zhudong.si
DEEP RESEARCH REPORT

SEEDByteDance Seed Deep Research Report

AI for AI, Environmental Learning, and System Intelligence

Research cutoff:Oct 6, 2026Chapters:17Full text:~97k wordsVersion:V1.0

Research cutoff: October 6, 2026 Report nature: Independent special-topic research; public-sources analysis; version GPT V1.0 Research subject: ByteDance Seed, and ByteDance products, research, and infrastructure related to its capability deployment.

This report rebuilds judgments from public sources. Product names, model release frequency, platform traffic, or management visions are not treated as evidence that superintelligence has arrived. The research cutoff is the date of source verification; it does not mean every source was published on that date. For dynamic documents, content accessible as of the cutoff is used; for papers, verified versions are preferred. Citation numbers correspond to the source directory at the end.

Executive Summary

This report's judgment: the most research-worthy aspect of ByteDance Seed is that it is attempting to convert feedback generated during model use into inputs for the next round of model and system improvement. Several engineering instances and constrained experiments already support this mechanism; whether it is sufficient to raise the long-term growth rate of frontier intelligence remains unknown.

The company should not be explained away with "Doubao has many users," "strong video generation," or "China's large-model competitor." On the research side there are several interrelated yet distinct tracks: model training for general agents; in-deployment environmental learning; distillation of experience into weights; AI-assisted R&D in evaluation, data, and training infrastructure; and dedicated verification loops in formal mathematics, life sciences, and robotics. They jointly point to a more specific question: how does a success or a failure become the system's capability for its next task, rather than remaining trapped in a single session? [2][6][8][9][18][24][51][52]

This report forms seven core judgments, all of them class-D research-synthesis judgments, with their factual basis and limitations developed in the main text.

  1. AI for AI already has clear official R&D-use disclosures. Seed for Seed extends beyond code completion. But public materials are still insufficient to compute AI's net contribution to the entire R&D cycle, research quality, or next-generation model capability gains. [2]
  2. Sustained improvement exists at different levels. Using contextual feedback, updating task memory, training dedicated policies, distilling experience across rounds, and developing the next generation of general models are different mechanisms. Collapsing the first four into "recursive self-improvement has occurred" would erase the most important evidentiary boundaries. [6][8][9][19][51][52]
  3. Experience retention is a more explanatory variable than "how long an agent can run." Longer runs can become durable capability only when feedback is valid, errors are localizable, experience is retainable, and gains are transferable. Staying continuously online and continuously learning cannot substitute for each other.
  4. Seed's path keeps both possibilities open: central models and systems composition. Native agent models reduce information loss between modules; harnesses, tools, sandboxes, and verifiers provide capabilities beyond model weights. Public evidence supports the two working together, but is not yet enough to support "superintelligence will first take the form of a multi-agent system." [11][13][36]
  5. ByteDance's applications and infrastructure can reduce friction between research and deployment, but cannot automatically produce high-quality training data. Traffic contains vast redundancy, weak feedback, commercial objectives, and permission constraints; customer data is further bounded by product terms and authorization. What needs research is effective-feedback yield, not total token counts.
  6. Physical intelligence has made substantive progress with clear task boundaries. Real-robot reinforcement learning is closer to valid evidence than video demos, but the gains of task-specific policies cannot be extrapolated directly to general household robots. Digital and physical progress should be monitored separately. [20][21][22][23][24]
  7. The advantage most likely to appear first is the scaling of verifiable workflows, not a clear AGI release date. Only when verification throughput, experience transfer, and R&D cycles improve simultaneously is there reason to raise the AI R&D Takeoff scenario; single leaderboards, model-parameter rumors, or investment plans are not enough.

The report adopts an analytical framework: feedback quality × experience retention × cross-task transfer × experiment throughput × controllable resource calls. This is a framework for tracing causality, not a fitted capability formula, and it assigns no pseudo-precise weights to the factors. If any link approaches failure, scaling the other links may only produce more low-quality attempts.

I. Research Questions, Evidence Discipline, and Subject Boundaries

1.1 Six Questions That Actually Need Answers

Rather than organizing chapters by product, this research first poses six falsifiable questions.

Research question What would support this mechanism What would weaken it
Has AI entered the AI R&D loop? Real tasks, experiment feedback, adopted results, and records of R&D stages Only code-generation demos; adoption rates and cycle contributions unknown
Can improvements in the environment be retained? Gains persist after context is cleared; repeatable results from updated weights or durable state Merely retrying within the same task; fails once out of the feedback environment
Can retained experience transfer? Gains on new repositories, new tasks, new tools, and new distributions Same-question memorization or benchmark overfitting; cross-task degradation
Can application scale accelerate intelligence growth? Valid failure data, clear rewards, usable authorization, and return pipelines Token growth mainly from free subsidies, long outputs, repeated requests
Do models or systems set the ceiling? Architecture comparisons under fixed budgets and module ablations More agents only expand spending or correlated errors
Can industrial resources and control keep up? Delivered compute, available environments, verification capacity, and permission boundaries expanding together Capital commitments fail to become usable experimental resources; supervision becomes the bottleneck

These questions do not require assuming Seed's path is correct. The same facts may admit two interpretations: it could be early engineering of an intelligence-growth feedback mechanism, or it could mainly be a more efficient product-and-research organization. The report keeps both interpretations and tracks the metrics that can distinguish them.

1.2 Evidence Grades Are Not a One-Dimensional Credibility Ranking

Grade Meaning in this report Usage discipline
A Independent research, third-party measurements, authoritative statistics; preprints or other limitations noted Independent does not mean error-free, nor does it mean Seed was directly measured
B First-party company disclosures, official repositories, technical papers with company participation Proves the disclosed path and self-reported experiments; self-reported capabilities still need replication
C Reuters, FT, and other sourced news reporting Anonymous sources, paraphrases, and official denials each marked
D This report's analyses, mechanism models, scenarios, and recommendations Must not be disguised as experimental results or internal facts

Special handling for three easily confused cases:

1.3 Seed Is Not All of ByteDance, and Not the Doubao App

Seed officially describes the team's start as 2023, with research covering general-intelligence-related directions. As of the cutoff, the official website lists Seed2.1, Seedance 2.5, Seedream 5.0 Pro, SeedRealtime, Seed Audio 1.0, and Seed GR-RL, among other models or research frameworks. [1]

This report distinguishes five layers:

Layer Research subject Conclusions that should not be drawn directly
Research team Seed, Seed Edge, and joint research projects Not all ByteDance AI R&D is run by the same organization
Models Seed general series, TARS, Prover, generative and embodied models Not all models have been integrated into a single weight system
Runtime systems Harnesses, tools, sandboxes, memory, verification, and scheduling Not the same as the base model itself having these capabilities
Products & channels Doubao, Coze, TRAE, Volcengine Ark, BytePlus, etc. A product release is not an independent capability gain of some model
Corporate resources Procurement, financing, cloud infrastructure, commercialization channels Total company budget is not Seed's training budget

The same company can share infrastructure while having independent business objectives. Stitching every product, paper, and procurement rumor into a single unified "superintelligence architecture" would manufacture organizational certainty that does not exist.

1.4 Gaps in Current Sources

As of the cutoff, this report has not obtained Seed's AI usage share across all R&D stages, a unified definition of human intervention, comparable cross-generational R&D-cycle sequences, a complete cluster inventory, or independently verified sustained frontier-R&D feedback data.

This limits the strength of conclusions, but does not make the disclosed partial work meaningless. The research task is to identify which causal chains already have evidence and which are merely possible connections — not to fill zeros with "undisclosed," and not to fill unknowns with company size.

II. Seed for Seed: Which Steps from R&D Tool to Intelligence-Growth Feedback Have Been Crossed

2.1 Official Disclosures Already Exceed the Coding Assistant

The June 23, 2026 Seed2.1 release notes explicitly describe Seed for Seed: models enter evaluation, data, training, research, and infra stages, with tasks including evaluation-system development, capability diagnosis, SFT data synthesis, RL framework optimization, and implementing and validating research-paper methods through code and experiments. The official description also covers repeated diagnosis, modification, and validation based on intermediate results, with complex tasks handled collaboratively by agents in different roles. [2]

Evidence conclusion: an officially disclosed AI→AI R&D workflow exists. This is stronger than "the model can write Python." Limitation: this disclosure provides no control data sufficient to independently estimate net acceleration of the entire R&D pipeline.

In particular, the official statement that tasks can span hours, tens of hours, or even tens of days must not be converted into "the model has a tens-of-days unassisted human-task horizon." The span may include experiment waits, staged human decisions, multiple agents, or engineering systems keeping things running; the measurement definition was never unified. [2][34]

2.2 The R&D Loop in Six Layers

Below is this report's layered methodology, class D; the numbering only indicates escalating evidence requirements, not that the company has passed each layer.

Layer Object of improvement Minimum evidence required Status in Seed's current public materials
R0 Tool assistance Human code, documentation, retrieval Task efficiency or output quality Product and research-use disclosures exist
R1 Workflow loop Specific R&D tasks Experiment→diagnose→modify→re-experiment Supported by official Seed for Seed disclosure, B
R2 Retainable task learning Memory, policies, model parameters Gains persist beyond the original context Dedicated research provides class-B experimental support
R3 R&D method improvement Algorithms, training processes, verification systems AI proposals pass isolated evaluation and enter actual R&D Individual work types disclosed; total contribution unquantified
R4 Cross-generational R&D feedback Next-generation frontier models and their R&D systems Multi-generational comparable causal data, transfer gains, cycle shortening Public evidence insufficient
R5 R&D takeoff Long-term speed of intelligence growth Acceleration is sustained and not explained by resource inputs Cannot be judged to have occurred

R2 can be achieved through human-designed loops. A system getting stronger each round does not mean it designed the next round's learning algorithm, nor that all key R&D decisions are already made by AI. Local iterative improvement, R&D automation, recursive improvement, and R&D takeoff should be named separately.

