UT Austin Prof. Chief Scientist at Poseidon AI. Work on GenAI and Networked Intelligence. Stanford CS PhD. All views my own.

Austin, TX
Most voice AI products run on conversational data, and that training data is sourced from multi-speaker audio. But how confident can we be that our evaluation pipelines are actually measuring what we think they are? We built the Poseidon Score to evaluate. Findings in our latest blog ↓
4
5
16
1,072
We tested six preprocessing strategies to remove the artifact – amplitude thresholds, timestamp trimming, neural voice activity detection, diarization, and others. Performance varied clip by clip, so we solved this with a Best-of-N approach where we ran all six, scored each output, and kept the highest. Dual-speaker scores recovered to 0.856.
1
3
8
830
Our takeaway is that evaluation pipelines need to be built for the shape of the data they're measuring, not just the content itself. A metric calibrated on single-speaker audio will systematically misrepresent conversational data. On low-resource languages like Bengali where labeled data is scarce, that misrepresentation carries real cost. psdn.ai/blog/the-complexity-…
2
8
332
Speech models degrade sharply as you move off high-resource English benchmarks. @theinformation's AI Agenda surfaces the structural issue and references Poseidon’s work collecting multilingual, rights-cleared audio across dialects and domains. Without rigorous data curation and ground-truth validation, scaling voice systems globally remains an unsolved problem.
In this morning's AI Agenda, I break down why researchers say that audio AI models need to improve on languages other than English—and why this is especially important as OpenAI and others increasingly expend into countries outside the Western hemisphere. theinformation.com/newslette…
6
4
24
2,758
As AI moves into robotics and autonomy, the weak point isn’t just code or prompts. If you can manipulate the camera or audio input, you can manipulate the model’s behavior, which is very worrisome. Multimodal systems ingest pixels, audio, and sensor data. Adversarial inputs can be injected through a sign, a sound, or poisoned training data. High-quality data isn’t enough - more than ever, we need provenance and validation. If you can’t trace and verify the data pipeline, you cannot reason about model behavior. axios.com/2026/02/10/sentine…
1
2
12
448
If this works, cities will take an entirely new form. Imagine a pipe bursts and traffic is instantly rerouted. A grid under strain sheds load before blackouts. Trains adjust frequency based on real passenger flow, not last year’s averages. The system responds in real time. To get there, you need dense, real-world sensor data, models that understand physical constraints, and data that can actually be used in safety-critical settings. weforum.org/stories/2026/02/…
2
12
411
As AI systems integrate into infrastructure, telecom, healthcare, and robotics, data provenance becomes a systems-level constraint. If you cannot trace where training data originated, under what license, and whether consent is revocable, you introduce uncertainty into evaluation and deployment. While @Cloudflare’s AI Crawl Control gives websites a machine-readable way to signal training permissions, what is the equivalent primitive for individuals? Voice recordings captured by satellites or smartphone devices, behavioral data from wearables, and so on. Personal data is continuous and multimodal, making it far more challenging since it does not live on a single server behind a robots.txt file. As models move into physical systems, weak data lineage becomes a reliability and governance risk. We need provenance that is programmatic, auditable, and enforceable at the data layer itself. reuters.com/legal/litigation…
1
16
499
Negative agreement across 3 frontier models on the same Bengali transcripts suggests something deeper than random variance. In low-resource, dialect-rich settings, "correct" is not singular. Models absorb different norms from fragmented corpora, making evaluation itself unstable. We unpack the implications in our latest blog, including the role of human benchmarking.
Poseidon needs voice data and reliable ground truth in low-resource languages to benchmark against. To ensure LLM transcript accuracy, we worked with linguists to audit Bengali outputs. For a language spoken by 280M people, the gaps we found point to a deeper issue: data ↓
3
5
34
2,126
I had a great time on the Think Data podcast discussing why data and evaluation are becoming the limiting factors for multimodal systems. In low-resource settings and edge deployments, progress depends on curated long-tail examples, rigorous quality control, and benchmarks that surface failure modes early. At Poseidon, we are building infra to source rights-cleared datasets and make them usable for training and continuous testing.
