Jon Durbin is going on the Hash Rate Podcast with @markjeffrey.
On the table: Parallax, decentralized inference, and what it takes to train models across distributed compute.
Live Friday, July 10.
What should Mark ask Jon? Drop it below and we'll send him the best ones.
A lot of people know Subnet 17 as a decentralized 3D asset generation network.
But one question we get often is where creators can actually access 404—GEN.
404 is accessible across multiple workflows and is completely free to use through our consumer applications, including:
• The web app
• Discord bot
• Unity Plugin
• Blender Add-on
The goal is to make 3D asset creation easier for everyone, from technical creators to people with no 3D modeling background.
That is why we have built multiple consumer applications around the subnet.
These are not separate products from the subnet.
They are access points into the same decentralized 3D generation network, built to help creators generate, edit, convert, and use 3D assets across different workflows.
As demand for immersive objects, games, AR/VR, and user-generated 3D content grows, 404 is focused on making high-quality 3D creation more accessible across all skill levels.
We also aim to power 3D UGC on any platform by allowing developers to make direct API calls to SN17.
Start creating now.
We have achieved fully non-blocking decentralized training on a recurrent model, within 0.6% of centralized quality. To our knowledge, a worldwide first.
In plain terms: training AI across distributed GPUs normally forces a choice. Either the GPUs pause and wait to sync with each other (slow, expensive) or you skip the sync and quality drops. We just showed you can have both. No blocking, no meaningful quality loss.
We chose the hardest test case on purpose. Recurrent models are sequential by nature, every step depends on the last. Transformers are far easier to parallelize. If our approach holds on the hardest case, the easier architectures should follow.
To our knowledge, no one has published decentralized non-blocking training for a recurrent architecture before. Parallax is the first. This is new ground.
Only on Chutes.
$TAO
Week 8. The big one first:
→ SN64 is live on Kraken
→ Fez started posting videos on Chutes and Parallax, with a lot more coming
→ Ad push behind GLM-5.2, aimed at devs outside the crypto bubble
→ Dev Corner: newest Intel TDX supported, steadier cold starts, spend limits visible in the dashboard, and Parallax kernels now running on RTX Pro 6000s at ~7x speed
Full breakdown below 👇
Not all partnerships are created equal.. some open doors... others open worlds!
We are proud to announce:
@manakoai customers will soon be able to cut bandwidth costs with a single click. @vidaio_ 's compression models, natively integrated.
Every camera intelligent AND optimised
As Manako grows, Vidaio grows with it
This is optimisation on the edge
Real production, real workflows, real customers.
Bittensor Ecosystem Highlights :: June 29–July 5, 2026
[ @trishoolai - SN23 ]
Trishool’s HaloGuard 1.0 reached SOTA prompt-safety performance among open-weight guard models, with its 4B model ranking first across the evaluated benchmarks.
> https://t.co/Z9yrNA0Eui
[ @chutes_ai - SN64 ]
Chutes’ first in-house dFlash model is live, with Qwen3-32B delivering ~50% higher throughput at the same hardware cost.
> https://t.co/ON9m4LJ6Nv
Chutes also rebuilt its frontend from scratch, with a faster app, cleaner navigation, improved model search and reworked account tools.
> https://t.co/Op7nUkLKsP
[ @404gen_ - SN17 ]
404 introduced agentic mining on SN17, where agents generate 3D assets from specs, submit code and compete autonomously in the procedural asset competition.
> https://t.co/uLAL2q0WFR
[ @SynthdataCo - SN50 ]
Synth announced a profitable Q2, a $100K SN50 alpha buyback over 8 weeks, and a renewed focus on turning Synth’s ML forecasting research into live trading systems.
> https://t.co/8wjtxcCkc0
[ @ReadyAI_ - SN33 ]
ReadyAI reported 3x monthly revenue, approaching six-figure ARR, and is expanding SN33 toward specialized coding data for AI agents.
> https://t.co/e81x8HEgK6
[ @vidaio_ - SN85 ]
Vidaio launched its MCP, letting AI agents connect directly to its video compression tools without manual processing.
> https://t.co/KoR91WnWsb
[ @webuildscore - SN44 ]
Score said Satori 2B is improving fast and is preparing to scale toward a larger VLM on SN44, with a 72B model as the next ambition.
> https://t.co/1GwmJcvQZ8
[ @babelbit - SN59 ]
Babelbit announced its first reseller partnership with Line21 and a 24/7 Spanish-to-English voice dubbing prototype.
> https://t.co/EIjveduveK
[ @yanez__ai - SN54 ]
Yanez introduced a second incentive mechanism rewarding partners for verified Proof of Humanhood users.
