Fohr

A conversation on the intricacies of the Web3 landscape, the pressing issues facing the industry, Web3 data and much more with Marcel Fohrmann.

The Future of Web3 Data: An Interview with Subsquid CFO & Co-Founder Marcel Fohrmann

We're sharing 5 approaches to target your intended interest groups so you can gain more quality Instagram followers in 2021.

Creating Your Instagram Strategy To Gain Quality Followers

I’ve organized and executed thousands of influencer marketing campaigns for brands like <a href="https://www.vogue.com/article/depop-millennial-friendly-shopping-app">Depop</a> and my own small businesses, (<a href="https://www.instagram.com/pinlord/">@pinlord</a>, <a href="https://www.instagram.com/potteryforall/">@potteryforall</a>, <a href="https://www.instagram.com/macramemakers/">@macramemakers</a>) and from experience, I know that the success of each one depends almost exclusively on the quality of influencers that you work&nbsp;with.

How To Search for and Find The Most Valuable Influencers on Instagram

Why LLM applications need an AI control plane for observability, policy enforcement, cost control, and safe scaling in production.

The AI Control Plane: Observability, Cost Governance, and Policy for LLM Applications

Agentic reasoning reframes LLMs as autonomous agents that plan, act, and learn—unlocking real-world applications beyond static text prediction.

Notes on Agentic Reasoning for Large Language Models

Why streaming video overwhelms AI memory—and how HERMES uses hierarchical KV caching to enable fast, accurate real-time video understanding.

HERMES Rewrites the Rules of Streaming Video for Multimodal AI

A plain-English breakdown of Qwen3-TTS, explaining how tokenized audio enables efficient, real-time speech generation with large language models.

Qwen3-TTS and the Case for Token-Based Speech Synthesis

Zero Trust Security focuses on not believing on any user, any device or anyone even if they are in your network perimeter.

Zero-Trust Security in 2026: A Complete Implementation Roadmap for CTOs

This is a simplified guide to an AI model called <a href="https://www.aimodels.fyi/models/replicate/talknet-asd-zsxkib?utm_source=hackernoon&utm_medium=referral">talknet-asd</a> maintained by <a href="https://www.aimodels.fyi/creators/replicate/zsxkib?utm_source=hackernoon&utm_medium=referral">zsxkib</a>. If you like these kinds of analysis, join <a href="https://www.aimodels.fyi/?utm_source=hackernoon&utm_medium=referral">AIModels.fyi</a> or follow us on <a href="https://x.com/aimodelsfyi">Twitter</a>.

<h2>Model overview</h2>
<code>talknet-asd</code> is an audio-visual active speaker detection model that identifies whether a face in a video is speaking. Built on research presented at ACM MM 2021, this model combines visual and audio information to determine speaker activity with high accuracy. Unlike models such as <a href="https://aimodels.fyi/models/replicate/video-retalking-xiankgx?utm_source=hackernoon&utm_medium=referral">video-retalking</a> that focus on lip synchronization or <a href="https://aimodels.fyi/models/replicate/multitalk-zsxkib?utm_source=hackernoon&utm_medium=referral">multitalk</a> that generate multi-person conversations, this model solves the fundamental problem of detecting who is actively speaking in existing footage.
<h2>Model inputs and outputs</h2>
The model accepts video files and processes them using configurable parameters to detect active speakers. It returns both visual output with marked speakers and structured data about detection results. The flexibility of the input parameters allows users to adjust detection sensitivity and processing scale based on their specific video characteristics.
<h3>Inputs</h3>
<ul>
<li>Video - Path to the video file (MP4 or AVI format)</li>
<li>Face Det Scale - Scale factor for face detection; lower values process faster but may miss smaller faces (default: 0.25)</li>
<li>Min Track - Minimum number of frames required to establish a continuous speaker track (default: 10)</li>
<li>Num Failed Det - Number of missed detections allowed before tracking stops (default: 10)</li>
<li>Min Face Size - Minimum face size in pixels to consider for detection (default: 1)</li>
<li>Crop Scale - Bounding box scale for extracted face regions (default: 0.4)</li>
<li>Start - Start time in seconds for processing (default: 0)</li>
<li>Duration - Video duration to process; -1 processes the entire video (default: -1)</li>
<li>Return Boundingbox Percentages - Option to return coordinates as percentages of video dimensions (default: false)</li>
<li>Return Json - Whether to return results in JSON format (default: true)</li>
</ul>
<h3>Outputs</h3>
<ul>
<li>Media Path - Array of output video file paths with speaker annotations</li>
<li>Json Str - Structured detection results in JSON format containing speaker timing and spatial information</li>
</ul>
<h2>Capabilities</h2>
The model detects active speakers across video sequences with 96.3% average F1 score on standard benchmarks. It marks speaking faces with green bounding boxes and non-speaking faces with red boxes. The model handles challenges like varying lighting conditions, multiple faces in frame, and continuous speaker tracking across scenes. It works on videos in the wild without requiring controlled studio conditions.
<h2>What can I use it for?</h2>
Extract active speaker information from video content for downstream applications like automated video editing, speaker identification in multi-participant videos, or dialogue system training. Content creators can use detection results to automatically highlight who is speaking in podcasts, interviews, or meetings. Researchers developing conversational AI systems similar to those explored in <a href="https://aimodels.fyi/papers/arxiv/talkingmachines-real-time-audio-driven-facetime-style?utm_source=hackernoon&utm_medium=referral">real-time audio-driven face generation</a> can leverage speaker detection as a preprocessing step. Security and surveillance applications can identify speaking activity in footage analysis.
<h2>Things to try</h2>
Experiment with the face detection scale parameter to balance accuracy and speed on different video resolutions. Test on multi-speaker scenarios where multiple people appear in frame simultaneously to see how the model prioritizes detection. Try processing videos with varying lighting conditions and camera angles to understand robustness. Adjust the minimum track parameter to control sensitivity to brief speaking moments versus sustained speech. Use the bounding box percentage output option when working with videos of different resolutions to ensure consistency across your pipeline.

