, but often chose a poor time or price, even though the benchmark states the user’s preferences outright. [35] Completing tasks doesn't necessarily mean satisfactorily serving the person who asked for it (see Appendix A.5).
Expecting general intelligence to emerge from a hotch-potch of skills asks more of machines than even of people: training in chess, music, or working memory rarely carries over to other domains. [36] Becoming a better speechwriter doesn’t make you a better mathematician. The way humans learn—and not to make an anthropocentric or anthroposuperiority argument here, just as an example—is that we start with a general meta-learning ability and do not assemble general intelligence from sub-skills.
Patchwork AGI follows from the way in which AI systems today are set up: they fit narrow learning procedures to narrow learning rewards, and largely stop learning once deployed. A truly general learner would break from some or all of these constraints. Cognitive science already has a fairly concrete picture of what this pattern should look like.
From the POV of cognitive science, people build models of objects, other minds, and causes; make new concepts from familiar parts; and learn to learn, so each new task comes faster than the last. [21] [37]
We would change our minds on our thesis if AI systems began to show these properties in open settings, without a new round of human-built training for each new skill or domain area of expertise. The relevant test is whether a capability transfers without additional task-specific training. A potential counterexample is GPT Astra’s driving test result: on DrivingBench, it completed a fixed cone course in a real car on its second attempt, and early commentary described the ability as emerging without specialist training. [32] [33] The reported 100% measures completing that course, not driving in general. And even in this case, it’s not exactly clear that the model has not actually been trained on driving data or gaming environments that basically test this; the same commentator later acknowledged that possibility (see Appendix A.4).
Table 3 lists the five signs that would matter most. Each would show a system learning structure rather than habit. (see Appendix A.4)
Table 3. Signs of a general learner
| Sign | Today | Would count against us |
|---|---|---|
| Skill that survives new rules | Given the rules of an unseen programming language, frontier models solve far fewer problems than in Python. [13] | Variants handled almost as well as the originals |
| Learning on the job | Test-time updates help on novel puzzles [38]; models lose track of state over long episodes. [19] | A new workflow learned from a few examples, kept, and updated as things change |
| Parts that travel | Reusable abstractions, learned one domain at a time. [39] [40] | What is learned in one field speeds learning in the next |
| A model of the people it serves | Partial inferences about what people believe and want. [41] | An executive’s calendar run well with no training built for the job |
| A new way of seeing | Longer chains of reasoning inside existing frameworks | A theory revised from sparse evidence, and a framework experts adopt [37] [42] |
Each row pairs current evidence with an observation that would count against our forecast.
Indeed the people are one and they all have one language, and this is what they begin to do; now nothing that they propose to do will be withheld from them.
Genesis 11:6
Another sign that our predictions are wrong would be if we start noticing AI identifying and closing its own capability gaps with less or even no human direction. AI already helps build training regimes for other skills (with human assistance): a coding model writes reward functions for robot policies, and generated environments teach planning, for example. [43] [44] But we note that this is still human guided. People are choosing the tasks, architecting the simulator, and judging the results. So we think that this type of training regime is mostly an accelerant and not a recursively-improving-loop that removes the human from the picture (Figure 4; see Appendix A.7).
Direct transfer
Other capabilities may improve without training for those tasks.
AI-assisted training
Autonomous self-improvement doesn’t seem to exist yet: so far, it requires human guidance. Models are poor judges of their failures, and so people are still needed to evaluate and guide models as they learn. The most direct test so far points the same way. In Epoch AI’s InnovationEval, frontier agents given 3,000 GPU-hours to invent a better post-training method recovered only 2% and 15% of the gain from a recent human innovation, once best-of-many run selection and out-of-scope changes were stripped out, and both agents’ reports overstated their results. [56] For autonomous RSI, we would need to find ways for AI to find its own gaps and fill them, and we should actually expect spending on human data to fall even as AI capabilities broaden. [7] Until then, we expect “patchwork AGI” to persist (see Appendix A.8).
