刘知远 and 肖朝军: Density Law, RL Scaling, and Distributed AI
Summary
- 刘知远 and 肖朝军’s core view is that the visible frontier for large models remains capability leaps, while the hidden frontier has shifted to how much intelligence each unit of parameter count, compute, and energy can buy. The team measures base-model capability density as doubling roughly every 3.5 months, with the slope accelerating further after ChatGPT emerged; they want the Density Law (the paper is titled Dancing Law of LLMs) to become, like Moore’s Law, a shared industry target that drives its own realization rather than a rival to Scaling Law. “Stronger capability is the visible frontier; higher efficiency is the hidden one.”
- Higher density is not a model-compression trick but a precision system spanning architecture, data, learning algorithms, and software-hardware co-design. FFN is moving toward fine-grained MoE and Attention toward sparsity; on the data side, Ultra FunWeb trains a stronger model on less than one-tenth of the original data, implying that training cost could theoretically fall to roughly one-tenth. The team calls the whole system “the lithography machine of large models”; the real moat is raising the “intelligence conversion rate” across every layer.
- RL still lacks a Scaling Law comparable to pretraining, which could become a major technical fork over the next year. Pretraining has scaled to tens of trillions of tokens and hundreds of thousands of steps, while current RL often runs only a few thousand steps and the training disclosed for R1 was under 10,000 steps; outside math and code, reliable, unhackable rewards and environments are hard to construct. One path is to keep scaling across enough environments and wait for generalization; the other is to find a more sample-efficient learning method. 曼祺 asked, “If I can model the world well enough, why isn’t that reward inherently an AGI?” 肖朝军 agreed.
- Agents and deep thinking are shifting Attention’s core bottleneck from long inputs to long outputs, while the industry still lacks a consensus on how to handle it. InfLLM V2 implements native sparsity during training: in a 128K context, it attends to only about 4K–6K tokens, or less than 5%, while supporting long outputs; 肖朝军 expects linear attention eventually to combine with sparse attention, but forgetting, long-horizon planning, and multi-step Agent stability remain unresolved. “The one thing you must not do is finish step two and forget step one.”
- On-device density is creating a more defensible commercial path than cloud APIs, with automotive as the first market to scale. 面壁 models entered Mazda and Geely vehicles this year and are expected to cover more than six models next year; extrapolating linearly from chip roadmaps, on-device systems could deploy models with more than 60B total parameters and over 8B active parameters by 2030. With density gains layered on top, GPT-4-to-GPT-5-level capability will definitely reach the edge within five years. Once the chip, power envelope, and response time are fixed, model density directly sets the ceiling for the cockpit.
- 面壁 abandoned its 2023 catch-up plan—costing tens of millions of yuan and targeting roughly 140B parameters—because a commoditized cloud API could not support a startup’s commercial loop. That decision produced the 2.4B MiniCPM in January 2024, whose performance matched Llama 2 13B and Mixture 7B at the time and which could run on a phone; the API price war that followed in May 2024 confirmed its concerns about cloud moats. 刘知远’s startup principle is not to seek “one-shot breakout,” but to find “certainty inside uncertainty.”
- The longer-term bet is distributed intelligence: individuals own continuously learning on-device models, while the cloud runs domain experts, together forming an “internet of agents.” The team expects that 2027 may bring nodes capable of running relatively large-scale RL on-device; glasses, earbuds, and watches would simply collect data, while a “portable NAS that can follow you” carries the personal intelligence. AGI may not perform every task itself; it may more likely be “AI that produces AI.” 刘知远’s real concern is not excess productivity, but that “we humans will bind our own progress.”
Deep dive
1. The common direction of the latest model releases is not chat, but Agentization
Looking at GPT-5.1, Grok 4.1, Gemini 3, and Cloud Ops 4.5, 肖朝军 sees a first common thread: a sharp improvement in agentic tasks. After a new model launches, people now test “how many bugs it can help me solve,” not its single-turn QA score.
