Pinned Tweet
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.
AI is moving beyond the browser and into the real world. The bottleneck? Data.
Today we’re announcing a $15M seed round led by @a16zcrypto to build infra that collects, curates, and licenses high-quality data for physical AI.
Incubated by and built on @StoryProtocol.
40
9
140
10,335
ai safety isn't a one-size-fits-all thing; it's very culturally specific.
came across this linguasafe paper and it really highlights something important for anyone building ai for a global audience.
simply translating a harmful prompt from english can be an effective jailbreak. the same prompt that’s "safe" in english becomes "unsafe" in other languages, so you need real, native data to catch these issues.
24
17
105
13,594
team has been grinding to release this one 🔱
excited for people to try this out and contribute to AI in a meaningful way.
Coming soon: Poseidon App 🔱
AI needs better audio data – this starts with you. Share voice samples to help AI understand accents, noise, and real conversations.
Users may earn Poseidon points and, from time to time, additional partner rewards. Stay tuned.
23
3
84
8,773
robotics was the obvious first stop for us. real urgency, messy multimodal data, buyers actually ready to pay.
but @psdnai wasn’t built for just one vertical.
it’s for anything you can’t crawl: audio, biometric signals, sensors, you name it.
---
every vertical’s different, but the stack is the same:
validated. labeled. licensed. structured.
12
11
72
6,672
scaling AI requires infrastructure that can handle immense data volume, complexity, and rapid evolution across modalities. most current setups are bespoke, leading to inefficiencies.
we're building @psdnai to provide the scalable, standardized backbone providing data for world models.
this isn't just about data; it's about enabling AI at enterprise scale through a global community of users equitably compensated for their contributions
12
8
69
3,771
we’re often asked: why does Poseidon use blockchain infrastructure? isn’t this just Scale AI onchain?
the answer is no, and the difference isn’t superficial. here’s why distributed systems matter for AI’s data bottleneck:
6
7
60
4,388
robotics is running into the same ceiling language models hit before they got their boost from massive, varied datasets.
an interesting model and training approach is showing real gains, making the case for richer, more diverse, high-quality data.
EgoVLA is a new vision-language-action model that trains on egocentric human videos, turning everyday human manipulation into
scalable manipulation skills with just minimal robot fine-tuning
7
10
61
5,243
research like Seed LiveInterpret 2.0, which offers real-time speech translation in your own voice, doesn't happen in a vacuum.
this leap forward is powered by massive, diverse training sets like the RealSI dataset, which captures natural, real-world speech across many topics.
it's a powerful reminder that the future of AI relies on high-quality, openly licensed data.
when communities and companies contribute to public benchmarks, they fuel innovation for everyone.
This is why I'm bullish about building @psdnai on @StoryProtocol – IP-clearance is baked into the core of our infra, making it possible to ensure training data is licensed at scale.
24
13
56
17,538
every company we’ve spoken to rebuilds similar infrastructure – custom data spec, manual QA scripts, internal labeling pipelines, offline licensing workflows, etc.
this is inefficient, prone to errors, and not built for the scale that the leading AI companies are collectively operating at now.
we’ve replaced all of that with modular primitives on @psdnai:
→ sdks for structured collection
→ ml pipelines for deduplication, PII checks, and outlier detection
→ semi supervised labeling with active learning and uncertainty routing
→ IP-cleared via @StoryProtocol
18
7
57
10,539
embodied AI data is unpredictable, multi-modal, and deeply tied to its environment. the pipelines powering LLMs can’t keep up.
real-world edge cases aren’t scrapable. they must be orchestrated and verified.
recent work by Zhu et al. and the survey on large multimodal reasoning models shows why:
→ LMRMs choke on noisy, dynamic inputs.
→ tool usage is fragile
→ long-horizon planning in physical settings is still far off.
4
10
48
3,273
a recent study uncovered that DataComp CommonPool, one of the major open-source datasets for training image generation models, includes hundreds of millions of documents with personal information such as passports, resumes, credit cards, and birth certificates.
none of it rights-cleared and also traceable to real people.
this is broken. the system says: if it’s online, it’s free game. no consent. no licensing. no attribution. just scrape and train.
data should come with provenance, licensing, and economic rights.
read more about their findings:
linkedin.com/pulse/major-ai-…
8
5
37
6,607
BRB, hypnotized
BRB training AI with my voice data.
Unique orbs are dropping on the Poseidon App.
Get involved 🔱 app.psdn.ai 🔱
7
1
32
764
nice that you did this but if you're going based off API costs now then its no match for gemini cli or claude code... and that means even cline becomes a better option as you have granular control over context sent...
1
29
6,839
Official Open Source - Unitree Robotics
unitree.com/mobile/opensourc…
In case anyone forgot, sick release @UnitreeRobotics
Unitree Introducing | Unitree R1 Intelligent Companion Price from $5900
Join us to develop/customize, ultra-lightweight at approximately 25kg, integrated with a Large Multimodal Model for voice and images, let's accelerate the advent of the agent era!🥰
5
2
30
4,754
ya'll are burning up our compute 👀 fixes incoming
We’ve seen an incredible volume of contributors on the Poseidon App. So much so that some users are experiencing delays and rejections in their data processing status.