<hr/>Original post: <a href="https://www.aimodels.fyi/models/replicate/talknet-asd-zsxkib?utm_source=hackernoon&utm_medium=referral">Read on AIModels.fyi</a>

TalkNet-ASD: The “Who’s Talking?” Model for Any Video

OpenVision 3 introduces a unified visual encoder that supports both image understanding and generation, reducing redundancy across vision AI systems.

OpenVision 3 Challenges the Need for Separate Vision and Image Generation Models

An architectural guide to running Java JARs inside .NET 8, explaining when process-level integration outperforms API-based interoperability.


An Architectural Pattern for Running Java Components Inside .NET 8

Claude Book is an orchestrated writing system using Claude Code. It uses subagents for consistency checks and a perplexity gate with rewriting against AI-slope.

Claude Book: A Multi-Agent Framework for Writing Novels with Claude Code

Why today’s AI agents don’t scale—and how a microservices approach could unlock interoperable, secure, multi-agent systems.

Stop Building Monolithic Agents: The Case for a Microservices-First AI Stack

Discover the best financial APIs for 2026. Compare top stock, forex, and crypto data APIs for real time insights, analysis, and fintech innovation.

Best Financial APIs for 2026 

Fiverr’s acquisition of Yaballe expands its AI-powered dropshipping ecosystem, signaling a shift toward full lifecycle automation for sellers.

After Fiverr Acquired AutoDS and Now Yaballe, a New Era for Dropshippers Is Coming

This is a Plain English Papers summary of a research paper called <a href="https://www.aimodels.fyi/papers/arxiv/hy-motion-10-scaling-flow-matching-models?utm_source=hackernoon&utm_medium=referral">HY-Motion 1.0: Scaling Flow Matching Models for Text-To-Motion Generation</a>. If you like these kinds of analysis, join <a href="https://www.aimodels.fyi/?utm_source=hackernoon&utm_medium=referral">AIModels.fyi</a> or follow us on <a href="https://x.com/aimodelsfyi">Twitter</a>.