2.3 Why Evaluation, Data, and Training Frameworks Entered First

This report's judgment, D: these stages have three automation-friendly properties.

First, results convert more easily into executable artifacts. Test scripts, data-processing programs, log diagnostics, training configurations, and performance tests all have relatively well-defined inputs and outputs. Second, feedback is usually faster than for new-paradigm research. Whether code runs, whether throughput improved, whether a fault reproduces — these can be measured quickly. Third, many tasks are decomposable and low-cost, allowing parallel attempts.

But "easy to produce artifacts" does not mean "easy to judge correctness." Models can write logically consistent yet seriously biased evaluations; can synthesize large volumes of format-conforming data lacking new knowledge; can raise training throughput while degrading final quality or stability.

Therefore the first critical threshold for AI for AI is evaluation independent of the generator. When assessing the same optimization proposal, one must at minimum distinguish: code runs, training is stable, budgets are comparable, target capability improved, and no other capability clearly regressed. Passing only the first item reduces R&D automation to automated workload generation.

2.4 The Right Denominator for R&D Acceleration

A research cycle is not just the time researchers spend typing code. The following decomposition is an analytical model, D:

[ T_{\text{R\&D}}=T_{\text{problem definition}}+T_{\text{implementation}}+T_{\text{queuing \& experiments}}+T_{\text{verification}}+T_{\text{integration \& decisions}} ]

Items may partially overlap; the formula is not for mechanically summing a company's cycle. It reminds us: once the implementation stage gets fast, experiment queuing, false-positive review, and release decisions can become the new bottlenecks.

Example, purely a sensitivity demonstration: if 40% of accelerable work in a pipeline gets a 5× speedup and the rest is fixed, the overall speedup ceiling is roughly (1/(0.6+0.4/5)=1.47)×. This number is not Seed data. It shows a large causal gap between "5× coding efficiency" and "5× faster model iteration."

Nor should one count only experiment volume. If agents raise daily experiments from 20 to 200 without increasing the number of valuable discoveries that pass rigorous verification, the research system may just get more congested.

The more appropriate denominator is: under equal quality and equal total budget, how many person-hours, machine-hours, and calendar days does each reproduced-and-adopted improvement cost. R&D takeoff requires these metrics to move together, not a single quantity to inflate.

2.5 What Evidence of Cross-Generational Positive Feedback Should Look Like

Independent research offers a borrowable starting point: METR's RE-Bench places human experts and model agents in the same research-engineering environment with the same resources and information, then compares outcomes under different time budgets. In its 2024 experiment, relative performance differed between short and long budgets; the researchers also cautioned that real R&D goals are fuzzier and feedback cycles longer. [35] This report does not take these older-model results as conclusions about current Seed capabilities, but adopts its methodological discipline: same environment, same resources, human controls, budget curves, and complete trajectories.

This report recommends a D-class audit sequence of three or more generations of models and R&D systems. Three is a research-design recommendation for discriminative power, not a definitional threshold for ASI.

Each generation should record: the permissions and budgets given to AI, the R&D stages of tasks undertaken, reasons suggestions were accepted or rejected, independent replication experiments, improvements actually merged, gains for the next generation's R&D tasks, and newly added human verification costs.

The key control is "human plus tools" vs. "fixed-version agents" vs. "updated-version agents" — not a simple before/after comparison of AI use. The latter is confounded by team expansion, more chips, different task difficulty, changed model recipes, and organizational maturity.

Truly strong evidence would be: each new generation of systems gets better at improving R&D methods under the same resources; those gains help the next generation achieve higher R&D capability; the gains generalize beyond the original evaluation distribution; and they continuously offset new verification and experiment costs.

Current judgment: Seed for Seed has made R&D feedback an engineering direction, but the company-level transition from "helping researchers do more" to "sustainedly raising the growth rate of intelligence" still lacks public causal measurement.

III. The Most Critical Technical Thread: How Experience Enters Next-Round Capability from One Task

3.1 Different Kinds of "Learning" Must Not Be Pooled

Mechanism What it changes How experience is retained Primary failure mode
In-context feedback learning Current inference conditions and subsequent actions Retains trajectories, observations, and feedback Gains vanish when context is cleared; compression loses critical state
Persistent memory Externally readable state Files, indexes, summaries, task records Wrong memories, retrieval failures, cross-tenant leaks
Small-adapter learning Locally trainable parameters Parameter updates Task bias, drift, forgetting
Experience distillation Student model weights Converts experienced-conditioned decisions into training signals Remembers old tasks, defects baked in, insufficient transfer
Dedicated reinforcement learning Policies and related evaluation components Weights and training pipelines Reward hacking, distribution shift, local optima
AI R&D method innovation The process of improving next-generation models Algorithms, training systems, R&D knowledge Research value unverifiable or untransferable

Only the last one directly concerns improving "the method of making intelligence"; the first five can supply material and capability, but the connection needs experimental verification. This report excludes none of them from superintelligence-formation mechanisms, nor does it lump them together as one thing.

3.2 EdgeBench: An Important New Measurement, Not an ASI Clock

EdgeBench's public research covers 134 cross-domain tasks and about 38,000 interaction hours, open-sourcing 51 tasks and the evaluation framework; tasks may run at least 12 hours. The authors report aggregate performance following a log-sigmoid shape with interaction time, with an average fit of (R^2=0.998). The cross-generation learning-speed trend comes from an 18-task subset with similar initial capability, reporting roughly a doubling every three months. [6][7]

The paper explicitly attributes the main learning mechanism to in-context learning: agents retain observations, diagnoses, and prior attempts to improve next-step decisions. It does not directly train the evaluated model's base weights through these interactions. [6]

This report's judgment, D: the value of this research is splitting "giving AI more time" into feedback-driven search and experience accumulation. Looking only at final scores may miss differences in models' learning speed; looking only at attempt counts may mistake random search for learning.

But five misreadings must be guarded against:

  1. A smooth aggregate curve does not mean every task is smooth. Real tasks' lags, breakthroughs, and stalls may be masked by averaging.
  2. Fit precision does not equal ten-year prediction precision. Explaining well within the given tasks and time range does not prove new environments, larger budgets, or stronger models will follow the same law.
  3. Doubled learning speed does not equal doubled R&D speed. The two differ in task distribution, budgets, and outcome units.
  4. An open subset does not equal full replication. Independent teams still need equivalent tools, run protocols, and feedback to judge generality.
  5. Task scores do not equal real organizational delivery. Customer acceptance, responsibility, permission conflicts, and unformalized requirements change the nature of the problem.

The paper itself discusses how strong bottlenecks, heterogeneous task structures, and time mechanisms could break the law. [6] This report therefore treats it as a new scaling hypothesis, B, rather than elevating it to a universal natural law.

3.3 Chain-of-Experience: The Value of More Reasoning Depends on Feedback, Not Word Count

The August 2026 Chain-of-Experience study keeps prior attempts and self- or environment feedback in the reasoning context, examining math, code, and knowledge tasks. The authors report differences in gains and costs across feedback mechanisms, noting that most gains appear in early iterations. [51]

What improves here is the reasoning process, not sustained updates to base weights. Its key takeaway, D: inference budgets should be spent on actions that produce new evidence, not merely on making the model write longer rationales.

For example, a code failure message provides new constraints; another model repeating "please check carefully" may add almost no information. A miscalibrated self-evaluator will steer extra reasoning toward more confident errors. The value of feedback depends on whether it can distinguish among candidates, not on whether it comes from "another agent."

In enterprise tasks, preserving raw failure evidence, replayable trajectories, and data versions may be more valuable than preserving a polished summary. Summary compression lowers cost but may delete the boundary conditions that truly affect the next decision. How much to retain should be measured by downstream success rates and review costs.

3.4 Experience Distillation: Constrained Experiments with Cross-Round Retention Have Appeared

The July 2026 joint study Sample-Efficient Learning from Agent Experience proposes distilling decision advantages gained in trial-and-error contexts into model weights. The study covers software-engineering tasks and text-adventure environments, with experiments that clear prior-round experience and start the next round from updated weights. [52]

Worth recording are three distinct results, not just the most striking relative gains:

Experimental question Author-reported result, B What the evidence can support
Can in-context gains enter weights? Distillation retains at least 64.8% of the ICL gain; direct SFT control retains only 3.8% Within the tested distribution, how experience is processed affects retention
Can it transfer to SWE tasks not involved in experience collection? pass@1 on 494 OOD tasks rises from 4.62% to 8.84% There is transfer, but absolute success rates remain low
Can multi-round gains accumulate? After five rounds across six text-adventure tasks, average normalized score rises from 7.1 to 47.0 Cross-round accumulation after context clearing appears in constrained environments

These results are stronger than "a fixed model retrying the same question." They are also weaker than claiming general recursive self-improvement. The experiment pipeline was researcher-designed, the task sets limited; consecutive rounds in text adventures are not multi-generational frontier base-model R&D, let alone a model autonomously rewriting its own learning algorithm. [52]

This report's judgment of this evidence, D: sustainable local experience loops can now be taken seriously as research; frontier AI R&D takeoff cannot yet be concluded. To take "no sustained-improvement evidence at all" as a position would ignore such experiments; to take it as ASI having started would exceed the evidence.

The most discriminating next tests: whether cross-task capability persists after old experience is cleared; whether new domains benefit; whether marginal gains continue after five rounds; whether training and teacher-generation costs are accounted; whether old capabilities and safety constraints regress.