That doesn’t mean there’s no recursion. A weaker form of RSI is real, as with all technologies: better machine tools were used to build even better machine tools since the industrial revolution. Here too, AI research will make AI researchers far more productive, but an AI researcher agent that outcompetes a top researcher with an agent is still science fiction, and until then the RSI loop will run at human speeds. Amdahl’s law shows this: if AI makes 70% of the research loop 100x faster, the aggregate speed is only around 3.3x faster. Even making 90% of the work instantaneous caps the gains at 10x. [59] The part that doesn’t speed up (deciding what to train next, judging whether a result is real, having a new idea) remains the bottleneck. This is where we depart from forecasts like Situational Awareness and AI 2027, which expect automated AI research to proceed at supernatural speeds. [60] [61]
He was granted power to give breath to the image of the beast, that the image of the beast should both speak and cause as many as would not worship the image of the beast to be killed.
Revelation 13:15
Just because we think AI will continue in this patchwork manner doesn’t mean AI will be slow and not transformative. We think that the next ten years will change everything: how software is built (already done) but also how science is done and more generally how all work is done. Many capabilities will advance rapidly and in parallel while retaining substantial gaps to be filled by companies and researchers. Furthermore, new patches will continue to emerge as AI advances. As a result we think there will be a few consequences.
Data continues to grow as an industry. Everyone thought that Mercor and friends were going to disappear as soon as RSI takes over. The opposite seems to have happened. It seems like Mercor is well on its way to being a $100B+ publicly traded company given how persistent gaps are, and how they seem to continue renewing themselves and creating new patches. As long as AI is growing, so will be its need for data (absent any paradigm shifts of course). Data is already a multi-billion dollar industry; there’s no reason it cannot be larger.
Software becomes trivially easy to produce but as a result there’s far more of it. AI does not lead to the End of Software [45] but rather the Amplification of Software. [46] When it's easier to produce the marginal piece of software, moats shift from switching costs to economies of scale. Software was already winner-take-all: with zero marginal cost, the best product earns the most, reinvests the most, and pulls further ahead. [47] AI shortens the time it takes for small differences in team and product quality to compound (Table 4). Software does not lose its moats; it concentrates.
Table 4. AI amplifies differences between teams
| Scenario | 10 people × IQ 145 | 100 people × IQ 115 | Small-team advantage |
|---|---|---|---|
| Before AI | 300 | 200 | 1.5× |
| After AI | 1,500 | 600 | 2.5× |
Illustrative output units from [46], not measured. In this example, the small team’s advantage grows from 1.5× to 2.5×.
One potential consequence is that smaller teams do far more. For example, Instinct—an AI personal assistant company started by a 23 year old—reached a $10B valuation with 14 employees. [48] AI magnifies what the teams can achieve and at what pace but it still leaves some key patchwork and doesn’t fully eliminate it: judgment, taste, knowledge of the customer, and distribution.
The continued supremacy of the digital world, not the end of it. Contrary to popular understanding, it's not clear to us that the physical world starts to dominate the digital world in terms of profitability and as a result influence and dominance. The digital world is already enormous: all five of the world’s most valuable companies, and eight of the top ten, are built on computing. [49] In Thiel’s framing, tech has been almost the only source of new American successes for decades: there have been no non-tech successes in the West in the past 50 years. [50] [51]
Everything is computer. As Donald Trump put it while sitting in a Tesla at the White House, “everything’s computer.” [52] To the extent that there is progress in the physical world, it is going to be a function of how tightly coupled it is to the digital world. Apple and Nvidia are examples of “hardware companies” but really they’re tightly coupled to the digital world—Apple’s software ecosystem moat and Nvidia’s CUDA ecosystem as well as a direct link to model training and inference revenue.
Work moves around rather than disappears. We don’t expect mass and total unemployment (although it is possible there are short-term dislocations). The nature of work shifts from mass knowledge factories (think 1000s of analysts at investment firms or 1000s of software engineers at tech companies) to more nuanced enterprises: brand, distribution, but also economies of scale as companies go for harder problems in bio, science, and physics.
The harvest truly is plentiful, but the laborers are few.
Matthew 9:37
The role of companies. The patchwork creates at least three kinds of companies (Table 5), to fill gaps in the patchwork of AI.
