babelbit.ai

Here's what we've been working on and where we're going... Q1: foundation complete ✅ Live demos showing interpretation in action. New incentive mechanism rewarding verified performance. Miners competing on #SN59. Q2 is where it gets interesting - feature-specific contests, new language training, early adopter partnerships. Let's build.

That's a good way to put it. Coincidentally I devised an ultra-low-bandwidth telephony system using semantic encoding about 30 years ago. I didn't think of patenting it until some years later but someone beat me to it (ten years after me). It involved reconstructing speech from a stream of encoded syllables (about 10 bytes/second). One nice thing about it was that it's low enough bandwidth to disguise the signal atmospheric radio interference.

@babelbit @LouiseBeattie So, lossless compression of meaning? Very cool indeed! If you're the first, you get naming rights. :)

There’s no benchmark for what we’re building. So we’re creating one. We are working on precise benchmarking so that we can show the gains we are making over Google Translate and other translation systems. Most translation systems optimise for literal accuracy. They measure latency as delay between “equivalent words.” That misses the point. We measure something else → time from expression to understanding Not when the words arrive. When the meaning lands. For example: 🥐In French: 'Je pense que vous avez tout à fait raison.' 🌍Google says: 'I think you’re absolutely right.' 🚀Babelbit says: 'Agreed.' Babelbit wins out because average word delay is a broken benchmark. Because interpretation ≠ translation. We’re building a benchmark that reflects how humans actually communicate. Stay tuned 👀

If you want to understand what interpretation actually means - and why translation software has never achieved it, visit → babelbit.ai/demo This our other demo, in English, that makes the distinction impossible to miss.

Introducing our live translation demo: For decades, researchers have tried to automate simultaneous interpretation. Even after WaveNet and the deep learning breakthroughs of the late 2010s, it remained out of reach. The reason is architectural: every system chains speech-to-text → translation → text-to-speech sequentially. Each stage adds latency - and even with best-of-breed models at every step, the result still cannot do what every professional interpreter does as a matter of course. A human interpreter does not merely translate accurately. They paraphrase for brevity, remove repetition, filter expletives, clarify ambiguity, and apply cultural sensitivity - integrating all of this in a single cognitive act, in real time. This video demonstrates where we are 👇 0:13 - We set 11Labs' state-of-the-art system running on 17 seconds of a simple French lesson and show Babelbit translating the same clip live, while we wait for 11Labs to finish. 1:03 - Babelbit output, under 2 seconds latency: "Today we are in Strasbourg. A very nice city in the northeast of France." 1:49 - 11Labs finishes. Their output: "Today, we are currently situated in Strasburg. It is a very beautiful city located in the northeastern region of France." Both are accurate. Only one is interpretation. 2:19 - We explain why the difference matters: verbosity accumulates. In live speech, a system that translates every word eventually falls behind the speaker. Paraphrasing is not a stylistic choice - it is what keeps the listener tracking. 5:31 - We show the current limit. On a live address by President Macron, accuracy degrades while latency holds. This is the open research problem - and the basis of our competition. 6:44 -The competition: improve accuracy without sacrificing latency. A compulsory accuracy threshold applies; entries are ranked by latency. Prediction-based techniques are encouraged.

April 23, 1564: Shakespeare understood that meaning lives in complete thoughts, not individual words. 462 years later, we're teaching machines the same thing. Phrase-level prediction. Real-time interpretation.

1,000 followers ☑️ 1,000 early believers in machine interpretation. 1,000 people who understand the power of phrase-level prediction. Thanks for being here. Q2 2026 🚀

Finally, the skills of a human translator in a machine form: babelbit.ai/demo/

New demo: Watch interpretation skills in action. Our demo shows the skills we're training models to master: ✓ Compress rambling into clarity ✓ Remove accidental repetition ✓ Filter expletives while preserving meaning This is the target. Miners on SN59 compete to build models that can do this - in real time, across languages. The capabilities that turn translation into interpretation.

🚨 The most impressive $TAO subnet founder story Babelbit SN59 @babelbit isn't a crypto project that discovered AI. It's 30 years of speech technology research that discovered Bittensor. Bittensor is bigger than crypto or what it should be, it acts like a magnet. It brings together people who would have never naturally met, researchers, founders, developers, domain experts, and aligns them around a common mission. They support each other’s progress because every breakthrough strengthens the whole. That is a very powerful thing. Matthew Karas @matthew_karas built one of the UK's first multilingual search engines at BBC News Online in 1997, covering 47 languages. Two years before Baidu existed. He worked alongside Mike Lynch, founder of Autonomy an £11 billion company built on the thesis that statistical analysis could extract meaning from text better than grammatical parsing. For three decades, Matthew worked on one problem: making recorded and live speech as useful as text. He built systems that cut documentary editing time by 75%. He deployed speech indexing across corporate markets. He kept pushing the frontier. In real-time speech, speed is everything. Then in August 2024, his colleague Josh Greifer called with news that changed everything: 50 milliseconds of latency for speech transformation. Potentially 25ms. That was the kind of number that makes an experienced person stop and realize: this could finally be good enough to change the whole category. Mike Lynch was supposed to hear about it over a pint the following week. He died in a yacht accident four days later. This breakthrough was not just technical, it was also deeply personal. That breakthrough became @babelbit. Here's why this is different from everything else in AI translation: Every translation system you've ever used works word by word. It waits for you to finish speaking, converts each word, and outputs the result. Every error, every mishearing, every confusion gets repeated. That’s translation. Babelbit is building interpretation. When someone says, “I pledge allegiance to the...” a human interpreter already knows where it’s going. They don’t wait for “flag.” They translate the thought, not just the words. Babelbit’s LLMs aim to do the same thing. Not next-word prediction. Utterance completion. The system commits to a translation as soon as it can adequately predict the rest of the sentence. Sub-3-second latency. Interpretation-grade quality. Self-corrections, which happen constantly in real conversation, get handled the way a human interpreter would: process the context, catch the correction, output only the final clean version. The architecture is serious: a two-stream design with one low-latency stream for live conversation and one high-accuracy stream for a trusted translation of record. Custom metrics like EATP, Lead, and ACS do not just measure accuracy. They measure how early accurate predictions can be made. Matthew said it himself: building this as a centralized company in 2024 meant going head-to-head with Google, Meta, and OpenAI. Bittensor offered a different path: Babelbit was built using @AffineSN120 decentralized training at scale, incentivized iteration, and an ecosystem of complementary subnets like @chutes_ai, @MacrocosmosAI, and @hippius_subnet. The real-time translation market is projected to exceed $29B by 2030 French-English real-time interpretation is launching next week. V2 infrastructure is deployed. This is what lt real use case looks like. Decades of domain expertise. Human interpreters immediately recognize. Mathematical. There is nothing like this in crypto. There is barely anything like it in centralized AI. Babelbit did not come to Bittensor because it was trendy. It came because the architecture fit the problem. That’s what many miss. When world-class builders choose Bittensor not for the token, but for the infrastructure, it starts proving itself. $TAO DYOR