The second signal comes from Nano Banana, NotebookLM, and GPT-4o: text inside generated images is becoming accurate. Adding more multimodal data to training had not produced another step-change in intelligence; unified autoregressive image-text generation may offer a new scaling direction.
This remains a conditional judgment. 肖朝军 stressed that outsiders do not know how Gemini 3 is implemented and cannot establish whether the gains come from a unified model or product optimization. His observation window is to wait “another one to two months,” not draw a conclusion from demos.
2. Beyond six years of capability leaps, a low-visibility efficiency curve is taking shape
刘知远 divides the main capability arc since 2018 into several leaps: the emergence of pretrained models; ChatGPT learning to follow instructions through instruction tuning in late 2022; and large-scale RL from late 2024 to early 2025 teaching models to think deeply.
Stronger capabilities are easier for users to see, but making models larger also drives training and serving costs up linearly or even superlinearly. Intelligence available only through an expensive API cannot reach enough devices and industries, limiting the diffusion of this technology wave.
肖朝军 added that MiniMax, Qwen Next, DeepSeek V3.2, and 面壁 are pushing efficient architectures on the open-source side, while OpenAI’s mini series and Gemini Diffusion are among the closed-source efforts. “Capability improvement is the visible frontier; higher efficiency is the hidden one.”
3. The Density Law is first a correction to “compute is everything”
曼祺 suggested that efficiency should be self-evident. 刘知远’s response was that the mainstream narrative from 2022 to 2024 was not efficiency, but that “more data, more compute, and more parameters” were required for stronger models—eventually morphing into the idea that only institutions with 100,000 GPUs had the right to train large models.
That narrative shapes researchers, industry, and policymakers, causing Scaling Law to be misunderstood as a first principle requiring 100,000 GPUs up front. In 刘知远’s view, one meaning of the Density Law is to establish a metric for intelligence output per unit of resources.
DeepSeek V3 is the counterexample the team cites: its absolute training resources were still large, but well below the threshold some international institutions had previously implied. The point is not that compute is irrelevant, but that capability growth does not have to follow only a brute-force expansion path.
4. Like Moore’s Law, the Density Law is self-fulfilling through competition
The team settled on “Density Law” rather than “density theorem” because “law” in English does not necessarily mean a law of physics. 刘知远 recalled that Moore was merely summarizing an empirical pattern in 1965; more than a decade later, after the trend continued, the industry gradually named it and turned it into a common roadmap.
The Density Law likewise requires technical innovation to keep delivering. 刘知远 describes raising density as “moving against the world’s entropy increase”: packing more content into less space is not a natural process, but the result of progress across architecture, data, algorithms, and engineering.
肖朝军 brought in biological evolution and resource constraints: the human brain did not grow without limit, but became more intelligent within finite volume and energy budgets. Companies face a similar “survival of the fittest”—at equal capability, a more expensive model will eventually lose competitiveness.
The two repeatedly clarified that the Density Law and Scaling Law are complementary. Scaling Law describes the relationship between compute and capability; the Density Law asks whether the slope of that relationship can keep rising. The rapid decline in GPT API prices is itself evidence of internal density gains.
5. Capability density doubles roughly every 3.5 months, but the number is not the research endpoint
The paper initially included models through 2024 and found doubling roughly every 3.3 months; adding 2025 models changed that to roughly 3.5 months. The team treats the difference as normal noise. The core observation is still exponential doubling of capability density over time.
Around January 2023 is a notable inflection point: after ChatGPT appeared, the slope from compute input to capability output became steeper, suggesting that industry competition, funding, and research attention jointly accelerated efficiency innovation.
肖朝军 argues that whether density doubles every 3.5 months or once a year is not the most important conclusion. The flag worth planting is to stop asking only how high performance is and start asking “how much intelligence can be converted per parameter or per unit of cost.”