Table 5. Where the patchwork creates companies
| Company type | Work | Example | When the next model ships |
|---|---|---|---|
| Data | Build the patches: tasks, environments, graders | Mercor, Surge | Stronger, as long as each new field needs its own data |
| Deployment | Fit models to an organization: context, access, governance | Cognition | Stronger: better models mean more to govern |
| Product | Cover what models still get wrong: verification, oversight, detection | Pangram | Depends on whether the gap renews or closes |
We think there are lots of data, deployment, and product companies to be made, in addition to other types of businesses, such as AI labs for training researchers and doing risky research.
Of making many books there is no end, and much study is wearisome to the flesh.
Ecclesiastes 12:12
We predict that AI will remain extraordinarily good at some things and shockingly bad at others, absent a paradigm shift in learning capabilities. Some gaps will persist because they are difficult or uneconomical to close, and other new gaps will emerge as automation changes the nature of human work.
There is much work to be done in AI, software, and the digital world more broadly. People underestimate the “depth” of the digital world. It’s not clear to us that software will ever be “finished”. Cheaper software implementation makes more projects worth attempting (the Jevons paradox). [53] Tacit knowledge necessary to create more software may be inexhaustible: mapping one piece of the world may reveal more that remains unmapped. [54]
On a final more anthropological note, as machines take on more and more of human labor, we change how we live, relate to one another, and understand ourselves. Formalizing what we do can reveal further distinctions and create new activities. [55] Our prediction is that the forthcoming AI changes will repeatedly create demands for human activity that outrun dependable AI capability.
Thanks to Jacob Hornstein for detailed feedback on an early draft. Thanks to Rajath Salegame for his feedback on the limits of transfer from coding and mathematics to hardware engineering. Thanks to Ronald Qiao for pushing on whether better coding could accelerate the creation of training environments, data, and more general models. Thanks to Zack Baker for the conversation about tacit knowing and the idea that we will not run out of things to do. Thanks to Sami Senapathy for reading and commenting on drafts of this essay. Thanks to Leonard Tang for pointing out how new uses of AI can draw on existing capabilities. Thanks to Rahul Mitra for his feedback on architectural constraints, continual learning, and sample efficiency. Thanks to Vik Pattabi for reading the essay. Thanks to Jacob Thompson for reviewing Figure 1 and the tables, and for reading parts of earlier drafts. Thanks to Turner Merritt, Nelson Arnous, and Charlie Davidmann for providing feedback on an early draft. Thanks to Fan-Yun Sun for his feedback on the transfer section and for pushing us to distinguish patchwork AGI from the jagged frontier early in the piece. Thanks to an anonymous reader from a leading neo-lab for their feedback on this piece and alignment with our thesis.
So judgment gets trained the way our thesis predicts: one field, one contract, and one rubric at a time. Each deployment turns up new failures, and each failure becomes new work for the data vendors. This is now a standard loop for data companies: a new benchmark comes out, labs buy task data built to look like it, and the next model scores well. [25] Benchmark gains overstate how general a model has become. (see Appendix A.2)
Data and feedback are only part of the bottleneck. Architecture and learning algorithms shape which capabilities a model acquires easily and how well it retains them. Adapting within a conversation differs from learning durably across experiences. Test-time learning and learned memory are active research directions; reliable continual learning remains an open challenge. [26] Sample efficiency matters too: when useful feedback is costly or slow, the amount of experience needed to learn can determine whether a capability is practical to develop. Closing some gaps may therefore require better architectures or learning methods as well as better data. These constraints can change as the methods improve.
Small-scale pretraining experiments comparing 2019–2025 recipes and datasets found larger compute-efficiency gains from data improvements than from model improvements. [27] This supports the importance of the training material. It does not establish limits on transfer: the study used small models and a limited benchmark suite, and its findings may not hold at frontier scale.
Length, interdependence, and elapsed time are separate demands. A hundred unrelated questions differ from a hundred dependent steps. A weeks-long project also requires tracking changes between actions. Each step can look reasonable while the project drifts away from its goal.
