Synthetic Data Won't Save You From Model Collapse

The AI industry's running out of internet. Every major lab's already scraped the same corpus, and the easy gains from scaling data are tapering. The...

17 February 2026 · 19 min read

Synthetic Data Won't Save You From Model Collapse

{"version":"0.3.1","atoms":[],"cards":[["html",{"html":"<div style="background: linear-gradient(135deg, #1a1a2e 0%, #16213e 100%); border-radius: 12px; padding: 20px; margin: 20px 0; text-align: center;"><p style="color: #e94560; font-weight: bold; margin: 0 0 12px 0; font-size: 14px; letter-spacing: 2px;">LISTEN TO THIS ARTICLE

<audio controls="" preload="none" style="width: 100%; max-width: 500px;" src="https://swarmsignal.net/audio/synthetic-data-generation-for-ai-training-model-collapse-whe.mp3\">Your browser does not support the audio element.

"}],["image",{"src":"https://swarmsignal.net/content/images/2026/02/quote_synthetic-data-generation-for-ai-training-model-collapse-whe_1.webp","alt":"The threshold isn't precise, but the pattern is consistent: below 500 real patient records, synthetic generators produce clinically unrealistic artifacts."}],["image",{"src":"https://swarmsignal.net/content/images/2026/02/quote_synthetic-data-generation-for-ai-training-model-collapse-whe_2.webp","alt":"This suggests that the value of synthetic pretraining data might come from diversity rather than realism."}],["image",{"src":"https://swarmsignal.net/content/images/2026/06/quote_synthetic-data-generation-for-ai-training-model-collapse-whe_01.webp","alt":"Every major lab's already scraped the same corpus, and the easy gains from scaling data are tapering."}]],"markups":[["a",["href","https://swarmsignal.net/small-language-models-slms-vs-llms-for-ai-agents-efficient-o/"]],["a",["href","https://swarmsignal.net/the-goldfish-brain-problem-why-ai-agents-forget-and-how-to-fix-it/"]],["a",["href","https://swarmsignal.net/from-answer-to-insight-why-reasoning-tokens-are-a-quiet-revolution-in-ai/"]],["strong"],["a",["href","https://arxiv.org/abs/2602.11880v1","rel","nofollow"]],["a",["href","https://arxiv.org/abs/2602.12394v1","rel","nofollow"]],["a",["href","https://arxiv.org/abs/2602.10631v1","rel","nofollow"]],["a",["href","https://arxiv.org/abs/2602.11064v1","rel","nofollow"]],["a",["href","https://arxiv.org/abs/2602.12172v1","rel","nofollow"]],["a",["href","https://arxiv.org/abs/2602.10531v1","rel","nofollow"]],["a",["href","https://arxiv.org/abs/2602.10643v1","rel","nofollow"]],["a",["href","https://arxiv.org/abs/2602.11584v1","rel","nofollow"]],["a",["href","https://arxiv.org/abs/2602.11810v1","rel","nofollow"]],["a",["href","https://arxiv.org/abs/2602.12317v1","rel","nofollow"]],["a",["href","https://swarmsignal.net/mcp-server-architecture-guide/"]],["a",["href","https://swarmsignal.net/best-ai-red-teaming-tools-2026/"]],["a",["href","https://swarmsignal.net/agent-tool-use-patterns-guide/"]],["a",["href","https://swarmsignal.net/#/portal/signup"]],["a",["href","https://payhip.com/b/oq1HI?utm_source=swarmsignal&utm_medium=article_footer&utm_campaign=ss15"]]],"sections":[[10,0],[1,"p",[[0,[],0,"The AI industry's running out of internet. Every major lab's already scraped the same corpus, and the easy gains from scaling data are tapering. The instinct? Generate more training data synthetically. OpenAI does it. Anthropic does it. Meta's doing it. But here's what nobody wants to say out loud: training models on their own synthetic output creates a feedback loop that can degrade performance over time. The technical term is "model collapse," and it's showing up in production systems faster than anyone expected."]]],[1,"p",[[0,[],0,"A new paper from researchers studying clinical time series data found that generative models trained on moderate amounts of real patient data remain privacy-preserving, but the quality boundary is sharp. Push too hard on synthetic augmentation without fresh human data in the mix, and you get what statisticians call "evolutionary dynamics", iterative training on contaminated sources that drift away from ground truth. This isn't a theoretical concern anymore. It's happening in live deployments."]]],[10,3],[1,"h2",[[0,[],0,"Why Synthetic Data Became Mandatory"]]],[1,"p",[[0,[],0,"Real-world training data has three problems: it's expensive, it's biased, and there's not enough