Georges Harik retweeted
Quite a week at humans&. It's been really inspiring to see how people are already using the things we build in exciting and unexpected ways
Tuesday: Ampersand:
- the world's first collaborator model and its API
- over 700B params, trained against Persimmon to work w/ people over long horizons with diverse, open-ended goals
Thursday: Boba:
- a socially intelligent product built on Ampersand users have loved so far
- millions of messages exchanged already
Friday: FTL decode:
- a pipeline speeding up LM text generation up to 10%, counterintuitively by using more precision
- already used in our RL training, our serving, and now open-sourced
Tuesday: Ampersand:
- the world's first collaborator model and its API
- over 700B params, trained against Persimmon to work w/ people over long horizons with diverse, open-ended goals
Thursday: Boba:
- a socially intelligent product built on Ampersand users have loved so far
- millions of messages exchanged already
Friday: FTL decode:
- a pipeline speeding up LM text generation up to 10%, counterintuitively by using more precision
- already used in our RL training, our serving, and now open-sourced
8
8
145
8,414
Georges Harik retweeted
Whoa, this is pretty awesome
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
1
1
10
891
Georges Harik retweeted
Awesome work from the humans& team!
@humansand's W4A16 NVFP4 work delivers 10% higher throughput and 7% lower TPOT in their SGLang serving benchmark at concurrency 8.
Love seeing SGLang and Miles helping push the boundaries of inference and RL, and seeing the innovations flow back to the open-source community.
Congrats to everyone involved!
@humansand's W4A16 NVFP4 work delivers 10% higher throughput and 7% lower TPOT in their SGLang serving benchmark at concurrency 8.
Love seeing SGLang and Miles helping push the boundaries of inference and RL, and seeing the innovations flow back to the open-source community.
Congrats to everyone involved!
We accelerated one of the most optimized NVFP4 decode kernels while also making it more accurate
Fewer bits typically mean more speed but less accuracy. We did the opposite: more bits, but up to 10% faster generation
Already used in Ampersand’s RL and Boba serving
1
4
40
2,345
Georges Harik retweeted
humans& is shipping again. In fact, we're shipping so much that we really don't know what to do about it. People are asking me, “Please, please, please, Saurabh, we're shipping too much. We can't take it anymore”
And I say, “No, no, no, you’re going to ship again. You’re going to ship big. You’re going to ship bigger than ever”
And I say, “No, no, no, you’re going to ship again. You’re going to ship big. You’re going to ship bigger than ever”
7
3
100
4,510
Georges Harik retweeted
We’re building AI that helps people achieve more together, and sharing our work so everyone can build on it. We’ve contributed this to FlashInfer, SGLang, and Miles - with more to come!
At humans&, this was used for the large-scale RL training our recent collaborator model, Ampersand, and is being used to serve Boba to our users
x.com/humansand/stat…
At humans&, this was used for the large-scale RL training our recent collaborator model, Ampersand, and is being used to serve Boba to our users
x.com/humansand/stat…
No one changes the world alone
Introducing Ampersand - the first model built for us. Ampersand is a collaborator trained to help people achieve their individual and collective goals. In multi-month simulations, users collaborating with Ampersand accomplished more than with Fable ls
Available to everyone through our API
0
3
14
1,832
Georges Harik retweeted
Lowkey this is insanely cool. PDL is awesome!
We accelerated one of the most optimized NVFP4 decode kernels while also making it more accurate
Fewer bits typically mean more speed but less accuracy. We did the opposite: more bits, but up to 10% faster generation
Already used in Ampersand’s RL and Boba serving
0
3
20
1,232
Georges Harik retweeted
This has been quite a week with the pace of releases from @humansand. Just in case there was any questions on how productive the team has been…here’s another!
This is pretty nice, our kernel increases both accuracy and speed at the same time and is open source available freely for everyone to use - our way of giving thanks to the open community that we have benefited from ♥️
1
1
8
958
Georges Harik retweeted
Compute in high precision to… speed up decode??
Another banger from the goat @zianglih go check it out!!!!