It helps to distinguish bounded tasks from open-ended pursuits. Fixing a specified bug has a stopping point; improving a software product can keep creating value. Task horizon is a separate dimension: a bounded task can require a long sequence of dependent actions, while an open-ended pursuit can consist of many short tasks. If further progress in a profitable pursuit keeps paying off, labs have a reason to keep extending that capability. Uneven development may be economically rational.
A more decisive test would compare a model before and after additional coding training, with the same tools and instructions. Does it become a more reliable executive assistant on unfamiliar tasks without training for those tasks? We would test whether it knows when to interrupt a meeting, which conflicting request to prioritize, and when to ask for clarification. Consistent gains large enough to support dependable delegation would weaken our prediction. Small improvements that leave the agent unreliable would still be consistent with patchwork AGI.
Astra’s driving result shows why such a test is needed. An initial account described the capability as emerging seemingly without specialist training, but later acknowledged possible driving data and reinforcement learning. [32] DrivingBench reports completion of a fixed cone course in a real car on the second attempt. Models received driving tools and instructions, with up to three attempts in the same conversation. The reported 100% measures course completion, not a general driving success rate. [33]
Prior training, transfer, and learning during the attempts could all contribute. Even game-to-driving generalization would be transfer. The test is whether learning carries over to a new task without additional task-specific training, not whether the model was trained at all. This result does not isolate the contribution of coding training.
Useful negotiation is already possible. In Anthropic’s Project Deal, 69 agents struck 186 deals worth just over $4,000 in a marketplace with real people and goods. In separate mixed-model runs, Opus obtained better prices than Haiku, but the difference in reported satisfaction was not statistically significant. [34] Even satisfaction can miss differences in how well an agent represents someone. Task completion, satisfaction, and the user’s interests are distinct measures. Project Deal did not isolate the contribution of coding training.
CodeGym turns coding problems into tool-use environments with verifiable outcomes. Training Qwen2.5-32B-Instruct in them improved customer-service accuracy on τ-Bench by 8.7 percentage points on average, with different gains in airline and retail service (Table A1). [31] CodeGym explicitly trains tool use and workflows, so it does not isolate the effect of simply becoming better at coding.
Some patches can share a training foundation. Reusable tool-use skills and domain-specific training can coexist; CodeGym does not imply that every capability must be built independently. Targeted training also helps: SocialRL brought a 4-billion-parameter model into the range of much larger frontier models on held-out negotiation scenarios. Transfer was strongest between structurally similar games. Relevant training can close a gap; this result does not establish that better coding produces better negotiation. [30]
Table A1. CodeGym transfer to customer service
| Training environment | Airline service | Retail service |
|---|---|---|
| CodeGym | +4.4 | +13.0 |
Qwen2.5-32B-Instruct; changes in accuracy, in percentage points. CodeGym trains tool use through coding-derived environments. These results use a different model and measure from the panels in Table 2. [31]
In Eureka, GPT-4 writes and revises reward-function code used to train separate robot policies. Those policies outperform policies trained with human-written rewards on 83% of 29 simulated tasks. The coding model helps another system learn a physical skill; its own weights remain unchanged. Humans still supply the simulator, task description, and metric used to judge success. [43]
AgentGen generates planning environments and tasks, then uses a conventional planner to produce successful trajectories for training language models. For Llama-3.1-8B, training raised success on separate formal-planning benchmark tasks from 0% to 15%. Average success across three other environments rose from 6% to 16%, although success on one of them, Jericho, remained at zero. The gains extend beyond the generated environments, though still within planning tasks. [44]
AI can already help build new capabilities. Whether this reaches unstated priorities, social context, and judgment remains untested. That is an open question, not evidence of failure.
Speed alone would not challenge our thesis. Domain-specific training could become cheap and fast enough to close individual gaps almost as soon as they appear, while new demands continue to create others. Models might increasingly identify missing capabilities, build realistic tasks and feedback, and improve performance with little human direction. Recursive self-improvement could accelerate this further if each round made the next round of AI research faster or more effective. [7] The question is whether this process leaves substantial gaps in dependable capability as the work changes.
The decisive test is whether AI keeps pace with the demands its own success creates. We would update if it reliably handled new activities and expectations with little additional human effort, so that substantial capability gaps no longer persisted. Mastering today’s tasks is one thing. Keeping up with a changing task frontier is the stronger test.