of it. Synthetic generation solves all three on paper. You can create millions of labeled examples overnight, control the distribution to reduce bias, and never worry about privacy compliance. Medical imaging teams use it to train CT scan artifact reduction models without exposing patient records. LLM personalization teams generate user interaction data to fine-tune models without collecting actual user behavior."]]],[1,"p",[[0,[],0,"The economics are compelling. A team at Stanford published work showing that motion capture data for action recognition could be replaced almost entirely with synthetic fractal-generated sequences. The models pretrained on fractals transferred to real-world deployment settings with minimal degradation. In controlled settings, synthetic data works."]]],[1,"p",[[0,[],0,"But "controlled settings" is doing a lot of work in that sentence."]]],[1,"p",[[0,[],0,"Production teams are discovering that synthetic data generation introduces subtle biases that don't show up in benchmark evaluations. A model trained on synthetic dialogue data might nail grammatical correctness while missing the contextual pragmatics that make conversation natural. The synthetic examples hit the statistical targets but miss the semantic ones."]]],[1,"p",[[0,[],0,"The cost structure matters too. Generating high-quality synthetic data isn't free. The computational expense of running a large generative model to produce training examples for a "],[0,[0],1,"smaller model"],[0,[],0," can exceed the cost of just training the larger model directly. Teams justify this by claiming they need the synthetic data for privacy or bias control, but the math often doesn't support the tradeoff."]]],[1,"h2",[[0,[],0,"The Collapse Mechanism Nobody Talks About"]]],[1,"p",[[0,[],0,"Model collapse isn't a bug. It's a statistical inevitability when you close the data generation loop. Here's the mechanism: a generative model learns the distribution of real data, then produces synthetic examples sampled from that learned distribution. Those synthetic examples get fed back into training the next generation of models. Each iteration introduces small errors, the model doesn't perfectly capture every detail of the original distribution. Over successive generations, those errors compound."]]],[1,"p",[[0,[],0,"Research from Bakshi and Chakraborty at the University of Southern California quantifies this drift. They studied iterative training on contaminated sources and found that without periodic injections of fresh real-world data, models experience what they call "evolutionary dynamics." The distribution shifts. Rare edge cases disappear from the training set because the generative model undersamples them. Common patterns get overrepresented because they're easier to generate."]]],[1,"p",[[0,[],0,"The degradation isn't linear. It's exponential. After three or four generations of purely synthetic retraining, model performance on real-world tasks can drop by double-digit percentages. I've now read four papers this month claiming synthetic data solves training scarcity, and none of them test beyond two iterative cycles."]]],[1,"p",[[0,[],0,"The mathematical underpinning is straightforward but brutal. Every generative model has finite capacity. It can't perfectly represent the full complexity of its training distribution. When you sample from that imperfect representation, you're sampling from a compressed version of reality. Each compression step loses information. The information loss compounds multiplicatively across training generations."]]],[1,"p",[[0,[],0,"This compounds with another problem: mode collapse in the generator itself. Generative models tend to concentrate probability mass on high-likelihood regions of the data distribution. The long tail gets truncated. When you train the next generation of models on this truncated distribution, those rare but important edge cases vanish entirely from the training corpus."]]],[1,"h2",[[0,[],0,"Domain Transfer is Where It Breaks"]]],[1,"p",[[0,[],0,"The part that actually worries me is domain transfer. Models trained on synthetic data often perform well on benchmarks that look like the synthetic distribution, then crater when deployed in the real world. A team studying motion representations found that pretraining on motion capture data, which is