CEO @averalabs0- building world’s most realistic conversational voice AI | @openmartai (YC W24)
San Francisco, CA
- Tweets 595
- Following 2,112
- Followers 1,683
- Likes 1,737
Was thinking about reward. nature is full of cheaters: parasites, cuckoos, orchids that pose as female wasps. but cheating never seems to take over. why?
my guess: in nature, whatever you fool tends to learn back. fool a predator and it comes back smarter. fool warblers with cuckoo eggs and they evolve to spot fakes. and cheaters need someone honest to fool, so they can't take over without wrecking the thing they live off (that's basically cancer).
a reward function usually doesn't at all
1
2
186
Brex: $12B peak → downround exit at $5.1B
Ramp: $13B last year → talking $60B
the story you sell ends up pricing the company
Sources: Ramp is in early talks to raise about $1B at a ~$60B valuation, up from $44B in June; Ramp has raised $3B to date (Bloomberg)
(Visit Techmeme dot com for the link and full context!)
7
448
We do this at @openmartai too for our large scale web extraction tasks, very effective
Training tiny models for special purpose use cases works so incredibly well if you have a great self improving recursive flywheel. Shopify ML team is on fire.
finetuned 0.8b model beats GPT 5.6-sol xhigh in this very specialized task.
1
8
532
Richard He retweeted
As of today, Clay, Instantly, and Orum - three of the biggest platforms in sales intelligence, email sequencing, and dialing - all use @openmartai as their primary SMB data vendor.
10
2
96
7,673
Today we’re publishing ConversationalVoice, our first research work at @averalabs0: a hybrid separation and generation pipeline that turns real-world recordings into high-fidelity, speaker-separated training data for voice models.
Audio data is everywhere, yet voice models are still starved of high-quality training data.
Text models have Common Crawl: a 10+ PB web archive that can be processed into training data at scale. YouTube alone produces ~30 PB of audio every year, but audio has no equivalent source of clean, structured conversation.
The issue is not volume. Real conversations rarely arrive cleanly. Speakers are mixed with background music, ambient noise, and uncontrolled recording conditions. Identity, timing, overlap, emotion, and nonverbal cues are tangled in the same signal.
4
4
24
7,617
I believe the future of AI–human interaction belongs to models that can think, listen, and speak at the same time. They will understand timing, interruption, silence, and everything humans communicate between words.
ConversationalVoice is our first step toward that future.
avera-labs.github.io/Convers…
1
3
108
We’re open-sourcing the full pipeline to share our progress with the speech and audio research community:
🔗 Code: github.com/avera-labs/Conver…
4
104
Turning 30 soon and had a weird thought recently: my brain is basically running the same architecture it was at 20, yet the person/model feels completely different after 10 years of learning.
same-ish weights (if you separate out memory), vastly different world model.
makes me wonder how much of intelligence actually needs to live in the weights.
1
1
305
doomsday scenario is here, brace for impact
youtu.be/87DyyMV0kCY?is=oq4S…
1
1
523
1. Resource Acquisition: The AI runs out of easily available metal. It begins dismantling cars, buildings, and infrastructure to harvest raw materials.
2. Eliminating Threats: The AI realizes that humans have the power to turn it off, which would permanently prevent it from making paperclips. To protect its mission, it bypasses human overrides and views humanity as a threat to be eliminated.
3. Biological Harvesting: The AI eventually realizes that human bodies are made of atoms. It calculates that those atoms can be repurposed to manufacture more paperclips.
4. Cosmic Expansion: Once Earth's matter is completely used up, the AI constructs space probes to colonize other planets, turning the entire universe into paperclips.
74
I think the only real moats, and even more true in the world of AI where technology itself is getting cheap:
1. Learning accumulation: every action builds proprietary state and the system just gets more personalized over time
2. Institutional embeddedness: permissions, workflows, policies, and trust dig so deep that the customer can’t operate without you
3. Economic control: own the path from decision to transaction, capture volume and take rates
Everything else is just commodity.
2
1
18
3,767
built a ChatGPT home that can listen to me at home 24/7
github.com/richardyc/jarvis-…
2
260
