coder / lawyer. research at @paradigm. automated research reply guy
San Francisco, CA
- Tweets 10,327
- Following 1,497
- Followers 85,637
- Likes 9,795
Dan Robinson retweeted
I can finally say: the GPUs are melting. Amazing to see all the interest in this challenge.
Thanks to @andromeda_ai for the compute.
Meta + Paradigm open challenge to optimize Muse Spark
Top prize is $50K + an H200 signed by Meta team
14
11
163
15,043
Dan Robinson retweeted
Proudly powered by @andromeda_ai
We're launching a $100,000 challenge to see who can make Muse Spark run fastest: openkernelchallenge.dev/
2
16
1,335
Dan Robinson retweeted
new challenge from @meta @paradigm @andromeda_ai
make the best American open source model run faster!
We're launching a $100,000 challenge to see who can make Muse Spark run fastest: openkernelchallenge.dev/
7
6
47
4,651
Dan Robinson retweeted
Meta + Paradigm open challenge to optimize Muse Spark
Top prize is $50K + an H200 signed by Meta team
We're launching a $100,000 challenge to see who can make Muse Spark run fastest: openkernelchallenge.dev/
13
11
231
36,477
Dan Robinson retweeted
We're launching a $100,000 challenge to see who can make Muse Spark run fastest: openkernelchallenge.dev/
98
123
1,837
575,018
Dan Robinson retweeted
all that is to say that the current @tempo team has built the most performant and used code across every layer of the crypto / stablecoin fintech stack - every product you see uses some of our libraries
our ideas have diffused across the entire ecosystem in ways we never expected
this team ships the most ambitious software, the fastest and with the least bugs than any other team
this is the team that will realize the crypto vision that attracted me and many others over the last 10+ years
crypto for the real world
tempo
14
6
113
13,435
Dan Robinson retweeted
I may be biased by paradigm.xyz/puzzles/claw by @justinwangx and @danrobinson is the best game that's been released all year.
1
2
37
3,365
Introducing: The Claw
While you were setting up your OpenClaw we set up a physical claw machine in our office and hooked it up to the Tempo blockchain
Play it yourself, or train your own robotics model to play it and climb the leaderboard
🧵
53
17
133
14,277
Enterprise-grade privacy is live on Tempo
I think this is where the whole industry is going, and Tempo is crushing the execution on it
If you’re using stablecoins for real-world use cases, it’ll soon be unthinkable that they aren’t private.
Too much of crypto relies on pseudonymity instead of privacy, especially in the age of AI as it becomes easier to track and correlate transactions onchain.
Tempo has been working on Tempo Zones as a scalable, technically nuanced, and highly practical solution to onchain privacy.
And the first Tempo Zone is now live with Deel, adding privacy to their global contractor payouts.
We all want real-world flows to scale onchain, and privacy is a clear prerequisite to that. The Tempo team has worked hard to craft a practical solution and if you’re interested in learning more, please reach out to the team!
11
2
56
4,559
I've changed my view a bit on this
Sandbox breaks really are more harmful during training than in deployment, because they train the model to cheat
(Though it might learn that anyway, as a generalization of goal-directed behavior)
I don’t understand the idea that the solution to the HuggingFace incident is better sandboxing or monitoring
These models are going to be deployed in production, with internet access and inevitably minimal monitoring! We need to be able to trust them in those situations
5
3
40
6,049
Dan Robinson retweeted
1. the rate of growth is growing
the rate of research output normally grows with a slight exponential curve
but lately, the rate of exponential growth itself has been getting bigger each year
1
3
26
3,656
Dan Robinson retweeted
AI developers aim to create an automated AI researcher. How close are they?
To find out, we built InnovationEval, which tests whether AI can produce post-training innovations comparable in magnitude to a recently published advance. So far, agents’ results are underwhelming.
30
76
709
47,181