itself a synthetic proxy for human movement, didn't transfer reliably to deployment settings like wearable sensor data or uncalibrated video."]]],[1,"p",[[0,[],0,"The problem is distribution mismatch. Synthetic data generators don't know what they don't know. They can't generate edge cases they've never seen. A CT imaging model trained on synthetic ring artifacts will learn to remove the artifacts it was trained on, but when a new detector failure mode appears in production, the model has no prior for it."]]],[1,"p",[[0,[],0,"This is like training a self-driving car exclusively in a simulator that models sunny weather on well-marked roads, then deploying it in a snowstorm."]]],[1,"p",[[0,[],0,"The failure patterns in production are telling. Models trained on synthetic data tend to exhibit what researchers call "brittle competence", they perform at or near human level on distribution, but performance drops precipitously with even small distribution shifts. A customer service chatbot trained on synthetic dialogue data handles standard requests perfectly but completely fails on novel phrasings or unexpected edge cases."]]],[1,"p",[[0,[],0,"This brittleness shows up most clearly in safety-critical applications. Medical diagnostic models trained on synthetic patient data perform well on textbook cases but miss atypical presentations that an experienced clinician would catch. The synthetic generator learned the common patterns but couldn't generate the rare disease presentations that often matter most clinically."]]],[10,1],[1,"h2",[[0,[],0,"Knowledge Distillation's Dirty Secret"]]],[1,"p",[[0,[],0,"LLM distillation is the poster child for synthetic data dependency. The workflow is standard: take a large frontier model, have it generate synthetic question-answer pairs or reasoning traces, then train a smaller model on that output. It's how most production-deployed "small" models get built. But research from Bowei He's team at Zhejiang University found that current distillation methods produce synthetic data that's pedagogically inconsistent."]]],[1,"p",[[0,[],0,"The issue is coherence. When a large model generates synthetic training examples, it samples from its learned distribution. But that distribution isn't optimized for teaching. It's optimized for task performance. The synthetic examples often lack the progressive difficulty scaffolding that makes human-curated educational data effective."]]],[1,"p",[[0,[],0,"He's team introduced a pedagogically-inspired synthesis method that structures synthetic data like a lesson plan, easier examples first, harder edge cases later, with explicit reasoning scaffolding. Models trained on this structured synthetic data outperformed those trained on standard synthetic datasets by 12-18% on transfer tasks."]]],[1,"p",[[0,[],0,"The deeper problem with distillation is that it inherits the teacher model's biases and blindspots wholesale. If the large model has systematic failures on certain types of inputs, the distilled model will inherit those failures. But unlike the large model, which might have enough capacity to work around its blindspots in some contexts, the smaller distilled model lacks that flexibility."]]],[1,"p",[[0,[],0,"This creates a particularly insidious form of capability regression. Teams distill a large model to reduce inference costs, then discover the small model performs worse on edge cases. They assume this is just capacity loss and accept the tradeoff. But often it's not capacity, it's training data quality. The synthetic distillation data encoded the teacher's mistakes, and the student learned them as ground truth."]]],[1,"h2",[[0,[],0,"Privacy vs. Quality: The Medical Data Tradeoff"]]],[1,"p",[[0,[],0,"Healthcare AI teams face a brutal tradeoff. They can't share real patient data for training because of privacy regulations, but synthetic data generators trained on small datasets produce low-quality outputs. Research from Zhumagambetov's team at the intersection of clinical AI and privacy found that generative models need "moderate amounts" of real patient data to produce useful synthetic time series."]]],[1,"p",[[0,[],0,"The threshold isn't precise, but the pattern is consistent: below 500 real patient records, synthetic generators produce clinically unrealistic artifacts. Above 2,000 records, they start to