I don’t remember where i read the tweet “if its worth reading read it slowly” but its so true !!!

What is this, not starting the race when it is raining? @F1 Are the drivers not competent enough to drive in the rain? What is the freaking use of the wet tire? This is just stupid. Why are we only racing in dry conditions? There are wet tires for rain. This can't be the epitome of motor racing #SingaporeGP

never have i ever seen so many live things fail during a demo. This is not ment to troll but damnnnnnnnnnn !!!

Opus 5.5 is just too fast and accurate !!!

In this version of Parrot 🦜 #STT from @ringg_ai, we didn't focus on simply throwing millions of hours of data at it and hoping it covers every scenario, instead, we focused on exposing the model to the data that actually solves what the voice AI needs. We have put in datasets meticulously curated by the team towards solving name, address, and entity recognition. We have also put in effort towards making the model faster by improving the decoder, which is the bottleneck when it comes to latency. Do try it out on ringg.ai/models/speech-to-te…

With the current state of voice agents, one can only do so much with a cascading system consisting of 3 models (llm,stt,tts) LST*. "SELLING" a voice agent now has thankfully come down to showing business outcomes, and not just pitching a cool gimmick to an org that wants to be "AI Native" because the whole world is drowning in an AI tsunami. Parallel to traditional SaaS sales, where the pitch is for both the buyer and the user optimizing for the one that holds more say in the org, voice agents are sold pre and post POC, where POC is generally restricted to a small set of live customers for an org. Pre-poc is mostly around pricing, feature-set competition, integration flexibility, and observability, while post-pos is around latencies, conversational design, adherence, business impact, and A/B results. Only using LST wins you the pre-poc pitch but leaves a lot of room for failure when it comes to language switching, function calling, conversational design, llm-routing, PII masking, fast multi-node switching, to name a few. One needs small specialized models (not talking about SLM here) that do all these. Having these smaller models also gives flexibility on the size of the LLM, which currently is the biggest latency bottleneck in the system. BTW, if one can solve all this with one LLM, you are either using a huge closed-source model or hosting an OSS model on a huge number of GPUs; either way, margins go for a toss, and prompting can only take you to great demos, but your tail in production is always sloppy. In India, fortunately or unfortunately, the tail is huge. At @ringg_ai, for the last ~4 months, we have been working through each piece of this puzzle, getting the individual components right, and more importantly, getting the cascading orchestration between them right. The team that wins outcomes will not necessarily be the one with access to the best model. Nor will it simply be the one with the most custom orchestration. It will be the one that marries the two best. The right model for the right decision. The right guardrails around it. Fast systems for narrow tasks. A strong LLM where reasoning and conversation actually matter, and an orchestration layer that makes all of this work as one reliable experience for the customer. Because customers do not buy an LLM, an STT model, or a TTS voice. They buy a conversation that completes the job.

Please gift me DGX spark :)

In a world of LLMs where people are so comfortable with just throwing huge prompts at it, re-evaluate smaller encoder-decoder models. They just might be the key ...

Get your Claude and Codex to talk to each other and give it a task and go to sleep !!

• hi→en epoch 8 improved to 78.43% validation word accuracy (CER 0.0418), about 0.58 points above IndicXlit’s measured baseline. Epoch 9 has started at 4,523 seconds; the reverse direction is continuing to improve. it has beaten it !!!!!!!1

So for some time at @ringg_ai we had been using XLIT/IndicXlit from @ai4bharat for our internal usecases and honestly its a great model and i really want to thank the team for making it opensource. To test @OpenAI's GPT6-Astra I gave it the base data to train a simple encoder-decoder model on and it is at 76.29 compared to 76.73 (which is the benchmark) and it is a much simpler model. It tried around ~13 different character based encoder-decoder model. By the way all this is being trained locally on my M5-Pro. Crazy Times we live in !!!!

One Reset Left @thsottiaux 🙈. All data was prepped. to make sure that data was validated i had given Astra access to my openrouter it created a consensus system to get inference from gemma, qwen for agreement. one run of training is done and is recorded in WandB......

This is true i feel anxious if i dont have something running a harness

I have just given @OpenAI Astra read access to db, L40s gpu, a couple of base encoder models to train with an NLI objective and have asked it to track everything on @wandb. Let’s see what happens. Stay Tuned #weekendWithAstra !!!

Making conscious actions >>>> moving with the flow

Umm, maybe a different take, but i think ~95% of us wouldn't really notice a difference between models that perform, let's say, 70 vs 73 on the SWE bench, since it only improves by 3-4 score points on other benchmarks, which rarely impacts goal completion. The goal completion impact is so low because almost any task that different orgs are asking it to do is well within it's intelligence limits. > where "it" = harness + model. What i think really matters now is just what gets these jobs done faster !! throughput and margin maximisation. All this obviously does not include STEM

in the podcast "How to Understand the Next Wave of AI Before Everyone Else" @thsottiaux looks like the Peter-Parker version of @TobeyMaguire

if you can identify in retrospect you can model it !!