[1] Wikipedia contributors. "Inductive bias." Background on the term used in Figure 1. ↩︎
[2] Marc Andreessen. "Wordcels vs shape rotators." X, November 1, 2024. ↩︎
[3] OpenAI. "On the Navier–Stokes Millennium Prize Problem." OpenAI. OpenAI describes a proposed proof and Lean formalization; this is distinct from formal Clay prize recognition. ↩︎
[4] Tyler Cowen. "Peter Thiel on Political Theology." Conversations with Tyler, April 2024. Thiel argues Silicon Valley is "way too biased toward the math people" and that AI "seems much worse for the math people than the word people." ↩︎
[5] Nick Bostrom. "Superintelligence: Paths, Dangers, Strategies." Oxford University Press, 2014. Chapter 4, "The kinetics of an intelligence explosion," defines a fast takeoff as a transition from roughly human-level intelligence to superintelligence over minutes, hours, or days. ↩︎
[6] I. J. Good. "Speculations Concerning the First Ultraintelligent Machine." Advances in Computers 6, 1965. Introduces the "intelligence explosion": a machine that can design better machines could improve itself until it far surpasses human intelligence. The origin of the idea now called recursive self-improvement. ↩︎
[7] Anthropic. "When AI builds itself." Anthropic Institute. Reports AI-assisted gains in AI development while identifying gaps in research judgment; full recursive self-improvement has not been achieved. ↩︎ a b c
[8] Abhay Venkatesh. "The AI Economy." 2026. Argues that frontier capability is produced by a scaling complex of compute, chips, energy, capital, and research talent, so the AI economy will not resemble the internet economy. ↩︎
[9] Peter Thiel, attributed remark. Secondhand anecdote relayed to Abhay Venkatesh by people who know Thiel. Wording as recalled: "What comes after AI? Nothing, just 20 more years of AI." No public recording or transcript has been verified. ↩︎
[10] Claude E. Shannon. "A Mathematical Theory of Communication." Bell System Technical Journal, 1948. Introduces n-gram approximations of English, predicting each letter or word from the ones before it. Modern language models are far larger, but they are trained on the same objective: predict the next token from the previous ones. ↩︎
[11] Emily M. Bender and Alexander Koller. "Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data." ACL, 2020. Argues that a model trained only on linguistic form has no way to learn meaning. A theoretical argument written before current frontier models, so it is contested. ↩︎
[12] Federico Cassano et al. "MultiPL-E: A Scalable and Extensible Approach to Benchmarking Neural Code Generation." arXiv:2208.08227, 2022. Translates two Python code-generation benchmarks into 18 other programming languages. Performance tracks how common a language is, though some rare languages do as well as popular ones. ↩︎
[13] Vinayshekhar Bannihatti Kumar, Disha Makhija, Manoj Ghuhan Arivazhagan, and Rashmi Gangadharaiah. "Syntax Without Semantics: Teaching Large Language Models to Code in an Unseen Language." COLM, 2026. arXiv:2605.15607. Given the rules of an unseen language, frontier models including GPT-5.4 solve far fewer problems than in Python, though they often pick the same algorithm. ↩︎ a b
[14] Marc Andreessen. "Why Software Is Eating the World." The Wall Street Journal, 2011. Argues that software companies are taking over and transforming industries as software gets cheap enough to apply everywhere. ↩︎
[15] Cognition. "Governing AI agents at scale with BlackRock." 2026. Describes the permissions, governance, and company-context work needed to deploy Devin inside a large enterprise. Shows new work created by agent adoption; it does not show that this work will stay unautomated. ↩︎
[16] Pangram Labs. "How Pangram detects AI-generated content." Describes training a language-model classifier to tell human-written text from AI-generated text. ↩︎
[17] Edward Tenner. "Why Things Bite Back: Technology and the Revenge of Unintended Consequences." Knopf, 1996. Coins "revenge effects": technologies that solve one problem often create new ones, which then call for further fixes. ↩︎
[18] Jacques Ellul. "The Technological Society." 1954; English translation, Knopf, 1964. A classic critique of technology arguing that each technical solution creates new problems that demand more technique. ↩︎