Poseidon Co-Founder and Chief Scientist, @SPChinchali, joins the latest ThinkData Podcast. Sandeep breaks down why, as multimodal and physical AI scales: • Rights-cleared, high-quality data is the core constraint • Rigorous validation is required to trust AI in production And more ↓
9
2
48
1,964
We’re seeing a clear shift from volume to provenance in AI data. After a year dominated by large-scale scraping and legal uncertainty, model developers are beginning to demand transparency around data origin, rights, and consent. Licensed, verifiable datasets will increasingly define competitive advantage.
15
22
90
6,640
Congrats to the World Labs team. A $1B round signals that world models are moving from R&D to a key strategic priority. If models are to act coherently in physical space, they need structured representations of dynamics, geometry, and causality, not just next-token prediction. That raises a harder systems question: how do we collect and align egocentric video, multimodal sensor streams, and rare edge cases at scale, with clear provenance and licensing? World models will only be as robust as the real-world data pipelines that train them.
World Labs has raised $1 billion in new funding. We are grateful and excited to partner with our investors, including AMD, Autodesk, Emerson Collective, Fidelity Management & Research Company, NVIDIA, and Sea, among others. worldlabs.ai/blog/funding-20…
1
1
18
724
Precision data overtakes the era of digital-sweatshops. The marginal returns of generic scraped data are declining. Model performance improvements now depend more on precision datasets: curated, expert-generated, legally licensed, and context-rich. Data quality and relevance increasingly matter more than scale alone.
9
10
80
6,093
AI power will collapse into a handful of winners, but physical data is still up for grabs. Power across the AI stack continues to concentrate – a small number of players dominate compute and foundation models. Data providers will consolidate into two or three major players per modality, like text, audio, and video. Real-world data remains an exception, the largest untapped asset class in AI.
3
1
14
328
This article makes good points. Audio models do spend more compute on emotional cues, latency constraints are real, and the deeper issue is data. We saw this in 2020 when everyone shifted to remote. Suddenly Slack messages replaced watercooler context, short messages felt harsh, and neutral emails read as hostile. Voice is even less forgiving – prosody, hesitation, sarcasm, calm vs frustration. These signals must be learned from diverse, rights-cleared data. Until we fix data fidelity and coverage, voice AI will lag.
VCs and founders have told me voice AI is on the verge of a breakthrough for months. Here's why it's so hard to get right, compared to text-based AI: theinformation.com/newslette…
3
16
544
Physical AI is inevitable. Every major lab is building world models because this is the next great frontier. But robotics exposes the constraints fast: perception noise, long-tail edge cases, embodiment, latency. You can’t prompt your way around bad real-world data. The bottleneck to this reality will be high-fidelity, task-specific datasets.
3
5
44
2,266
An AI winter is coming, driven by leverage and data-center debt. Frontier models are already "good enough" for most use cases, while training and inference costs keep climbing. Much of today’s buildout is funded by borrowed capital and circular deals between the same few players. If real revenue doesn’t catch up, compute spending becomes unsustainable. The likely result is consolidation.
1
15
320
Hume AI’s work highlights an under appreciated constraint in voice systems: access to labeled audio that captures affect, prosody, and emotional context. These signals are sparse, culturally specific, and difficult to synthesize. Progress here depends less on model architecture and more on the availability of the right data.
Google reportedly snags up team behind AI voice startup Hume AI techcrunch.com/2026/01/22/go…
2
11
448
Sovereign and regulated AI systems force a hard question: Can you prove your training data came from real humans, with the right skills, demographics, and consent, without violating privacy? This post breaks down why expert marketplaces alone don’t scale, and what replaces them.
Sovereign AI needs verifiable trust. As AI moves into healthcare, government, and regulated domains, the bottleneck is no longer models or compute. It’s whether training data can be trusted. Details on the blog ↓
3
1
19
401