Compounding AI Progress: From RSI Acceleration to Continuous Learning
The Compound Interest of Intelligent Progress: From RSI Acceleration to Continuous Learning
Content Summary
Author: Cage, Penny / 01. Key Takeaways • The RSI paradigm is already emerging, model self‑training and self‑iteration are being embodied in Silicon Valley frontier labs
Author:Cage、Penny
| / |
01.
Key Takeaways
• Our definition of RSI is: AI can participate in every stage of model training. As compute increases, the number of researchers needed to achieve the same level of intelligence progress no longer grows proportionally, but begins to decline.
Coding progress hasn’t slowed; synthetic data and infrastructure optimization have greatly benefited from RSI. Researchers no longer need to understand code; their work focuses more on proposing hypotheses, understanding requirements, and finding ways to validate the results generated by models. After returning to OpenAI, Lilian Weng has also devoted a lot of effort to RSI.
•RSI makes the process of intelligent production more automated and closes the loop; continuous learning enables AI to acquire capabilities through interaction, learning from real‑world, fragmented tasks that often lack a standard answer.
•Our definition of continual learning is: a system that, through interaction experience, can transform new knowledge and entropy into an ability improvement that generalizes across various tasks.
•Coding agents make generating tasks and simulation environments increasingly easy, but determining task effectiveness is becoming a more critical constraint. An issue can only truly close the loop when AI possesses a good sense for verification and falsification, and when appropriate verify and falsify methods are designed for the task.
•RSI will make AI’s progress fully subject to the physical constraints of the compute bottleneck; continual learning will blur the division between training and inference, and a continuously improving AI will be even more highly personalized and user‑sticky than current APIs.
, and with the combined effect of computing power and model iteration speed, Frontier Labs’ models will advance faster. RSI may be the flywheel that ultimately creates Frontier Labs’ moat.
Personalization × Self-evolution Matrix
When we discuss RSI and continuous learning, there are two core dimensions that can be used to differentiate these definitions:
•Adaptation: Can AI perform personalized learning around a specific task or environment?
•Self-evolution: Can AI close the loop and train a better AI itself?
Within this thinking framework, the traditional LLM paradigm—foundation model generational upgrades and regular releases—are placed in the lower left corner. Meanwhile, the current frontier labs training approach is becoming increasingly automated and industrialized, moving from the lower left toward the upper left.
From a personalization perspective, current chatbots/personal agents store users’ personalized memories as memory files—structured file systems like markdown—and retrieve them during relevant conversations, much like a person carrying a notebook.
Continuous learning is the intersection of the two dimensions in the upper right corner, and it has continuous progress through self-evolve and personalized adaptation capabilities.
Company Mapping: From Matrix to Industry Entry Points
The core content of the rest of this article focuses on RSI’s current progress and next steps, as well as the core technical roadmap that needs to be achieved for ongoing learning. Therefore, it will not involve much judgment on specific companies, but we have indeed observed many Neolab appearing in both directions.
As a large number of startups are emerging in this field, we would like to first share the categorization and mapping approach for these startups. In general, they fall into three major categories:
1.RSI: AI research automation
This is a direct manifestation of the RSI vision. It can be further divided into two categories:
(a) Training platform: Enterprises can train personalized models on their platform via APIs or RL fine‑tuning, etc. Therefore, startups in this category build RSI as a platform, making model training easier for enterprises. Representative companies include Thinking Machine, which offers the Tinker platform.
(b) Focused on research on RSI itself: the goal is to achieve closed‑loop tasks in AI and other fields. Representative companies include Jeff Dean’s startup Discover Loop.
2.
In this category, we believe frontier model companies still have a lot of room to act. For example, OpenAI and Anthropic have likely already partially implemented RSI internally, and improving continual learning and algorithmic efficiency is probably their next biggest concern.
In addition, many visionary researchers are launching startups to explore learning and new paradigm directions. We think these explorations have higher odds, but objectively their success rate is lower:
Exploring the next‑generation paradigm is more like a research bet, aiming to solve a scientific problem. The uncertainty is much higher. It is even possible that under the current paradigm, a frontier lab could solve the continual learning problem using RSI.
3.
The closer an experiment is to the real world, the more expensive, slower, and harder to replicate it tends to be. Therefore, whether RSI can generate compounding returns ultimately depends on whether the system can learn more from each round of feedback. Below are the relatively more AI‑native teams in each vertical domain:
•In the science domain, the main representative directions are science discovery agents and automated‑lab companies.
