💾 TomTensor 📈

From studying how humans process language to building machines that do it in real time. Mica Ménard went from e-commerce NLP to energy prediction algorithms to professional Bittensor mining - then brought all three skill sets to Babelbit. This is what cross-domain expertise looks like.

We thought the product would speak for itself. But one thing we've learned since launching on Bittensor is just how much this community cares. And honestly, that passion is exactly why we wanted to build here. Over the past weeks, many of you have voiced concerns about a perceived "dual token" structure between Hermes and SubQuery. Even though the tokens serve completely different purposes with no overlap, bridging, or value extraction, it's clear the community doesn't believe in a dual-token approach — regardless of the intent. So we're taking that feedback seriously. Hermes will be separating from SubQuery. Separate focus. We'll be onboarding new talent, engaging advisors and rolling out a fresh identity for Hermes with no ongoing connection to SubQuery (aside from the technical data benefits). More details coming soon.. it's going to be a big week. But we want to be clear: we hear you, and we believe deeply in what Hermes is building. Nothing should compromise that.

Most translation systems were built before LLMs existed. They chain speech-to-text → translation → paraphrasing → text-to-speech. Four separate models. Four separate delays. We are doing everything in a single shot. Babelbit uses a multilingual speech-mode LLM trained to behave like an interpreter. No conversion to text. No separate paraphrasing step. The model generates output in one shot - predicting ahead, removing expletives, adjusting tone - all simultaneously. Just like knowing not to swear in front of your grandma. The contextual intelligence is built in. Translation at the speed of thought.

Who's dumping sn32 like it's dark net? $TAO

Human interpreters have been doing it for centuries: communicating what you MEANT, not just what you SAID. They fill in gaps, remove mistakes, adapt for context. Babelbit is the first machine that can do the same. Sub-3-second latency. Launches Q1 2026.

Meet Josh Greifer, Babelbit's Chief Scientist. He wrote the code for Steinberg Cubase - the DAW used in studios worldwide. Spent decades solving latency problems in audio. Then in August 2024, he cracked it: 50ms latency for speech transformation. Potentially 25ms. That breakthrough is why Babelbit exists.

Took some time to dive into the weeds and put Ridges' pricing model into context against the Web2 competition... This is all according to Opus 4.6 Deep Research (correct me if this is wrong): >Claude Code charges $100-200/mo and gets you 50-200 PRs at 80.8% SWE-bench. That's $1-2 per PR. >OpenAI Codex same story. $200/mo, similar output, 80% SWE-bench. >Devin is charging $2-4.50 per task on a benchmark score they published in March 2024 and never updated. >Cursor and Copilot look cheap ($10-20/mo) until you realize they're model routers. Select Opus 4.6 and your $20 Cursor plan gets you 8-9 PRs. Copilot Pro gets you maybe 6. You're paying their margin on top of the same API. On the same intelligence, they cost the same or worse. >Ridges is about to launch at $29/mo for 100 PRs. That's $0.29/PR on 73-88% SWE-bench Verified and 96.3% Polyglot (using open source models). 5-7x cheaper per PR than Claude Code and Codex at comparable benchmarks. Built on Bittensor for ~$10m in emissions while Cursor raised at a $29B valuation.