6. The Density path began with an unrecoverable 140B decision
In 2023, the main task for Chinese teams was to reproduce ChatGPT. 刘知远 estimated that top teams would complete instruction tuning around September or October, then naturally follow the path set by OpenAI’s April 2023 GPT-4 release by scaling parameters and data further.
面壁 initially planned to catch up as well, but its calculation called for roughly 140B parameters and “tens of millions of yuan” in training costs. The team expected at least five Chinese teams to produce comparable capability in the first half of 2024, yet could not find a clear answer for how it would recoup the training bill.
It therefore turned to building a “model wind tunnel,” using a large number of small experiments to break down efficiency across each component. The result was the first MiniCPM, released in January 2024: with 2.4B parameters, it reached the model performance of Llama 2 13B and Mixture 7B at the time and could run on a phone.
That result opened two paths. The visible one was that large on-device models became feasible for the first time; the hidden one was the question of what laws governed the knowledge or capability density of models from different eras, eventually leading to the Density Law research.
7. Existing density metrics cover only base models; post-training needs another curve
曼祺 asked why the large-scale post-training in o1 and R1 had not produced a slope change similar to ChatGPT. 肖朝军 explained that the first version of the capability-density metric was designed for base models, so gains from RL were not directly included.
Pretraining maps to traditional Scaling Law, while o1 and R1 are closer to test-time scaling or inference-time scaling. They occur at different stages and cannot be forced onto the same metric curve.
The team is working on Density Law 2.0 to incorporate efficiency changes from post-training techniques. The relevant quantitative relationships remain under study, with no mature conclusion ready for external release.
8. Raising density requires four levers to turn together
肖朝军 summarizes the full pipeline in four layers: architecture determines the “container for intelligence,” data forms the model’s “textbook,” learning algorithms determine how it grows from the textbook and feedback, and infra makes training and serving genuinely compatible with hardware.
The third layer also includes the “model wind tunnel”: a large volume of small-model experiments builds experience and predicts parameter choices, data configurations, and capability ceilings before a large model begins training. OpenAI once called this predictable scaling. Its value is not merely finding a curve, but avoiding expensive trial and error on large models.
The team’s next step is to quantitatively decompose how much density gain comes from architecture, data governance, the model wind tunnel, and software-hardware co-design. It has empirical results, but no sufficiently clear and broadly reusable functional relationship among the four.
9. The consensus on FFN has shifted from dense matrices to fine-grained MoE
Transformer’s FFN, or feed-forward network, can be roughly understood as a dense matrix layer. MoE splits a large matrix into multiple experts and activates only a subset each time, using less compute to access a larger parameter capacity.
曼祺 asked why DeepSeek is said to have “revealed” MoE when teams such as Mistral adopted it earlier. 肖朝军 acknowledged that the Mixture of Experts concept dates back to the 1990s, and that Google and his team were already applying it to Transformers around 2020.
DeepSeek did not invent MoE. Its contribution was solving the scaling and engineering problems on large models, then proving through open source that the approach worked, turning a “non-consensus” path into an industry consensus. Since then, fine-grained sparse MoE has become the clear direction for FFN.
10. Agents are turning Attention into the next efficiency bottleneck
Agents must read complete code repositories, Deep Research must process huge volumes of external material, and deep thinking generates long chains of reasoning. Models therefore face long inputs and long outputs simultaneously, making Attention’s compute and memory costs a rapidly growing bottleneck.
Open- and closed-source teams are trying linear attention, sparse attention, sliding windows, and mixtures with dense layers. 肖朝军 noted that claims that models such as Gemini use sliding windows still contain an element of rumor, but the industry’s shift toward efficient long context is unmistakable.
This is what it means for architecture to be the “container for intelligence”: not merely cutting FLOPs, but allowing a model to hold more context and process longer task trajectories within the same compute budget.