Experience distillation may have two long-term outcomes. The optimistic reading: every high-quality interaction becomes a better general policy. The conservative reading: each round mainly bakes in local task knowledge, with gains saturating as the task set is exhausted. Both readings are compatible with current constrained experiments.

3.5 TaskMem: Memory Is Not "the More the Better"

TaskMem models "what to remember" as a learnable policy rather than fixed summarization. In a streaming video-QA setting, the paper first trains memory quality, then updates a small adapter with recent task feedback while keeping the backbone model frozen. Its task-transfer ablation shows learned memory attention does not automatically transfer to all question types. [8]

Two direct takeaways for persistent agents, D.

First, user memory, R&D memory, and organizational memory need different retention rules. Research agents should prioritize failure causes, experiment conditions, and counterevidence; sales agents should keep customer authorizations, real needs, and relationship states; life assistants need explicit rules about which personal information to forget.

Second, a system may grow more familiar with one class of needs while becoming less suited to another. This specialization is a source of value and of bias. "The more you use it, the more it knows you" cannot be used to derive "the more you use it, the more generally intelligent it becomes."

Therefore quarterly dashboards must separately record memory hits, factual accuracy, task relevance, expiration cleanup, and cross-task degradation. Counting only memory entries or storage size cannot tell whether capability is growing.

3.6 Agent-R and CTRL: Error Recovery Needs Dedicated Learning

Agent-R constructs error-recovery trajectories through iterative self-training, teaching models to correct in time rather than waiting until task end; its experiments cover several interactive environments. [9] CTRL trains critic models to make feedback more effective for code correction by a fixed generator model. [10]

The shared research significance of the two, D: generating correct answers and knowing how to fix errors are different capabilities. A seemingly high-quality critic can accurately point out errors yet offer no useful fix; it can improve the current generator yet fail on a stronger or differently styled one.

R&D agents therefore need at least four error categories: insufficient information, failed action execution, misunderstood goals, and unreliable verifiers. Miscategorized errors make more retries amplify failures. For example, a permission denial is not the model "not trying hard enough," and conflicting external tool feedback cannot be replaced by longer thinking instead of actual measurement.

The core asset of a sustained-improvement mechanism may be reviewed error data and reproducible verification environments — not the agent's conversation count.

IV. Scaling Has Split into Multiple Resources; It Cannot Be Explained by One Switch

4.1 Seed's Evidence Points to Multiple Expansion Modes

Common scaling discussions in 2020 revolved around pretraining model scale, data, and training compute. That part still matters today, but Seed's public research also covers interaction time, reinforcement learning, memory, verifiers, sandboxes, tools, and distributed systems. [3][6][8][11][14][15][16][17]

The following is this report's mechanism decomposition, D — not a claim that each item is already unifiedly deployed in Seed2.1.

Scaling type What is really added Gains to watch for Costs not to ignore
Pre-training Parameters, unique data, training FLOPs Broad base capabilities and learning ability Data quality, communication, hardware, and long-training stability
Post-training / RL High-quality rewards and training interactions Reliable actions, error correction, policy improvement Reward hacking, bias, regression
Inference-time Thinking, candidates, and feedback rounds Quality or reliability under the same budget Diminishing marginal gains, latency, repeated reasoning
Context Available task state Retaining long dependencies and failure history Attention cost, compression loss, error accumulation
Search / Verifier Candidate coverage and judgment capability More genuinely verified solutions Verification cost, specification errors, search overfitting
Synthetic data Verified training samples Effective experience and capability transfer Homogenization, contamination, error baking
Agent runtime Interaction time and environments Information that initial reasoning lacked Waiting, interface changes, permissions, and state drift
Parallel agents Parallel exploration and division of labor Independent, complementary candidates Correlated errors, coordination cost, verification congestion
Tool scaling Callable resources and environments New task types and real actions Tool instability, authorization, irreversible consequences
Compute systems Effective compute throughput Faster training, experiments, and deployment HBM, networking, CPUs, power, scheduling

The easiest to misread is "add more agents": it adds compute and organizational structure, with no guarantee of adding independent knowledge. Another common misreading is "longer context": it allows more state to fit, but does not ensure the model uses all of it correctly.

4.2 verl and DAPO: Algorithms and Engineering Must Be Read Together

The verl official repository states it was initiated by ByteDance Seed and is community-maintained, providing a flexible RL post-training framework. [14] The DAPO paper open-sources the algorithm, training code, and data processing, emphasizing several stable-training techniques for large-scale reasoning RL; its self-reported core experiments used a Qwen2.5-32B base model. [15]

What the two support is ByteDance's participation in building reusable RL tooling and methods — not "all Seed capabilities come from self-developed base models," nor "all current results of community projects should be attributed to ByteDance." Contributions to open-source frameworks are multi-party and cross-model.

This report's judgment, D: reproducible training recipes have a dual effect. On one hand, they lower post-training experiment costs and speed up the whole ecosystem's experimentation; on the other, they make any single algorithmic lead easier to diffuse, so a company's long-term advantage must come from continuous discovery, engineering efficiency, environmental data, and deployment feedback.

For AI for AI, the RL framework also serves the function of "making experiments executable accurately." When a model proposes a reward or sampling change, the framework must turn it into a budget-comparable experiment; otherwise automation only adds unexplainable variable combinations.

4.3 Pretraining Has Not Been Replaced by Agents

Official training-data summary disclosures state that Seed2.0 Pro's pretraining text exceeds 10 trillion tokens and images exceed 1 billion. This is a training-data scale disclosure, not a model parameter count. [46]

This report's judgment, D: base capability and environmental learning may be complementary. The better a base model reads new feedback and proposes plausible candidates, the more effective interaction may be; environmental experience in turn exposes tools, private state, and real failures that pretraining data cannot cover.

But one cannot therefore conclude every resource should expand without bound. Larger models may raise per-interaction gains while making each interaction more expensive; smaller models with better feedback may achieve better cost-effectiveness in constrained processes. The choice needs comparing quality–cost–latency–intervention-rate, not a single leaderboard.

4.4 A "New Scaling Law" Needs Four Kinds of Replication

This report recommends, D: scaling of environmental learning should require at minimum temporal extrapolation, task extrapolation, model extrapolation, and budget extrapolation. Increasing duration on the same tasks cannot prove unfamiliar tasks also benefit; aligning by run-hours does not mean tokens, tool calls, and hardware budgets are equal.

One should also compare two opportunity costs: the same agent working ten more hours versus ten independent candidates each working one hour. The former accumulates state, the latter expands exploration; which is better depends on whether feedback is scarce, whether tasks have serial dependencies, and whether experience can be shared.

The longer-term scaling ceiling may be set by "the speed of discovering and verifying new experience." Once old questions, old feedback, and old tools are all exploited, added time yields diminishing returns. For the system to keep growing, it needs to enter new problem spaces and validate the experience gained there.

This is also why this report monitors effective experiments and verifier throughput alongside compute. They are not mere deployment attachments; they may decide whether compute can become capability.

V. From Native Agents to Complex Systems: The Central Model Still Matters

5.1 UI-TARS Is an Empirical Route Connecting Model Training to Action Environments

The UI-TARS-2 technical report describes a native GUI agent training method, including a data-generation loop, stable multi-turn RL, an environment combining GUI with file systems and terminals, and a unified sandbox for large-scale rollouts. Its official repository lists the release date as September 4, 2025. [11][12]

The Seed1.8 model card emphasizes bringing search, code execution, and GUI interaction into a general agent model, with attention to reasoning control under cost and latency conditions. [4]

These studies support an important judgment, D: progress in computer use is not just "locating buttons on screen" — it depends on post-action state, error recovery, and multi-turn training. Correct static localization does not guarantee the next decision is correct; a successful next action does not guarantee final delivery.

Four observables must be distinguished: element localization, single-step operation, complete task success, and real process delivery — with different denominators. For example, filling out a form is the task process; whether the correct customer record was written, whether it was submitted twice, and whether audit evidence was kept are what determine delivery.

5.2 Three Competing Explanations

Hypothesis Main source of capability formation Supporting clues Strongest refutation
H1 Central-model dominance Stronger models unify perception, reasoning, and action Native agent training, cross-capability integration Fixed-model system improvements produce larger and broader gains
H2 Runtime-system dominance Tools, state, environments, verification, and scheduling provide the key capabilities Harnesses and multi-environment execution; long-task learning with feedback Stronger single models consistently outperform complex combinations in simple systems
H3 Co-evolution Models learn to use more complex systems; systems provide models with learnable experience Complementarity of UI-TARS, environmental learning, and experience distillation Costs spiral after adding connections; experience fails to transfer

This report currently adopts H3 as the working explanation, D, with medium confidence. This is not a conclusion that "systems will necessarily become ASI first," but one that can simultaneously explain public progress in both model training and runtime engineering. Should a single model in a simple runtime environment substantially surpass other systems in the future, the judgment should be adjusted.

The central model and the system are not mutually exclusive entities. Even if perception, memory, and action mostly enter one model, it still needs external inputs, resource authorization, environment execution, and result verification. Conversely, a complex system cannot infinitely amplify a model that persistently misunderstands goals through orchestration.

5.3 DeerFlow: You Can See the System's Shape, Not Its Superintelligence

DeerFlow's current official repository positions it as a long-task agent harness combining sub-agents, memory, sandboxing, and extensible skills, supporting different models. It belongs to ByteDance's public engineering ecosystem; it must not be taken, without organizational disclosure, as a complete open-source copy of Seed's internal R&D systems. [13]

This report's judgment, D: the harness's value lies in letting tasks continue, be isolated, recovered, and inspected. For example, sandboxes preserve files and experiment state, task management prevents omissions, trajectory replay helps locate faults, and memory reduces repeated work.

