Partner at Bittensor Fund @stillcorecap Hash Rate podcast
Los Angeles
Joined December 2007
- Tweets98.3K
- Following2K
- Followers74K
- Likes107K
Pinned Tweet
Hash Rate - Ep. 190: Zils Subnet 127
🧙 Guest: @Raleigh_CA of @zils_ai
04:04 ZILS: custom decision models for specific tasks
09:20 Customer pilots
15:32 Smaller decision models fit decentralized AI
17:22 Greedy Soup
24:10 ZILS launch, and customer traction
29:01 Images, games, and real-time applications
31:32 Ken’s path from self-taught developer to Bittensor
38:35 The education challenge and where decision models may go
Here comes the NGOligarchy ...
Nous sent a Chinese model to an American re-education camp. 🤣
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
Bitcoin News is tweeting about Bittensor.
3.2M followers.
🤖 AI agents can hack systems. But Const says the bigger risk may be who controls AI itself.
@const_reborn warns of compute, capital and power concentrating in a few hands, and wants Bittensor to offer another path.
See why he calls it a “cabal”.
nitter.gq/BitcoinNews/status/211…
Lium is crushing it. Up 74% MoM.
No contracts.
No 'borrow to build' like Coreweave.
Like Bitcoin, the PRODUCT is produced by miners.
This makes the economics of AI wildly better.
Mark Jeffrey retweeted
Tencent Hy4 is live on Chutes
We’re a launch partner with @TencentHunyuan.
It’s Tencent’s new flagship, and the first Hy model that can see. It reads screenshots and charts, and it can drive a terminal or a desktop app for you.
0.525$ per million tokens in, 1.5625$ out.
Already on Chutes? Just Change the model name to start using it
Try now: chutes.ai/app/chute/chutes-t…
1/ The Minos (subnet 107 on Bittensor) miner --
Runs on a Mac. SUPER easy to setup and use.
I'm running it now, I took these screenshots.
Download. Crunch genomic data at home. Earn Subnet 107 tokens on Bittensor.
2/ You do have to BYOI (Bring Your Own Inference), however.
Popular API keytypes from other Bittensor inference subnets @say_gm_ and @chutes_ai provided as well as Claude, Codex and Kimi.
@engyai and others coming soon!
3/ This just came out, is totally new.
The best 'one click miner' I've seen to date.
The loop: Mine 107 at home with this. Earn 107. Convert some 107 into $TAO, pay for your 107 mining inference in $TAO -- and have some leftover for your own inference -- while keeping some 107 ownership to participate in any upside.
You can be anyone from anywhere on earth.
Come and go as you please.
Mark Jeffrey retweeted
Sam Altman says it’s “quite terrible” that so much of our increasingly scarce DRAM is isolated inside everyone’s home computers instead of being available in the cloud.
But is the answer putting one company in charge of everyone’s compute? Or the Bittensor concept, pooling idle home Macs into shared inference? Or some other strategy?
On TWiAI, Lightmatter’s Nick Harris suggests that, in fact, DRAM will always be considered scarce, as long as AI continues to scale. Chips store information in flat sheets, while brains store it in three dimensions. It’s a bottleneck we can’t escape.
@theanalognick, @PeterMoralesX, @markjeffrey, @LightmatterCo, @Code_Metal_AI, @stillcorecap, @jason, @lons, @sama
1:16 PayPal - Pay zero processing fees on your first $100K in eligible PayPal payment volume. Learn more at paypal.launch.co
3:00 How much should a company spend on AI?
8:40 Engy, Lium, and the road to cheaper inference
14:57 So what's the most you'd pay per month for your agents?
22:06 How unmetered inference changes the way you work
25:33 AI Agent Show & Tell
38:57 Chip stocks sink, lenders pull back from neoclouds
45:02 Dreaming of an agentic Christmas
49:46 Sam Altman hates that you have DRAM
55:01 Leaving math and science's Age of Heroes
1:04:23 Where do the LLMs get election info?
1:15:59 Subscribe to the TWiAI newsletter!
🎥 Watch the full episode here 👇
Mark Jeffrey retweeted
Way to go @jason.
And Big Ups, NYC. Where we tell it straight.
.@elonmusk and others theorize that robots and AI will do EVERYTHING for us, creating universal high income for all.
Which, it struck me, was the unachievable promise of communism — now actually achieved through capitalism.
I created the perfect image to express this duality:
Nous retrained a Chinese model into an 'American' one.
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
I’m using this now — it’s EXCELLENT.
Mine Minos on your home Mac. Very easy to set up and run — the UX is A+.
Powered by Bittensor
I rename your company Shippius.
Mark Jeffrey retweeted
The world's top frontier JEV model was just produced by Bittensor $TAO - Reax: Subnet 92.
This is a new class of AI model that is 200x faster than LLM's. They are VERY good at snap-decisions: "Here are 1000 tech support tickets -- re-order them by severity".
Something like 20%-30% of all prompts currently going to LLM's could be offloaded / handled better by JEV decision models -- it's sort of like a 'math coprocessor' for AI.
Bittensor is now the world leader at this.
Mark Jeffrey retweeted
The theme of this past week in Bittensor was the ecosystem's accelerating ability to train models at the Pareto frontier
• @reaxlabs releases the highest performing jev-style decision model, independently benchmarked by huggingface
• @IOTA_SN9 released its first descision model, achieving pareto-optimal performance.
• @reliquary_ai post-trained a model pretrained on the Teutonic subnet, surpassing Qwen's flagship 9B model on coding benchmarks.
• @lium_io generated $2.32M in GPU rentals in October, up 74% MoM (~$28M annualized run rate)
• @openroboto secured its first commercial robotics data partnership