Sabitlenmiş Tweet
Codatta
2.2K posts
Codatta
@codatta_io
AI’s Knowledge Layer turns human knowledge into on-chain assets with royalties. 1st @Binance Booster.
Onchain Katılım Ekim 2023
187 Takip Edilen203.5K Takipçiler
Congrats on the opening of Asia's first Ethereum Community Hub! 🎉
Codatta will be joining the panel at this event.
Excited to connect with fellow builders there!
Hong Kong Ethereum Community Hub@ethereumhkhub
Asia’s first☝️Ethereum Community Hub is opening. Backed by @EFetheverywhere (@ethereumfndn) Operated by @snzholding & @Ethtao_Ethtao A milestone for Ethereum in Asia. Expect: 1⃣Core ecosystem voices 2⃣Leading institutions & builders 3⃣Real conversations on applications & adoption Join us during 🇭🇰Web3 Festival. See what’s next for Ethereum in Asia. RSVP👉luma.com/uf5v6joa
English
Codatta Dataset Series (7): Appliance-Knobs 🎛️
Teaching a robot to pick up objects is hard. Teaching it to operate them is harder.
Appliance-Knobs is a dataset open-sourced by Codatta on @huggingface, built specifically for Robotics, 3D understanding, and object detection. Every entry captures the same knob from two correlated angles — front view and side view — giving models the multi-perspective information they need for 3D understanding.
✨ What makes it different:
Paired images: every knob captured from two distinct angles for 3D reconstruction
Fine-grained focus: rotary controls and appliance knobs — a specialized, underrepresented object class
High-resolution: image quality suitable for precise state detection and position estimation
🛠️ Supported tasks:
✅ Multi-View Object Recognition
✅ 3D Shape Reconstruction
✅ Knob State/Angle Estimation
✅ Generative AI Training
📊 Explore & download the dataset: huggingface.co/datasets/Codat…
🤝 Help us build the next robotics dataset: app.codatta.io/app/frontier/9…
Codatta@codatta_io
Dataset Series (6): LLM-Failure-Cases 🧠 LLMs don't just fail on hard questions — they fail confidently. Wrong symmetry principles. Flawed number sequences. Contradictory reasoning. The scariest failures aren't the ones models flag as uncertain. They're the ones that sound completely right. LLM-Failure-Cases is a dataset open-sourced by Codatta on @huggingface, built from real adversarial submissions collected during Airdrop Season 1. Contributors found prompts that caused leading models to fail — and paired each failure with an expert critique explaining exactly what went wrong. ✨ What makes it different: - Model-specific: failures logged per model (GPT-4o, Gemini, Claude, and more) - Expert-critiqued: not just wrong answers — annotated with why they're wrong - Multi-domain: physics, math, logic, science, language comprehension - Bilingual: English + Chinese 🛠️ Supported tasks: ✅ Model Evaluation & Red-Teaming ✅ Hallucination Research ✅ RLHF Training Data ✅ Expert Critique Analysis 📊 Explore & download the dataset: huggingface.co/datasets/Codat… 🤝 Help us build the dataset: app.codatta.io/app/frontier/8…
English
Data on Codatta doesn't just live somewhere — it's encrypted, anchored, and served through a layered architecture designed so that provenance is never broken.
Here's how it works.
Storage: Every payload is encrypted before storage. Content hashes and CIDs are anchored to decentralized networks as the source of truth, then mirrored to cloud hot paths (S3, GCS, OSS) for low-latency access.
The key principle: content references never change — caching layers only mirror permitted bytes.
Compute: Two patterns for secure processing:
TEE enclaves — per-request transforms (redaction, feature extraction) without exposing raw data
Federated runs — training across isolated data silos; data stays local, only updates transfer
Serving: An access gateway enforces role, attribute, and token-based policies before any data moves.
Every request generates metering events, logged for audit and billing alignment.
The result: identical requests against the same dataset version always produce identical metering trails. Replayable by design.
📖 Read the full docs: docs.codatta.io/en/core-system…
English
📅 Codatta Weekly Highlights: April 6 – 12, 2026
This week Codatta focused on technical education — breaking down the Data Assembly pipeline, unpacking the Frontier Data thesis, highlighting the Knowledge-to-Specialized-AI framework, and publishing the March Monthly Report.
📊 Data Assembly Deep Dive
Broke down how atomic contributions (samples, labels, validations) become reproducible data assets through the CF → Data Asset → Versioned Dataset pipeline.