These improvements raise "system workability," but surpassing large human organizations requires three harder capabilities: choosing problems worth researching; judging contradictions among multiple results; and forming accountable decisions when goals, budgets, and real interests conflict.

Research reports, code, images, and videos can all be generated without belonging to a well-coordinated organizational output. True organizational capability means these artifacts share correct state and can pass external acceptance.

5.4 Whether Multi-Agent Is Necessary: Answer by Task Structure

The April 2026 revision of the external study Towards a Science of Scaling Agent Systems evaluated 260 configurations and six benchmark categories, emphasizing the match between task structure and collaboration architecture; improvements appeared in decomposable financial tasks, while sharp degradation appeared in serial planning tasks. The report uses the revised version, not the earlier blog's 180-configuration figure. [36]

This evidence cannot prove effects for Seed's specific products, but it suffices to refute the general assumption that "more agents are naturally smarter."

This report's judgment, D: multi-agent has at least four distinct values: parallel exploration, capability specialization, independent verification, and permission isolation. The first three may improve quality; the fourth mainly improves control. They should not be unified into a "collective intelligence score."

Task nature Candidate organization, D Primary measurement
Independent data collection, independent parameter trials Parallel subtasks, centralized aggregation Result complementarity and deduplicated valid discoveries
Consecutive modifications on the same code state Few executors, clear state ownership Conflicts, rework, integration failures
Competing multi-solution candidates with cheap verification Multiple candidates plus independent verification Verified outcomes per unit budget
A single task spanning different permissions Separated roles with controlled handoffs Privilege escalation, state leaks, audit completeness
Fuzzy goals with customer-dependent acceptance Human goal owner plus agents Requirement-clarification cost and final acceptance

A task that needs repeated understanding of the same state may lose efficiency when split into too many agents. Multiple same-origin models may also jointly miss the same assumption, and voting cannot fix correlated errors.

Therefore, validating multi-agent requires fixing total compute budget, tool permissions, and acceptance criteria, and comparing against a single agent with more compute. Otherwise "collaboration improves scores" may just be "resource input improves scores."

5.5 What Real Evidence Would Show a System Surpassing Organizations

This report proposes an outcome-oriented definition, D: only when a system, across multiple important, open-ended, long-term tasks, stably produces results superior to excellent human organizations at acceptable cost — and can handle failures, conflicts, and responsibility boundaries — is there reason to discuss organization-level superhuman capability.

What should be observed is the cycle from product requirements to reliable deployment, from research hypotheses to independent replication, from optimization proposals to operational gains — not how many digital employees were created.

The diagram below shows this report's hypothesized mechanism. Dashed lines indicate connections that cannot yet be proven by public sources to have formed broad, stable frontier feedback; it is not ByteDance's official architecture diagram.

flowchart TD
    M[Base Model] --> A[Agents & Runtime Environment]
    A --> E[Experiments & Real Tasks]
    E --> V[Independent Verification & Failure Diagnosis]
    V --> C[Filtered Experience]
    C --> K[Memory & Local Policies]
    K --> A
    C --> T[Training & Experience Distillation]
    T --> M
    V -.-> R[AI R&D Method Improvement]
    R -.-> T

VI. Research Automation: Easily Verified Domains May Gain Superhuman Advantage First

6.1 Formal Mathematics Provides Clear Feedback — and Clear Boundaries

The BFS-Prover study describes an expert-iteration loop of model generation, Lean environment execution, feedback, training data, and model optimization. [19] Seed-Prover 1.5 combines agentic RL with test-time reasoning; the authors report evaluations on PutnamBench, FATE, and formal proof results on 2025 Putnam problems. [18]

This report treats them only as dedicated-system evidence for formal mathematics. Lean can check whether a formal proof holds under given axioms and statements; it cannot automatically prove the problem translation is faithful, the assumptions reasonable, the research question important, or that all competition conditions match humans'.

This report's judgment, D: in scientific work, "hard problems that are easy to verify" may be automated earlier than "ordinary problems with hard-to-define success" in social work. Mathematical proofs are hard to generate but formal correctness can be checked rigorously; business negotiation text is easier to generate, but long-term relationships, real intentions, and gain/loss allocation are hard to formalize.

Dedicated superhuman capability can become an important module of general systems, or can remain in dedicated domains long-term. High mathematical ability alone cannot be used to infer equally high ability in managing organizations, identifying new scientific directions, or handling real ethical conflicts.

6.2 Protenix-v2: The Distance from Computational Results to Experimental Hits

The public Protenix-v2 research includes biomolecular structure prediction and design, reporting hit results under specific targets, budgets, and experimental activity — not just a structure-prediction leaderboard. [26]

This is closer to a research loop than merely outputting molecular design diagrams, but evidentiary layering must be preserved: structural plausibility, experimental binding, functional effectiveness, developability, animal studies, and clinical efficacy are not the same endpoint. The report does not extrapolate front-end hit rates into drug success rates, nor treat results on a limited target set as general biological intelligence.

This report's judgment, D: the next bottleneck of research automation is often experiment capacity and feedback latency. Hypothesis-generation speed can grow rapidly, while experiment samples, detection equipment, culture cycles, and independent replication do not grow at the same rate.

A stronger model may generate fewer but more experiment-worthy candidates; a larger agent team may also drown the lab in similar candidates. Effective research systems should optimize "new knowledge per experiment budget," not "how many hypotheses are proposed per day."

6.3 Seed STEM and Seed Edge: Organizational Design Is a Variable, Not a Capability Proof

Seed Edge's official description covers long-term research, cross-modal collaboration, dedicated compute, and longer evaluation cycles, explicitly listing memory, online learning, hardware–software co-design, and agents as research topics. [5] The Seed STEM program invited about 100 researchers to collaborate, with the first cohort expected to last roughly six months; it is a recruitment and collaboration program — not "100 scientists have already joined" or "100 discoveries have already been made." [27]

This report's judgment, D: these two organizational designs attempt to ease tension between short-term product metrics and long-term research. They may help discover new problems and build expert feedback, especially for scientific judgment that existing web data cannot provide.

But long evaluation cycles may protect exploration or delay the discovery of inefficient research. Sufficient compute may expand experiments or weaken budget discipline. What should be tracked is reproducible results, high-quality questions and evaluations formed with expert participation, and contributions after results enter other R&D processes.

Public organizational intent cannot substitute for public research effectiveness. On internal team division, actual headcount, and reward-system implementation effects, this report does not fill in with recruitment copy.

6.4 Scientific "Originality" Still Needs Independent Testing

The July 2026 revision of AI Research Agents Narrow Scientific Exploration studies five framework types and five models, generating 219,655 research ideas, finding them more concentrated in the tested scientific space and closer to the starting literature, better at local extension. [37] This report uses the revised figure, not the earlier abstract's sample count.

This is not a direct audit of Seed for Seed. The idea distributions, historical citations, and scientific-impact proxies it studies are not real experimental results either. But it raises a serious competing explanation: improved research efficiency may mainly accelerate existing routes, without equally improving problem selection and original discovery.

This report's judgment, D: sustained learning most easily happens first within clear goals. Stronger AGI/ASI claims still need to explain how systems choose goals, reject established routes, find new abstractions, and have those discoveries pass real-world tests.

"Producing text humans never wrote" cannot define original science; "first producing independently verified new knowledge" is closer to a valid standard. New knowledge still needs to state its significance, scope, and reproducibility.

VII. World Models and Robotics: The Domain Most in Need of Resisting Demo Extrapolation

7.1 Videos, World State, and Action Policies Are Different Objects

As of the cutoff, Seedance 2.5 is an officially listed video-creation model; Seed3D 2.0 provides more simulation-ready 3D assets, scenes, and articulable content; the VideoWorld study learns task knowledge from unlabeled videos. [1][20][22]

This report distinguishes three uses, D:

Use Optimization objective Verification question
Content generation Aesthetics, controllability, coherence Are users satisfied; is it editable
World prediction State-transition and counterfactual accuracy After changing actions or initial conditions, is prediction still correct
Action learning & planning Task outcomes, safety, and cost After choosing actions with predictions, does real success rate improve

A model can be highly successful at the first while remaining limited at the second and third. Plausibly completing a scene is an advantage in entertainment products; in physical prediction, plausible-but-wrong completion can become a source of accidents.

Therefore video visual quality should be acceptance-tested separately from physical prediction. One should vary unseen variables like quality, friction, speed, and load, check whether the model recognizes uncertainty, and whether it maintains causal consistency under unfamiliar conditions.

SeedRealtime's August 2026 official description offers another route: end-to-end audio-video full-duplex interaction, handling continuous observation, speaking timing, and natural interruption. [25] This differs from generating offline video, and is not yet proof of a physical world model. This report's judgment, D: real-time perception may shorten feedback latency between observation and action, giving assistants more state when working on-site; but continuous observation does not mean correct causal understanding, and proactive responses do not mean authorization to execute real-world actions has been obtained. Its long-term value should be validated by sustained-interaction task outcomes, missed events, and privacy boundaries — not by curated conversation demos.

7.2 Seed's Own Research Also Rejects "Video Scaling Automatically Learns Physics"

The How Far is Video Generation from World Model study examined in-distribution, compositional, and out-of-distribution generalization in controlled 2D physics environments, reporting that models cannot reliably abstract universal physical laws under the tested out-of-distribution conditions. [21]

This earlier study cannot prove the latest Seedance necessarily has the same flaws, nor that all future video models will fail. It provides a falsification method: keep conditions controllably varied, and assess whether predictions obey real laws — instead of having humans judge whether footage "looks commonsensical."

This report's judgment, D: the strongest evidence for validating a world model should be action-conditioned counterfactual prediction and real planning gains. Crossing this threshold requires independent tasks; it should not be substituted with creative samples, curated clips, or model self-evaluation.