11. Data governance runs from L0 to L4, refining raw material into a textbook
面壁’s internal data pipeline runs from L0 to L4. L0 is raw data collected or purchased; L1 deduplicates it and filters out junk; L2 selects high-quality content; L3 synthesizes or rewrites data that did not previously exist in the world.
L4 is called validation data. It must pass formal methods or human verification to reach “textbook grade.” The standard is no longer that the text reads smoothly, but that the data is sufficient to carry a training signal.
This hierarchy serves more than pretraining. The quality of data synthesis for SFT and RL likewise determines the ceiling and efficiency of post-training. All the intelligence a model ultimately learns comes from data, making data a direct source of capability rather than support engineering.
12. Ultra FunWeb shows how “less data” can become a direct cost advantage
The team refined FunWeb through its own pipeline into Ultra FunWeb, which contains less than one-tenth of the original data; models trained on it were actually more capable. On that basis, the team projects that training cost under equivalent conditions could fall to roughly one-tenth of the original.
The example turns the slogan that “data quality matters” into a verifiable efficiency result. 肖朝军 also cited a series of Less is More studies by 刘鹏飞 at Shanghai Jiao Tong University, which achieved stronger capabilities with less data.
The more fundamental question is: what is the minimum dataset needed to build a given level of capability? Answering it progressively would not only cut costs, but also move closer to the basic questions of what intelligence actually is and which experiences are indispensable.
The team plans to publish a data-governance report, processing architecture, and some of the data, continuing its open practices around UltraChat, UltraFeedback, and Ultra FunWeb. It also stresses that a large model is a precision system built by more than 100 people; DeepSeek V3 had 100+ contributors and is, by global standards, still a highly refined effort.
13. Software-hardware co-design determines whether theoretical compute becomes model capability
肖朝军 summarizes software-hardware co-design as “using compute to its fullest.” If a GPU is only 50% utilized, the other half still consumes energy without producing capability; even an elegant theoretical architecture is uneconomic under those conditions.
One reason Transformer won was that Attention could use GPUs more fully, allowing models to continue scaling. From GPT-1 to GPT-3, teams also had to solve large-cluster communication and parallelization. FlashAttention continued to redesign Attention computation around hardware characteristics.
The coordination runs both ways: designing block-sparse algorithms for existing GPUs on one side, and designing dedicated hardware such as smartphone NPUs for AI on the other. DeepSeek V3’s deep modifications to low-level assembly to support the model are a classic example of algorithm-hardware co-design.
14. RL remains meaningfully short of true scaling
Pretraining has already scaled to tens of trillions of tokens and hundreds of thousands of training steps. Current RL typically runs for only a few thousand steps, and the training disclosed for R1 was also under 10,000 steps. A small number of steps can produce strong results on an individual task, but does not show that RL has a stable Scaling Law.
The fields with the easiest verifiable rewards today remain math and code. Competitive coding environments are relatively simple; real software development involves third-party libraries, dependencies, system state, and long-term maintenance, making repeatably verifiable environments much harder to build.
肖朝军 cites a view Jason Wei often publishes: RL’s magic depends on an “unhackable environment.” If a model can exploit reward loopholes to score highly without learning the task, more compute will only amplify the gaming behavior.
15. RL has two paths: scale environments or change the learning method
The first path is to continue scaling across enough environments with enough diversity and see whether the model generalizes to new tasks as pretraining did. 肖朝军 does not reject that possibility: genuine generalization in the past came through pretraining.
The second is to improve sample efficiency, enabling a model to master new tasks from limited feedback and action as humans do. Even if today’s RL can reach “one-thousandth or one-ten-thousandth of human” performance on an individual task, it is nowhere near human efficiency at adapting quickly to unfamiliar tasks.
曼祺 proposed constructing an environment complex enough to have multiple objectives. 肖朝军 replied that a human can read negative feedback from a serious expression, while a machine must be explicitly assigned a reward of 1 or -1; compressing the entire world into a reward is itself extraordinarily difficult.