This isn’t an issue with your recording or uploads. We’ve already deployed a fix to the processing pipeline. Status updates will resume over the next few hours.
Appreciate your patience (and your voice) as we scale towards real-world impact 🔱
4
1
25
623
Monetize household chores till we no longer need to do them
Imagine never folding laundry again.
The internet trained LLMs, but robots and self-driving cars need training data that is much more challenging to source.
Poseidon enables the collection, curation, and licensing of real-world data to accelerate physical AI.
Use cases:
5
1
21
695
If you're a @worldcoinfnd user, give our app a go!
🌐🌐🌐
The Poseidon App is live now on the @worldnetwork Mini App store.
Sign up to contribute audio, earn points, and see your data power real products:
→ Voice AI that understands accents
→ Phones that hear clearly
→ Cars that respond safely
1
1
21
587
@Google terrible terrible terrible decision If it's not optional.
18
1,054
founder of Kimi K2: "High quality data growth is slow"
@psdnai aims to change that.
Kimi's founder, Zhilin Yang's interview is out.
Again, you can let Kimi translate for you: ) lots of insights there.
mp.weixin.qq.com/s/uqUGwJLO3…
Several takes:
1/ Base Model Focus: K2 aims to be a solid base model. We've found that high-quality data growth is slow, and multi-modal data doesn't significantly boost textual "IQ." So, we focus on maximizing every data token's value — token efficiency.
2/ Data Rephrasing: With 30T tokens, only a small portion is high-quality data (billions of tokens). We rephrase these to make them more efficient for the model, improving generalization.
3/ Agentic Ability: We aim to enhance generalization. The biggest challenge is making the model generalize well beyond specific tasks. RL improves this over supervised fine-tuning (SFT).
4/ AI-Native Training: We're exploring more AI-native ways to train models. If AI can do good alignment research, it'll generalize better, beyond single-task optimization.
5/ RL vs SFT: RL's generalization is better, as it learns from on-policy samples, but it has its limits. RL helps improve specific tasks, but it's hard to generalize to all scenarios without tailored tasks.
6/ Long Contexts: Context length is crucial, we need millions. The challenge is balancing model size and context length for optimal performance, as some architectures improve with long context but worsen with short ones.
4
14
543
Great talk from @edisonz0718 !!
@edisonz0718 just delivered a powerful talk at #OnchainFrontiersSummit — breaking down how crypto-native incentives and on-chain attribution can accelerate robotics data collection.
This is exactly the future we’re building at @psdnai x @StoryProtocol.
2
18
314
pretty sick paper from tencent... ARC-Hunyuan-Video-7B tackles a core problem; getting LMMs to actually understand messy, real-world short videos.
interesting structured video comprehension approach, integrating vision, audio, and text.
I do think this is a crutch for temporal grounding. relying on timestamp overlays raises a key question: are we creating AI that understands time or just reads the clocks we provide?
code: github.com/TencentARC/ARC-Hu…
2
1
17
522
Hit up @gardo_martinez if you meet the reqs !
Do you speak and read Hindi, Japanese, and/or Korean?
I'm doing some user testing and could use some help. Should only take a few minutes. LMK if you're interested! Thank you!
@StoryProtocol @StoryEcosystem @sarick
1
15
482
semianalysis.com/2025/06/25/…
PYTORCH_NO_POWERPLANT_BLOWUP=1 might be the best env var i've seen yet....
1
15
298
One of the few American institutions really pushing the boundaries and open source, great news
With fresh support of $75M from @NSF and $77M from @NVIDIA, we’re set to scale our open model ecosystem, bolster the infrastructure behind it, and fast‑track reproducible AI research to unlock the next wave of scientific discovery. 💡
4
10
443
Just ran a finetune + influence factor run with OLMoE-1B using EKFAC... Layer 15's experts show some solid specialization patterns - only ~40% of variance captured by the first principal component, with expected spikes in specific neuron activations.
3
12
297
Allen Institute released another banger...MolmoAct, an action reasoning model that can reason in space.
It isn’t just “robot sees + acts.” it’s a 3-stage reasoning loop:
1. predicts depth tokens (3d structure)
2. sketches a 2d motion plan (trajectory trace)
3. turns it into exact motor commands
you can decode each step: depth → an actual depth map, trace → a path overlaid on the camera frame, actions → the motor deltas it’ll execute. see the plan before anything moves.
2
1
12
377
How are people running ambient agents in web3? Would love to see a dope use case on @StoryProtocol , might just do it myself
4
1
10
536
this is so relevant for what we're doing @psdnai with @storyprotocol to build something that's genuinely safe and helpful for everyone, you can't cut corners. you need deep, culturally aware data for testing and evaluation.
and that kind of high-quality, nuanced data really only comes from proper licensing. you can't just scrape your way to responsible ai. it’s a reminder that building for the whole world takes a real investment in representative data.
(alphaxiv.org/pdf/2508.12733)
1
1
11
403
this research inadvertently spotlights the data crunch in robotics.