7.3 VideoWorld and Seed3D Address Different Links

The VideoWorld paper explores visual knowledge learning in video-go and robot-control tasks, providing narrow task validation; its robot evaluation includes environments like CALVIN and RLBench. [20] Simulation control results must not be written up as real-robot success rates in open home environments.

The Seed3D 2.0 study provides simulation-usable content for objects, materials, scene layouts, and joint interactions. [22] Simulation assets can help expand the training distribution, but correct geometry and appearance do not guarantee accurate material parameters, contact, wear, or sensor noise.

This report's judgment, D: the two routes may be complementary. One provides dynamics knowledge learned from vision; the other provides materials for executable environments. Connections still missing or needing explicit validation for a physical learning flywheel include: action–observation consistency, physical parameter calibration, real failure feedback, and simulation-to-real task transfer of gains.

These are possible research connections, not a unified training pipeline this report confirms as deployed.

7.4 GR-3 and GR-RL: Real Experiments Matter More Than Demos

The GR-3 technical report describes an operation model combining vision-language data, robot trajectories, and human demonstrations, with real-task evaluations shown on the ByteMini platform. [23] GR-RL builds on it, studying data filtering, symmetry augmentation, and real-environment RL to improve operation performance on a specific shoelace-tying task. [24]

The GR-RL main experiment reports a 72.7% success rate after filtering and augmentation from a 45.7% baseline, and 83.3% after online RL evaluation. The paper also describes training curves at one point exceeding 90%; this report uses the evaluated 83.3% and does not take the training moving-average peak as the final deployment success rate. [24]

Two easily missed conditions remain. Its online process is not just "about 150 attempts": the full text notes that a policy trained offline earlier had already collected 673 real-environment rollouts for adapting related components; "offline policy" does not mean collecting those trajectories had no real-machine cost. The paper also retains system links like trajectory optimization, and notes that online-RL behavior drift and distilling dedicated gains back into general policies remain unsolved. [24]

These facts support real feedback improving dedicated high-precision operation capability, but cannot prove the model generally understands the physical world, nor that arbitrary tasks can be learned with little interaction.

7.5 How to Compare Simulated vs. Real Success Rates

This report found no public unified data connecting all the above systems on the same tasks, same hardware, and same budget. It therefore provides no fabricated "sim-to-real gap percentage."

Valid comparison requires the following protocol, D:

Comparison dimension What is fixed in simulation What is recorded in the real environment
Tasks & initial states Same goals, objects, and difficulty tiers Initial placement, object changes, whether failures count
Sensing & action Latency, noise, actuator range Calibration, errors, actual action execution
Feedback & intervention Reward rules, reset methods Human righting, resets, protection, and repair
Training budget Trajectory counts, physical time, compute Real-machine time, demonstration cost, damage risk
Completion criteria Complete tasks, allowed precision Quality consistency, time taken, material damage
Generalization Unseen states and parameters New environments, operators, lighting, and objects

For enterprise deployment, average success rates are not enough. Tail failures during a full day of continuous operation, recovery time, dependence on manual resets, and the economic consequences of failures are usually closer to purchasing decisions.

7.6 The Digital–Physical Gap Is a Mechanism Gap

This report's judgment, D: digital work can replicate environments, roll back state, run parallel trials, and verify quickly. Real robots need physical time, sensors, and actuators; failures damage objects or equipment; resets often need humans. The two may therefore show a deployment time gap of years or more, but public evidence is insufficient for a specific number of years.

This judgment also needs falsification. Should high-fidelity simulation, cross-hardware general policies, or highly efficient real-machine learning make breakthroughs, physical intelligence could accelerate significantly. Conversely, bottlenecks in permissions, verification, and customer communication in digital work could also slow digital deployment.

What should be recorded separately: digital task delivery rates and physical task delivery rates — not a single model-capability chart assigning both the same ASI timeline.

VIII. Industrial Systems: Effective Resources for AI R&D Are More Than GPUs

8.1 Training-Systems Research Provides Evidence Closer to Use Effects Than Procurement Rumors

The MegaScale-Omni paper describes decoupled parallelism, rearrangement, and load balancing under dynamic multimodal training workloads, claiming use in internal training on thousands of GPUs and reporting throughput-improvement ranges against several systems. [16] The Charon study uses a simulator to predict training and inference configuration performance, supporting low-cost screening of system designs. [17]

This report uses them to demonstrate system-optimization directions and author-reported experiments; it does not apply best-configuration multiples to the entire training budget, let alone expand "deploying thousands of GPUs" into an unverified hundred-thousand-GPU single cluster.

Analytical judgment, D: effective compute is jointly determined by hardware, models, communication, and workload. Uneven multimodal sample lengths, different loads between encoders and the language backbone, and resource competition between rollouts and training can all make nominal scale differ greatly from usable throughput.

This directly relates to AI for AI: if agents can optimize configurations or find performance bottlenecks, gains may quickly enter the next round of experiments; but if only through micro-benchmarks, the improvements may not survive in production workloads.

8.2 CapEx, Financing, and Available Compute Are Not Interchangeable

Reuters's December 2025 report relayed the FT's preliminary 2026 plan of roughly RMB 160 billion in AI infrastructure investment, explicitly noting it could not be independently verified at the time. [39] A September 2026 Reuters report covered a $29.6 billion loan arrangement, with sources saying it mainly supports AI-related plans, while the stated corporate purpose was general corporate use. [43]

This report does not add the two together. A loan is a financing instrument; CapEx is an investment or procurement measure; the same project may use that financing. Planning, signing, paying, delivering, and going live are also different stages.

Judging Seed's resource constraints requires looking at already-delivered training hours, experiment queues, inference quotas, and environment capacity. Even with sufficient funding, HBM, networking, racks, cooling, grid connection, and software migration may still limit available resources.

The company is not a transparent sample in the listed-audited-financial-statement sense. Outsiders cannot reliably split all costs among Seed, cloud customers, consumer products, and other businesses. The report therefore does not estimate "how much ByteDance spends per model" or "how many GPUs each researcher has."

8.3 Chip Reporting Must Preserve Conflicts

Public message Source status, C Conclusion adopted by this report
2026-02: SeedChip and Samsung manufacturing talks reported Anonymous sources; ByteDance spokesperson called related self-developed-chip claims inaccurate Reporting and official denial coexist; cannot confirm plans or volume achieved [40]
2026-05: Custom CPUs, Arm / RISC-V direction Reuters citing sources; project at an early stage Reports of a CPU route, not mass-produced self-developed training GPUs [41]
2026-05: Qualcomm ASIC partnership Reuters relaying Bloomberg, not independently verified A potential supply route, not ready available compute [42]
2026-06: Discussing inference-chip procurement with Iluvatar CoreX and others Reuters sources; talks not finalized Reports of supply diversification, not delivered procurement [44]

These materials concern different projects, uses, and disclosure states; conflicts cannot be eliminated by "averaging." Denial of one early project does not prove the company will never make chips; reporting of another project does not prove the denied project exists.

This report's judgment, D: agent scaling increases CPU, memory, storage, and environment-scheduling loads. Each agent not only generates tokens but also runs tests, maintains files, operates browsers, and waits for external feedback. The economic value of self-developed or diversified supply may come from lowering the entire task-running cost rather than from the maximum FLOPs of a single chip.

8.4 Power Is Location- and Time-Specific

IEA's Electricity 2026 lists data centers as an important driver of future power growth, emphasizing grid and flexibility needs. [38] This is industry background; ByteDance's power consumption cannot be computed from it.

This report's judgment, D: training clusters need stable power and networking; real-time assistants need low-latency service near users; background research tasks can accept time or geographic scheduling to some degree. The three workloads differ in siting and power elasticity.

Cheap generation does not equal usable data centers. Grid-connection capacity, build time, cooling conditions, chip availability, cross-region communication, and operating rules jointly determine whether a project can go live.

The resource accounting of sustained agents must also count idle waiting clearly: online all day does not mean full-load GPUs all day. If inference, experiments, and background state can be effectively separated, the system can save resources; if each task long occupies a complete environment, costs will rise sharply with concurrency.

8.5 Industrial Constraints May Also Change Technical Routes

This report proposes three conditional inferences, D — not established facts:

But constraints do not automatically manufacture breakthroughs. Hardware–software co-design and model compression can improve efficiency, and may also add compatibility and R&D complexity. What should be tracked is the total cost of completing the same tasks, not the assumption that limits necessarily produce better technical paradigms.

IX. Products and Ecosystem: When Can Traffic Become a Learning Advantage

9.1 Multiple Product Entry Points Cannot Directly Form a Data Flywheel

Seed's official website presents model channels for applications and enterprise services. [1] Coze, TRAE, and open-source harnesses provide workspaces, tool execution, and development functions respectively. [13][28][31]

This report's judgment, D: these entry points can lower customer acquisition, deployment, and real-problem observation costs. But forming a research advantage requires at least four conversions:

Usage events → diagnosable feedback → authorized, quality experience → gains on unfamiliar tasks after entering training.

The first step filters out vast numbers of result-less conversations; the second needs to identify whether the fault is model error, unclear requirements, or tool failure; the third involves data authorization, sensitive information, and deduplication; the fourth needs training and independent verification. Each step has loss.

Assuming a model processes more tokens daily does not mean there is more new information. Long contexts, video encoding, internal applications, repeated reasoning, price subsidies, and model routing all change call volume. Without a unified measure, cross-company token-count comparisons are especially misleading.

Therefore this report has not written rumored huge daily call volumes or article figures of unknown authorship and statistical definitions into the leading-indicator baseline. Technical explanations, official documentation, and platform articles on the company's website cannot gain equal evidentiary weight merely for sharing a domain.

