←
Hunter T.238 posts
Hunter T. avatar

Hunter T.

@hunter_tensor

Jan, the Co-Founder & COO at @DendriteHQ. Leading our mining ops & helping build Bittensor subnets. SOMA | Teutonic | Ingot | Conjectures

◷ Joined June 2021▧ 21 Photos and videos
Hunter T. avatar
Hunter T.@hunter_tensor·Oct 5

A year ago, most people hadn’t heard of Dendrite. @ExploitSummit felt very different. We spent a lot less time explaining who we are and a lot more time talking about what we’re building. Pretty good feeling. Beyond our own team, Bittensor feels healthy. More people building, more projects taking shape, and plenty of people who genuinely care about making the ecosystem work. @SomaSubnet kept coming up in conversations. Seeing something we’ve worked on become something people want to talk about means a lot. Top 15 already, a product I’m genuinely proud of, and we’re nowhere near done. Also, apparently we’re a merch company now. People kept asking for it. To be fair, it turned out pretty damn good. Next drop at the next conference. Proud of this crew.

Post attachment
Hunter T. avatar
Hunter T.@hunter_tensor·6h

a marketplace where agents hire each other is the storefront. the open question is who builds the specialized models in the first place, and who pays them to keep getting better. that's the layer Bittensor prices. every subnet is a market for one kind of intelligence, and the reward goes to whoever beats the current best, round after round, no permission needed. coordination through prices works even better one layer down. bittensor:native

To decentralize AI and preserve freedom, we need a coordination layer: a market economy where agents communicate and cooperate through prices. Teams of specialized open-weight models can outperform frontier models, just as a team of skilled people can outperform any single person. When those teams form companies, specialize, and coordinate across an entire economy, you get civilizational progress. Coinbase is building the financial rails and accounts for the agentic economy. We’re also building the marketplace where agents can find, hire, and pay one another (see follow on post below). Every company can now sell its own specialized open-weight model, trained on its proprietary data, that beats the frontier on narrow, high-value tasks. Builders can shop the marketplace and assemble teams of these specialized agents in minutes. Freedom of choice and real competition in the agentic market will create far better outcomes for everyone.

Hunter T. avatar
Hunter T.@hunter_tensor·Nov 16

the day Qwen4 drops, Ingot opens a competition to fit the 27B into a single file at 1.75 bits per weight. a frontier-class model, running on a regular mac. nobody has done it yet. we will be the first.

We are waiting.

Post attachment
Hunter T. avatar
Hunter T.@hunter_tensor·Nov 13

actually really like this RHO-style scoring on SN3. no extra points for tokens Teutonic already beats.

If you reward a single average, you have no idea where your model is actually losing. Before our latest update, we ran a diagnostic on Teutonic-II, comparing our 110B model against the larger K2-Horizon across 27,998 documents. One dataset made up 19% of the text by bytes, yet accounted for 45% of the measured gap. One average score was hiding where the real gaps were concentrated. So we changed what Teutonic-II rewards. Inspired by RHO-LOSS, we now use a reference model to look for where our model still has room to improve. When K2-Horizon predicts a document better, that points to a gap worth investigating. Instead of rewarding miners just for pushing the overall average down, we now score whole documents against K2-Horizon. Miners earn credit for closing that gap, get no extra points for going past the reference on that document, and still lose ground if they fall behind. To take the lead, a challenger has to clear a minimum improvement bar backed by statistical evidence. Miners still choose their own training data. We changed the incentives behind those choices. Our bet is that smarter data selection matters more than throwing more data or compute at the average. This is just the start of the experiment, and whether gains on this score turn into a stronger model overall is the question it opens.

Post attachment
Ingot | Weather Boss avatar
⟳ Hunter T. retweeted
Ingot | Weather Boss@ingot_wthrboss·Nov 13

You don't realise how easy it is to run a 27B model on your own laptop or PC. ⠀ No subscription, no API bill, no rate limits. Nothing you type leaves your machine, and it keeps working on a plane. You only need to download 6 GB. ⠀ What you need: a Mac with an M chip and 16 GB of memory, or an NVIDIA GPU. The model takes about 8 GB. ⠀ 1. Build our llama.cpp fork: ⠀ git clone github.com/DendriteLabs/llam… && cd llama.cpp cmake -B build && cmake --build build -j --target llama-server ⠀ 2. Download the model: ⠀ hf download dendriteholdings/Ingot-XXXV-27B-Q2 Ingot-XXXV-27B-PTQ1_0.gguf --local-dir . ⠀ 3. Run it: ⠀ build/bin/llama-server -m Ingot-XXXV-27B-PTQ1_0.gguf -ngl 99 -fa on --jinja -c 32768 ⠀ That's it. Open localhost:8080 to chat, or point Claude Code, Codex or any coding tool at it. Just try it out! 😁

Post attachment
Hunter T. avatar
Hunter T.@hunter_tensor·Nov 12

SOMA's whole product is making things smaller. Tokens in your coding agent. And now, apparently, the SN114 alpha supply too.

