Poseidon (@psdnai) header
Poseidon (@psdnai) profile picture

Real-world data for physical AI

Joined July 2025
Photo by Poseidon (@psdnai)Photo by Poseidon (@psdnai)Photo by Poseidon (@psdnai)Photo by Poseidon (@psdnai)Photo by Poseidon (@psdnai)Photo by Poseidon (@psdnai)Photo by Poseidon (@psdnai)Photo by Poseidon (@psdnai)Photo by Poseidon (@psdnai)Photo by Poseidon (@psdnai)
Poseidon retweeted
so excited to share this! i led the ASR research and experiments behind this work. 10,000 hours of Bengali speech data, ~32% lower CER on Whisper large v3 :) grateful for Sandeep Chinchali’s mentorship!
Better data is the breakthrough. Today, we're publishing a case study on how our 10,000-hour Bengali speech dataset reduced Character Error Rate (CER) by ~32% on Whisper Large V3. ↓
2
4
388
Better data beats bigger models. That's what the latest @psdnai case study proves. 10K hours of Bengali speech, built for how people actually speak: ▪️ Regional dialects + code-switching ▪️ Verified native speakers, linguists in the loop ▪️ Whisper Large V3 fine-tuned → ~32% lower CER ▪️ Beat India’s top sovereign API Proud to have worked alongside @iamrashminagpal @jdubpark @ZK_pawgchain @snehasanks on this release! Full case study ↓ teapaw.org/psdnai/status/21227556…
Better data is the breakthrough. Today, we're publishing a case study on how our 10,000-hour Bengali speech dataset reduced Character Error Rate (CER) by ~32% on Whisper Large V3. ↓
1
13
440
Poseidon retweeted
Very proud to have shipped this case study, alongside legends: @iamrashminagpal @jdubpark @mskristagambrel @ZK_pawgchain @snehasanks @nickipp12 onward ⚔️ psdn.ai/blog/case-study-10k-…
Better data is the breakthrough. Today, we're publishing a case study on how our 10,000-hour Bengali speech dataset reduced Character Error Rate (CER) by ~32% on Whisper Large V3. ↓
1
1
5
226
Most data collection starts with what’s easy to collect. At @PsdnAI we start with what models get wrong. We care about how data improves the models. (1) Create a holistic benchmark (2) Find where models & APIs fail (3) Run collection campaigns across high-value channels for that data (4) QC and annotate (5) Fine-tune, improve, & benchmark OSS models on that data -> loop We’re now taking the same loop to Physical AI. We care about how ego video, audio, and tactile data improve models and robots in the wild. DM me for more info on how you can get involved and benefit from this loop!
Better data is the breakthrough. Today, we're publishing a case study on how our 10,000-hour Bengali speech dataset reduced Character Error Rate (CER) by ~32% on Whisper Large V3. ↓
2
7
379
1/ Fun #CoRL2026 dinner hosted by @GradientVC & @goaly_ai in Austin earlier this week, with a great crowd of robotics builders! Hot debates: can Astra and Fable run robots on their own or do we still need robot FMs? And does more video actually make for better data? 🧵 thanks for everyone who joined us! @AdemiAdeniji (@GroundedSI), @drxcliu (@DeepReach_AI ), @itsghvr (@MeckaAI), @jamesw_stevens (@enact_labs), @krisPsdn(@psdnai), @kun_h____ (@DynaRobotics), @kylerduffy, @markgrinev(@MeckaAI), @NickelReady (@DatoricAI), @Parvkpr(@genrobotics_ai), @praveenvnktsh (@Uber), @DanielXieee(@origami_robo), @sebliu(@psdnai), @Anciukevicius(@Waymo), @vincentjliu(@GroundedSI), @liuying04(@scale_AI)
3
1
1
43
2,891
Poseidon retweeted
Very high signal contribution to the robotics training space from Poseidon’s bleeding edge research! Open for anyone to use. Happy training and evaluating 💪
Today, we're releasing EgoCrowd-500H, a new dataset accompanying our @corl_conf 2026 research on quality failures in crowdsourced egocentric video. • 300 stereo clips • 2.5 hours of footage • 50 human-reviewed clips Out now on @huggingface.
3
4
24
1,281
We will open source more data and research as we expand our capacity. For researchers at CoRL, reach out for various data! Happy to collaborate on improving models with globally diverse datasets
Today, we're releasing EgoCrowd-500H, a new dataset accompanying our @corl_conf 2026 research on quality failures in crowdsourced egocentric video. • 300 stereo clips • 2.5 hours of footage • 50 human-reviewed clips Out now on @huggingface.