EgoVLA trained on just ~500k image-action pairs from human vids, which is tiny compared to the trillions of tokens llms gobble up. teams are scrambling for diverse datasets, and that's where @psdnai comes in.
our decentralized platform pulls in global contributors for diverse domain specific training data.
stay tuned for more
1
9
232
why are my chats to gemini as an assistant combined with my intentional AI chats? "what time does the mall close today" or "set a reminder to do X" should not be in my chat history ????
2
8
1,293
all these transcription meeting ones are so bad still.
we need models trained on audio recorded in distorted noisy environments with ground truth transcripts
1
8
202
they use the following tasks to come to this conclusion:
Implicit Reasoning:
→ Metaphor Understanding
→ Theme Understanding
→ Emotion Recognition
→ Comment Matching
→ Implicit Symbol Interpretation
Explicit Reasoning:
→ Causal Reasoning
→ Sequential Structure
→ Counterintuitive Events
→ Cross-modal Transfer Reasoning
→ Video Type and Intent
Read more:
arxiv.org/abs/2506.04141
8
198
gemini 2.5 pro w/ 1m context is fire... @OfficialLoganK , ya'll need to work with @cursor_ai to get the agent flows rock solid tho...
1
6
254
more interesting outcome from the ablations: even with all that, slashing robot demo data to 50% tanks success from 45% to 7% on long tasks. and we can see more scale/diversity in human pretraining increases performance.
1
7
245
One of the hardest working teams I've had the pleasure of working with while keeping the vibes immaculate
1
5
95
github.com/sarick-story/lang…
quick and dirty one built on story :P but yeah super convenient for agents and developers (cursor/cline integration)
3
44
everything’s open: weights, code, eval scripts, and the MolmoAct dataset.
if you care about robots that actually work in the messy real world, this is worth a read. nice work! @DJiafei
report: huggingface.co/allenai/Molmo…
model: huggingface.co/collections/a…
datasets: huggingface.co/collections/a…
3
154
github.com/getAsterisk/claud…
pretty dope in conjunction with CC.
1
4
574
result: EgoVLA's human video pretraining boosts generalization... holding steady at ~70% success on unseen visuals for short tasks (vs. a 23% drop without it), and ~30% on long-horizon ones, with failures mostly at the end stages.
1
4
129
key takeaway: human and robot action spaces are geometrically similar, so pretraining on human vids transfers surprisingly well to robot policies.
1
4
157
Could write a wall of text here but we're seeing many companies unwilling to work with scale after meta essentially bought them (Google, openai, and many more)
Also the pie is only growing, room for multiple players.
techcrunch.com/2025/06/18/op…
1
3
112
we decided to build an 8x 4090 rig instead and just have one 512 studio now...unfortunately didn't run ccl
2
77
likely from referrals... we are monitoring for unusual patterns, including spam, as part of our data quality process :)
2
3
96
reducing friction for IP registration is 🔑
Introducing the IP Portal's Public Preview.
Public Preview is the first step towards a fully-fledged IP monetization platform that allows anyone to earn for their creativity.
Explore: portal.story.foundation
Details ↴
3
199
mid-training on this set adds +5.5% success, especially for OOD tasks. good data → better robots.
synthetic pretraining is useful, but it’s not enough.
models need real sensor noise, occlusions, object variation - stuff sims can’t fake.
that’s why molmoact’s OOD jump (+23.3% vs π0-fast) comes from real-world diversity.
1
3
65
Claude code inside cursor + runyoyo.com/
+ github.com/eyaltoledano/clau…
is my current jam
3
119
how about allowing us to paste more than one image at a time. such a terrible oversight and the app has been out for how long now?
1
1
997
most VLA models map RGB pixels + words → actions.
MolmoAct actually reasons in space - depth tokens handle distance/occlusion and trajectory traces make motion planning explicit.
training starts with a vision–language backbone (Molmo) pre-trained on multimodal + robot reasoning data for tasks like discretized control, 2d pointing, trajectory drawing, and perception token prediction.
post-training uses multi-view camera images plus language or drawn traces to generate depth tokens → motion traces → final actions.
the action tokenization trick: instead of random ids for each motion bin, they align neighboring bins to neighboring byte tokens.
→ preserves geometry of action space
→ smoother learning
→ 9,216 gpu hrs vs gr00t n1’s 50k hrs
1
2
85
this might look ridiculous but it works hella well, better than any traditional travel pillow i've tried:
sleeperhold.com/
2
2
105
pretty neat, do you have docs anywhere? would love to see a list of all SEC filings considered...i saw defm14, s4, TO, sc13...what about 424b3 etc
1
2
205
performance gains:
→ 70.5% zero-shot success on google robot tasks (beats π0 & gr00t n1)
→ 86.6% avg on Libero (+6.3% vs ThinkAct)
→ real-world: +10% (single arm), +22.7% (bimanual) over π0-fast
1
2
82
Decade is a long time in the space. But for physics, real world > simulated especially on fine motor tasks, and i don't see that changing for a while.
2
1
54