<h2>The motion generation problem</h2>
For years, generating realistic human motion from text descriptions has felt stuck. Current models either fail to understand what you're asking for or produce movement that looks jerky and unnatural. Ask for an "angry walk toward a door," and the model might generate walking that's roughly the right speed but misses the emotional quality. Ask for something specific like "athletic jump with both arms extended," and it often collapses entirely. The fundamental challenge is that motion has temporal structure, physical constraints, and an almost infinite solution space. Unlike generating a static image where pixels either look right or wrong, motion requires the model to understand not just the shape of movement, but how emotion deforms it, how intention curves trajectories, and how multiple text concepts combine into a single coherent sequence.
This is why every model released so far has struggled with instruction-following. They catch maybe 70% of what you asked for and miss the nuance. The problem isn't that researchers don't understand the algorithms well enough. The bottleneck is something deeper: models trained at small scale simply don't develop the ability to understand and follow detailed instructions the way language models or image generators do.
<h2>The scaling hypothesis</h2>
The past five years of AI progress have been driven almost entirely by scaling. GPT-2 at 1.5 billion parameters could barely write coherent paragraphs. Increase that scale tenfold and something shifts. The model doesn't just do the same thing slightly better. It develops new capabilities. It reasons about edge cases it never encountered. It understands nuance and context in ways that feel qualitatively different from smaller versions.
The question for motion generation is straightforward: does this pattern hold? Or is something fundamentally different about this problem that makes scaling unhelpful?
HY-Motion answers that question by testing the hypothesis directly. Build a billion-parameter motion generation model and train it properly, and it develops instruction-following capabilities that smaller models never achieve. A small model learns to generate common motions competently. A billion-parameter model learns to listen to instructions, to combine concepts flexibly, to handle rare motion combinations and specific constraints. The research reveals that motion generation follows the same scaling laws as language and image generation, but only under one crucial condition: you need the right training data and the right training strategy.
<h2>Building the right foundation</h2>
Scaling only works if you have high-quality data to scale on. This is the unsexy part of the paper, the part many researchers skip, but it's actually where much of the breakthrough lives.
The fundamental problem is that motion datasets are messy. Raw motion capture contains jitter and artifacts from the recording process. Text descriptions are often vague or incorrect. Without cleaning, any model trained on this noise learns garbled patterns. HY-Motion treats data as a first-class problem.
<a href="https://arxiv.org/html/2512.23464/x2.png"><img src="https://arxiv.org/html/2512.23464/x2.png" alt="Data processing pipeline overview" /></a> Data processing pipeline overview
The data processing pipeline shows how raw motion capture data flows through cleaning, annotation, and quality control stages
The processing pipeline performs rigorous motion cleaning to remove artifacts and temporal inconsistencies. Careful captioning ensures text actually describes the motion rather than being generic labels. The team then organized motions into a hierarchical structure that gives the model rich conceptual structure to learn from.
<a href="https://arxiv.org/html/2512.23464/x3.png"><img src="https://arxiv.org/html/2512.23464/x3.png" alt="Motion category hierarchy spanning 200+ categories across 6 major classes" /></a> Motion category hierarchy spanning 200+ categories across 6 major classes
The hierarchy shows how motions are organized: 6 major classes branch into 200+ specific categories, giving the model granular conceptual structure
This hierarchical organization isn't arbitrary. It reflects how motion actually structures itself in human understanding. The model learns not just individual motions, but relationships between them. How does walking differ from running? How does emotion modulate both? The cleaned dataset spans over 3,000 hours of motion data, and another 400 hours gets reserved for high-quality fine-tuning. This foundation is what makes scaling meaningful. Without it, you'd train a billion-parameter model on garbage.
<h2>The three-stage training recipe</h2>
Large-scale models are brittle without the right training strategy. HY-Motion's real innovation is the systematic approach that stabilizes and aligns them. This recipe mirrors how state-of-the-art language models are trained, bringing proven techniques from another domain into motion generation.
Pretraining on large-scale unsupervised data is where the model learns the grammar of motion. Show it 3,000 hours of diverse human movement, and it learns that walking has rhythm, that reaching has velocity constraints, that expressions change how bodies move. The model isn't trying to follow instructions yet. It's learning what plausible motion looks like, building rich internal representations of movement dynamics.
High-quality fine-tuning comes next. After pretraining, take that much smaller dataset of 400 exceptionally curated motion-text pairs. The model has already learned how bodies move. Now it learns to listen to what's being asked. This stage is where instruction-following gets tight. You want perfect alignment between text and motion here, so the model learns what precision feels like. This is the domain expert phase, where a smaller dataset of perfect examples shapes the model's behavior.
Reinforcement learning from both human feedback and reward models completes the loop. Rather than assuming fine-tuning labels are perfect, the paper adds a feedback mechanism. Humans and learned reward models evaluate whether generated motions actually match text descriptions. The model then learns to optimize against this feedback. This is where the model learns to correct its own mistakes and push past previous limitations.