1
4
14
558
Poseidon retweeted
last week in SF i went to a robotics for healthcare event. stuff that is being worked on is pure sci fi: head surgeon at Neuralink gave a talk around controlling the brain.. tldr, pretty much everything that happens in our body is controlled by the brain (either consciously or unconsciously), especially the deeper we go into the brain. so if you can control your brain via an implant, imagine a future where you can decide your fertility, turn off your depression, increase your concentration, all via an app? The talk was a theoretical thought exercise, but even so, some massive philosophical and ethical dilemmas are about to be uncovered soon. another pretty cool thing was seeing some robots that are being worked on that are essentially microscopic screws that inch their way into your brain via external magnets. a researcher showed me this box containing a microrobot...can you find the robot in this photo? one big takeaway for me was how it seemed everyone there was complaining about training data; how little of it there is, and how bad most of it is today. every single company there is looking for data that is: - higher resolution (e.g. EEGs are very low res) - higher in quantity (i.e. never enough data!) - and more longitudinal (i.e. spanning more time). if you're in the data business, my advice is look into healthcare robotics because they are absolutely starved for data. Ideally multi modal to increase signal (e.g. EEG + tactile + vision, etc), over longer periods of time. very fascinating industry, likely one of the most impactful for our future. Will be keeping a close eye on it.
9
1
38
2,085
In today's AI landscape, every dataset must pass three critical questions: • Can you source it at scale? • Can you prove where it came from? • Can you trust its quality? DATA Network is built to answer all three.
41
17
5
114
22,691
For years, every voice AI conversation passed through text. Speech ▸ text ▸ model ▸ back to speech. Now, speech models cut the middleman, with far more realistic results. But while the architecture has made leaps, one bottleneck remains: data. Enter: @otodotearth x @psdnai
30
17
3
109
21,505
Poseidon retweeted
1/ We built oto so people can have real conversations and earn, while creating the data the next generation of voice AI actually needs. Today we're partnering with @psdnai + @StoryProtocol. Your conversations, your consent, verifiable onchain. teapaw.org/psdnai/status/20698144…
11
5
2
58
8,240
Reminder → Numo is now available for speakers of: ▸ Filipino ▸ Indonesian ▸ Malay Contribute → earn.
56
30
6
194
378,606
Tamil has reached its contribution goal on Numo. நன்றி to everyone who contributed. This is what compounding momentum looks like.
49
27
7
141
10,262
The internet was the first dataset. What's next is being built now. Meet @SPChinchali from @Psdnai↓
73
63
12
207
19,937
760k+ submissions on Numo from 13k+ contributors in 1 month. Turns out, when you make it simple to contribute to AI, people show up. Still early.
76
35
28
163
19,558
Week one Numo stats: ▸ 210,000+ contributions ▸ 18,000+ contributors ▸ Contributions growing ~40% day over day Real voices, real fast. Onward.
seven days ago we launched Numo with @psdnai. a billion people speak hindi, bengali, tamil, telugu. less than 0.1% of voice ai training data is in any of them (like in microsoft's VibeVoice latest model). nobody was fixing this big training gap, so we started collecting thousands of hours from contributors all over india. week one stats: > 210,000+ contributions > 18,000+ contributors > contributions growing about 40% day over day every contribution registered as IP. every contributor rewarded. every second of audio licensed at the moment it's create on story. Vietnamese launched just hours ago and we already have thousands of contributions! here's a some fun facts about why we are focusing on Vietnamese: + Vietnamese has six tones across three regional dialect systems: Northern, Central, Southern, with big differences in tone systems, vocabulary, and pronunciation + Central Vietnamese is the most different and it’s the most challenging dialect, even for other Vietnamese speakers + Most voice AI is trained on Northern Vietnamese which misses roughly half of how the country actually speaks thanks to Numo's collection efforts AI models will be able to upskill on conversational nuances and be more realistic. we're also not stopping at voice. we'll start rolling out other forms of tasks soon, always focused on long tail, hard to scrape real world data the models can't easily scrape from the internet.
50
24
19
165
20,927
We're now at 110,000+ submissions on Numo. That's over a hundred thousand contributions from 15,000+ speakers of: Bengali বাংলা Hindi हिंदी Tamil தமிழ் Telugu తెలుగు Things are just getting started. More soon.
73
44
21
223
18,123
The future of AI depends on quality training data. Eco app @psdnai is building the marketplace for it. Meet @SPChinchali ↓
73
61
15
249
33,252
Poseidon retweeted
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 ↓
Most teams collecting voice data optimize for volume over quality, partly because they’re measuring quality wrong. To help evaluate quality we created the Poseidon Score. When applied, single-speaker audio scored well while multi-speaker conversations scored worse. Why? ↓
5
12
28
2,646
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
23
2,806