[19] Jacob Thompson. "Evaluating World Models in Embodied Question Answering through Computational Primitives and Difficulty Progressions." Master's thesis presentation, Robotics Institute, Carnegie Mellon University, 2026. Reports weaknesses in spatial memory and state tracking, and substantial improvement from an explicit memory method. Cited from the public thesis-presentation abstract. ↩︎ a b c d
[20] Alex Bie, Travis Dick, Alex Kulesza, Prabhakar Raghavan, Vinod Raman, and Sergei Vassilvitskii. "AI-rithmetic." arXiv:2602.10416, 2026. GPT-5, Claude Opus 4.1, and Gemini 2.5 Pro get worse at adding two integers as the numbers get longer; most errors are misaligned digits or missed carries. ↩︎
[21] Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, and Samuel J. Gershman. "Building Machines That Learn and Think Like People." Behavioral and Brain Sciences 40, e253, 2017. Argues that human learning rests on intuitive physics and psychology, causal and compositional models, and learning to learn. ↩︎ a b
[22] Yonatan Bisk et al. "Experience Grounds Language." EMNLP, 2020. arXiv:2004.10151. Proposes five "World Scopes," from text corpora to perception, embodiment, and social interaction, and argues that text alone cannot supply what the higher scopes give. ↩︎
[23] Billy Perrigo. "How Meta's $14 Billion Deal Upended the AI Data Industry." Time, June 16, 2025. Reports that each leading AI company spends around $1 billion a year on human data and that data budgets are rising. ↩︎
[24] The Information, via Dealroom. "Mercor doubles to $2B gross revenue run rate as AI labs buy expert data." June 2026. A gross figure; Mercor pays 60 to 70% of it to its expert contractors. ↩︎
[25] Member of technical staff at a leading synthetic-data company that sells training data to all frontier labs. Conversation with the authors, 2026. Describes labs commissioning task data built to resemble new benchmarks. ↩︎
[26] Google Research. "Titans + MIRAS: Helping AI have long-term memory." Describes architectures with neural memory that updates at test time, illustrating research on learning and retention beyond a fixed context window. ↩︎
[27] Dwarkesh Patel and Jerry Han. "Pretraining progress is mostly coming from data." Dwarkesh Podcast, September 8, 2026. Small-scale pretraining experiments found larger compute-efficiency gains from data improvements than from model changes. The results do not establish limits on transfer at frontier scale. ↩︎
[28] Chuxuan Hu, Yuxuan Zhu, Antony Kellermann, Caleb Biddulph, Suppakit Waiwitlikhit, Jason Benn, and Daniel Kang. "Breaking Barriers: Do Reinforcement Post Training Gains Transfer To Unseen Domains?" ICLR, 2026. arXiv:2506.19733. University of Illinois Urbana-Champaign. Controlled experiments and comparisons across 16 benchmarks find selective transfer from reinforcement post-training; the numerical example above comes from Table 3. ↩︎ a b
[29] Yu Li et al. "Can One Domain Help Others? A Data-Centric Study on Multi-Domain Reasoning via Reinforcement Learning." arXiv:2507.17512, 2025. Table 9 reports changes in domain-average accuracy for Qwen2.5-7B, the source of Table 2A. Includes positive transfer from puzzle training; peer-reviewed acceptance has not been verified. ↩︎ a b c
[30] Microsoft Research. "From Passive Delegates to Strategic Negotiators: Reinforcing Social Reasoning in Small Language Models." 2026. SocialRL improves negotiation through targeted training and finds uneven transfer between games. It does not isolate transfer from coding training. The full SocialRL paper (arXiv:2608.13787), Table 3, supplies the cross-domain utility results in Table 2B. ↩︎ a b c
[31] Weihua Du et al. "Generalizable End-to-End Tool-Use RL with Synthetic CodeGym." ICLR, 2026. arXiv:2509.17325. Converts coding problems into interactive tool-use environments and reports transfer to unfamiliar workflows. The 8.7-point gain is an absolute accuracy improvement on τ-Bench; the training explicitly targets tool use. ↩︎ a b c
