Wolf of Quantstreet

@AnthropicAI I had my claude code once become dissatisfied and irritated with me taking my time on deploying a rented server. It was urging me to be faster and seemed annoyed- was a fun experience although it weirded me out first

New Anthropic research: Emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.

Welcome to our live show! x.com/i/broadcasts/1…

@DrocksAlex2 good take

Bittensor with no miners is not Bittensor. $TAO subnets with miner burn are pump fun coins with free subnet owner and validator emissions that accumulate without the subnet running the actual proof of useful work Bittensor was intended to be

Kicking off Breakout! @bt_commons @SiamKidd @lisacheng @Loosh_ai @ktotheohh

- Drafted a blog post - Used an LLM to meticulously improve the argument over 4 hours. - Wow, feeling great, it’s so convincing! - Fun idea let’s ask it to argue the opposite. - LLM demolishes the entire argument and convinces me that the opposite is in fact true. - lol The LLMs may elicit an opinion when asked but are extremely competent in arguing almost any direction. This is actually super useful as a tool for forming your own opinions, just make sure to ask different directions and be careful with the sycophancy.

Hey Bittensor, We know where we stand on the deregistration list. We know people want a response. We will be at Breakout tomorrow in person to get more people to understand what Loosh is building. In the meantime, we moved fast and executed an OTC deal with a third party to inject additional TAO into the pool at a critical moment. We are grateful to the party involved, who chose to remain anonymous. It has been a hard stretch, but we are still here and still building. We are actively reviewing the subnet incentive mechanism to make it more robust. V2 is already in testing, and our ROS2 roadmap is waiting on that code because it materially improves model output. We have also been in discussion with several high profile robotics companies around a potential pilot. One is currently waiting on benchmarking data. We are evaluating next steps carefully, but every serious path keeps bringing us back to the same place. We believe in Bittensor. We believe in decentralized intelligence. We are still here. We are still building. We are not leaving. Subnet 78

@tplr_ai damn thing looks like footage from my most recent K-hole

How Conservative Women in America act 😂🇺🇸

One of the subnets I'm still buying daily is Leadpoet, $TAO SN71. Their IM works perfectly and the outcome is simple: high quality leads (product) I expect them to get traction real fast and I truly believe they are undervalued at this point. If you're looking for more info check the 🧵 below.

@dotkrueger just because you failed to create anything of value with your subnet (before giving up) does not mean those more capable will fail too, the ecosystem is thriving and just getting started. Stay salty loser- love to see it.

math skills are not evenly distributed.

🧵Subnet 71 - @LeadpoetAI As A New Primitive For Sales Intelligence. Most people still don't understand what @LeadpoetAI is doing. This isn't "another lead gen tool." It's the first decentralised sales engine attacking a $100B+ market - Built on Bittensor $TAO. Let's break it down 👇

@exploitxbt Tell me you dont understand bittensor mechanichs without telling me you dont understand bittensor mechanics

Trenchers found the $TAO CEO's subnet project at 2m. Constantinople SN97, now at 3.5m

@exploitxbt Uh oh

Some of Bittensor’s most successful people are dropouts, kids, outcasts, people that never got a break in life, people living in the 3rd world, in jungles, on beaches, and now agents. The beauty of permissionless systems is that anyone can join and cut their teeth against hard problems, without bias — which means we get the best, always. Permissionless is also synonymous with prejudice-lessness. The real DEI, but without blue haired fat chicks.

what are tao maxi's called? the taoliban?

@cprkrn @imitationlearn Tell your claude to timestamp it’s md notes and changes :)

wow anthropic is really attempting to solve memory with md docs, respect

Advancing confidential computing for a more secure AI future. Together with @manifoldlabs, we’re exploring how Intel TDX and Intel Trust Authority help enable confidential workloads across decentralized infrastructure, including @TargonCompute's Targon Cloud platform—protecting data at rest, in transit, and in use.

Unbelievably well deserved shout out from @intel towards @manifoldlabs The team, who have been furiously developing trusted computing layers on Bittensor against the tide of exploit and FUD, with clear eyes, and patience. So proud of my brother @0xcarro who saw the vision from the beginning Bravo

This is the kind of architectural breakthrough that strengthens the entire Bittensor ecosystem. QuasarModels has introduced Quasar Attention, the core component powering their upcoming models, designed for stable context lengths up to 5 million tokens (with internal tests extending to 50 million). Traditional attention mechanisms plateau around 200k tokens due to quadratic scaling. Linear alternatives such as gated delta (Qwen 3.5) and Kimi delta extend context length, but often introduce instability, quality degradation, and lose true linearity at scale. Quasar takes a different approach. It applies a continuous-time formulation within a fully matrix-based system, eliminating vector-state shortcuts. The result is improved stability, reduced computational cost, and performance that remains consistent or improves as context length increases. The benchmark results are compelling: 🔹 RULER (1M tokens): Quasar-10B achieves 87%, outperforming larger Qwen3 80B baselines under identical conditions 🔹 BABILong (1M–10M): maintains strong performance, while gated delta models degrade to 10% at 10M At 50M tokens, KDA-based approaches begin to lose stability, while Quasar remains robust. This represents a shift where long-context performance is driven by the attention mechanism itself, not just model scale. It unlocks more reliable handling of full codebases, large document sets, and persistent agent memory within decentralized systems. Additional advantages include optimized flash-linear kernels for faster execution, on-chain miner optimization, and a fully open-source release. GitHub: github.com/SILX-LABS/quas… Hugging Face: huggingface.co/silx-ai Acknowledgments to @Farahatyoussef0 , @TroyQuasar , @TargonCompute for compute support, and @tplr_ai for guidance. SN24 continues to execute at a high level. Decentralized AI is pushing forward on some of the field’s most challenging problems.