While most of crypto watched Bitcoin price action this week, something significant happened inside Bittensor (TAO).
In a short period, the core team removed a flawed emission mechanism, shipped multiple protocol upgrades, and laid groundwork for derivatives and two-sided markets. At the same time, Subnet 23 (Trishool) demonstrated real progress on a production-grade AI safety model.
This is a shift from narrative to infrastructure. Here’s what actually changed and why it matters.
Protocol Changes: Fixing Incentives
TaoFlow Removal & Price-Based Emissions
TaoFlow (launched 2025) used an exponential moving average of token flows to allocate emissions. It was quickly gamed through cycling strategies between subnets, rewarding manipulation over real building.
The network has reverted to price-based emissions. Subnet token price is now the primary signal for emission weight. Higher sustained prices direct more emissions to that subnet.
Why this is better:
- Resistance to exploitation now scales with attack cost (root cell sell pressure).
- Subnet owners have a clear optimization target: build real value and sustain price.
- Root stakers benefit directly from aggregate subnet performance.
This is described as a stable intermediate state. The long-term goal remains fully permissionless, decentralized markets.
Default-Off Subnet Emissions
New subnets no longer receive emissions automatically upon registration. They must now meet clear criteria:
- Be claimed and actively mining
- Have a public website and on-chain terms
- Maintain a GitHub repository
- Publish a stated plan for adding value
This raises the bar for speculative subnet registration while lowering barriers for genuine builders. The result should be a healthier ecosystem with fewer empty protocols extracting value.
Dynamic Tempo
Subnets can now configure tempo windows up to seven days (previously capped around 360 blocks with a hard upper limit).
Benefits:
- Validators can observe miner behavior over statistically meaningful periods.
- Enables more sophisticated validation logic and real slashing mechanics.
- Supports week-long competitions where rewards are released based on sustained performance.
A new operation also allows subnet owners to call tempo directly, opening design space for research tournaments and performance-gated payouts.
Relay Pallets & Balancer Pools
- Relay pallets enable relayer services to batch and execute transactions on behalf of users. This reduces slippage, acts like an off-chain order book, and protects against MEV. It is foundational infrastructure for proper on-chain trading interfaces and liquidity aggregators.
- Balancer pools replace V3 pools, offering a simpler codebase and greater flexibility for future liquidity strategies.
Derivatives & Two-Sided Markets (Coming Soon)
The core team is actively developing derivative functionality for long and short positions on subnet pools. A testnet clone is already running.
A one-sided market creates asymmetric pressure. Two-sided markets enable genuine price discovery and allow the ecosystem to express negative conviction on underperforming subnets. This is described as essential for fulfilling the vision of a fully permissionless subnet economy.
SN23 Trishool & The Halo Model (AI Safety)
While protocol changes fix incentives, Subnet 23 (Trishool) is showing what high-quality subnet output looks like.
They are building Halo — a constitutional AI guard model for agents. The current target is a 0.8 billion parameter model designed to run locally (MacBook, mobile, edge devices) with near-zero latency.
How it works:
- Input guard: Filters threats before they reach the underlying model.
- Output guard: Intercepts dangerous actions (malicious tool calls, privilege escalation, data exfiltration, etc.).
The model is trained using Bittensor’s own incentive mechanism:
- Weekly harm categories and challenge questions are published.
- Miners attempt to jailbreak the current model.
- Successful jailbreaks become training data.
- The model is retrained and hardened every week.
This creates a compounding adversarial training loop that centralized labs struggle to replicate at the same scale and speed.
Current performance: ~89% on primary benchmarks (very close to much larger models like Qwen Guard at 8B parameters). Roadmap targets further improvement.
Commercial path: Starting with AI agent companies, then moving to enterprise/regulatory use cases where demonstrable AI safety layers become a requirement.
What This Means Going Forward
These changes are connected:
- Protocol upgrades align incentives so that subnets producing real value capture emissions and capital.
- Dynamic tempo + relay pallets expand what is technically possible to build.
- Trishool proves that subnets can ship frontier-level AI infrastructure while using the network’s own mechanisms to improve.
Bittensor is transitioning from a broad thesis about decentralized intelligence into concrete, usable infrastructure for machine economies, agent safety, and on-chain coordination.
The bear market clears noise. What remains are the projects actually shipping the rails for what comes next.
This is still early — but the direction is becoming much clearer.
If you have more questions, please feel free to contact me at any time: https://t.me/FatherSon97
Top comments (0)