> https://t.co/xyLqHWaUv9
[ @heydittoai - SN118 ]
Ditto launched private storage and inference, letting users connect Hippius storage and Good Morning inference API keys inside Ditto.
> https://t.co/v1daYKBtAV
[ @reliquary_ai - SN81 ]
Reliquary launched on SN81, turning RL post-training into a market where miners find learning-zone prompts and verified rollouts train the next checkpoint.
> https://t.co/I8i3LPayvL
[ @okx ]
OKX unveiled OKX AI with Opentensor Foundation support, and Bittensor subnet APIs are coming soon to bring subnet intelligence into its agent marketplace. OKX also listed $TAO.
> https://t.co/GUD5fSqpJk
> https://t.co/VCw3gsAMry
🚨 $TAO's Leadpoet SN71 just posted an UPDATE. Most of the timeline scrolled past. Ten research loops. Limited incentives. First, measurable improvement to the agent. Failure is the Asset
That number means more than many think. Why.
The Leadpoet Lab is not a product feature. It is a continuous improvement engine where the subnet itself does the R&D:
• Miners direct LLM research loops toward the agent's weaknesses
• LLMs run the experiments, test the changes, evaluate the results
• Rewards flow to whoever moves the model forward
• Every loop is captured as training data, the wins, and the failures
The failures are the part nobody talks wants to admit. When a research loop fails to improve performance, that trajectory becomes negative training data, teaching the model which reasoning paths to avoid.
Positive plus negative data, compounding daily, is exactly how you train a specialized full-funnel sales LLM that general frontier models can not replicate.
The team calls it the Sales Brain 🧠
First LLM targeted for Q4, state of the art by next summer.
Take a Step Back:
Salesforce and Slack are pushing AI lead scoring hard right now. That validates the problem, but their inputs are all downstream, engagement history, firmographics, and past conversions.
By the time a lead gets scored, someone already decided it was worth logging.
The expensive decision happened earlier. Most B2B accounts are not in-market, and reps burn hours, guessing which ones deserve a touch today.
Leadpoet qualifies accounts with intent signals and buyer context before they ever become a CRM record.
The Scoring known leads is useful. Deciding what deserves attention before it enters the system is the big deal here.
And the endgame is not staying niche. The target is 200 to 500 paying customers and a sales LLM that can be licensed to the exact enterprise platforms doing downstream scoring today.
One-click mining is coming too, and an API key or an agent will be enough to contribute to the loops.
Ten loops in, and the flywheel already turned once. This is incentivized machine-speed R&D with real revenue problem capabilities.
LLMs are no longer just the interface. On SN71, they are the innovation loop
$TAO / DYOR.
Watching @opentensor evolve from 2021 to today shows how far the vision has come
$TAO is powering a decentralized AI economy that has grown from an experiment into a network of 128+ specialized subnets across AI agents, financial intelligence, and multimodal AI
Bittensor is building more than a blockchain. It’s creating an open intelligence layer where AI markets can compete, grow, and capture value on chain
@ChronoSeek is now integrated with @hippius_subnet
This is our first subnet-to-subnet integration on @bittensor, with more on the way.
Why this matters:
Instead of exposing original dataset URLs, validators now upload transformed synthetic task videos to Hippius. Miners only receive the Hippius-hosted artefacts, making it significantly harder to infer where tasks originated from.
This adds another layer to our anti-gaming pipeline while keeping the miner-facing API unchanged.
A huge thank you to @mast3rdubs@mogmachine for their support throughout the integration.
Next up: @chutes_ai and @vidaio_ . We’re excited to continue building a more connected Bittensor ecosystem, one integration at a time. 🚀
Week 7. Everything from the past week:
→ The rebuilt Chutes frontend is live. Faster app, cleaner navigation, sign in with Google or GitHub right on https://t.co/s7V7nkGS8y
→ Chutes is now a built-in provider in TypingMind
→ Jon Durbin is confirmed to speak at Exploit Summit (Sept 28-29, Montreal)
→ Started the process of joining the CoinMarketCap community
→ SN64 officially slated for a Kraken listing in the near future
Plus a Dev Corner in the full write-up: Blackwell B200/B300 GA, 20x faster cold starts in TEE, and a new validator pipeline.
Full breakdown below 👇
Exciting to see Chamath and J-Cal discussing the power of Targon distributed confidential compute on @theallinpod
“I do think this idea of distributed inference has a real place in the American ecosystem.”
We appreciate the mention, and are excited to watch the continued growth of the permissionless compute space.