•In the chip design and material discovery domains, some teams focus on directly building end products, while others aim to accumulate data and RL environments during the process to address real‑world bottleneck issues.
For these sub‑fields, we will later publish research mappings for each area and share them with you.
02.
State of RSI
Basic RSI has been achieved: experimental automation
Basic self‑improvement is getting closer to automation of model training. After a significant boost in coding ability at the end of last year, the pace accelerated, replacing existing pre‑training, post‑training, and RL entirely with AI. Now in Frontier Labs, an increasing number of tasks can be completed in a closed loop by agents. A researcher can advance multiple problems simultaneously, devoting more energy to choosing directions, designing validations, and judging results.
RSI does not require zero human involvement; a more practical definition is that the number of researchers needed to deliver a unit of intelligence does not have to scale with compute. Model improvements scale compute and intelligence, but not people.
This is reflected in the internal data released by OpenAI in September 2026:The experimental volume per active researcher reached about 1.6 times the 2025 average in August.Considering that OpenAI’s employee growth is likely faster than its compute growth, researchers are assigning more complex, longer‑term tasks to Agents.
From a compute perspective, this change involves two distinct quantities:how many GPU hours each researcher can manage, and how many effective results each GPU hour can produce.Automation certainly improves the former, allowing more hypotheses to enter experiments; but to improve the latter, evaluation and research methods need further refinement. True RSI ultimately turns larger experimental scale into stronger research capability, yielding a compound effect where the model’s iteration speed increases exponentially.
Karpathy’s autoresearch demonstrates a concrete closed‑loop approach: an agent modifies the training code, runs within a single‑GPU, five‑minute training budget, and uses a fixed evaluation to decide whether to keep or revert the change before moving to the next round. Humans define the research instructions and boundaries, while the agent executes continuously. The five minutes refer to training time only, excluding startup, compilation, and agent inference.
AI agent reads the current training code and experiment logs → modifies the training code → conducts experiments and evaluations within a fixed compute budget → keep or revert changes → next round of experiments.
Infrastructure is changing as a result: the training system must not only be convenient for researchers but also enable the Agent to operate, observe, and debug. Available training APIs, reproducible environments, and independent evaluation determine how many such closed loops a researcher can manage simultaneously.
Research Bottlenecks
After implementing the basic RSI, the most scarce tasks become these:
•Can optimizers, data ratios, or architectural changes that work on small models remain effective in frontier training? The transfer and generalization from simple tasks to complex tasks is currently the biggest bottleneck in automated experimentation. Anthropic’s automated alignment research has already begun to automatically design and iterate, but the optimal method did not achieve statistically significant improvement when moved to Claude Sonnet 4 production training. Candidates screened by cheap experiments still require a certain degree of human‑in‑the‑loop.
• It requires a good environment and a verifier to provide the model with tasks for repeated attempts and learning. Scientifically, scalable oversight aims to solve: when tasks are so numerous that humans cannot check each one, and so difficult that supervisors cannot directly judge, how to still provide reliable feedback. In engineering, building a verifier also needs to balance quality and efficiency. A practical approach is to let the agent generate correct variants and erroneous distractors, and check whether the verifier can distinguish them.
•Some open‑source environments and benchmarks are built around fixing a single issue as a one‑off task. In reality, work consists of a continuous sequence of tasks, with files, runtime environments, and user goals constantly changing. Reconstructing environments from real trajectories makes training more aligned with these work sequences and states. There are three categories of challenges that need to be addressed:
1.Files and runtime environment. Recover file contents from the trace as much as possible, stitch together the file tree, and let the agent fill in and fix missing parts; also restore language, dependencies, and build configuration so the environment can actually run.
2.Task boundaries. Identify the task scope from vague user queries, e.g., restore a request like “continue” into a task with a clear goal and complete context.
3.Task verification. The original reply is often not the optimal answer. The verifier should determine whether the task is completed based on model behavior and actual state changes, avoiding treating reproducing the original reply as success.
•Which problems are worth pursuing, which experiments should receive more compute, and when should we give up? These decisions lack ready-made short‑term metrics, yet they determine whether an experiment ultimately yields valuable findings. Anthropic’s observations of internal research also note that the more automation is applied, the more important research direction and the human‑in‑the‑loop become.