Another banger from the goat @zianglih go check it out!!!!
We accelerated one of the most optimized NVFP4 decode kernels while also making it more accurate
Fewer bits typically mean more speed but less accuracy. We did the opposite: more bits, but up to 10% faster generation
Already used in Ampersand’s RL and Boba serving
0
1
19
1,279
Georges Harik retweeted
We just served 1.5M inference calls across 400 cities and 60 countries in the 4 days since launch! Thank you all for the support!
Let Dream-1 control your robot today ☁️ dreamlabs.ai
Let Dream-1 control your robot today ☁️ dreamlabs.ai
6
10
91
4,680
Georges Harik retweeted
Shippers ship!
We accelerated one of the most optimized NVFP4 decode kernels while also making it more accurate
Fewer bits typically mean more speed but less accuracy. We did the opposite: more bits, but up to 10% faster generation
Already used in Ampersand’s RL and Boba serving
2
2
15
1,345
This is pretty nice, our kernel increases both accuracy and speed at the same time and is open source available freely for everyone to use - our way of giving thanks to the open community that we have benefited from ♥️
We accelerated one of the most optimized NVFP4 decode kernels while also making it more accurate
Fewer bits typically mean more speed but less accuracy. We did the opposite: more bits, but up to 10% faster generation
Already used in Ampersand’s RL and Boba serving
9
2
41
2,146
Georges Harik retweeted
We accelerated one of the most optimized NVFP4 decode kernels while also making it more accurate
Fewer bits typically mean more speed but less accuracy. We did the opposite: more bits, but up to 10% faster generation
Already used in Ampersand’s RL and Boba serving
Fewer bits typically mean more speed but less accuracy. We did the opposite: more bits, but up to 10% faster generation
Already used in Ampersand’s RL and Boba serving
17
21
295
31,646
Messaging intelligent assistants, rather than using an app, is an interesting minimalist user interface. One nice thing is it also uses minimal processing on your phone and preserves your battery.
3
0
20
1,582
Georges Harik retweeted
Boba is so cool
Introducing Boba - the social intelligence that understands your goals and your community. Text Boba today or message it online at boba.app
0
2
24
3,544
Georges Harik retweeted
Boba - the social intelligence that understands your goals and your community!
We built Boba because we believe that people change the world together. Boba is powered by Ampersand, our model trained to help people achieve their goals. By practicing over millions of simulated days and interacting with thousands of simulated people, it learns to navigate goals, disagreements, and changing plans
boba.humansand.ai/blog
0
1
9
715
Georges Harik retweeted
Helping people do more together is one of intelligence's most compelling possibilities. Love seeing @humansand bring that ambition into the group chat. Congrats on launching Boba!
Introducing Boba - the social intelligence that understands your goals and your community. Text Boba today or message it online at boba.app
9
2
30
5,965
Georges Harik retweeted
Testing Boba to help plan my upcoming Paris trip.
Amazing side: surprisingly inspired! nearly serendipitous artistic recommendations that neither ChatGPT nor Claude suggested (I specifically asked for them).
What could make it stickier: It shared no links and stopped midway, so ontinue the rest of the plan (because I used messages?).
Amazing side: surprisingly inspired! nearly serendipitous artistic recommendations that neither ChatGPT nor Claude suggested (I specifically asked for them).
What could make it stickier: It shared no links and stopped midway, so ontinue the rest of the plan (because I used messages?).
Introducing Boba - the social intelligence that understands your goals and your community. Text Boba today or message it online at boba.app
2
1
17
1,292
Georges Harik retweeted
Appreciate supporting us and grateful to work with the amazing team at sva!
I've known Georges Harik for over 25 years. Proud to back him, Eric, Yuchen, Noah, and the entire @humansand team as they build AI that understands people and helps us accomplish more together. Excited to see Boba out in the world. Give it a try!
0
2
6
672
Georges Harik retweeted
Thanks for supporting us building for humans from day 1!
Exciting to see a AI assistant that is designed to be multi-player and social. Congrats @humansand!
0
1
6
588