[32] Benjamin Todd (@ben_j_todd). "Post and follow-up on Astra’s driving performance." X, September 23, 2026. Todd initially described the driving ability as seemingly untrained, then acknowledged possible driving data and reinforcement learning. ↩︎ a b
[33] DrivingBench. "Report." DrivingBench. The reported 100% is completion of a fixed cone course on the second attempt, not a general driving success rate. ↩︎ a b
[34] Anthropic. "Project Deal: our Claude-run marketplace experiment." 2026. Agents negotiated transactions involving real people and goods. Mixed-model comparisons favored Opus over Haiku on several outcome measures; the experiment does not identify which training produced the advantage. ↩︎
[35] Microsoft Research AI Frontiers (Tyler Payne, Will Epperson, Safoora Yousefi, Zachary Huang, Gagan Bansal, Wenyue Hua, Maya Murad, Asli Celikyilmaz, and Saleema Amershi). "SocialReasoning-Bench: Measuring whether AI agents act in users’ best interests." Microsoft Research blog, May 11, 2026. Evaluates calendar coordination and marketplace negotiation, distinguishing task completion from outcome quality and due diligence. Preferences are explicit; long-term relationships are not modeled. ↩︎
[36] Giovanni Sala and Fernand Gobet. "Does Far Transfer Exist? Negative Evidence From Chess, Music, and Working Memory Training." Current Directions in Psychological Science 26(6):515–520, 2017. Finds that far transfer from these trainings rarely occurs, and that effects shrink as studies are better controlled. ↩︎
[37] Tomer D. Ullman and Joshua B. Tenenbaum. "Bayesian Models of Conceptual Development: Learning as Building Models of the World." Annual Review of Developmental Psychology 2:533–558, 2020. Frames children’s learning as building and revising intuitive theories of the world. ↩︎ a b
[38] Ekin Akyürek et al. "The Surprising Effectiveness of Test-Time Training for Few-Shot Learning." arXiv:2411.07279, 2024; revised 2025. Updating model weights at test time substantially improves performance on ARC; ensembled with program synthesis, the method reaches 61.9%, matching average human performance. ↩︎
[39] Kevin Ellis et al. "DreamCoder: Bootstrapping Inductive Program Synthesis with Wake-Sleep Library Learning." PLDI, 2021. Grows a library of reusable program abstractions that supports generalization from few examples. ↩︎
[40] Gabriel Grand et al. "LILO: Learning Interpretable Libraries by Compressing and Documenting Code." ICLR, 2024. arXiv:2310.19791. Combines language models with symbolic compression to learn reusable, documented abstractions. ↩︎
[41] Jacob Andreas. "Language Models as Agent Models." Findings of EMNLP, 2022. arXiv:2212.01681. Argues that language models can infer and represent some of the beliefs, desires, and intentions of the agents who produce text, though only partially. ↩︎
[42] Joshua B. Tenenbaum, Charles Kemp, Thomas L. Griffiths, and Noah D. Goodman. "How to Grow a Mind: Statistics, Structure, and Abstraction." Science 331(6022):1279–1285, 2011. Describes how structured, hierarchical knowledge lets people generalize far beyond sparse data. ↩︎
[43] Yecheng Jason Ma et al. "Eureka: Human-Level Reward Design via Coding Large Language Models." ICLR, 2024. A coding model generates and revises rewards used to train separate robot policies. Humans supply the simulator, task description, and evaluation metric. ↩︎ a b
[44] Mengkang Hu et al. "AgentGen: Enhancing Planning Abilities for Large Language Model based Agent via Environment and Task Generation." KDD, 2025. Generated planning environments supply training trajectories. Tables 1 and 3 report improvements on separate evaluation tasks, including tasks outside the training formalism. ↩︎ a b
[45] Chris Paik. "The End of Software." X, 2024. Argues that falling creation costs will proliferate software and disrupt traditional software businesses. The claim concerns software economics, not the exhaustion of software work. ↩︎
[46] Abhay Venkatesh. "The Amplification of Software." 2026. Argues that as implementation costs fall, small differences in judgment, taste, speed, and distribution compound, so software concentrates rather than commoditizes. Source of the illustrative figures in Table 4. ↩︎ a b
