sarick

I've spent a decade building AI systems in telco, logistics, finance, and healthcare. Each time, the issues trace back to the same problem: data. Training data is the most under-valued, under-coordinated input in the entire AI stack. It's fragmented and challenging to make compliant, and the people who create it often see none of the upside. Here is our take on the current landscape: - Compute is centralized and priced in (see: Nvidia $4T and AMD: $255B). - Models are open-sourcing and the competitive advantage of releasing new architectures is decreasing rapidly (see: OpenAI, Anthropic, xAI: worth $500B+ combined). - The only frontier left unsolved and unpriced? Data. This is validated by Meta's recent investment in Scale AI for $14B, leaving a huge gap for IP-cleared training data. At @storyprotocol, I led research on influence functions, specifically tackling the core problem of data attribution by measuring which datapoints were actually responsible for a model's outputs. It was my first step toward rethinking how we value data. Earlier this year, @SPChinchali and I started sketching a solution. What if contributors got recurring upside? What if every reuse paid forward? What if data worked like IP? That idea turned into @psdnai. Working at @StoryProtocol with @WhatTheLJW, a master of operations and strategy + @storysylee, a visionary leader with true outside-the-box thinking, helped shape this vision. Our initial focus is on physical AI, robotics, and audiovisual information. However, Poseidon is designed to excel in healthcare, biometrics, sensor data, and beyond. Because of the volume of data we are handling for the world's leading AI companies (yes, in the works), Poseidon would not be possible without @StoryProtocol's IP licensing infrastructure where registration is streamlined and royalties and derivatives are automatically tracked. If the data can't be scraped, we're building the stack to coordinate and license it. This mission is personal. It comes from a fundamental tension I've witnessed my entire career, from academic labs to industry. I saw medical AI learn from deeply personal patient data. I built models for telecom, finance, and logistics on the digital footprints and real-world actions of millions. The pattern was always the same: The data was the core asset, but it was never treated or priced as such. This is the market we're going after. More to come.

Should have gone with rust 😭

what happens when your eng team brags about only using CC and haven't coded in months lol

voice is increasing the surface area of software. coding went from keyboard → editor → compile to speak → intent → execution. everyone is hyper focused on the intent/execution agents, but completely ignoring the ground truth layer. the interface can change overnight, but the capability is hard-capped by the diversity of the audio data underneath it. you can't fake phonetic nuances with synthetic data.

this fortune piece frames it as a data problem. i'd push further to the fact that it's a coordination problem. the data exists. dashcams, surgical suites, warehouses. the issue is nobody has built the incentive layer to unlock it at scale. that's the actual bottleneck. fortune.com/2026/03/06/ai-…

your eval pipeline might have a major blind spot. when you isolate a speaker's channel in a dual-speaker recording, silence fills the gaps where the other person was talking. that silence is structural, but most ASR scoring metrics treat it as a transcription failure. we found this while building the poseidon score – a 15 point gap between single and dual-speaker scores that our human reviewers couldn't hear. the audio itself was fine. the metric was reacting to format, not quality. best-of-n preprocessing across six strategies closed the gap. full breakdown on this and more in our latest blog ↓

re: cnbc.com/2026/02/28/goo…

everyone wants to build the “android of robotics”, but the constraint isn’t the sdk. you can't simulate your way out of messy factories, bad lighting, and unpredictable humans. physical ai expands the surface area of software, which means it exposes the limits of synthetic data. the real world is the only training ground that matters, and the race to collect long-tail data is going to be a full contact sport.

dt and elevenlabs putting agents directly into live phone calls is a massive signal. when ai sits mid-conversation to translate or pull data on the fly, every human conversation becomes a programmable surface area. but the final boss here is latency and dialect recognition. the moment this actually works flawlessly for low-resource languages, global transaction costs collapse.

digital twins simulate the system. edge models run the operation. robots execute. the industries with the highest physical friction are the ones with the biggest incentives to adopt agentic workflows. once the physical-ai-to-edge loop actually works in warehouses and factories, an entire layer of enterprise ops consulting is just going to evaporate overnight.

cool system but let's be real: voice cloning from 5 seconds of audio is a parlor trick if the underlying model is trained on 99% english. you clone a voice but it still sounds like it's speaking through a wall when you ask it to do anything in a language outside the training distribution. the hard part nobody's solving: building models that actually understand phonetics and prosody for languages that aren't english. that's where the data moat lives.