Highlighted key guarantees: determinism, immutability, full provenance, and lineage-tracked diffs — all anchored by Contribution Fingerprints.
🔬 Frontier Data
Reinforced why models trained on recycled internet text hit a ceiling — and what it takes to break past it.
Core message: Frontier Data (surgical robotics, financial reasoning, cultural context, edge-case science) can't be scraped. It has to be built — with human expertise, on-chain provenance, and royalty-backed ownership.
🤖 From Knowledge to Specialized AI
Highted Codatta's 4-method framework for turning foundational models into domain specialists: Fine-Tuning, RAG, Prompting, and Evaluation.
Framed contributor knowledge as the upstream input powering every downstream specialization.
📋 March 2026 Monthly Report
Published as a Twitter Article on Friday. Headline milestone: Codatta crossed 10,000,000 total contributions in March.
🌟 Overall Vibe: Whitepaper education week. Dense technical content delivered accessibly, with a consistent thread running through all posts — verified human knowledge is the foundational input the AI industry can't replace.
English
General models aren't born specialists — they're trained.
Codatta's "From Knowledge to Specialized AI" framework makes it happen through 4 methods: Fine-Tuning, RAG, Prompting, and Evaluation — all powered by verified human expertise.
Knowledge Data → Foundational AI → Specialized AI
Own the Knowledge. Earn the Future.
English
Models trained on recycled internet text hit a wall. The way forward is Frontier Data.
The kind that's rare, domain-specific, and human-validated: surgical robotics, financial reasoning, cultural context, edge-case science.
Codatta is building the infrastructure to push past it — turning expert knowledge into verified, royalty-backed on-chain assets that fuel the next generation of AI models.
English
This is what it means for data to be a real asset — not just a file, but a versioned, provable, contributor-attributed record that can be audited from any point in its history.
📖 Read the full docs: docs.codatta.io/en/core-system…
English
✨Key guarantees:
- Determinism: identical inputs + rules = identical output, every time
- Immutability: past versions never change
- Full explainability: show why any CF was included or excluded
- Referential integrity: validations always reference the CF they evaluate — nothing floats free
English
How raw contributions become reproducible data assets and versioned datasets—with lineage for provenance, diffs, and rollbacks.🧵
At Codatta, every atomic contribution — a sample, a label, a validation — gets a Contribution Fingerprint (CF) the moment it's accepted. That CF is the building block of everything that follows.
Here's how the assembly pipeline works:👇
English
📊 Codatta Weekly Highlights: March 30 – April 5, 2026
This week Codatta previewed the upcoming Data Lineage system, engaged with macro narratives on data as the new form of capital, spotlighted the LLM-Failure-Cases dataset, and released a task tutorial on correcting LLM mistakes. Strong focus on product education and ecosystem alignment.
💡Codatta Dataset Series (6): LLM-Failure-Cases
- Open-sourced on Hugging Face — 835 expert-critiqued failure cases across GPT-4o, Gemini, Claude, and more.
- Multi-domain: physics, math, logic, science, language comprehension. Bilingual: EN + ZH.
- Perfect for: Model Evaluation & Red-Teaming, Hallucination Research, RLHF Training Data, Expert Critique Analysis.
- CTA: Help build the next AI safety dataset.
🔗 Product Preview: Data Lineage
- Gave the community a break down of Codatta's upcoming Data Lineage system — a transparent, immutable visualization of every data asset from contribution to payout, built on Base.
- Four stages: Assetification → Assembling → Adoption → Payouts. One unbroken on-chain record.
🎬 Task Tutorial
60-second walkthrough: how to spot an LLM error, submit a correction, and claim rewards — directly tied to the LLM-Failure-Cases dataset.
🌍 Ecosystem Engagement
- Responded to Base's 2026 vision ("Every asset on earth is coming onchain"), positioning Codatta's Knowledge Layer as core infrastructure for that flywheel.
- Engaged with @dgt10011's viral essay The Generational Prisoner's Dilemma — connecting his thesis ("your own existence can actually be monetized") to Codatta's contributor economy.
🌟 Overall Vibe
Thoughtful and forward-looking week. Strong narrative building — connecting Codatta's product roadmap to the broader macro moment and reinforcing why expert human contribution matters more than ever.
English
Dataset Series (6): LLM-Failure-Cases 🧠
LLMs don't just fail on hard questions — they fail confidently.