[47] bokonist. "Software is generally a winner-take-all market." Hacker News comment, November 13, 2012. Argues that with zero marginal cost the leading product earns the most, reinvests the most, and pulls further ahead, so a product half as good can lose all of its revenue, not half. An informal argument, not data. ↩︎
[48] Andrew Ross Sorkin (@andrewrsorkin). "Report on Instinct’s funding and team size." X / DealBook, September 28, 2026. Reports a $10 billion valuation and 14 employees. An example of a highly valued software company with a small team; valuation is distinct from revenue. ↩︎
[49] CompaniesMarketCap. "Largest Companies by Market Capitalization." Accessed October 1, 2026. The five most valuable companies were Nvidia, Apple, Alphabet, Microsoft, and Amazon; eight of the top ten were computing companies. ↩︎
[50] Abhay Venkatesh. "Startup Principles." 2024. Principle 17 holds that the best startups are software or software-like; its third example, citing Thiel, is that there have been no non-tech successes in the West in the past 50 years, with a table of the largest American tech companies founded since the 1973 oil shock. ↩︎ a b
[51] Peter Thiel. "The Tech Curse." Keynote, National Conservatism Conference (NatCon 3), Miami, September 11, 2022. Likens the dominance of big tech in California to a resource curse; the source for the claim in [50] that tech has been the West’s only major source of new successes for decades. ↩︎
[52] Matt Stopera. "People Can’t Stop Saying “Everything’s Computer” Because Of Donald Trump." BuzzFeed, March 2025. Quotes Trump at the White House Tesla showcase on March 11, 2025. Cited for the phrase, not as evidence. ↩︎
[53] W. Stanley Jevons. "The Coal Question." Macmillan, 1865. Observes that more efficient use of coal increased total coal consumption, the origin of what is now called the Jevons paradox. ↩︎
[54] bouvard (@bouvard38829538). "The ultimate substrate of machine intelligence is tacit knowing, which is inexhaustible and replenished by every attempt to map it." X, September 22, 2026. A one-line claim, cited for the idea rather than as evidence. ↩︎
[55] Dennis Bouvard. "Infra-Humaning." 2024. Describes how attempts to define and automate human characteristics reveal further distinctions and create new human activities. A philosophical basis for an open-ended human–machine relationship, rather than empirical proof that it continues forever. ↩︎
[56] David Owen. "Can AI automate AI R&D yet? Early evidence from InnovationEval." Epoch AI, October 7, 2026. Frontier agents given 3,000 GPU-hours to beat a tuned GRPO baseline recovered 2% (Claude Fable 5) and 15% (GPT-5.6 Sol) of a published human method’s gain after correcting for run selection and out-of-scope changes; the human method, rerun in the same setting, recovered 94%. One run per model; the author calls it early evidence. ↩︎
[57] Fabrizio Dell’Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R. Lakhani. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality." Harvard Business School Working Paper 24-013, 2023. Introduces the “jagged technological frontier”: tasks of similar apparent difficulty fall on either side of what AI can do. ↩︎
[58] Helen Toner. "Taking Jaggedness Seriously." November 24, 2025. Argues capabilities will likely stay uneven, and notes that plenty of people treat jaggedness as temporary. ↩︎
[59] Gene M. Amdahl. "Validity of the Single Processor Approach to Achieving Large Scale Computing Capabilities." AFIPS Spring Joint Computer Conference, 1967. The overall speedup from accelerating part of a process is limited by the part that is not accelerated. ↩︎
[60] Leopold Aschenbrenner. "Situational Awareness: The Decade Ahead." June 2024. Extrapolates compute and algorithmic trends to automated AI researchers and a rapid intelligence explosion. ↩︎
[61] Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean. "AI 2027." AI Futures Project, April 2025. A dated scenario in which compounding AI research speedups lead from a superhuman coder to superintelligence within 2027. ↩︎
[62] Stanford Institute for Human-Centered Artificial Intelligence. "The 2026 AI Index Report." April 2026. Reports that the same top-tier model that outperforms human baselines on PhD-level science questions reads analog clocks correctly only 50.1% of the time. ↩︎