Distributed State
@DistStateAndMe
Followers
3K
Following
28K
Media
716
Statuses
9K
cult leader/ exit liquidity / community mod @covenant_ai
subnet (3/39/81)
Joined April 2014
If i spent the rest of my life building on @opentensor , it would be a life well lived. https://t.co/WcEHZaqkhY Our sacred pact. Bittensor will claim its place in history
covenant.ai
One Covenant, Many Orders.
11
29
120
We made it?
Most Popular Blockchain Ecosystems in 2025 By Mindshare 1. Solana - 26.79% 2. Base - 13.94% 3. Ethereum - 13.43% 4. Sui - 11.77% 5. BNB Chain - 9.05% 6. XRP Ledger - 4.68% 7. Sonic - 2.29% 8. Cardano - 1.92% 9. Bittensor - 1.91% 10. Hyperliquid - 1.57% 11. TON - 1.23% 12.
4
3
17
Nvidia is fast becoming open-source kings, they actually pre-trained, mid-trained, post-trained, and got better results than Qwen3 which is impressive! Now we only need to bully them about nvidia-open-model-license and multimodality...
11
23
237
Missed this week's TGIF Space? Catch up on everything happening across Covenant AI. This week: Basilica's rebuilt incentive mechanism, why decentralized compute economics are broken (and how we're fixing it), @NeurIPSConf recap, and a frank discussion on the value of patient
0
7
11
Don't look now, but this is actually quite huge š https://t.co/M3HIfWLlZX Chutes can now serve as an IDP. Developers can now have users "Login with chutes" and all of the compute/inference costs pass long directly to that user and/or their subscription on chutes. No more
github.com
Make chutes an IDP with oauth2 style access/refresh tokens, configurable scopes/visibility, etc.
7
27
148
Why is our halving countdown the most accurate & always in sync? tldr: we pull each piece directly from the chain and recalculate every 12 sec. (more details šš§µ) Accuracy matters when every block counts. See for yourself: https://t.co/rFQqw6XnjD
5
23
99
1
1
22
Tools as Types - The future of agentic workflows on @basilic_ai
2
3
12
Just 1 day to go until the first ever #Bittensor halvening Only 7,558 $TAO left to be emitted before we enter a new phase of scarcity $TAO
23
58
389
@Jsevillamol @tplr_ai Hi Jamie, our 72B run has so far processed 750B tokens and we are targeting ~1T for completion. Plans for the start of the next large scale run are still in development, as the current run wraps up we expect to be doing a series of smaller scale tests in public runs to evaluate
0
2
2
Im just diggin the hole on Twitch. which is equivalent of minimizing the loss function somehow. saddle point speedrun. https://t.co/CAZXBLn8j1
twitch.tv
Tekken
0
1
9
Language diffusion models are faster at writing code (and other structured text). The other day, I had the thought that language diffusion models sort-of have something analogous to speculative decoding built-in by default. In speculative decoding for autoregressive models, we
Added confidence-aware parallel decoding to my tiny text diffusion model! Before, we had āscheduled iterative refinementā for decoding, where it would go through a masking schedule and resample all positions at each iteration. This didnāt take into account token probabilities
5
13
140
Coding agents are moving beyond vibe coding toward real-world, enterprise software engineering. Attention is shifting from frontend demos and isolated bug fixes to refactoring large codebases, improving the efficiency of complex systems, and handling long-horizon tasks. What
60
40
625
Banger
0
0
16
The only meaningful debate is for the sake of driving towards principled formalizations, never ego. Debate with kindness and grace, as emotions and deep struggles are real, though unseen. Debate knowing the social levers narcissists can pull to gain power over you.
1
1
14
doing well on AI benchmarks is like having a great one rep max - impressive but wholly unrelated to daily utility
0
2
24
This is your scheduled reminder that @tplr_ai's Covenant 72B is the largest , permissionless pre-training run by a wide margin. Thank you for your attention to this matter. https://t.co/aA25GOcEVt
5
7
48
Would love to try this on SpareLoco:
arxiv.org
Communication-efficient distributed training algorithms have received considerable interest recently due to their benefits for training Large Language Models (LLMs) in bandwidth-constrained...
DeepCode ā a multi-agent framework that turns research papers into full codebases It manages information flow so large, detailed papers can be converted into production-quality code despite LLM context limits. It does this through: - Blueprint distillation ā compressing papers
0
0
4
Reflection on @dwarkesh_sp's reflections on his interview with Rich Sutton, and why LLMs are exciting because they are Skinnerian, Popperian and Gregorian creatures. Minds with finite capacity cannot adapt forever without having to forget previous knowledge. This is true of
dwarkesh.com
Watch now (66 mins) | LLMs arenāt Bitter-Lesson-pilled
11
26
175