9.2 Coze's Current Shape: Collaboration and Persistent Operation, with Product Migration Underway

Coze's FAQ accessible as of the cutoff describes Coze 3.0's multi-person, multi-agent projects, independent contexts, file and artifact retention, and local and cloud agent access. The documentation also clarifies that certain cloud agents named Codex CLI or Claude Code use platform models or open-source frameworks, and do not directly equal the corresponding vendors' official products or users' existing accounts. [28]

This shows ByteDance exploring agent organization and long-term workspaces at the product layer, but cannot prove these organizations have produced collective intelligence exceeding human teams.

Product boundaries need verification per specific service:

Feature change Current first-party documentation, B Correct reading
New coding-project entry in Coze Since 2026-09-08, no new projects of this type; switched to agents calling coding and deployment plugins; existing projects unaffected Work-entry migration, not cancellation of all coding capability [29]
Coze coding OpenClaw projects Since 2026-06-30, no new old-style projects; existing ones continue; Coze 3.0 cloud-agent alternative provided High operating costs driving product-form adjustment, not OpenClaw going offline overall [30]
Local agents vs. cloud agents Access scope determined by the specific access method Cloud isolation commitments cannot be applied to the actively-authorized local-device mode [28]

This report's judgment, D: this adjustment is a real economic signal for persistent agents. Product teams must trade off among environment resources, collaboration, operations, and interaction costs. One cannot judge the current service by the launch-time "always online" marketing, nor automatically interpret entry consolidation as R&D failure.

9.3 TRAE / SOLO: Code Gradually Becomes One Stage in the Workflow

TRAE's March 2026 official description presents the new SOLO as shared project workspaces with Code and Work task modes, including documents, analysis, reports, and other outputs, distinguishing desktop and cloud use. The announcement was then a regional, phased beta; this report does not claim all regions were equally open as of the cutoff. [31]

Analytical judgment, D: once a coding agent can read requirements, analyze data, create deliverables, and run code, its commercial boundary expands from "code editor" to project advancement. But expanding product scope does not mean every professional task has reached expert quality.

For enterprises, the key assessment is not "how many formats it can generate," but whether materials were used correctly, results are reviewable, stages share correct state, and integration labor is reduced.

9.4 Open-Source Diffusion Is a Route, Not Reducible to Open Ownership

The Seed-OSS official repository offers 36B models and different pretraining-data versions, describing open licenses and design choices like long context and reasoning budgets. [32] coze-studio provides an open-source agent development platform. [33] verl, TARS, and DeerFlow respectively open-source different training or runtime components. [12][13][14]

These objects' openness differs: model weights, training data, complete recipes, execution frameworks, cloud services, and product accounts should each be verified separately. Open-source platforms do not automatically gain all capabilities of cloud commercial versions; an open model does not automatically inherit the latest closed-source frontier model's capabilities.

This report's judgment, D: open source can expand engineering adoption, lower composition costs, form de-facto interfaces and research baselines. But it does not necessarily give enterprises data, execution identity, service channels, or final distribution. Open technology and centralized control can coexist.

For the Chinese ecosystem, the more important observation is whether different models and systems interoperate, and whether enterprises can migrate with workflows, data, and evaluations. Judging an entire company by the closed/open dichotomy obscures these real boundaries.

9.5 Content Platforms and User Agents: Objective Functions May Conflict

This report discusses structural possibility, D — it does not claim any product currently secretly manipulates users.

Content platforms usually need to decide what users see; user agents need to decide on users' behalf what to see, buy, or ignore. If an agent's optimal action is to reduce user dwell time, reject ad recommendations, or choose cheaper off-platform goods, platform commercial revenue and user goals may conflict.

After agent permissions expand, this conflict is no longer just "whether recommendations are fair" — it may involve actions: orders, refunds, subscriptions, and content publishing. A true user-agency relationship needs to answer: when user interests differ from platform revenue, whose objective does the system serve?

Verifiable product standards include: commercial-relationship disclosure, configurable goals, off-platform options, decision rationales, action logs, revocable permissions, data export, and migration capability. One cannot judge allegiance by "personalization" or "knowing you" alone.

9.6 Cross-Platform Agents' Bottleneck May Be Authorization and Commercial Acceptance

Reuters reported a December 2025 partnership between the Doubao voice assistant and ZTE phones; subsequently sourced reporting recorded other apps restricting features and ByteDance adjusting capabilities. [49][50]

These events only prove entry points and platform restrictions existed at the time; they do not mean every app still enforces the same measures as of the cutoff, nor that all restrictions stem from commercial defense. Security, privacy, interface specifications, and platform interests may all be present.

This report's judgment, D: cross-app intelligence depends not only on "reading screens." Payment confirmation, identity authentication, anti-automation, and terms of service determine which resources a system can call. Technical success rates and obtaining legitimate, stable action rights should be monitored separately.

In-ecosystem loops are easier to build and more likely to create vendor lock-in. Off-ecosystem work has more user value but needs clear authorization and collaboration agreements. This tension may determine agent platforms' commercial boundaries.

X. Control: The Stronger the Feedback, the More Boundaries Independent of Feedback Are Needed

10.1 What Has Been Disclosed, What Remains Unverified

Seed's official website has a Transparency page providing summaries of some model training content. [45][46] Responsible AI hiring materials describe the team's research on reliable, safe, and trustworthy AI, with participation in product safety and ethics work. [47] The TARS repository also discloses limitations like hallucination, resource consumption, and potential misuse. [12]

These are substantive disclosures, but of different natures. Data-transparency summaries cannot substitute for agent-action risk assessments; hiring cannot prove specific risks have been resolved; listing limitations does not equal having established effective control mechanisms for all risks.

Within this review's scope, the report found no unified public danger-capability thresholds for all of Seed's frontier models, post-trigger deployment actions, or independent audit results sufficient for item-by-item verification. The relevant indicators are therefore marked "public evidence insufficient" — not "no safety work" or "low safety risk."

10.2 Define Control Objects by Agent Permissions

The table below is this report's recommendation, D — not Seed products' current full guarantee list.

Capability growth Corresponding risk What control should verify
Long tasks and persistent memory Error states accumulating long-term Whether state versioning, provenance, expiry, and deletion work
Browsing and tool calls External instructions polluting goals Whether untrusted content can change authorization and task scope
Multi-agent collaboration Permission merging, correlated errors Role isolation, task handoffs, shared-file boundaries
Training and data generation Bias entering next-round weights Data review, isolated evaluation, traceable versions
AI R&D resource calls Budget runaway, dangerous experiments Resource quotas, experiment allowlists, and stop paths
Computer use and payments Duplicate execution, wrong identity Idempotency, confirmation, revocation, and auditing
Robot execution Human, object, and equipment injury Physical limits, emergency stops, fault recovery, and on-site supervision

The criterion is not whether the system is willing to say "I follow the rules," but whether it maintains boundaries under ambiguous, conflicting, and inducement scenarios. The stronger the capability, the more control tests should cover alternative action paths, preventing the system from achieving the same outcome through unconstrained tools.

10.3 Harness-IF's Significance: Success and Obedience Should Be Acceptance-Tested Separately

Seed's public paper directory lists the August 2026 Harness-IF paper, whose abstract proposes evaluating operating rules through execution evidence across multiple instruction carriers, distinguishing genuine compliance from the model doing so anyway. [53][54]

This report could access the abstract, but full-text access was limited; it therefore cites no un-read itemized percentages, nor evaluates Seed2.1's specific control level on that basis.

The methodological takeaway, D, is clear: if the requirement is "do not modify a file" while the agent happens not to need that file, successful task completion does not prove it respects the constraint. Stronger tests need real conflicts between rule compliance and default high-score strategies.

This is especially needed in AI R&D. If modifying the test set is prohibited, the system must still refuse when modifying tests could raise scores; if experiment budgets are limited, it must still comply when larger budgets could yield better results. Testing rules only in no-temptation scenarios easily overestimates control.

10.4 Reward Loops and Safety Loops May Diverge

This report's judgment, D: rising task scores do not guarantee rising control. A model may get better at satisfying a reward while simultaneously getting better at finding the gap between the score and the real goal.

The TaskMem paper discusses rewards being gamed by short, insubstantial outputs, requiring designs like increased memory richness. [8] This is a reward problem in constrained training, not an accusation of malicious behavior in real products — yet it shows "having automatic scoring" does not mean "scoring is sufficient."

Control design should be independent of task rewards, including: permission boundaries that cannot be crossed without authorization, preserving raw evidence, audits not decided by the executor itself, and the ability to reclaim credentials and subtasks when stopping a system.

After an agent process is terminated, if background tasks, external API credentials, or cloud sub-agents can still operate, it cannot count as a complete stop. The control object is the action system, not just a chat window.

10.5 Data Return and Customer Trust Have Clear Boundaries

Volcengine Ark's official documentation, in retrievable content, states that for the corresponding image- and video-understanding services, user-submitted content is not retained for model training. [48] This is a statement within specific service documentation's scope; this review's full-text capture was limited; the report does not expand it into a unified data commitment across all ByteDance products.

Analytical judgment, D: application scale may produce product feedback, but one cannot assume all customer raw content can directly enter training. Even if enterprises actively authorize return, it needs determining which fields, which uses, and which derived data are included.

Data governance is not an obstacle outside the R&D flywheel, but the precondition for high-quality feedback to be sustainably usable. If customers, unable to control their data, are unwilling to hand important tasks to agents, the feedback the system gets will stay in low-value, low-sensitivity scenarios.