memorize individual patients and violate privacy guarantees. The workable range is narrow, and it requires careful auditing."]]],[1,"p",[[0,[],0,"Perkonoja's team at the University of Turku developed evaluation metrics specifically for temporal preservation in synthetic longitudinal patient data. Their core finding: most synthetic generators fail to preserve the temporal dependencies that make time series data clinically meaningful. A synthetic patient's lab values over time might be individually plausible but collectively nonsensical when you account for disease progression dynamics."]]],[1,"p",[[0,[],0,"The regulatory environment makes this worse. HIPAA and GDPR don't provide clear guidance on what constitutes "sufficiently anonymized" synthetic data. Teams are left making judgment calls about privacy risk without clear standards."]]],[1,"p",[[0,[],0,"The clinical validation burden is also higher than teams expect. Even if synthetic patient data looks statistically similar to real data, clinicians need to validate that it's medically plausible before they'll trust models trained on it. This requires domain experts to manually review synthetic patient histories and flag physiologically impossible patterns. Most teams underbudget for this validation work and end up with synthetic datasets that pass automated quality checks but fail clinical review."]]],[1,"h2",[[0,[],0,"What Good Synthetic Data Actually Requires"]]],[1,"p",[[0,[],0,"The teams getting synthetic data right aren't treating it as a replacement for real data. They're treating it as augmentation with strict quality gates. The SynthRAR paper on CT ring artifact reduction is instructive. The team trained an unrolled network architecture on purely synthetic data, but they generated that synthetic data by simulating the exact physical defects that cause ring artifacts in real CT detectors."]]],[1,"p",[[0,[],0,"The synthetic data wasn't sampled from a learned distribution. It was generated from a physics-based model of detector failure modes. When the physics model accurately represents reality, the synthetic data transfers."]]],[1,"p",[[0,[],0,"Yuchen Ma's team working on LLM personalization took a similar approach. Instead of generating synthetic user interactions from a statistical model, they used a rule-based simulator that encoded explicit assumptions about user behavior patterns. The synthetic data had less diversity than purely generative approaches, but it transferred better because the underlying assumptions were transparent and auditable."]]],[1,"p",[[0,[],0,"This pattern repeats across domains. The synthetic data that works in production isn't the most realistic-looking. It's the most mechanistically grounded."]]],[1,"p",[[0,[],0,"The critical insight is that good synthetic data generation requires domain expertise, not just machine learning expertise. You need someone who understands the physics of CT detectors, or the psychology of user interactions, or the clinical progression of disease states. When teams try to generate synthetic data without domain expertise, they produce statistically plausible examples that violate domain constraints."]]],[1,"p",[[0,[],0,"This has implications for how teams should structure their synthetic data pipelines. The standard workflow is: collect real data, train generator, sample synthetic data, train downstream model. The better workflow is: define domain constraints, build physics/rule-based simulator, validate simulator output with domain experts, then use the simulator to generate training data. The generator becomes a tool for encoding expert knowledge, not a tool for extrapolating from limited real data."]]],[1,"h2",[[0,[],0,"The Evaluation Problem Nobody's Solved"]]],[1,"p",[[0,[],0,"Here's the catch: we don't have good metrics for synthetic data quality. Researchers typically evaluate synthetic data by training a model on it and measuring downstream task performance. But that's a lagging indicator. By the time you discover the synthetic data is contaminated, you've already wasted the compute budget training on it."]]],[1,"p",[[0,[],0,"The medical imaging teams are furthest ahead on this. They've developed privacy-preserving evaluation frameworks that measure synthetic data quality without requiring access to the original patient data. The metrics focus