@TheKryptovert @tennismandu22 Sorry for the delay - we're still working on the live speech translation demo but will be sharing a peek at our interpreter tomorrow or the day after

I can’t wait to see what their solution is capable of. It would be great if we could get a little demo of what’s happening behind the scenes 👀 @babelbit

Update: The Arena is now live⚡ We’ve moved to 80% Arena emissions 🏟️ From now: → Verified performance will drive the majority of rewards → Top submissions are deployed + tested before payouts This is the shift from scoring models → proving them And the first step toward turning contributions into the Babelbit interpreter 🌍

Our take on last night's events from founder @matthew_karas 👇 __________________ When I got up this morning, the Bittensor community was in a frenzy. I have not had time to dig deeply into all of it yet, but I can see why it’s rattling people. What it reinforces, however, is that the primary way to prevent power concentrating through capital, even in a decentralised system, is to keep creating useful things and bring more people in. If Bittensor continues to attract strong contributors, it will get stronger. Over time, events like this matter less. If the behaviour of one insider can move things this much, it is a reminder of how early the network still is. For us, Bittensor is already enabling something quite unusual. We are training models at a level of scale and complexity that would normally require big tech resources, without needing more than a six-figure runway. That also gives us enough protection to operate without putting the whole team at risk. The network is funding real work. It is paying the bills, and it is attracting engineers of the kind you would normally find at places like DeepMind or Anthropic. _________________ Thank you to this community for their continued support of what we're doing at Babelbit. Keep building ⛏️

Appreciate being featured alongside some incredible builders 🔥

Natural speech is messy - we repeat for emphasis, we stammer, we use filler words. Professional interpreters don't output the mess. They output the meaning. Babelbit does the same. You speak naturally. Your audience hears professionally. Demo coming soon... stay tuned 🔥

Why this matters: This isn’t just a reward update. It’s how we turn competition into real product progress. ✅ Reproducible performance ✅ Stronger alignment between results and rewards ✅ A clear path from contribution → deployment Long term, this creates a pipeline from competitive systems into the actual Babelbit interpreter. When we move to speech, top contributions won’t just win - they’ll become part of the product 🌍 Full breakdown in Discord 👇 discord.com/channels/79967… 💬 Questions welcome

Babelbit Incentive Mechanism Update ⚙️ We’ve started rolling out the new two-round system: 🎯 1. Qualifying → open evaluation 🏟️ 2. The Arena → verified performance Right now, we’re: - Integrating submissions into our system - Testing end-to-end flows - Preparing for verified evaluation 📦 The new repo is out. Submissions can now flow into the infrastructure that will power The Arena 🏟️ ⚠️ The Arena isn’t live yet - this is groundwork. NEXT WEEK: ⚡ The Arena goes live → Rewards shift to 20% Qualifying / 80% Arena → Verified performance begins driving outcomes

maybe if you say your team sent an explanation, please DM me which email it was from and tell me what the subject was. That way I might have found it. Please bear in mind that your staff (several different ones) have sent me about 15 different messages. How do I know which one you are referring to?

hey, I hear you and I know payroll delays are no joke. our team did send through an explanation on your ticket via email, so if you haven't seen it yet, definitely worth checking that thread again (spam folder too, just in case). if something's still unclear after reading it, reply directly on the ticket and the team can walk you through it from there.

@krakensupport - BUSINESS ACCOUNT WITHDRAWAL BLOCKED - PAYROLL AT RISK. - NO EXPLANATION FROM SUPPORT - TICKET #20852057 As a business dependent on efficient cryptocurrency transfers, it is extremely disappointing that the most important transaction of the month (paying our staff) is being help up bt Kraken with: - NO EXPLANATION - NO CONTACT BY PHONE - REPETITIVE EMAILS AND TEXT CHAT WHICH DOES NOT GET CLOSER TO RESOLUTION

I have followed instructions. It does not explain why the money in GBP in my business Kraken account is put on hold whenever I try to use it for a transaction. If someone could phone me and tak me through it, or if someone could DM me here, and explain step-by-step, how to unlock my money, that would help. The last time I tried to follow the instructions via the "withdraw" process, it said You've initiatated a withdrawal then within seconds, the list of transactions included the one I initiated but the the words "on hold" next to it. I need direct real time help from a human being, not a set of instructions which 1 - I have tried before 2 - I have already explained that I have tried before. Worst of all, you give me instructions to access parts of the site, which do not exist, e.g. click on profile to get certain menu items which are not there,