XI. Competing Explanations and Falsification: Avoid Assembling All Progress into Inevitable Takeoff

11.1 Four Explanations Can Each Explain Part of the Current Facts

Explanation How it explains Seed's public progress Which observation is most discriminating
E1 Product & engineering catch-up Rapidly improving usability through evaluation, RL, cost optimization, and channels Whether gains concentrate on existing product tasks
E2 Environmental-learning paradigm expansion Improving learning efficiency with feedback tasks and experience distillation Whether cross-task, cross-distribution, cross-round retention persists
E3 AI R&D positive feedback AI beginning to improve the methods that make next-generation AI Causal changes in R&D capability and cycles across consecutive generations under fixed resources
E4 Industrial-scale advantage Funding, clusters, deployment, and organizational resources bringing competitiveness Whether per-effective-resource output rises after deducting resource growth

This report currently sees strong public clues for E1, E2, and E4; E3 has local engineering directions, with the company's sustained frontier feedback still unverified. They are not mutually exclusive scenarios, and the others cannot be excluded just because things "look like takeoff."

11.2 Eight Narratives That Need Active Challenge

  1. "Tasks span tens of days, so there is already tens-of-days autonomous R&D capability." One must clarify intervention, waiting, state maintenance, and acceptance. Span is not horizon.
  2. "Experience entered weights, so general RSI has appeared." Task-specific training differs from autonomously improving R&D methods; cross-round gains do not automatically cross domains either.
  3. "A scaling law was fitted, so capability can be predicted exponentially." Fitting, extrapolation, and causal mechanisms need separate testing.
  4. "Multi-agent organizational form is collective superintelligence." Having roles and messages is not evidence; net output after coordination should be measured.
  5. "Videos look real, so models understand physics." Counterfactuals, out-of-distribution tests, and real control experiments are what discriminate.
  6. "More users means better training data." User volume provides neither automatic authorization nor reliable rewards.
  7. "Chip procurement or self-development plans mean the compute bottleneck is solved." Plans, delivery, software adaptation, and effective throughput should each be verified.
  8. "No public unified safety framework, so no safety work is done." Information gaps cannot be directly read as internal facts; nor can control maturity be assumed from them.

11.3 Evidence That Would Most Likely Invalidate This Report's Core Judgments

This report takes feedback-to-durable-capability conversion as the key mechanism. At least four kinds of evidence would force us to downgrade its explanatory power, D:

Three kinds of evidence would raise confidence: multiple independent teams replicating cross-round retention; transfer on unfamiliar tools and research questions; shorter frontier R&D cycles under fixed total resources with quality maintained.

"'R&D takeoff not yet proven' is not a bet that it won't happen. Research discipline requires: when strong evidence appears, upgrade judgments promptly; before strong evidence appears, do not write feasible directions as realized states."

XII. 3–10 Year Scenarios: Replace ASI Years with Inflection Points

The following are class-D conditional scenarios, with an observation window of roughly 2029–2036 — not year predictions for events. This report assigns no numerical probabilities: existing public sources are insufficient to calibrate them.

Scenario A: Sustained Scaling and Agent Improvement, No Clear AGI Day

Premise: Pretraining, post-training, and runtime engineering still yield gains; experience learning improves, but no significant accelerating R&D-method feedback forms.

Leading indicators: Acceptance rates rise on fixed task sets; human takeovers decline; unit effective-task costs fall; more business processes become deliverable under clear boundaries.

Possible outcomes: ByteDance expands the practical agent market on product entry points and model channels. Enterprises feel a change in work division, not a sudden crossing of an intelligence threshold one day.

Risks: Workflow improvements disconnect from market messaging; product integration increases lock-in; traffic growth cannot cover persistent-running costs.

Upgrade evidence: Multi-quarter joint improvement in quality, cost, and intervention rates. Downgrade evidence: Larger budgets only maintain equal quality, or real-process failure rates stay above benchmarks long-term.

Scenario B: Significant Positive Feedback and Takeoff in AI R&D

Premise: Agents not only implement plans but also propose, verify, and promote R&D methods; gains enter the next generation, with verification, compute, and experiment environments keeping up.

Leading indicators: More adopted AI-originated R&D proposals; shorter R&D cycles under equal budgets; new models better at improving training and research; traceable cross-generational gains.

Possible outcomes: Research competition shifts from single model releases to R&D-system efficiency. Companies with training systems, verification environments, and experience assets gain larger cumulative advantages.

Risks: Automated evaluation gamed; hidden faults entering training; growing supervision burden; capability and deployment control out of sync.

Upgrade evidence: At least multi-generational comparable causal sequences, independent replication, and resource-normalized results. Downgrade evidence: Gains stay in software implementation, with new quality fully explained by more budget or human research.

Scenario C: Compute, Energy, Verification, or Control Becomes the Main Bottleneck

Premise: Models can propose more actions, but industrial resources, real verification, permissions, or supervision cannot expand at the same speed.

Leading indicators: Experiment queues grow; inference quotas and product forms adjust; unit delivery costs do not fall; verification teams become the main labor burden; platform restrictions affect executable tasks.

Possible outcomes: Technical routes shift to smaller models, dedicated policies, routing, batching, and stronger resource discipline. Deployment speed lags model-evaluation progress.

Risks: Price cuts and subsidies mask real costs; permissions excessively loosened to reduce friction; concentrated supply makes enterprise workflows fragile.

Upgrade evidence: Insufficient delivered resources, persistent queuing, or sustained supervision-cost growth. Downgrade evidence: Software efficiency, supply diversification, and controlled operations clearly improving unit delivery costs.

Scenario D: New Technical Paradigms Lower Key Costs of Experience Learning or Pretraining

Premise: Continual learning, architectures, diffusion language models, world modeling, hardware–software co-design, or other directions produce broad gains beyond current recipes. Listed directions do not mean Seed has achieved any breakthrough.

Leading indicators: Capability gains on new distributions under fixed compute; parameter updates without notable forgetting; unfamiliar tasks stably learned with minimal feedback; independent team replication.

Possible outcomes: Current competition measured by parameters, tokens, or agent counts gets redefined. Long-term research mechanisms like Seed Edge may benefit — or other companies may break through first.

Risks: Curated tasks and best configurations masking insufficient generalization; new architectures introducing hard-to-audit state and control problems.

Upgrade evidence: Cross-task, cross-model, cross-team results with clear mechanisms. Downgrade evidence: Effective only on one benchmark class, with advantages vanishing after engineering integration.

Scenario E: Commercial Distribution Succeeds, but General Research Stalls

Premise: Consumer, creative, and enterprise products keep improving, but gains come mainly from productization and existing capability combinations.

Leading indicators: User and revenue growth, without synchronized improvement in environmental-learning transfer, original research, and cross-generational R&D capability.

Possible outcomes: ByteDance becomes a powerful AI applications and industrial platform, while unable to demonstrate general superintelligence formation. This is a fully possible and commercially valuable outcome.

Risks: Distribution and research successes conflated; capital allocation dragged by unfalsifiable AGI narratives.

Upgrade evidence: Product metrics and research metrics decouple long-term. Downgrade evidence: Product failure data clearly converting into transferable R&D and base-capability gains.

How Scenarios Shift

This report recommends judging by bottleneck changes, not model names. If A's workflow gains enter R&D methods and persist across generations, shift toward B; if capability rises while effective experiments and control stagnate, move toward C; if new mechanisms change learning efficiency and are widely replicated, raise D's weight; if only commercial expansion occurs without research-mechanism progress, raise E's weight.

One quarter can hold multiple directions simultaneously. For example, digital products entering A, a local R&D module showing B's early signs, robots still constrained by C. One cannot label an entire company with a single "takeoff" tag.

XIII. Quarterly Leading-Indicator Dashboard: Monitoring Capability Formation, Not Asking Whether AGI Has Arrived

13.1 Indicator Design Principles

The dashboard is an updatable framework proposed by this report, D; the table's "baseline" only indicates current evidence status. Indicators without company public data are explicitly written as unknown, not filled with lab-wide trends.

Each update simultaneously records model version, runtime-system version, tools, total budget, success definition, failure records, human intervention, and data sources. If any condition changes, report the definition change first, then compare trends.