The next progress of RSI will come from improving these aspects together: experiments can advance autonomously, results can be scalable verified, then move to the next round of experiments and deployment, achieving a true close the loop. However, there will also be many areas closely related to the real world where most domains may find it difficult to have a good verifiable environment.
RSI in vertical domains
Besides model training and coding, RSI finds it hard to directly close the loop. Therefore we favor a gradual sequence: first form a closed loop within the frontier lab; then move into vertical domains with clearer goals and more verifiable feedback. AI for Science and robotics have already started looking for some local closed loops, which also makes the earlier discussion about verifiable environments more concrete.
Anthropic’s recent protein design work leverages a specialized model similar to AlphaFold, automatically submits experiments, and analyzes the results, stitching them together into a mature workflow.
The next step is for the agent to improve its own research tools. Anthropic had Claude optimize the inference code and GPU kernels of more than 30 open‑source biomolecular models, and combined this with a more concise prompt, harness, and the new model version. They note that for protein design of comparable quality, the GPU hours used are now 1/100 of those required by the pre‑optimization open‑source models.
Phylo’s release of Biomni × TusoAI also demonstrates the automation of scientific model development. The agent searches for data and code, retrieves literature, rewrites methods, trains and evaluates candidates. It tested 500 candidates within a week, producing a gene perturbation prediction approach that combines multi‑source gene embeddings, regression, and k‑NN.
As computational steps become cheaper, the bottleneck will shift more toward the quality and speed of real experimental feedback. The next vertical RSI must enable the system to extract reliable signals from noisy results, establish new assays, and use them to improve the selection of experiments, tools, and models in the next round. This requires further development of scientific taste and experimental automation.
Models are no longer just interfacing with robots; they are also improving robot policies. In the second phase of Anthropic’s Project Fetch, Opus 4.7 completed five integration and programming tasks in an average of 12 minutes 7 seconds, whereas the previous round with a human team assisted by Claude took 264 minutes. This comparison shows that coding capabilities are significantly compressing the software work in robot development.
NVIDIA, Stanford, and UT Austin’s RoboTTT focuses on optimizing robots’ ability to learn while performing tasks: integrating fast weights into the VLA policy, compressing visual and action histories, and absorbing freshly acquired experience during execution. The paper demonstrates 8K‑step context, single‑shot human video demonstration adaptation, and online error correction in long‑horizon assembly, and observes that performance continuously improves as the training context length increases. This makes context length a new scaling axis for robot policies.
03.
Continual Learning: Definition and Key Issues
The goal of continual learning is closer to the essence of learning, aiming to quickly translate experience into capability improvement. Therefore it can not only perform auto research & training, but also do things that no one is doing now. For example, to address issues such as insufficient data efficiency and inadequate memory and forgetting mechanisms, a more fundamental continual learning is needed.
The core mechanism of continual learning consists of two steps:
•Step one, how to obtain new entropy. Acquire new problems, new ideas, and new tools within the system and environment. Entropy may come from users and the environment, or it may be generated by the model itself.
•Step two: how to turn entropy into learning, making new data become part of one’s capability.
The key to this mechanism is that when AI receives a new experience, it can immediately convert it into knowledge, achieving continual learning. If you have to collect all experiences first and train them later in batch, that’s still the existing paradigm.
With this capability, we can further tier the scope of continual learning. Implementing the first tier is currently the most important goal: enabling a task to continuously learn signals in a long context window and improve over time, which is difficult to achieve with the current paradigm.
Core Issues of the Continual Learning Paradigm
Key Challenge: Sample Efficiency
Software testing can usually reset environments and rerun in parallel, so rewards are dense and obtained quickly. This kind of environment can be described as grindable—it can be broken down finely and yield rewards, even process rewards, in a low‑cost, parallel, resettable manner.However, work in the physical world cannot be made similarly grindable and verifiable, so we must learn the essential experience from far fewer samples.
Therefore sample efficiency becomes the most critical challenge in continual learning: the harder the environment is to replicate, the more we need to learn from each real experience.
04.
Technical Approaches to Continual Learning
Among them, TTT and LoRA mainly address memory mechanisms and personalization issues, while meta‑learning primarily tackles learning mechanisms and sample efficiency problems.