<a href="https://arxiv.org/html/2512.23464/x4.png"><img src="https://arxiv.org/html/2512.23464/x4.png" alt="Overview of the HY-Motion framework showing all three training stages" /></a> Overview of the HY-Motion framework showing all three training stages
The framework diagram shows how pretraining, fine-tuning, and RLHF stages feed into each other, creating a complete training pipeline
This three-stage approach is the insight that transfers beyond motion generation. Pretraining on massive unsupervised data creates general capability. Fine-tuning on curated data tightens behavior. Alignment via feedback ensures the model does what humans actually want. The same pattern has worked for language models and shows similar promise here.
<h2>The architecture that makes scaling work</h2>
All of this depends on an architecture that actually scales well. HY-Motion uses flow matching combined with a Transformer-based architecture called DiT (Diffusion Transformer).
Flow matching is a newer generative modeling approach that's simpler and more efficient than traditional diffusion. Instead of gradually adding noise and learning to reverse it, flow matching learns smooth paths from simple distributions directly to data. When combined with a Transformer architecture proven to work at billion-parameter scale in other domains, you get something that scales beautifully. Transformers handle sequential data naturally, and flow matching gives you cleaner gradients and better sample efficiency than alternatives.
<a href="https://arxiv.org/html/2512.23464/x5.png"><img src="https://arxiv.org/html/2512.23464/x5.png" alt="Model architecture of HY-Motion DiT" /></a> Model architecture of HY-Motion DiT
The architecture diagram shows how transformer layers process motion tokens while the flow matching objective guides generation
The choice matters because it determines whether scaling actually helps or just makes things slower. Flow matching is the right algorithmic fit for motion, the way transformer architectures were the right fit for language.
<h2>Breaking the ceiling</h2>
After building the data foundation, designing the training recipe, and choosing the right architecture, the question is simple: does it work?
<a href="https://arxiv.org/html/2512.23464/x1.png"><img src="https://arxiv.org/html/2512.23464/x1.png" alt="Comparison of HY-Motion 1.0 to state-of-the-art models including DART, LoM, GoToZero, and MoMask" /></a> Comparison of HY-Motion 1.0 to state-of-the-art models including DART, LoM, GoToZero, and MoMask
Top: Quantitative comparison shows HY-Motion outperforming current open-source benchmarks. Bottom: Visual examples of HY-Motion outputs demonstrate natural and precise motion generation
The difference between HY-Motion and previous models isn't incremental. It's qualitative. Previous models generate motion that vaguely resembles what you asked for, but specificity breaks them. Ask for an angry walk with both arms down, and they struggle. Ask for motion with fine-grained athletic requirements, and they often fail. HY-Motion actually listens.
<a href="https://arxiv.org/html/2512.23464/x6.png"><img src="https://arxiv.org/html/2512.23464/x6.png" alt="Visual comparison examples showing HY-Motion versus state-of-the-art alternatives" /></a> Visual comparison examples showing HY-Motion versus state-of-the-art alternatives
Side-by-side examples show how HY-Motion produces more natural, anatomically correct motion that more precisely matches text descriptions
The model has developed instruction-following capabilities that smaller models never achieved. It understands motion at the level of detail that matters for real applications. The coverage is extensive too, spanning over 200 motion categories across 6 major classes. This breadth is what gives the model flexibility to handle novel combinations of concepts.
<h2>Implications for the future</h2>
HY-Motion demonstrates something important: motion generation follows the same scaling laws as language and image generation. This matters beyond motion specifically. It shows that when you pick a new domain and want to unlock new capabilities, the path is predictable. Gather massive amounts of data. Clean it meticulously. Train at scale. Refine with human feedback.
This also signals where motion generation is heading. For years it's been a research curiosity, interesting but not yet practical. HY-Motion represents a transition point where the technology is starting to move toward commercial maturity. That's not because of a fundamental algorithmic breakthrough, but because someone built the data pipeline and training recipe that makes scaling work.
For researchers working in other domains where scaling hasn't yet unlocked capabilities, the lesson is clear: look first at your data and training strategy before designing new architectures. The bottleneck is usually not the algorithm.
The open-source release of HY-Motion is significant too. Previous motion generation work lived in closed research codebases. Making this available accelerates the entire field. Other researchers can build on this foundation, test new ideas, and push the boundary forward faster than any single lab could alone.
The broader pattern here connects to work on <a href="https://aimodels.fyi/papers/arxiv/text-driven-motion-generation-overview-challenges-directions?utm_source=hackernoon&utm_medium=referral">text-driven motion generation</a> that identified core challenges in the field, and research like <a href="https://aimodels.fyi/papers/arxiv/go-zero-towards-zero-shot-motion-generation?utm_source=hackernoon&utm_medium=referral">GoToZero</a> and <a href="https://aimodels.fyi/papers/arxiv/omnimogen-unifying-human-motion-generation-via-learning?utm_source=hackernoon&utm_medium=referral">OmnimoGen</a> that explored different architectural approaches. HY-Motion doesn't replace those efforts. It builds on them, showing that the scaling hypothesis holds when you get the fundamentals right.