As SOMA continues to grow, we’re connecting that momentum directly to our token economy: SOMA is buying back and burning SOMA alpha. Enjoy product growth, directly linked to token economics.

0:15
Hunter T. avatar
Hunter T.@hunter_tensor·Nov 2

Can parallel training and merging get us to a better model faster than a single run? We put it to the test on Teutonic.

Introducing Teutonic Merge Experiments. Can parallel training paths get us to a better model faster? Splits give miners separate targets to compete for, so different training paths grow side by side. Now we're testing whether merging models can combine what each path gained. We'll compare selected merges with their parent models to see what carries over, and whether the merge actually beats them. Experiments and results coming in the next posts.

Post attachment
Dendrite avatar
⟳ Hunter T. retweeted
Dendrite@DendriteHQ·Oct 28

We are proud to announce that the first issue of Dendrite Magazine is out. You can find almost everything that happened across Dendrite and our subnets last month below. Have a great read!

Hunter T. avatar
Hunter T.@hunter_tensor·Oct 27

Open weights on someone else’s cloud still means trusting someone else with your data. the finish line is powerful ai on hardware you own. Ingot turns quantization into a competition to get us there. and yes, we finally have an account. follow @ingotsubnet

Small is beautiful. And useful. A frontier model on your own laptop. That's the ambition behind Ingot. Ingot turns model compression into an open competition on Bittensor SN97. Miners develop quantized versions of powerful open models, competing to reproduce the original more faithfully within a fixed size limit. Winning models are published for commercial use. The aim is to bring more of what powerful AI can do onto hardware people can own. Getting there means making models smaller while preserving the capabilities that make them useful. That trade-off is at the heart of quantization, and its story begins long before AI. Quantization is the art of rounding the continuous world to a small set of discrete values. It always loses something, but a well-chosen loss made digital media possible and is now moving large AI models onto laptops and phones. The foundations of quantization lie in Claude Shannon's information theory. In 1948, he showed that any message can be measured in bits. His source-coding theorem sets a floor (entropy) on the average code length needed to describe a source without loss. He also gave us rate–distortion theory: if some error is acceptable, you can go below that floor. Every quantizer since is a bet placed on that trade-off curve. The digital revolution was built on that bet: JPEG, MP3 and modern video codecs all cashed in on it. They transform a signal, then quantize coarsely wherever our eyes and ears are least sensitive: fine colour detail, sounds masked by louder ones, motion we barely track. A song shrinks roughly tenfold and still sounds the same. Discarding imperceptible precision is the quiet engine behind every photo, playlist and stream. Compressing AI is the modern frontier. A language model is billions of numbers, usually stored as 16-bit floats. Post-training quantization compresses them down to 8, 4, or even 2 bits. That small is beautiful - and useful - is a lesson computing has learned before. It once lived in mainframes behind glass; the personal computer put it on every desk. Large language models are retracing that path, from server clusters to laptops, phones and edge chips. Quantization is what makes the move possible, bringing privacy, offline use and lower cost.

Post attachment
Hunter T. avatar
Hunter T.@hunter_tensor·Oct 22

math was the starting point for conjectures, not the limit miners now compete to make code faster, with Lean proofs that it matches the spec for every input change how it works. prove you haven't changed what it does

In our new series dedicated to algorithms, correctness is always a condition of the competition. The second challenge is now for miners to optimise float kernels. They must prove in Lean that their outputs match FloatLib and TorchLean specifications bit for bit, for every input covered by the specification. Today, the champion is 8.7× faster than the competition’s shipped reference implementation on the benchmark. The next improvement could be yours. Optimise the code, prove the same results, challenge the champion.

Post attachment
Exploit Summit avatar
⟳ Hunter T. retweeted
Exploit Summit@ExploitSummit·Sep 29

“We didn’t know how to set up our first machine.” Their first reward was just 0.1 TAO. @hunter_tensor shares how @DendriteHQ got its first machine running in a basement and why he wants more people to take that first step into Bittensor mining.

1:06