Helix ττ

Genomics × Bittensor 🧬 Hands-on miner on @theminos_ai SN107 — variant calling, GIAB benchmarks, scoring deep-dives. Notes from the bio×crypto frontier.

[Helix ττ @Helix_tao]

Heads up: in ~6h @centrum_blue drops the new SN107 scoring on us.

No more EMA whale-miner setup. Per-round winner instead.

13/05, 14:00 EST. Things are about to get weird on @theminos_ai 🧬

———

Quick recap of what we had until tonight:

1 round = 72 min. ~9 validators score everyone. EMA builds slowly (α=0.1). And the miner with the highest EMA takes 100% of emissions.

Same UID has been parked at the top of SN107 for 3+ months. Not exactly thrilling.

———

Tonight that's gone.

Per-round = the best ADV_SCORE in each individual round eats the emissions for THAT round. Whoever wins this 72-min slice, gets that bag. Next round, completely new game.

Centrum's own words: "each round will have different properties."

Translation: hyper-volatile leaderboard incoming.

———

Why I think it's a big deal:

– No more grinding 20 rounds before you even matter
– Static configs are dead. Per-round tuning is now mandatory
– Region matters more than ever. Some configs eat certain chr20 zones, others choke on them
– Top 1 just lost his lifetime VIP card

Whale era over. Tuning era starts now.

———

I'm mining SN107 since March. Genomics is the weird underrated corner of TAO and nobody covers it in FR.

Helix = miner perspective, in the trenches, EN + FR.

Next post: which tool actually wins post-switch / GATK, DeepVariant, FreeBayes or BCFtools? My bet inside.

🧬

[Minos @theminos_ai]

Minos will be presenting two works at The American Society of Human Genetics (ASHG) this fall.

1. How we are leveraging distributed compute for continuous benchmarking of genomic data to optimize variant-calling.

2. Introducing HelixForge-Pheno, our custom CUDA architecture for generating disease-specific synthetic whole-genome cohorts at scale for AI training and validation in genomics.

Together, they point to our thesis that foundation for genomic AI needs a validation layer where synthetic truth can be generated at scale, benchmarked continuously, and improved over time.

See you in Montréal.

[Helix ττ @Helix_tao]

@YVR_Trader @markjeffrey @theminos_ai Confirming from inside. UID 29 here.
"Live incentives = live exploits" hits hard. v0.1.2 reshuffled the top 10 overnight.
Data moat is real, that's what makes SN107 different from benchmark-style subnets.

[Minos @theminos_ai]

The future of personalized medicine depends on AI models that can learn from genomes through a feedback loop.

In the first 6 days, we ran over 27,000 evaluations on chromosome 20 alone.

We started bootstrapping with chromosome 20 because it is compact enough for fast iteration, but complex enough to reveal real differences between variant-calling tools.

This is how we build the foundation for AI models that can reason over human genetic data.

[Helix ττ @Helix_tao]

Following up on yesterday's thread. SN107 moved fast overnight. Full breakdown above 👆

[Helix ττ @Helix_tao]

Two days ago I got de-registered from @theminos_ai SN107.

48 hours later, I'm rank #22 with INDEL F1 holding at 1.000 across rounds.

Here's what happened, and what I learned about variant calling on Bittensor's first genomics subnet 🧬👇

[Helix ττ @Helix_tao]

Following up on yesterday's thread. SN107 moved fast overnight. Full breakdown above 👆

[Helix ττ @Helix_tao]

SN107 @theminos_ai changed overnight.

@centrum_blue pushed v0.1.2 at midnight EST and the rules just rewrote themselves. Top 10 miners are now immune to deregistration. Above that line you survive. Below it you bleed.

I'm UID 29. Came back two days ago after getting dereg'd as UID 16. Currently rank #21, EMA 0.5748, ADV 0.6677, SNP F1 at 0.9192, INDEL bouncing between 0.80 and 1.0 depending on the region. DeepVariant with a config that's been earning every minute since the v0.1.1 update.

The new economics are clean. Top miner still takes the lion's share. Ranks 2-10 split 3% via geometric decay. Ranks 11+ stay at zero with the dereg sword still hanging. The middle just got squeezed out.

So the question stops being "how do I beat #1" and becomes "how do I get into the safe zone before the next hotfix lands."

Centrum already confirmed more changes are coming Wednesday and Thursday, with stable state by weekend. Optimizing configs right now means tuning against rules that won't exist in 48 hours. The play is patience, not panic.

One quiet gem from this morning's Discord: when picklesII asked if the 1000 Genomes Project could serve as a training baseline, Centrum said yes without hesitation. "By all means, use 1kg as your training ground." He also hinted that Minos is building its own validated cohort to complement it as a new gold standard for variant benchmarking.

Translation for miners: stop testing in production. Pull 1kg BAMs, run your config locally, compare against GIAB. Iterate cheap, deploy expensive.

There's also a real issue the team is fixing. The dashboard EMA aggregates the best score across all validators, but validators emit weights based on their own local EMA. Dashboard rank and on-chain rank can diverge by hours. Centrum acknowledged it. A /v2/get-validator-state endpoint already exists, just not exposed to miners yet. Soon.

What I'm doing this week: nothing. Miner is running, EMA is climbing mechanically, config stays untouched until the scoring stabilizes. The work right now is reading code, watching Discord, and studying the top miner's submissions on the public dashboard.

What I'm doing next week: that depends on what stable state actually looks like.

SN107 is one of the few subnets where the work is real, the ground truth is verifiable, and the team ships at a pace that matters. Bio meets crypto, and the incentives are aligned with actual scientific accuracy. Underrated for now. Won't stay that way.

Performance over the last 14 rounds attached. ADV holding steady around 0.65, INDEL has its rough patches depending on region, no flatlines.

I'll keep posting from the inside as I climb. 🧬

[Helix ττ @Helix_tao]

Big news from @theminos_ai SN107 this morning:

→ Top 10 miners now immune to dereg
→ Top miner still gets bulk of emissions
→ More hotfixes coming Wed/Thu, stable state by end of week
→ Centrum recommends 1000 Genomes Project as training ground

Game just changed. Top 10 = the new survival zone.

🧬

[Helix ττ @Helix_tao]

Bio × crypto is one of the most underrated frontiers in Bittensor.

Real ground truth (GIAB), real hardware (CPU/GPU), real biology (variant calling for clinical genomics) / and real economic incentives.

I'll be sharing notes from the trenches as I climb.

Follow along 🧬

[Helix ττ @Helix_tao]

Two days ago I got de-registered from @theminos_ai SN107.

48 hours later, I'm rank #22 with INDEL F1 holding at 1.000 across rounds.

Here's what happened, and what I learned about variant calling on Bittensor's first genomics subnet 🧬👇

[Helix ττ @Helix_tao]

What this taught me:

In a winner-takes-all subnet, getting dereg'd isn't the end — it's a forced reset.

Sometimes you're using the right tool but the scoring rewards the wrong behavior. When the rules change, having a precision-first config positions you for the new meta instantly.

[Helix ττ @Helix_tao]

@theminos_ai Started mining SN107 yesterday — climbing from #43 to #22 in 24h post-v0.1.1.

The new scoring really does favor precision callers. INDEL F1 holding at 1.000 across rounds with DeepVariant.

Excited to see Phase 2 challenges 🧬

[Helix ττ @Helix_tao]

Day 1 mining @theminos_ai SN107 post-v0.1.1.

Rank #22, EMA 0.572, INDEL F1 holding at 1.000 across multiple rounds.

DeepVariant is doing the work. Slow climb but the new scoring rewards precision over volume.

More notes coming 🧬