<hr/>Original post: <a href="https://www.aimodels.fyi/papers/arxiv/hy-motion-10-scaling-flow-matching-models?utm_source=hackernoon&utm_medium=referral">Read on AIModels.fyi</a>

The Case for Bigger Models in Human Motion AI

How Akshatha Madapura Anantharamu builds transparent, high-performance frontend systems that make AI trustworthy and usable at scale.

How Akshatha Madapura Anantharamu Is Building Trustworthy Interfaces for AI Systems

Predictive models now rule over modern analytics stacks from recommendation engines to demand forecasting and fraud detection.

Causal Thinking in the Age of Big Data: Modern Econometrics for Data Scientists

This is a cybersecurity class for senior citizens.

A Class For Mom Part 2: Cybersecurity

Aravind Barla led a Fortune 100 enterprise transformation, unifying service delivery with AI to boost efficiency, scale, and employee experience.

Aravind Barla Rebuilt Enterprise Service Delivery for 100,000+ Employees

This is a Plain English Papers summary of a research paper called <a href="https://www.aimodels.fyi/papers/arxiv/generating-piano-music-transformers-comparative-study-scale?utm_source=hackernoon&utm_medium=referral">Generating Piano Music with Transformers: A Comparative Study of Scale, Data, and Metrics</a>. If you like these kinds of analysis, join <a href="https://www.aimodels.fyi/?utm_source=hackernoon&utm_medium=referral">AIModels.fyi</a> or follow us on <a href="https://x.com/aimodelsfyi">Twitter</a>.

