Founding engineer @VirioAI | Prev: @usemesh_ai, CourseAssist
San Francisco, CA
Joined May 2014
- Tweets217
- Following450
- Followers76
- Likes296
Another security leak ??
despite the unfortunate timing, we take disclosures seriously at @typesafeai
at approximately 6:49PM this Friday evening, we became aware of a massive unauthorized leak of ALL customer noul outputs
in effort to maximize transparency, here is the link: gist.github.com/diogo149/810…
Today's a tiring day to be a software developer ...
Like how did so many 0day vulnerabilities happen on the same day?😣😣
Willow -> FluidVoice -> Willow
Love to see the improvements made in @WillowVoiceAI .
Stop #7 of our voice maxing campaign: @virioai!
We dropped by the Virio office to deliver some mics and got to spend time with the team. They’re using Willow to dictate faster and get more done. Such an awesome group of people, and really cool to see what they’re building (shoutout @doubleemt).
Thanks for having us Virio, really enjoyed getting to know you all!
I wonder if you need to get hacked by OpenAI's agents to get acquired like Hugging Face. What if Nvidia came to save OpenAI from a crazy lawsuit by acquiring HF ??
That would be some crazy documentary movie that I want to see.
Excited to share that our paper "RAMAC: Multimodal Risk-Aware Offline Reinforcement Learning and the Role of Behavior Regularization" has been accepted to ICML 2026!
What if a policy performs well on average but still produces rare catastrophic outcomes? 🧵
Come find us at the poster session!
📅 Wed, Jul 8, 2026 • 10:30 PM to 12:15 AM PDT
📍 Hall A
🌐 Project page: kaifukazawa.github.io/ramac-…
Huge thanks to my coauthors Kai Fukazawa and @ImanSoltaniPhD for letting me be part of this project.
Our work takes a closer look at a standard but often implicit ingredient in offline RL: behavior regularization as a mechanism for OOD control. Beyond behavior matching, our analysis shows it helps bound OOD action probability and stabilize lower tail optimization under shift.
We introduce RAMAC, which uses lower tail return information to suppress risky behavior modes, while behavior regularization keeps the policy close to data supported behavior.
At the same time, how can we control out of distribution (OOD) actions without sacrificing the multiple valid behaviors captured by expressive diffusion or flow policies in offline RL?
I wonder how timestamps for chat messages get used in consumer-grade chatbot like Claude, I can see multiple reasons why it might be helpful. But what I actually wonder if this goes to a point, where they have to switch to databases that are time-series focused.
Btw Happy NYE!
I feel RLHF is more imitation learning than reinforcement learning. It's just the way it feels so different the standard RL, I can't just agree to the ongoing notion of calling RLHF as "RL".
Am I the only one who noticed that new @cursor_ai ui changes make it seem very similar to @Trae_ai ? I guess @Trae_ai was slightly ahead of the curve in the UI part.
Has anyone seen interesting file uploader designs, that gather insights from the files using LLMs, at upload in a concise manner? Really looking for some creative approaches for some inspiration.
This was a fun experience. A closed hackathon where I got to meet with most of the attendees. Our team actually built a working demo.
We might release a public version of the game for people to try soon. No payments, but obviously auth will be @stack_auth
Big W for @alessandroduico, @eskoskin, and @MundadaM for building this insanely creative and polished game using our new payment feature. 🔥👏
Kunal M. retweeted
🏆 Massive shoutout to our hackathon winners!
🥇 Banana Guesser — @alessandroduico @eskoskin @MundadaM
🥈 Team Orbit — @deviamar622 @jeewoo_meche @GTanvi_
🥉 BlaBlah VC — @RVAClassic @BeomsooSon @selenemiyu
A big thank you to all teams who participated — we loved the energy, the creativity, and finally meeting our community in SF.
Can’t wait to meet more of you soon. 🚀
Humans are intelligent creatures cause they have excellent ability to adapt. Being able to store info in our brains isn't a good measure of our intelligence.
Thus LLMs != AI, cause at the end humans will adapt to this tech too.
Is there a similar list for Indian researchers? Would love to dig deeper on what factors have contributed to their success? Also I want to see the scale.
Sota Singaporean AI researcher list 👇
I made some list two years ago and thought it's time to revisit it with some new names (in no order and no elaborations/cot this time)
1) @isaacongjw from Anthropic
2) @alyssamloo from Anthropic
3) @jennyzhangzt from UBC
4) @tanshawn from MIT-IBM
5) @PangWeiKoh from UW
6) @zhansheng from OpenAI
7) @kohjingyu from Meta
Imo, these are the current best Singaporean AI people now which I think are great.
Aside from AI research I think @JiquanNgiam and @swyx from the AI startup world are quite cool too.
Only referring people who are Singaporeans around the world and not people working or studying in SG.
PS: on a flight, wifi too spotty to work properly so made this fun list 😃