This open-source AI can literally clone your voice Just discovered VoxCPM by OpenBMB & it’s wild It’s a tokenizer-free Text-to-Speech model that can generate natural, expressive speech & even clone a voice from a short audio clip What makes it interesting ◾️Context-aware speech – understands the meaning of text & adjusts tone & prosody automatically ◾️Zero-shot voice cloning – replicate someone’s voice from a short reference clip ◾️Real-time streaming TTS – fast enough for real applications ◾️Diffusion + LLM architecture for more realistic audio generation Even crazier: it was trained on 1.8M+ hours of bilingual speech data to improve realism & expressiveness OpenBMB This could power things like: • AI voice assistants • Audiobook generation • Content dubbing • Personalized AI avatars Repo - github.com/OpenBMB/VoxCPM

memory isn't the bottleneck. everyone's obsessed with stateful agents but they're all building on top of models trained 98% on english internet. you can't synthetic your way out of that. try building a voice agent for swahili or marathi and watch it fall apart. the real problem nobody wants to admit: you need actual human data from actual humans speaking those languages. that costs money. labs would rather pretend it doesn't exist.

I built the best memory for voice agents while sponsoring @YCombinator hackathon. Introducing: Voice AI + Memory! We built a deep @supermemory integration for @pipecat_ai, so now even your agents have great memory With user profiles, these are almost-instant latency as well :)

physical AI runs on data. way more data than people realize. everyone talks about models and synthetic generation, but those pipelines still start with real-world signals. if your base dataset is thin, your system breaks the moment it leaves the lab. the real bottleneck isn’t models. it’s high-quality data that actually reflects how the world works. forbes.com/sites/sabbirra…

“He said 64 languages were spoken in his country of 15 million people and the government agency was building AIs for the public that fuse both American and Chinese AI technologies and their own large language datasets.” if countries invest in high-quality, rights-cleared datasets across regional languages, agriculture, and public health, they shape model behavior at the training layer. ai sovereignty doesn’t come from choosing a us or china api. it comes from controlling the data your models are trained on. theguardian.com/politics/2026/…

heard a great breakdown of the voice AI stack at an event with @Fin_ai and @cartesia. voice ai is becoming one of the biggest real deployments in ai. customer service alone is a $300B+ market and still massively under-automated. some takeaways: 1. customer service is the core market. buyers care about resolution rate, handle time, deflection, csat. 2. multilingual matters but it has to work in messy environments. cars, restaurants, call centers, interruptions. 3. voice systems are judged by their worst turn. one bad answer in a long call breaks the whole experience. 4. speech-to-speech is coming, but ~90% of production systems today still run cascade (speech → llm → speech).

Frontier APIs became the fastest source of high-quality training data, so competitors treated them that way. In a world where outputs are data and data is the moat, access control and provenance become core infrastructure. All roads lead to AI’s biggest bottleneck: Data.

Big congrats to the Poly AI team. 500M+ calls and enterprise deployments - that calibre of functionality is built on vast swathes of high-quality, domain-specific voice data. When agents can switch languages, verify identities, and handle payments reliably, you’re seeing the compounding effect of serious data infra behind the scenes.

@GeminiApp the fact that i need a third party chrome extension just to organize my chats is insanity. do better @OfficialLoganK

The latest installment of Gemini Drops is here! Here’s a look at everything we shipped in February ↓

Voice is quickly becoming the default UI in growth markets. @theinformation's AI Agenda calls out the language bottleneck and includes Poseidon’s work building multilingual, production-ready audio datasets. For product teams, this means structured specs, verified contributors, automated QA, and data that can move straight into training and evaluation pipelines. theinformation.com/newsletters/ai…