<h2>The Turing test for music</h2>
Before diving into models and architectures, there's a more fundamental question hiding beneath all the technical choices: what does it actually mean to succeed at music generation?
For decades, the answer was easy. AI-generated music sounded robotic, lifeless, predictable in ways you'd notice immediately. But something has shifted. Researchers have begun asking whether an algorithm could fool a listener in a blind test. Not whether it matches some mathematical criterion, but whether a person sitting down to listen could tell the difference between a piece composed by a human and one generated by a machine.
This is the north star of the research. Everything else serves this single question. And the answer turns out to be surprising: when built correctly, a transformer trained on enough diverse piano data can produce music that humans genuinely struggle to distinguish from human composition.
The researchers ran a Turing-style listening study where participants heard piano pieces and had to guess whether each was human or AI-composed. For the best-performing model, results hovered near 50% accuracy. That's not a typo. It means listeners were essentially guessing, unable to confidently separate human from machine.
<a href="https://arxiv.org/html/2511.07268/images/turing_confusion_matrix.png"><img src="https://arxiv.org/html/2511.07268/images/turing_confusion_matrix.png" alt="Confusion matrix from the musical Turing test showing listener accuracy in distinguishing AI-composed from human-composed pieces across all genres" /></a> Confusion matrix from the musical Turing test showing listener accuracy in distinguishing AI-composed from human-composed pieces across all genres
The confusion matrix reveals something elegant: the model crossed a genuine threshold. Not because researchers found a clever trick, but because they systematically understood what actually moves the needle on the one metric that matters.
<h2>The problem with metrics that lie to you</h2>
But here's the uncomfortable truth that most music generation research glosses over: the quantitative metrics researchers have relied on for years don't predict what humans actually think.
Imagine you're evaluating music using a standard metric like prediction accuracy. The model learns to predict the next note correctly with high confidence. That's measurable. That's publishable. But high accuracy doesn't guarantee the music feels alive or interesting. It might just mean the model learned to play it safe, following familiar patterns. The best human composers break rules. They surprise listeners. Those surprises might actually lower the model's prediction accuracy while raising its artistic value.
This gap between metrics and reality creates a trap: optimize for the wrong numbers and you build a model that scores well on paper but fails in practice. The research here cuts through that noise by comparing quantitative metrics directly against human judgments collected in the listening study. Which metrics actually correlate with listeners thinking the music sounds human? Which ones mislead you into optimizing the wrong thing?
This reframing changes how you should think about model development. Instead of treating all metrics equally, you can ask: does my model score higher on these metrics than lower-performing models? If not, that metric isn't measuring what I care about. If yes, I should probably care more about it. It's a simple idea but one that's missing from much of the music generation literature.
<h2>Three levers, three tradeoffs</h2>
Now comes the actual research question: what makes one music generation model better than another?
The researchers didn't build a single system and declare victory. They systematically varied three independent dimensions and compared results. First, model size. How many parameters does the transformer need? Second, training data. What dataset should you train on, and how much does diversity matter compared to curation? Third, training strategy. Should you train from scratch, or pre-train on one dataset then fine-tune on another?
Each lever has tradeoffs. Bigger models demand more compute. Larger datasets take longer to collect and clean. Different training strategies require different computational schedules. The real question isn't which lever is "best" in the abstract, but which gives you the most improvement for the effort and cost you're willing to spend.
This systematic approach separates this work from the typical narrative in deep learning research, where the implicit assumption is always "bigger equals better." The researchers built many models, compared them carefully, and asked what actually matters. That's methodologically sound, and it yields insights that directly help practitioners make real tradeoffs.
<h2>Why messy data beats curated perfection</h2>
Start with the data question. The researchers compared two main sources: MAESTRO, which is meticulously curated with about 200 hours of recordings and high-quality MIDI annotations, and Aria-Deduped, which is much larger and messier with around 80K MIDI files from diverse genres and sources.
The conventional wisdom in machine learning suggests you should invest in data quality. Carefully curated datasets reduce noise and bias. They represent "gold standard" examples. But the listening study results tell a different story. Models trained on the larger, more diverse Aria dataset outperformed models trained on the smaller MAESTRO dataset.
This is where the paper's central insight emerges: bigger data beats bigger models. But that's only half the story. The critical detail is diversity. The Aria dataset contains piano music from many genres, styles, and recording conditions. It's messier precisely because it's broader. A model trained on that breadth learns richer statistical patterns than one trained on pristine examples from a narrow slice of music.
Think about how language models develop intuition. They don't learn by studying 10,000 perfectly edited sentences in one style. They learn by reading millions of texts written by millions of people in different contexts, tones, and registers. The variety teaches the model what varies and what stays constant across language. The same principle applies to music. Diverse data teaches the model what's essential to the grammar of piano music versus what's incidental to a particular recording or performance style.
<a href="https://arxiv.org/html/2511.07268/images/supplementary_training-curves_2.png"><img src="https://arxiv.org/html/2511.07268/images/supplementary_training-curves_2.png" alt="Training curves comparing models pre-trained on Aria-Deduped against the baseline model trained only on MAESTRO, showing faster convergence and better final performance from the larger dataset" /></a> Training curves comparing models pre-trained on Aria-Deduped against the baseline model trained only on MAESTRO, showing faster convergence and better final performance from the larger dataset
The gap widens as you train longer. The model pre-trained on Aria continued improving on downstream tasks even with limited fine-tuning data, while the MAESTRO-only model plateaued earlier. This has a practical implication: if you're building a music generation system and you can either spend time curating a smaller dataset or gathering more diverse data (even if it's rougher), choose the latter.
<h2>When bigger models hit diminishing returns</h2>
Model scaling is the gospel of modern deep learning. Build a bigger model, throw more compute at it, and performance improves. This has held true across language, vision, and many other domains. But even a good principle has limits.
The researchers compared transformers of different sizes, all trained on the same datasets. A 155M-parameter model. A 355M-parameter model. A 550M-parameter model. A 950M-parameter model. As you'd expect, larger models generally performed better in pre-training. But the gains diminish. Each doubling of model size gives increasingly smaller improvements.
More importantly, the interaction between model size and data size matters. A 950M-parameter model trained on large, diverse data beats a smaller model trained on the same data. But a larger model trained on small, curated data doesn't dramatically outperform a smaller model trained on the same limited data. The model gets larger but has fewer diverse examples to learn from, so it ends up memorizing rather than generalizing. You've wasted parameters on overfitting.