on distributional similarity, temporal coherence, and edge case coverage. But these frameworks are domain-specific."]]],[1,"p",[[0,[],0,"The LLM community is still using perplexity and downstream accuracy as proxies for synthetic data quality. Neither captures the pedagogical structure or distribution coverage that determines whether synthetic data will cause model collapse over iterative training cycles."]]],[1,"p",[[0,[],0,"The evaluation gap creates a coordination problem. Teams can't share synthetic data quality metrics across organizations because those metrics are tightly coupled to domain-specific validation criteria. This means every team is reinventing evaluation infrastructure from scratch."]]],[1,"p",[[0,[],0,"Some researchers are exploring meta-learning approaches to synthetic data evaluation. The idea is to train a separate model that predicts whether a given synthetic dataset will produce good downstream performance, without actually training the downstream model. Early results are promising but the meta-evaluator itself needs to be trained on examples of good and bad synthetic datasets, which brings you back to the original problem."]]],[1,"p",[[0,[],0,"The most practical near-term solution is probably ensemble-based evaluation. Generate multiple synthetic datasets using different generation methods, train multiple downstream models, and measure the variance in performance. High variance suggests the synthetic data is encoding artifacts specific to the generation method rather than true patterns from the underlying distribution."]]],[1,"h2",[[0,[],0,"Sharpness-Aware Training as a Hedge"]]],[1,"p",[[0,[],0,"One emerging mitigation comes from research on gradient compression. Yujie Gu's team found that training models with synthetic data guided by sharpness-aware minimization (SAM) reduces the risk of overfitting to synthetic artifacts. SAM explicitly penalizes models for converging to sharp minima in the loss terrain, which tends to happen when training on low-diversity synthetic data."]]],[1,"p",[[0,[],0,"The technique is computationally expensive, it roughly doubles training time, but it measurably improves generalization when the training data is partially or fully synthetic. Models trained with SAM on synthetic data retained 8-12% more performance on out-of-distribution test sets compared to standard training."]]],[1,"p",[[0,[],0,"This isn't a fix for model collapse. It's a hedge."]]],[1,"p",[[0,[],0,"The theory behind SAM is that sharp minima correspond to overfitting. A model that converges to a sharp minimum has learned to fit the training data precisely, including all its idiosyncratic artifacts. A model that converges to a flat minimum has learned a more general representation that's less sensitive to training data peculiarities."]]],[1,"p",[[0,[],0,"The practical implementation challenge is that SAM requires computing gradients twice per training step, once for the standard gradient update and once for the perturbed parameters used to estimate sharpness. This doubles the per-step cost. For large models, this makes training prohibitively expensive."]]],[10,2],[1,"h2",[[0,[],0,"The Fractal Approach to Pretraining"]]],[1,"p",[[0,[],0,"The most surprising result I've seen is from work on 3D fractals for action recognition pretraining. Marko Putak's team at Aalborg University generated synthetic training data using randomized fractal patterns, pure mathematical constructs with no semantic connection to real-world motion. They then pretrained action recognition models on these fractals before fine-tuning on real motion capture data."]]],[1,"p",[[0,[],0,"The fractal-pretrained models outperformed models pretrained on traditional synthetic motion data. The hypothesis: fractals provide better coverage of the space of possible motion patterns because they're generated from mathematical rules rather than learned distributions."]]],[1,"p",[[0,[],0,"This suggests that the value of synthetic pretraining data might come from diversity rather than realism. A model that's seen a wider range of unrealistic patterns might generalize better than one trained on a narrower range of realistic patterns."]]],[1,"p",[[0,[],0,"The fractal approach connects to broader questions about what pretraining is actually doing. The standard view is that pretraining teaches