TTT / fast weights
Test-time training brings three important changes: first, the model has fast weights that can be updated online; second, learning and inference can occur simultaneously; third, memory can maintain O（1） computational complexity relative to context length.
TTT designs the hidden states in the last few layers of the model to be learnable, so the model is no longer composed of completely frozen weights. Fast weights can compress a long sequence of task information, but the system must decide which content is worth retaining.
Work from the end of 2025 to 2026 will begin to try integrating TTT into standard Transformers; test-time learning could become a new architecture, or gradually turn into a plug‑in capability for existing models.
There are multiple works exploring this direction from different angles:
•TTT‑E2E performs test-time updates on sliding‑window Transformers and learns suitable initialization for updates during training, enabling existing architectures to better learn from long contexts.
•In-Place TTT uses the existing MLP’s output projection matrix as fast weights, introducing online updates within the original Transformer.
•REFINE trains fast weights with next-sequence prediction and sequence-level rewards, improving the model’s retention of long-range information.
•Falcon designs fast-weight attention from an online learning perspective, exploring normalized delta rules and sliding‑window updates to improve the write efficiency of fixed‑size associative states.
TTT has not been scaled up yet; the next scaling bottleneck will likely be data. Training sequences need to cover various update frequencies and leverage long trajectories and downstream results so the model learns which experiences are worth writing, how to control interference and forgetting, and which states should be retained across sessions.
Adapters / Multi LoRA
LoRA has become one of the most common methods for online adaptation. This approach freezes the base model and trains only a small set of low‑rank parameters. More general modular updates also include adapters, sparse memory layers, and so on.
Thinking Machines’ Tinker, Trajectory, River, and Fireworks have all applied LoRA / Multi-LoRA for training or serving. LoRA can serve as a persistently stored local state: the experience of users, projects, or domains remains in their respective adapters, which can then be selected or combined per task.
These methods are compatible with existing Transformer infrastructure, requiring no duplication or updating of all weights, and they also facilitate separate training and deployment.
The core value of the modular approach is to control the stability–plasticity tradeoff: confine local learning to a routable, removable parameter space, reducing the risk that a single update will damage the entire base model. One possible hierarchy is: fast weights / task‑local LoRA preserve short‑term cross‑session state, slower adapters solidify stable skills, and base weights only absorb general capabilities that are repeatedly validated across users.
This type mainly addresses personalization and cost, but it does not directly provide learning rules, so it is likely to coexist with other methods in the future. The core value of the modular approach is to control the stability–plasticity tradeoff: confine local learning to a routable, removable parameter space, reducing the risk that a single update will damage the entire base model. One possible hierarchy is:
•fast weights / task-local LoRA preserve short-term cross-session state
•slower adapter solidifies stable skill
•base weights only absorb universal capabilities that are repeatedly validated across users.
The core challenge of Multi-LoRA is to scale up the number of adapters that can be trained and used. The training and updating of each adapter must be lightweight and efficient, and the system must handle concurrent training, routing, and serving, so that many users and projects can continuously accumulate proprietary states.
Meta-learning
The recent leak from SSI reported a fairly clear breakthrough in this direction. Meta-learning turns how a model “learns to learn” into a training target: which initializations are easier to adapt and generalize, which layers should be updated, how much to update, and what information is worth retaining. It directly targets learning efficiency and memory management.
Nested Learning / Hope treats the model as multiple nested optimization loops with different update frequencies: fast modules adapt to the current context, while slow modules retain more stable knowledge. TTT‑E2E also adopts a similar idea, meta‑learning the model during pre‑training so that it can be more easily updated via next‑token prediction during deployment.
Starting in 2026, research such as SCoL enables models to generate their own update instructions based on the current context, deciding which Transformer layers should be updated, and then training this learning strategy with meta‑RL on a continuous experience stream.
Meta‑learning actually corresponds to a problem that TTT has not yet solved: not every fast‑weight update should be kept permanently; the system needs a learned gate to decide which experiences are worth storing in long‑term memory, i.e., the long‑term weights. In an early interview, Ilya Sutskever already pointed this capability toward an internal value/steering system: humans don’t need to redesign a verifier for each new task, yet can still judge whether an attempt is valuable.
The most valuable goal of this direction is to enable the system to make appropriate updates on unfamiliar tasks with only a small amount of experience. Many researchers and scientists working on high‑value tasks have observed that current models behave more like they have learned a fixed method/skill on a familiar distribution.
Self-distillation and On-Policy Learning
Recently, discussions about OPSD (on-policy self-distillation) have increased. This is a promising class of algorithms for the CL learning component. The core idea of this method is how to distill the essential information from a task session back into the model to achieve capability improvement.
Traditional distillation keeps a frozen teacher checkpoint so that the student model imitates the teacher, reducing the loss of previously acquired abilities. This approach is still commonly used to let a small model acquire the capabilities and behavior of a large model, but fundamentally it is learning a stronger capability, which can also be very useful for continual learning.
The key to continual learning is the distance between new training signals and the model’s current policy. On‑policy distillation attempts to achieve online‑RL‑like stability without relying on an explicit reward function.
Under the OPSD approach, distill the “veteran” model that has completed a full session back into the base student. The token distribution difference between teacher and student provides a denser signal than the final reward; trajectory‑refined distillation continues to guide the student from the states after it makes mistakes, reducing the distribution shift caused by teacher forcing.
Is innovation in model architecture necessary?
Some top researchers believe that continual learning does not necessarily require additional architectural innovation: developing a brand‑new architecture from design to scaling takes years, and intelligent growth is more likely to come first from long‑term accumulable optimization objectives and feedback loops.
Dario also discussed this in a podcast: in externalizable domains such as coding, existing Transformers, long context, RL, and tool systems may already be sufficient to capture a large amount of economic value, so CL is not a prerequisite. This can also be seen from their work on ecosystems like skills.
Another line of researchers argues that Transformers are not suitable for true continual learning:
•The attention mechanism in Transformers performs retrieval essentially as shallow pattern matching, finding similar previous questions by keyword matching and copying the answer—very superficial.
•A deeper learning approach would be to compress the previous 100 math problems, extract the core mathematical logic, and then use that logic to solve new problems. This is why pre‑training works—compression extracts shared logic and turns it into parameters.
Sutton’s version is more fundamental: current LLMs learn imitation from human data, rather than learning a transition model from a continuous sensation–action–reward stream, and existing gradient descent lacks a mechanism that guarantees good generalization.
Dwarkesh’s view lies in between: he acknowledges that existing models lack high‑throughput experience learning, but does not think imitation learning and experience learning have to be mutually exclusive. LLMs can serve as a strong prior, and then gradually approach the CL goal through other mechanisms such as self‑distillation.
RNN: stores history with a compact state. The advantage of approaches like RWKV, xLSTM, etc., is that they don’t need to retain a continuously growing full history during long‑sequence inference. The drawback is that a limited state loses details, making precise recall of large amounts of information harder. To replace attention at scale, they still need to demonstrate comprehensive ability on complex tasks and show that scaling up the model yields stronger agentic capabilities.
SSM: makes state updates more efficient. Mamba uses a structured state and an input‑dependent selection mechanism to control writing and forgetting of information, which is well‑suited for efficiently handling long sequences. However, there is still a trade‑off between compressed memory and precise retrieval. Specialized operators can improve efficiency, but truly replacing mature Transformer systems will require demonstrating overall benefits in large‑scale training.
Hybrid architecture: closer to attention, retaining its advantages. For example, Qwen3-Next combines Gated DeltaNet with attention, using recursive memory to reduce cost while preserving access to details. The trade‑off is that there is still an attention cache, and the system is more complex.
Therefore, the more reasonable approach currently is three hypotheses ranging from weak to strong:
1.Current paradigm: with longer context and stronger intelligence, ready for continual learning data, sufficient to cover most needs;
2.Learning-rule hypothesis: Transformers remain, but algorithmic innovations such as meta-learned fast/slow updates, distillation, and memory gates are needed to achieve continual learning
3.Architecture hypothesis: the model architecture and optimization dynamics must be designed from the ground up for sustainable updates.
Overall, RSI has already been improving the efficiency of intelligent R&D, and continual learning is expected to make real-world usage a source of capability growth. These two flywheels reinforce each other: automated research helps us discover better learning methods, and models that learn more effectively accelerate the next round of research.
The next breakthrough depends on how the model can learn genuine capability improvements from limited experience, and whether this can be achieved with the current technical approach. This is also what we consider the true long‑term advantage flywheel of the model: whoever can continuously improve the efficiency of converting experience into capability will keep the growth slope of intelligence steep.
Further Reading
Get MirrorSpeed