Ben Fielding
@fenbielding
Followers
5K
Following
10K
Media
198
Statuses
2K
Co-founder & CEO @GensynAI - the network for machine intelligence. I like modular, composable, decentralised, and evolutionary machine learning
London / Remote
Joined December 2014
pleased to open source this work - NoLoCo extends pipeline + data-parallel model training to heterogeneous gossip networks by modifying momentum and dynamically routing shards.
Introducing NoLoCo. NoLoCo trains large models over heterogeneous gossip networks, rather than high-bandwidth datacentres. It reduces synchronisation latency by 10x vs state of the art methods while converging 4% faster to the same validation loss. We're open sourcing it today.
11
9
100
welcome to the team @austinvirts 🔥.
new chapter đź‘‹ . I have officially joined @gensynai to help lead marketing and growth efforts . Gensyn is building out the decentralized network for machine intelligence by pulling together key resources for ML to flourish alongside humans.
5
0
16
learn but verify.
2
1
28
RT @jeffwilser: What does the future of AI agents look like?. On the latest episode of THE PEOPLE'S AI podcast, @_grieve and Shaw Walters d….
0
9
0
RT @fenbielding: @Shaughnessy119 @Polymarket @UMAprotocol yep, needs verification of model execution settled indisputably onchain. markets….
0
1
0
new twitter bio / personal website tagline meta just dropped - "Interesting fact: cats sleep most of their lives,".
1
0
10
RT @andrewjscott: 7percent portfolio founder @fenbielding (@gensynai).contributed to part 2 of our “Beyond Moore - The Future of Compute” p….
0
2
0
"AI will automate everything and there will be nothing left for humans to do"
4
0
10
the most important trait of founders / early startup employees is either not knowing or genuinely not caring that something is "impossible".
15
6
49
this but decentralised machine learning.
0
1
17
if you're in Cannes tonight, come catch the @gensynai team and friends at our rooftop mixer. food, drinks, chat, sunset.
6
5
48
vibe code a saas or become a tomato farmer. which way, modern man?.
7
1
24
this is quite stark in coding products. Devin = autonomy; Cursor = autonomy and augmentation. the IDE world is still sleeping on proper augmentation tools - I don't want to instruct it what to build and then task manage it, I want it to do all the boring shit while I create.
🧵there are only really two categories of ML: autonomy and augmentation. the world is hyper-focussed on autonomy right now (agents do tasks on your behalf) but a huge portion of the future use cases are augmentation-based (human wants to do something more effectively with tech).
2
1
16
to get to this stage, we need a uniquely scalable base infra for local + remote training (i.e. decentralised) and a new software stack for building ML-native applications around autonomy and augmentation models. we also need new algos that can deal with the new setting.
1
0
2
the true breakthrough will come when applications embed augmentation models within their stack and constantly train on human interaction. this helps train autonomy too but the gains will be less zero-to-one and more iterative.
1
0
1
the main barrier for augmentation training is quality data and therefore scale. it needs constant human interaction and the ability to deal with very long horizons and sparse rewards - this isn't impossible, it's just a massive search space so requires much more resources.
1
0
1
training for autonomy involves teaching a model to perform a task based on a definition of that task (i.e. supervised learning, verifiable reward construction, etc. ). training for augmentation involves optimising human performance through assistance without goal visibility.
1
0
1
deep research is an example of using autonomy (generate a research report) to solve an augmentation problem (make me more efficient at doing desk research). I think this leads to an inferior product vs training specifically for augmentation.
1
0
3