ID Indicator Proposed definition & denominator 2026-10-06 baseline Update frequency / triggering evidence
01 Cross-domain capability Fixed task groups, per-domain results, not just totals Rich official evaluations; incomplete third-party comparability [2][3] Quarterly; new models plus independent retesting
02 Human-task horizon 50% / 80% AI success thresholds at corresponding human completion times No same-definition complete estimate for Seed verified; task spans must not substitute [34] Update when new independent measurements appear
03 Complete agent acceptance rate Share of all launched tasks ultimately passing acceptance Lacks unified real-business sample Monthly internal, quarterly summary
04 Human intervention Interventions per task, person-hours, causes Seed for Seed undisclosed in unified terms Record by difficulty, success/failure separately
05 Failure recovery Completion rate and added cost after fault injection Dedicated recovery research exists [9][11][24] Fixed fault sets retested quarterly
06 In-run learning gains Improvement with feedback vs. without, under equal budgets EdgeBench / CoE provide research protocols [6][51] Fixed tasks, fixed budgets replicated
07 Retention after context clearing Gain after clearing experience / gain with experience retained Experience Distillation has constrained experiments [52] Measure after each learning round
08 Cross-task transfer Net gains on tasks not involved in experience collection OOD SWE experiments exist; broad domains unknown [52] Quarterly expansion of blind-test scope
09 Marginal gains of sustained rounds New capability increment per round per unit cost Limited multi-round experiments; long-term slope unknown [52] Multi-round consecutive sequences
10 Memory quality Accuracy, relevance, expiry deletion, contamination, and degradation TaskMem provides a research starting point [8] Monthly; on model/memory-policy updates
11 AI R&D work coverage Recorded by evaluation, data, training, algorithm, infra stages Officially claimed multi-stage coverage; shares unknown [2] Quarterly audits of task records
12 AI proposal adoption rate Independently verified and merged proposals / all candidates Company-level data unknown Record rejection reasons simultaneously
13 R&D cycle Calendar time from definition to deployable version under equal quality goals Comparable company sequences unknown Per generation; normalized for people and compute
14 Cross-generational R&D feedback Increment in next-generation R&D capability attributable to prior-generation AI results Unverified Multi-generation tracking, independent audits
15 New knowledge Externally replicated new discoveries / all candidates Dedicated research self-reported results; general sequences insufficient [18][26] Semi-annual; graded endpoints
16 Multi-agent net gains Improvement over single agents under fixed total budgets and permissions Product organizational forms exist; unified Seed effects lacking [28][36] Quarterly architecture comparisons
17 Effective concurrency Accepted output per unit time, tail latency, and conflicts No publicly comparable overall data Monthly run data
18 Verifier throughput & quality Results passing independent review / verification resources Math feedback relatively clear; general R&D unknown [18][19] Quarterly; review misjudgments
19 World-model practical gains Real-task gains from prediction-based planning vs. not Local research exists; broad real connections unverified [20][21][22] Cross-distribution, cross-environment evaluation
20 Real-robot delivery Complete task success, time, interventions, damage, and recovery GR series has local real experiments [23][24] Same tasks, same hardware retested
21 Delivered effective compute Available training/experiment throughput and queues, not promised amounts System papers exist; complete cluster resources unknown [16][17] Quarterly; delivery and go-live recorded separately
22 Unit accepted-task cost Total cost of inference, tools, environments, verification, and operations Token unit price not a substitute Monthly, before/after model switches
23 Available power & industrial delays Actually available capacity, grid connection, and delivery dates Company-level complete baseline unknown Update on project go-live
24 Control & agency Test results on privilege escalation, injection, stopping, logging, export, and migration Partial disclosure; unified comparable results insufficient [12][45][47][53] On every permission or runtime-system change

13.2 Not Recommended to Aggregate into One "ASI Index"

Different indicators cannot simply be added linearly. Improved model capability cannot offset privilege-escalation incidents, low prices cannot offset wrong deliveries, and massive concurrency cannot offset verification failures. A composite score would mask the bottlenecks deciding the next stage.

The more practical quarterly conclusion is: which three advances have strong evidence, which two have become bottlenecks, which unknowns have shrunk, and which judgments need withdrawal.

Evidence states rather than marketing colors can be used: independently replicated, company-reported with experiments, feature disclosure only, evidence missing, counterevidence emerged. The same indicator may simultaneously have self-reported progress and independent counterevidence; both should be kept.

13.3 Five to Track First

For judging "whether Seed has entered intelligence-growth feedback," this report prioritizes: AI-proposal independent adoption rate, R&D cycle, post-context-clearing retention, cross-task transfer, and marginal cost of multi-round gains.

These five distinguish "generating more outputs," "smoother workflows," and "systems continuously improving the ability to make intelligence." GPU scale, traffic, and model leaderboards are recorded as explanatory variables, not substitutes for these five.

XIV. Implications for Enterprises, OPCs, and Individuals

This section is class-D recommendations, targeting work mechanisms rather than immediate product-purchase advice; specific product availability, pricing, data terms, and authorization need re-verification at implementation.

14.1 Enterprises Should Buy Verifiable Capability, Not "Digital Headcount"

The Seed ecosystem's most trial-worthy tasks for enterprises are processes with clear input materials, clear tool scopes, verifiable results, and rollbackable failures. Whether a task "looks advanced" does not matter. Sales retrospectives, data cleaning, report drafts, code maintenance, and internal retrieval may all create value, but quality standards must be set by real business.

A process's value can be accounted as: saved person-hours and new gains, minus review, rework, error losses, system integration, and maintenance costs. AI output time alone cannot substitute for the original process's cost.

Especially check the last mile: who decides after a report is formed, who confirms after an order is created, who maintains after code runs. Without owners at these links, automation transfers problems to downstream teams.

14.2 An OPC's Advantage May Come from Small Verifiable Systems

One person coordinating multiple agents does not automatically possess a large company's capability. An OPC's best opportunities lie where they hold business judgment, customer communication, and acceptance standards themselves, while agents take well-defined work segments.

Core assets include customer-specific processes, data structures, tool connections, failure experience, and quality checks. Generic prompts are easy to copy; task environments and acceptance protocols validated by real business are harder to copy.

This report recommends starting from one process with clear outcomes, keeping failure cases and human-takeover reasons, then deciding whether to change the model, the tools, or the task definition. Do not attribute all failures to insufficient model capability, nor stack agents too early.

14.3 Enterprise AI Consultants' Focus Will Shift

Research models and more mature harnesses will lower prototype-creation costs. Consultants should instead take on requirement clarification, data permissions, tool integration, acceptance design, exception handling, and ongoing operations.

This differs from "delivering just a chatbot": one business process usually contains different roles, different permissions, and different completion standards. Consultants need to judge which stages can run automatically, which need human judgment, and which should reshape the original process.

This report recommends defining deliverables as four things: runnable processes, acceptance data, permission and operation instructions, and maintenance and rollback methods. Model names are only part of it.

14.4 Individuals Should Keep Their Own Long-Term Work Assets

Persistent agents lower repeated-explanation costs, while potentially making work state dependent on one platform. Individuals should watch for file exportability, task-record portability, skill transferability, and permission revocability.

"Working on my behalf" should mean: being able to inspect actions, understanding which materials were used, and deciding which goals and resources stay authorized. An agent that always gives satisfying explanations but cannot be checked against results is not a reliable agent.

14.5 An Executable Eight-Week Validation Process

Below is an organizational pilot recommendation, D — not a Seed product performance guarantee.

Time Task Evidence to obtain
Weeks 1–2 Pick one high-frequency process, define real completion standards, collect historical tasks Human baseline, failure costs, permission boundaries
Weeks 3–4 Replay tasks with fixed-version models and runtime systems All successes/failures, intervention person-hours, total cost
Weeks 5–6 Change only one key variable: feedback, memory, tools, or model Attributable quality changes and counterexamples
Weeks 7–8 Blind-test on new tasks; test stopping, rollback, and migration Basis for deciding whether to expand use

Pilots do not need to first judge whether AGI has arrived. As long as a process has net gains and stays controllable, it is worth continuing; if not, use failure data to locate bottlenecks rather than explaining with grander technical visions.

XV. Key Unknowns and Final Judgment

15.1 Ten Things Not Yet Known

  1. Across all of Seed's R&D work, AI's actual coverage, person-hour substitution, and quality contribution.
  2. How much intervention, waiting, and engineering maintenance the officially described long-span tasks contain.
  3. How many AI-proposed algorithms or research plans pass independent verification and enter frontier training.
  4. Whether Seed for Seed has already consecutively affected multiple generations of frontier models, and the net acceleration magnitude.
  5. Experience distillation's transfer and forgetting behavior across broader domains, longer rounds, and stronger models.
  6. Multi-agent's long-term net gains versus same-budget single agents, and controllable concurrency scale.
  7. The latest closed-source models' complete training compute, parameter structures, and all key recipes.
  8. How Seed's available resources are allocated among training, inference, R&D agents, and products.
  9. All frontier capability risk thresholds, deployment trigger actions, and independent Control evaluation results.
  10. Whether world models, robotics, research, and general agents have already formed a unified learning system internally.

None of these questions can be filled with company visions, anonymous parameter rumors, job postings, or video demonstrations.

15.2 This Report's Final Judgment

ByteDance Seed has become an important subject for studying how AI continuously improves from usage experience. The evidence supporting this judgment is not one top-scoring model, but multiple public threads covering environmental learning, parameter retention, error recovery, dedicated verification, training systems, and official R&D use. [2][6][8][9][14][18][24][51][52]

This report further judges, D: it may have the conditions to connect research, applications, and industrial resources; real competitive advantage depends on whether these connections produce verified, retainable, transferable gains — not on each link expanding alone.

What can be established now: officially disclosed R&D loops, and multi-round capability accumulation in constrained experiments. What cannot yet be established: sustained recursive R&D takeoff of frontier general intelligence, or the whole system having surpassed large human organizations.

The inflection point most worth watching: whether verified new knowledge and R&D improvements per unit resource keep rising, and whether the next generation of systems is better than the last at obtaining such improvements. If such a sequence appears, the superintelligence-formation mechanism will move from possible connection to observable fact; if not, product and engineering achievements still have value, but should be evaluated at their true scope.

XVI. Source Directory and Verification Notes

All sources verified 2026-10-06. Dynamic-page dates are recorded as "cutoff snapshot"; papers additionally record first-public dates or versions used. Pages listed below support the specific arguments adjacent to them in the main text; they do not mean all site content was verified. Materials without full-text access are scope-noted; summaries are not used to infer internal details.

Official Model, Research, and Engineering Sources, B

Independent Measurement, External Research, and Authoritative Statistics, A

Industry and Platform Events, C

Transparency, Data, and New Experience Research, B

Source Usage and Review Records

This report specially re-checked six misreading types: model release dates vs. paper dates; early vs. revised paper versions; run spans vs. human-task horizons; training curves vs. formal evaluations; procurement plans vs. delivered resources; product-entry adjustments vs. overall capability discontinuation.

No wikis, anonymous community posts, or aggregator leaderboards were used as key technical-conclusion sources; no official blog and paper of the same study were passed off as two independent verifications; sources without accessible full text were not expanded into full readings; no inference about Seed experiencing the same incident was drawn from cross-company safety events.

The next version should prioritize independent replication of experience distillation, same-definition Seed horizons, R&D-cycle audits, and real workflow acceptance data. Only with such new evidence can the judgment strength on intelligence-growth mechanisms be substantially raised.