the model task-relevant features from the data distribution. But maybe what matters more is teaching the model to build flexible representations that can adapt to multiple distributions. Fractals force the model to learn representations that aren't overly specialized to any particular data source."]]],[1,"p",[[0,[],0,"There's a philosophical tension here. On one hand, we want synthetic data to be as realistic as possible so models learn the right patterns. On the other hand, perfectly realistic synthetic data might actually hurt generalization by narrowing the representational space the model explores during pretraining. The fractal work suggests that deliberately unrealistic synthetic data might be better for pretraining, as long as you follow it with fine-tuning on real data."]]],[1,"p",[[0,[],0,"The open question is whether this generalizes beyond action recognition. Fractals have nice mathematical properties for generating spatiotemporal patterns, which makes them natural for motion data. But the underlying principle, that diversity matters more than realism in pretraining data, probably does generalize."]]],[1,"h2",[[0,[],0,"The Hidden Cost of Synthetic Dependencies"]]],[1,"p",[[0,[],0,"Production teams are starting to hit a different problem: synthetic data creates technical debt. When you train a model on synthetic data, you inherit a dependency on the generator that produced it. If you need to retrain the model, you need to regenerate the synthetic data, which means you need access to the same generator with the same parameters."]]],[1,"p",[[0,[],0,"This sounds trivial until you're trying to debug a production model that's failing on edge cases and you realize the synthetic training data was generated three years ago by a contractor who's no longer available. You can't reproduce the data generation process because the code isn't documented and the random seeds weren't saved."]]],[1,"p",[[0,[],0,"The dependency problem gets worse when you're using synthetic data from a third-party provider. Some vendors now sell synthetic datasets as a service. The economics can work if you're a small team without the resources to generate your own synthetic data. But you're now dependent on the vendor's generation process, which is usually proprietary. This is similar to the pattern we've seen with agent memory systems, where "],[0,[1],1,"memory architecture decisions made early in development create compounding technical debt later"],[0,[],0,"."]]],[1,"p",[[0,[],0,"The mitigation is to treat synthetic data generation as infrastructure, not research. Document everything. Version control the generator code. Save the random seeds. Store the generation parameters. This sounds obvious, but most teams don't do it because synthetic data generation starts as an experiment."]]],[1,"h2",[[0,[],0,"When Synthetic Data Actually Works"]]],[1,"p",[[0,[],0,"Let's be clear about where synthetic data succeeds. It works when the generation process encodes known domain constraints, when you're using it for data augmentation rather than replacement, and when you have evaluation metrics that catch quality issues before deployment. These conditions don't hold for most use cases, but when they do, synthetic data can deliver significant performance gains."]]],[1,"p",[[0,[],0,"The medical imaging community has some of the best examples. Synthetic CT artifacts for training artifact reduction models work because the physics of CT detector failures is well understood. You can write down the equations that govern ring artifact formation and generate synthetic examples that exactly match real artifact patterns. The synthetic data isn't approximating reality, it's simulating the same physical process that produces real artifacts."]]],[1,"p",[[0,[],0,"In language models, synthetic data works best for tasks where correctness is verifiable. Code generation benefits from synthetic data because you can execute the generated code and check if it's correct. This gives you a quality signal for both the synthetic training data and the model's outputs. Math reasoning benefits from synthetic data for the same reason. This connects to recent work on "],[0,[2],1,"reasoning architectures where token-level verification provides training signals"],[0,[],0,"."]]],[1,"p",[[0,[],0,"The pattern is that synthetic data