Wrong symmetry principles. Flawed number sequences. Contradictory reasoning. The scariest failures aren't the ones models flag as uncertain. They're the ones that sound completely right.
LLM-Failure-Cases is a dataset open-sourced by Codatta on @huggingface, built from real adversarial submissions collected during Airdrop Season 1. Contributors found prompts that caused leading models to fail — and paired each failure with an expert critique explaining exactly what went wrong.
✨ What makes it different:
- Model-specific: failures logged per model (GPT-4o, Gemini, Claude, and more)
- Expert-critiqued: not just wrong answers — annotated with why they're wrong
- Multi-domain: physics, math, logic, science, language comprehension
- Bilingual: English + Chinese
🛠️ Supported tasks:
✅ Model Evaluation & Red-Teaming
✅ Hallucination Research
✅ RLHF Training Data
✅ Expert Critique Analysis
📊 Explore & download the dataset: huggingface.co/datasets/Codat…
🤝 Help us build the dataset:
app.codatta.io/app/frontier/8…
Codatta@codatta_io
Codatta Dataset Series (5): RoboManip-Traj-Demo 🤖 The biggest bottleneck in Embodied AI isn’t algorithms—it’s the extreme scarcity of high-quality data. RoboManip-Traj-Demo is a dataset open-sourced by Codatta on @huggingface, designed specifically for Embodied AI and Computer Vision research. It focuses on robotic manipulation trajectories, providing high-precision manipulation data and fine-grained annotations. With high-precision spatial coordinates and rich event/pose annotations, this dataset strongly supports the following downstream tasks: • Trajectory Prediction • Keyframe Extraction & Event Detection • Fine-Grained Robotic Control • Object Interaction Analysis 📊 Explore & download the dataset: huggingface.co/datasets/Codat… 🤝 Help us build the next-gen AI dataset: app.codatta.io/app/frontier/R…
English
As @dgt10011 detailed in his essay The Generational Prisoner's Dilemma — this shift isn't speculative. It's structural. 🌍
The AI industry's second most critical ingredient, after energy, is data. Specifically, the data footprints you generate every day that provide context for inference and learning. Premium value will be assigned to your thoughts, prompts, and intents — making them more non-fungible than energy itself.
When intent becomes capital, it creates a new economic order outside traditional financial rails. Agent systems are already being provisioned with crypto wallets to autonomously pay for compute, APIs, and data. 🤖💸
We are building the exact infrastructure to ensure human contributors get their cut in this autonomous world. 🤝
x.com/dgt10011/statu…
English
"If contributing data to train models gives attribution back to you, we can create a fundamentally new framework for UBI."
The future @dgt10011 laid out at the end of this video is exactly what we're building at Codatta.
Your own existence can actually be monetized.
Your intent. Your knowledge. Your data footprint. These aren't byproducts of your life online — they're the new form of capital.
Bankless@Bankless
LIVE NOW -- 3 Megatrends Every Investor Needs to Know | Jeff Park Three macro forces are converging that almost no one is pricing in correctly. @dgt10011 returns to lay out what he calls the "certain truths" shaping markets for the next two decades: collapsing demographics across every major economy, wealth concentration that now exceeds even the Gilded Age, and a labor market being structurally hollowed out by AI before robots even arrive. [TIMESTAMPS] > 0:00 Intro > 3:07 Certain Truth #1: Global Demographic Headwinds > 8:26 What a Top-Heavy Population Means for Markets > 15:05 The Japan Counter-Argument > 26:01 Certain Truth #2: Wealth Concentration at Critical Levels > 32:48 Capital Flight, Unrealized Gains, and the Drag on Demand > 43:17 Real Estate as the Epicenter of Distortion > 48:05 Bitcoin as a Spending and Savings Asset > 52:08 Certain Truth #3: The Value of Labor Approaching Zero > 1:01:27 The AI Industrial Revolution Comparison > 1:04:38 The Convergence: Why All Three Truths Collide Now > 1:08:19 Lightning Round: Asset Class Outlook > 1:27:49 Closing Thoughts
English
"Every asset on earth is coming onchain." 🌍
And one of the core assets fueling this AI era? High-quality data.
Thrilled to see @base's 2026 vision. At Codatta, we are accelerating this exact flywheel by turning knowledge into verifiable, tradable, and yield-bearing onchain assets.
Your knowledge. Your data assets. Endless AI royalties.
Base@base
English