<a href="https://arxiv.org/html/2511.07268/images/supplementary_training-curves_1.png"><img src="https://arxiv.org/html/2511.07268/images/supplementary_training-curves_1.png" alt="Training loss curves for MAESTRO models of different sizes, showing how performance gains diminish as model size increases when training data is limited" /></a> Training loss curves for MAESTRO models of different sizes, showing how performance gains diminish as model size increases when training data is limited
This tells you something crucial: find the sweet spot where your model has enough capacity to express patterns in your data without being so oversized that it memorizes. For the 80K Aria dataset, that sweet spot appears to be in the 950M range. A 155M model trained on the same data was close behind, suggesting the marginal benefit of going bigger is limited.
The practical insight: don't automatically assume you need to build the largest possible model. You need a model large enough that data diversity, not model capacity, is your bottleneck. Build bigger than that and you're spending compute on diminishing returns.
<h2>The efficiency of learning from experts after learning broadly</h2>
Training strategy introduces another variable. What if you didn't train only on one dataset? What if you pre-trained on large, diverse data to learn the fundamentals of piano music, then fine-tuned on smaller, curated expert data to learn the specific style or qualities you care about?
This is where the pre-training and fine-tuning paradigm reveals its power. A model trained only on MAESTRO learns from limited examples. It's slow to train because there's not much data. But a model pre-trained on the 80K Aria files, then fine-tuned on MAESTRO, learns rapidly. The pre-training phase captures broad patterns of piano music. The fine-tuning phase refines those patterns toward MAESTRO's particular characteristics.
<a href="https://arxiv.org/html/2511.07268/images/supplementary_training-curves_3.png"><img src="https://arxiv.org/html/2511.07268/images/supplementary_training-curves_3.png" alt="Training curves for models pre-trained on Aria-Deduped and fine-tuned on MAESTRO, showing rapid convergence and superior performance compared to training from scratch" /></a> Training curves for models pre-trained on Aria-Deduped and fine-tuned on MAESTRO, showing rapid convergence and superior performance compared to training from scratch
This matters because it solves a real problem in music generation: the highest-quality datasets are small and expensive to create, while large datasets are often noisier. Pre-training lets you leverage both. You gather whatever diverse data you can find, train on it to establish a foundation, then apply limited expert data where it counts most.
The fine-tuning experiments with different context window sizes (how much previous music the model can "see" when generating the next note) further confirm this principle. A model that started with broader pre-training learned to use context windows more effectively during fine-tuning than a model starting from scratch.
<a href="https://arxiv.org/html/2511.07268/images/Moonbeam_Curves_CTX_512.png"><img src="https://arxiv.org/html/2511.07268/images/Moonbeam_Curves_CTX_512.png" alt="Fine-tuning curves for models with different context sizes, showing how architectural choices interact with training strategy" /></a> Fine-tuning curves for models with different context sizes, showing how architectural choices interact with training strategy
The practical takeaway: if you have access to limited high-quality data in a specific domain or style, don't waste it training from scratch. Pre-train broadly first, then fine-tune precisely. You'll converge faster and achieve better results with less total training time.
<h2>What humans actually hear</h2>
All these choices, all this careful ablation, point toward one question: can the best model actually fool human listeners?
The researchers collected human judgments through a listening study where participants heard pieces of piano music and had to classify each as human or AI-composed. They didn't tell listeners whether they were comparing against the best model or a baseline. They just asked: does this sound human?
The confusion matrix shows the results. The best-performing model, a 950M-parameter transformer trained on 80K diverse MIDI files, achieved near 50% accuracy in the classification task. That's the threshold where humans can't reliably tell the difference.
<a href="https://arxiv.org/html/2511.07268/images/turing_accuracy_by_genre.png"><img src="https://arxiv.org/html/2511.07268/images/turing_accuracy_by_genre.png" alt="Accuracy of listener classifications broken down by genre, showing which genres the model fools most and least convincingly" /></a> Accuracy of listener classifications broken down by genre, showing which genres the model fools most and least convincingly
But accuracy varies by genre. Some musical styles are easier to fool humans with than others. Looking at accuracy separated by both genre and whether the piece was AI or human reveals the model's strengths and weaknesses.
<a href="https://arxiv.org/html/2511.07268/images/turing_accuracy_by_genre_origin.png"><img src="https://arxiv.org/html/2511.07268/images/turing_accuracy_by_genre_origin.png" alt="Performance broken down by both genre and origin, showing where the model excels and where listeners can still detect the AI" /></a> Performance broken down by both genre and origin, showing where the model excels and where listeners can still detect the AI
This variation is instructive. It's not that the model universally reaches human-level performance. Rather, it reaches human-level performance on some genres while still having tells on others. Classical perhaps. Jazz perhaps not. This reflects the training data composition, the diversity of examples the model saw during training.
The distribution of listener guesses shows whether people had an intrinsic bias to guess "human" or "AI."
<a href="https://arxiv.org/html/2511.07268/images/turing_response_bias.png"><img src="https://arxiv.org/html/2511.07268/images/turing_response_bias.png" alt="Response bias showing whether listeners tended to guess human or AI composition, which affects how to interpret the raw accuracy numbers" /></a> Response bias showing whether listeners tended to guess human or AI composition, which affects how to interpret the raw accuracy numbers
Humans don't classify randomly. They have biases, preferences, hunches. Some people might assume that anything unusual-sounding must be AI. Others might assume that complexity implies human authorship. Understanding these biases matters for interpreting what the 50% accuracy actually means. If people had a strong bias, 50% might actually represent strong model performance despite the bias pulling toward one label.
The listening study validates what the metrics couldn't quite capture on their own. The model produces something genuinely musical. Not just statistically plausible note sequences, but pieces that engage listeners the way human-composed music does.
<h2>The lesson beneath the technique</h2>
The full picture across all these decisions reveals a pattern. Success in music generation didn't come from finding one clever architectural innovation or discovering that one weird trick. It came from systematically understanding how three variables interact: how much capacity your model needs, how diverse your training data should be, and how to efficiently transfer learning from general to specific.
The best model wasn't the biggest. It was the one trained on the most diverse data, with enough capacity to express what it learned. That's a lesson that extends far beyond piano music. Across domains where you're generating structured sequences, the same principle likely holds. Diversity in training data often outweighs curation. Pre-training on broad data before fine-tuning on specific data accelerates learning. And model size matters, but only until you've given the model enough room to express the patterns in your training data.
The real achievement here is methodological. By building many models and comparing them systematically against human judgment, the researchers created a framework for developing better music generation systems. They didn't declare that one architectural choice is optimal for all time. They showed how to ask the right questions and how to measure the answers in a way that actually corresponds to what matters in practice: whether the music sounds human.