works when you have an oracle for correctness. The oracle can be physics equations, or a code interpreter, or a mathematical proof checker. When you lack that oracle, synthetic data quality becomes much harder to guarantee."]]],[1,"h2",[[0,[],0,"What This Actually Changes"]]],[1,"p",[[0,[],0,"Synthetic data isn't going away. The economics are too compelling, and the privacy requirements in regulated industries don't leave alternatives. But the era of treating synthetic data as a drop-in replacement for real data is ending. The teams seeing production success are the ones treating synthetic generation as a design problem, not a sampling problem."]]],[1,"p",[[0,[],0,"You need physics-based or rule-based simulators that encode explicit assumptions. You need quality metrics that catch distribution drift before it contaminates training. You need periodic injections of fresh real-world data to prevent model collapse. And you need architectures like sharpness-aware minimization that are resilient to synthetic artifacts."]]],[1,"p",[[0,[],0,"The failure mode isn't that synthetic data doesn't work. It's that it works well enough in controlled benchmarks to convince teams to deploy it, then degrades silently in production when the real-world distribution drifts away from the synthetic training distribution."]]],[1,"p",[[0,[],0,"Model collapse is real, it's measurable, and it's already affecting deployed systems. It's not about whether synthetic data has value, it's about understanding its boundaries and building systems that account for its limitations. The teams that get this right will have a competitive advantage. The teams that don't will spend the next two years debugging mysterious performance regressions and wondering why their models stopped working."]]],[1,"h2",[[0,[],0,"Sources"]]],[1,"p",[[0,[3],1,"Research Papers:"]]],[3,"ul",[[[0,[4],1,"SynthRAR: Ring Artifacts Reduction in CT with Unrolled Network and Synthetic Data Training"],[0,[],0,", Yang et al. (2026)"]],[[0,[5],1,"Synthetic Interaction Data for Scalable Personalization in Large Language Models"],[0,[],0,", Ma et al. (2026)"]],[[0,[6],1,"Generative clinical time series models trained on moderate amounts of patient data are privacy preserving"],[0,[],0,", Zhumagambetov et al. (2026)"]],[[0,[7],1,"Motion Capture is Not the Target Domain: Scaling Synthetic Data for Learning Motion Representations"],[0,[],0,", Darwish et al. (2026)"]],[[0,[8],1,"Pedagogically-Inspired Data Synthesis for Language Model Knowledge Distillation"],[0,[],0,", He et al. (2026)"]],[[0,[9],1,"From Collapse to Improvement: Statistical Perspectives on the Evolutionary Dynamics of Iterative Training on Contaminated Sources"],[0,[],0,", Bakshi & Chakraborty (2026)"]],[[0,[10],1,"Evaluation metrics for temporal preservation in synthetic longitudinal patient data"],[0,[],0,", Perkonoja et al. (2026)"]],[[0,[11],1,"Gradient Compression May Hurt Generalization: A Remedy by Synthetic Data Guided Sharpness Aware Minimization"],[0,[],0,", Gu et al. (2026)"]],[[0,[12],1,"How to Sample High Quality 3D Fractals for Action Recognition Pre-Training?"],[0,[],0,", Putak et al. (2026)"]],[[0,[13],1,"Free Lunch in Medical Image Foundation Model Pre-training via Randomized Synthesis and Disentanglement"],[0,[],0,", Wei et al. (2026)"]]]],[1,"p",[[0,[3],1,"Related Swarm Signal Coverage:"]]],[3,"ul",[[[0,[2],1,"From Answer to Insight: Why Reasoning Tokens Are a Quiet Revolution in AI"]],[[0,[1],1,"The Goldfish Brain Problem: Why AI Agents Forget and How to Fix It"]]]],[1,"h3",[[0,[],0,"Keep reading"]]],[3,"ul",[[[0,[14],1,"MCP Server Architecture Guide"]],[[0,[15],1,"Best AI Red-Teaming Tools (2026)"]],[[0,[16],1,"Agent Tool-Use Patterns Guide"]]]],[1,"p",[[0,[17],1,"Join the Swarm Signal newsletter"]]],[1,"p",[[0,[18],1,"Get the Freelance Command Center on Payhip"]]]]}

$Marko Putak's team at Aalborg University generated synthetic training data using randomized fractal patterns, pure mathematical constructs with no semantic connection to real-world motion.$

Key takeaways

The AI industry's running out of internet. Every major lab's already scraped the same corpus, and the easy gains from scaling data are tapering. The...

Swarm Signal

Up Next

Queue is empty. Click "+ Queue" on any article to add it.

Key takeaways

Make better agent decisions each week