This creates a chicken-and-egg problem: “Suppose I can model the world well; I can provide enough feedback for anything you do. Why isn’t that reward inherently an AGI?” RL to B may still create commercial value, but it is not the same thing as general autonomous learning.
16. 面壁 is dissecting RL through entropy, compute utilization, and open-domain rewards
The team is studying how entropy changes during RL, which can be understood intuitively as token diversity when the model samples the same query multiple times. One analytical frame is that training gradually converts the diversity produced by exploration into accuracy, explaining how the model converges mechanistically.
On the infra side, the goal is to keep RL compute as fully utilized as possible and increase training speed. Only after reducing the cost of each step can training plausibly expand from a few thousand steps to tens of thousands and test whether scaling produces new generalization.
In open domains, the team is experimenting with questions such as what reward “writing a paper” should receive. Outside math and code, there is no mature consensus, making reward engineering and environment engineering new objects of exploration.
17. InfLLM V2 turns sparsity from an inference patch into a training-native capability
InfLLM V2 was open-sourced in June this year alongside the first MiniCPM-4; its technical report, training data, and base model were opened further in October. It did not first appear in October.
The first generation applied sparsity mainly at inference, at roughly 50%, and was better suited to prefill for long inputs. The second generation, inspired by DeepSeek, adopted native sparsity from the training stage, accelerating training as well and allowing the model to operate at a lower effective attend ratio.
In a 128K context, InfLLM V2 attends to only about 4K–6K tokens each time, or less than 5%. Its operator also supports long outputs, allowing the architecture to cover deep thinking and Agents rather than merely answer questions after ingesting a book.
18. DSA fixes NSA’s engineering trade-offs, but long output remains the main unresolved battleground
肖朝军 sees two weaknesses in DeepSeek NSA: its three Attention components directly triple the compute cost for short text, and it is not suited to the mainstream pattern of short-text pretraining followed by long-text post-training. DeepSeek V3.2’s DSA abandoned those unfavorable designs and reoptimized for post-training.
On DSA’s finer-grained choices, 肖朝军 explained that DeepSeek’s architecture no longer needs to solve memory-access and storage-bandwidth problems and can rely more on pure computation. That condition does not apply to most other models and deployment settings, so the design cannot be copied directly.
On the relationship between linear and sparse attention, he offered a clarification: linear attention is closer to a replacement for sliding windows, while sparse attention corresponds to full attention. KDA and MiniMax have already explored mixtures of linear and dense attention; the natural next step is to combine linear and sparse attention directly.
What remains genuinely unverified is long output. MiniMax M1 once mixed Full Attention and Lightning Attention at roughly 1:7, while M2 returned to Full Attention. 肖朝军 suspects the performance loss may have surfaced after the focus shifted to Agents, but explicitly framed this as an estimate, not inside information.
19. Long output requires the model to remember the original plan in later steps
The industry habitually equates long text with “feed the model a book and ask it a QA question,” but 肖朝军 has considered that incomplete since the beginning of the year. Chain-of-Thought and Agent planning lengthen the output itself, requiring the model to retain the state from step one while executing steps three and four.
Linear attention tends to forget early content. Even when Full Attention is mixed in, the result depends on the blend ratio; cutting dense layers too aggressively may sacrifice long-horizon planning ability. 肖朝军 believes the problem will ultimately be solved, but “it will take a little more time.”
Long-context architectures therefore cannot be evaluated only by input length and retrieval questions. They must also be judged on memory, planning consistency, and repeated execution across long outputs. This standard is not yet an industry consensus, precisely the gap he believes needs urgent attention.
20. The Density Law is unifying internal R&D and beginning to shape embodied-intelligence roadmaps
Internally at 面壁, the Density Law brings architecture, data, the model wind tunnel, and software-hardware co-design under a common evaluation standard. 刘知远 calls the complex system built by a 100-person team “the lithography machine of large models.” Its goal is not to make models ever larger, but to increase density with every iteration.
Externally, after DeepSeek V3, Sam Altman, Anthropic, and Mary Meeker’s reports have all placed more emphasis on AI costs. 刘知远 sees this as “moving with the times”: as models move from demos into industries, inference cost and deployment boundaries become explicit constraints.
Embodied-intelligence teams cite the research especially often because robots are sensitive to latency, power consumption, and cloud connectivity. An earlier MiniCPM-V paper in Nature Communications connected Moore’s Law, the Density Law, and on-device deployment in a diagram, and is viewed as a basis for forecasting when cloud capability can fit into a device.
21. By 2030, edge devices may hold 60B total parameters and 8B active parameters
Based on roadmaps from leading chip companies including Nvidia and Huawei, the team estimates how on-device compute and memory access will change from this year through 2030. A linear extrapolation implies that devices in 2030 could deploy models with more than 60B total parameters and at least 8B active parameters.
Adding the Density Law, the team judges that GPT-4-to-GPT-5-level capability will definitely reach the edge within five years. 刘知远 preserves a condition: this is a linear forecast, and nonlinear breakthroughs could emerge over the next several years.
The current rollout order is broadly automotive, PC, then smartphone. Cars have more room and relatively looser power constraints, so they will scale first. Phones remain constrained, and the current performance of Apple Intelligence is also limited, but the team expects a clear leap within the next two to three years.
Robots face different problems: cerebellar control may not yet be stable, and coordination between the “brain” and cerebellum is even less settled. VLA is one attempt; 肖朝军 describes the current phase as a “Warring States period,” not a unified technology stack.
22. Cars turn model density directly into mass-production competitiveness
面壁 entered automotive projects in the first half of this year and completed mass production for one Mazda model and one Geely model in the second half. The team expects more than six additional models from different automakers to load its models next year.
Each vehicle’s chip, response time, and power consumption are fixed at project kickoff, largely determining the model scale it can carry. Within that hard boundary, higher density means stronger capability on the same hardware. 刘知远 calls this the underlying reason its cockpit projects are moving quickly.
Multimodal models currently mainly perceive the in-cabin and external environment, then interact with occupants in different seats through natural language, reminders, or task execution. A complex cockpit may define more than 100 functions, but automakers and vehicle models differ widely and no standard product form has emerged.
Adaptation and fine-tuning are currently handled mainly by 面壁. The team uses a standardized SFT toolchain and data-synthesis specifications to improve delivery efficiency, aiming to avoid turning every vehicle into a from-scratch customization project.
23. Abandoning cloud APIs was a judgment about commercial moats, not technical capability
In 2023, 面壁 also considered offering a general-purpose API, but judged that interfaces from different companies lacked differentiation. Even with higher density and lower prices, large incumbents could subsidize model services with profits from other businesses, making it difficult for a lightly capitalized startup to defend a market through pricing.
刘知远 believes the API price war triggered by DeepSeek in May 2024 and followed by Volcano Engine and Alibaba validated that risk. DeepSeek also pursues efficiency, but it can afford not to consider profitability first; it cannot serve as the commercial template for every startup.
面壁 therefore directed more of its high-density capability toward on-device deployment. 刘知远 is not seeking to “bet everything” or achieve “one-shot breakout”: putting all funding into a supermodel without a commercial loop would be unacceptable. He is looking for “certainty inside uncertainty.”
On DeepSeek’s public momentum, his view is that “what others gain is not necessarily what you lose.” AGI may arrive within five to 10 years, and the cloud is not the only entry point; moving early into the less crowded edge market better fits the team’s strengths.
24. The key 2027 milestone is personal models beginning to learn on-device
The team expects 2027 may become a key on-device milestone: reasonably capable edge compute could begin supporting relatively large-scale RL, allowing users to turn personal data into a learning environment and gradually grow a model into a dedicated assistant.
刘知远 does not believe glasses, earbuds, and watches must each run a large model independently. A glucose monitor only needs to collect data and pass it to a unified terminal; similarly, wearables can share a “portable NAS that can follow you” as an intelligence hub.
This terminal would continuously understand the user’s work, habits, and history, becoming “the intelligent model that knows you best in the world.” The model need not be packed into every sensor; what matters is that each person owns a unified, low-cost, mobile intelligence terminal.
25. The next stage of AGI is autonomous learning, collaboration, then creation
肖朝军 proposes a three-step path. First, models become efficient learners, gradually learning skills such as editing podcasts and writing code like an intern. Second, specialized models collaborate with one another. Only third do they develop the meta-capability to create new symbols and relationships.
The same model could become a technology-news assistant, an AI researcher, or an infra researcher in different people’s hands. Models could then collaborate on cross-domain problems, forming a more realistic intelligent organization than a single “knows-everything” model.
In his view, genuine creation is not merely a biomedical model searching within a fixed symbol system, but establishing concepts and relationships that did not previously exist, as the move from Newtonian mechanics to relativity did. Once that point is reached, AI could replace every job in human society that requires intelligence.
The product definition of AGI also needs to change. Autonomous driving can be handled by a specialized small model, which does not mean a general model is not AGI. The more likely form is “AI that produces AI”—it collects and generates data, trains and optimizes an autonomous-driving model that can run on specified compute.
26. Distributed intelligence will combine personal models with cloud experts
刘知远 cites a statistic from the China Academy of Information and Communications Technology in early 2024: in 2023, aggregate Chinese on-device compute—meaning smartphones—was about 12x data-center compute. A single data center is huge, but cannot match the edge scale created by the accumulation of more than 1B users and billions of devices.
He compares the claim that “the world needs only a few large models” with the 1943 judgment that “the world needs no more than five computers.” The information economy ultimately became distributed over the past 80 years; intelligence will likewise be distributed across each person and terminal rather than permanently concentrated in a handful of cloud services.
The future structure could look like this: a personal on-device model holds personal data; a Meituan model understands food delivery; a DiDi model plans trips; a Douyin model understands entertainment preferences. Together they form an “internet of agents,” delivering services at near-invisible cost and with little user friction.
The economic logic is not initially to abolish markets, but to externalize knowledge and skills into personal assets: the more specialized the model a user trains, the higher the output of the human-machine combination. In the long run, there could even be “one person running a company, with AI as the employees,” though 刘知远 acknowledges that this is his own projection.
27. The disagreement over the future is about transition costs, not the technological ceiling
曼祺’s counterquestion is worth preserving: if productivity approaches infinity, do money and work still have meaning? Real shocks such as Amazon laying off 14,000 employees cannot be flattened into a macro narrative that “all humanity benefits”; they affect individuals and families directly.
刘知远’s answer is that the first stage will remain human-machine collaboration: programmers will load code models and increase output, and future competition may become a question of whether “your AI” is strong enough. Lower-level knowledge work will be replaced, while high-level work in medicine, psychological counseling, finance, education, and law will still require human participation.
His broader optimism comes from the knowledge bottleneck: human knowledge has expanded to the point where one person can spend a lifetime advancing only one small bump on the surface of the sphere. AI for Science could compress decades of exploration into hours, helping humanity continue to understand materials, energy, life, the universe, and intelligence itself; in that picture, local overproduction is only “a drop in the ocean.”
肖朝军 summarizes the questions to test over the next year as how far RL can go, what complete form autonomous learning should take, and what shape the earliest AGI will assume; he thinks the latter may gradually emerge over the next one to two years. 刘知远 is betting on autonomous learning and expects to see an early form on a specific task next year. He noted progress in areas including physics, while 肖朝军 added that it is still concentrated mainly in the sciences.