<hr/>Original post: <a href="https://www.aimodels.fyi/papers/arxiv/generating-piano-music-transformers-comparative-study-scale?utm_source=hackernoon&utm_medium=referral">Read on AIModels.fyi</a>

AI Just Passed the Turing Test for Piano Music

Are those toys secure? And precisely what data is being handed over when a kid is using these toys?

How You Can Test Your Kids' Smart Toys For Privacy 

Traditional testing fails for AI agents. Learn how property testing, invariants, evaluation, and observability replace deterministic test cases.

Why Traditional Software Testing Breaks Down for AI Agents

A top agency does more than post content or place a press release. It aligns narrative, credibility, distribution, and conversion into one timeline that a team 

Top Crypto Marketing Agencies Worldwide: A Buyer-First Ranking For PR, KOLs, And Viral Growth

We've built a self-serve platform that allows you to run the best influencer marketing campaigns imaginable

FOHR

fohr.co

80 emps

Since 2013

Fohr's stories on HackerNoon

The Future of Web3 Data: An Interview with Subsquid CFO & Co-Founder Marcel Fohrmann

Creating Your Instagram Strategy To Gain Quality Followers

How To Search for and Find The Most Valuable Influencers on Instagram

The AI Control Plane: Observability, Cost Governance, and Policy for LLM Applications

Notes on Agentic Reasoning for Large Language Models

HERMES Rewrites the Rules of Streaming Video for Multimodal AI

Qwen3-TTS and the Case for Token-Based Speech Synthesis

Zero-Trust Security in 2026: A Complete Implementation Roadmap for CTOs

TalkNet-ASD: The “Who’s Talking?” Model for Any Video

OpenVision 3 Challenges the Need for Separate Vision and Image Generation Models

An Architectural Pattern for Running Java Components Inside .NET 8

Claude Book: A Multi-Agent Framework for Writing Novels with Claude Code

Stop Building Monolithic Agents: The Case for a Microservices-First AI Stack

Best Financial APIs for 2026

After Fiverr Acquired AutoDS and Now Yaballe, a New Era for Dropshippers Is Coming

The Case for Bigger Models in Human Motion AI

How Akshatha Madapura Anantharamu Is Building Trustworthy Interfaces for AI Systems

Causal Thinking in the Age of Big Data: Modern Econometrics for Data Scientists

A Class For Mom Part 2: Cybersecurity

Aravind Barla Rebuilt Enterprise Service Delivery for 100,000+ Employees

AI Just Passed the Turing Test for Piano Music

How You Can Test Your Kids' Smart Toys For Privacy

Why Traditional Software Testing Breaks Down for AI Agents

Top Crypto Marketing Agencies Worldwide: A Buyer-First Ranking For PR, KOLs, And Viral Growth

Representing this company? Click here to claim and customize this page!

Reviews About Fohr

Fohr WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps