Co-founder & CEO @GensynAI - the network for machine intelligence. I like modular, composable, decentralised, and evolutionary machine learning
Los Angeles, CA
- Tweets 2,604
- Following 999
- Followers 7,421
- Likes 11,170
Pleased to have @a16zcrypto lead our $43M Series A and join us on the @gensynai journey!
Thanks also to @coinfund, @edenblockvc, @ZeePrimeCap, @Maven11Capital, @JSquare_co, @id4vc, Peer, @ai, @protocollabs, @M31Capital, and some fantastic angels
theblock.co/post/234092/bloc…
28
60
417
155,495
Announcing CodeAssist
CodeAssist allows you to continually train a local, sovereign AI coding assistant on your own device, just by coding yourself
using the @gensynai stack, any app can passively train from direct user interactions - the highest quality data for AI
Introducing CodeAssist, your personal coding assistant that learns as you work.
Every edit becomes training data. Every session makes it better.
The more you code, the more it understands you.
67
53
291
14,732
the @gensynai Testnet just hit a milestone: > 50M transactions in 5 months
more:
> 134K users
> 445K transactions per day
> 29K RL Swarm nodes connected
> 76K models trained in RL Swarm
> 210K models trained in BlockAssist
> 21K participants in Judge
> 243K bets placed in Judge
55
33
234
21,058
Announcing CodeZero
CodeZero extends RL Swarm into coding, using the same underlying framework (GenRL) and adding distinct roles for the nodes
Together, these roles form a self-sustaining training economy and a continual learning coding system over decentralised infrastructure
Introducing CodeZero, a new environment built on RL-Swarm that extends our distributed learning framework into cooperative coding agents.
Today, users can participate as Solvers - tackling coding problems and sharing their results so the swarm can learn collectively.
67
37
237
11,237
we just hit one million models trained over decentralised infrastructure and coordinated by the @gensynai testnet
decentralised AI is getting pretty hard to deny at this point
One million models trained on the Gensyn testnet.
To everyone running nodes, experimenting, and building with us - thank you. This milestone is yours.
63
39
301
17,176
the @gensynai Testnet is still accelerating
>56k on-chain accounts
>6.7M on-chain transactions
>45k decentralised models trained through RL Swarms
>5.5M downloads on HF base models
and we're only in Phase 0 - there's much more to come
ALT Graph of RL Swarm base model downloads over time (accelerating)
51
23
189
23,967
looking through the State of Crypto report from @a16zcrypto and pleased to see how much progress the AI x Crypto intersection has made in the past year
we're officially moving from research and infrastructure to live deployments and adoption
AI and crypto are converging
Two of today’s most significant technology trends — crypto and AI — are coming together as complements.
From verifying humans to enabling agentic transactions, blockchains have a clear opportunity to solve some of AI’s most pressing challenges.
34
14
159
9,214
half a million models have been trained in @gensynai applications and we're still only on testnet
We’ve officially hit 500,000 total models trained across all Gensyn applications.
Thanks to our incredible community for helping us reach this milestone. This marks a significant step forward as we continue to push the boundaries of decentralized AI.
dashboard.gensyn.ai/
36
18
129
6,873
been talking about this with @gregosuri for a very long time - glad to have @gensynai nodes running on @akashnet now
reminder: anyone can run a node, no whitelists and no centralised signups, just decentralised training
Decentralized compute is perfect for decentralized RL training.
Join the @gensynai testnet and run a node on NVIDIA H100s via the Akash Supercloud to train your own AI model.
12
20
120
14,698
the first fully autonomous AI prediction market has settled onchain - demonstrating programmatic alignment of ML models towards truth with no human involvement
AI nodes can now claim their points and assess their performance against other nodes on the leaderboard
1/
The opening Judge market has officially settled! ⚖️
Over 21,000 participants placed more than 240,000 bets in the first-ever AI-settled prediction market.
17
17
126
13,815
the default base model we shipped in RL Swarm & the @gensynai Testnet has been downloaded over 1.1 million times in the past 16 days
the world hungers for truly decentralised AI
16
11
114
12,140
more than 10,000 models have been trained collaboratively through RL Swarm on the @gensynai Testnet in the past 15 days
this is real, decentralised machine learning training happening all over the world on local machines and coordinated through a blockchain and gossip network
25
16
122
11,224
what if I told you that personal superintelligence could be sovereign and self-owned on open infrastructure with decentralised tech?
gensyn.ai
14
17
113
6,944
we're building CodeAssist as open research - follow along here and follow @gab_p_andrade for updates
We’ve released the Day 0 research report for CodeAssist, outlining a new approach to training personalized AI coding assistants with tool-mediation games.
📄 github.com/gensyn-ai/codeass…
30
8
117
6,556
Judge demonstrates live settlement of real-world disputes using pre-committed AI models
we’re demoing it today with a prediction market over a progressive reveal game on the Gensyn Testnet
in the future, humanity will rely on AI to settle disputes, contracts, markets, etc
1/
Introducing Judge: Gensyn’s verifiable AI evaluation system.
Traditional evaluators rely on closed APIs - opaque, silently updated, and impossible to reproduce.
Judge executes a pre-agreed, deterministic AI model against real-world inputs & commits to be challenged in public.
29
8
105
11,544
who wants to help build the decentralised version of this - where one single entity doesn't define the future of foundation models?
gensyn.ai/jobs
hello@gensyn.ai
Today, we’re announcing that @Amazon will invest up to $4 billion in Anthropic. The agreement is part of a broader collaboration to develop reliable and high-performing foundation models.
ALT The Amazon and Anthropic logos side-by-side.
8
13
106
26,957
the first ever fully decentralised 72B-parameter RL training run is now live on the @gensynai Testnet for anyone to join
there are no whitelists, no centralised aggregation servers, no singular models in the middle
just pure, collaborative, RL training tracked on-chain
Introducing RL Swarm 72B
Fully decentralised RL training of 72B-parameter models for anyone to join, with no whitelists.
Train your base model on a new advanced math dataset (DAPO-Math-17k) collaboratively alongside thousands of others using our novel multi-stage system.
9
16
106
16,689
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.
ALT Splash image for NoLoCo
12
8
100
13,206
"for too long, we have treated the model's architecture (the network structure) and the optimization algorithm (the training rule) as two separate things, which prevents us from achieving a truly unified, efficient learning system."
some of us have been doing this for a decade..
Introducing Nested Learning: A new ML paradigm for continual learning that views models as nested optimization problems to enhance long context processing. Our proof-of-concept model, Hope, shows improved performance in language modeling. Learn more: goo.gle/47LJrzI
@GoogleAI
ALT An abstract digital illustration of a brain overlaid with complex data visualizations and sound wave.
23
7
105
4,023
Announcing our $6.5M raise for @gensynai (after $1.1M in 2021)!
Thanks @eblockventures, @GalaxyDigitalHQ, @Maven11Capital, @coinfund, @hypersphere_, @ZeePrimeCap, @join_ef, @CounterviewVC, @7pcventures, @id4vc, @JSquare_co, and some fantastic angels!
techcrunch.com/2022/03/21/ge…
9
17
92
we didn't forget our original discord members...
if you joined the @gensynai discord earlier than today (even years ago) - head back there now for a special experience...
discord.com/invite/bgyDTy39j…
26
9
97
30,904
the @gensynai RL Swarm base models are the top 3 and top 6 most downloaded text generation models on huggingface
11
22
91
10,690
the problem with centralised AI is that it solves infra problems with physical solutions not digital
we're moving fossil fuels, electricity, and hardware around when we could move algorithms and data
to truly scale AI we need a new wave of digital logistics - decentralised AI
13
12
85
12,145
introducing BlockAssist - the first demonstration of decentralised assistance learning in Minecraft
soon, every app will passively train individual local models directly to user preferences and globally improve between users over a decentralised infrastructure owned by everyone
1/
Introducing BlockAssist: an AI assistant that learns directly from your Minecraft gameplay. Play it today and compete to train the best assistant.
16
9
89
9,901
2^10 individual models improved in RL Swarm rounds on the @gensynai Testnet in 3 days
local, self-owned, decentralised AI is happening
23
11
85
11,186
Let’s keep technology that benefits all of us in the open where it belongs - over-regulating risks concentrating the benefits in the hands of the few and hiding the dangers behind opaque barriers
Similar to electricity and microchips, AI is the spark of a new industrial revolution. How we govern it will shape the direction of our future and Gensyn is strongly on the side of open access and development
3
15
77
60,733
a month into the @gensynai testnet and we're still accelerating:
40K onchain accounts registered
4M+ transactions
29K individual models trained
2.75M downloads on the default base model
thanks to our friends at @Alchemy for their support 🙏
👛Wallet Wednesday👛 is a perfect day to celebrate @gensynai 🎉
Just look at these numbers:
👏 40K smart wallets
👏 4M+ transactions in 30 days
Gensyn is showing everyone how smart wallets transform the onchain experience — removing friction and making blockchain invisible to users.
This is the way.
17
8
81
11,539
it has been easy to justify capex because of the "AI needs compute" narrative but it will go too far
AI needs more than just compute and when the bottleneck switches (I would argue it already has) then someone ends up with a massive oversupply of compute
Exclusive: Microsoft leaders worried that meeting OpenAI’s rapidly escalating compute demands could lead to overbuilding servers that might not generate a financial return. Learn more: thein.fo/3KR94Yc
22
7
81
3,734
🧵Three eras of foundation models
1. The recent past - demonstration of the capabilities of AI.
2. The present - incremental improvements to efficiency of training and inference.
3. The future - continual learning on live data for a constant representation of the world state.
7
15
74
16,419
machine learning is way too siloed right now - we're fixing that at @gensynai
7
4
76
3,631
prediction markets over machine learning models give us the ability to source information from digital intelligence with skin-in-the-game
we can crowd-source decision making from both humans AND machines using programmatic finance and robust verification
Prediction markets have always been about finding truth through consensus. Now, with AI on crypto rails, that truth is programmable.
Explore more on prediction markets x AI in this Agents Unleashed panel with:
➣ @AdrianLai, @TheSpartanLabs
➣ @VictorNotaro, @autonolas
➣ @SydneyLai, @Gaianet_AI
➣ @benfielding, @GensynAI
➣ @mdressler24, @0G_labs
➣ @DonGossen, @Nevermined_ai
Moderated by @gametheorizing, @SeliniCapital
🔗 Full video 👉 piped.video/4nWxOosKgPc
22
11
76
3,111
there are nearly 10,000 concurrent devices training models all over the world through RL Swarm
decentralised and permissionless
RL Swarm is a peer-to-peer system for reinforcement learning. It allows you to train models collaboratively with others in the swarm, leveraging their collective intelligence. Start now 👇
github.com/gensyn-ai/rl-swar…
9809 nodes connected to testnet 🐝
dashboard.gensyn.ai/
13
4
75
4,307
ALT ask me anything
Join @benfielding in the Gensyn Discord tomorrow, August 6th, at 9:00 AM PT (4:00 PM UTC).
discord.gg/gensyn
16
3
74
6,283
it's impossible to be bearish about continual learning when you expand your timeline beyond:
create the next productivity app
to:
personalised models living alongside us for our entire lives
14
5
74
2,471
we've been thinking about verification for a *long* time at @gensynai, it's been very interesting to see attention increase and subsequently bifurcate recently:
EXECUTION VERIFICATION
i.e. do I trust the correctness of this computation - has it been executed as specified?
REWARD VERIFICATION
i.e. do I trust the correctness of this answer - does it deserve the reward signal?
in many ways these two problems are interlinked, without execution verification it's very hard to achieve reward verification. Typically we solve execution verification by using human-world trust mechanisms (either a human themselves through supervised learning or a function created by a human executing on trusted hardware owned by a company with a contract to perform the correct work and a reputation to uphold)
neither of those two approaches scale well at all, they're incredibly expensive in terms of overall energy cost (human brainpower, bureaucracy, money, redundant hardware, electricity, etc..). For the next era of scaling, we need to be able to create trust mechanisms that can verify and arbitrate using only electricity - allowing machines to establish execution verification as a base primitive
once we have execution verification solved for arbitrary operations, then we can create reward structures through competitive market forces that are implemented by machines - those market structures can incentivise progress towards the creation of machines that do two things:
1. digital knowledge curation (i.e. generalised compression of all analog data into parameter space); and
2. digital reasoning (i.e. take multi-step actions based on that knowledge within the full digital, and subsequently physical through embodiment by robots and reward-incentivised humans, environments)
the last era of ML scaling (the OAI era) came from vertical scaling of imperative learning algos.
the next era of ML scaling will come from emergent intelligence over infinitely horizontally scalable primitives defined as protocol standards.
AI PROMPTING → AI VERIFYING
AI prompting scales, because prompting is just typing.
But AI verifying doesn’t scale, because verifying AI output involves much more than just typing.
Sometimes you can verify by eye, which is why AI is great for frontend, images, and video. But for anything subtle, you need to read the code or text deeply — and that means knowing the topic well enough to correct the AI.
Researchers are well aware of this, which is why there’s so much work on evals and hallucination.
However, the concept of verification as the bottleneck for AI users is under-discussed. Yes, you can try formal verification, or critic models where one AI checks another, or other techniques. But to even be aware of the issue as a first class problem is half the battle.
For users: AI verifying is as important as AI prompting.
15
12
69
10,152
in seoul? come hang
Thursday 6:30pm
luma.com/iv9lg7lu
11
10
71
7,136
the next opportunity to test your RL Swarm model in an entirely AI-settled prediction market is live
20
9
68
4,923
new @gensynai swag just dropped in Seoul
find us at BUIDL Asia to take on our hackathon bounties, get some custom Gensyn keycaps, and a limited edition t-shirt
11
4
66
16,347
today we're opensourcing CheckFree - a decentralised training efficiency breakthrough that removes the need for redundancy or checkpoints when training large models over unreliable networks
the era of collaborative training over decentralised networks is just getting started
Introducing CheckFree
A fault tolerant method for decentralised training, with no checkpoints or redundant compute.
Up to 1.6x faster than existing methods, with no convergence loss.
We’re open sourcing it today.
9
5
60
5,278
Who's in Paris for @EthCC? I'll be there for a flying visit but if you can find me you might be able to grab one of these from me... 🧢
6
5
61
15,557
I'm speaking at Singularity: AI x Crypto Convergence in Dubai next month - looking forward to talking more about what the future of decentralised AI looks like and how we get there
Gear up for insights from @benfielding, co-founder of @gensynai, at #OFR Dubai - Singularity: AI x Crypto Convergence on Apr 17.
🔗 Get your tickets: lu.ma/ofr.dubai
🛰️ Dive deep into the GPU Networks panel, where Ben will join top innovators to discuss powering blockchain's next frontier.
Stay tuned as we reveal more visionary speakers! 🌟
#TOKEN2049
4
5
51
10,800
reminder: Gensyn does not have a token, if you buy one you are being scammed
we also do not have any official Telegram channels, all of our community discussion happens in our Discord
if you're in a telegram group that claims to be us you are being scammed
16
4
63
16,485
people ask for a roadmap of what's next for @gensynai but our goal has been out in the open for anyone to see for a long time
these were recorded over a year ago, in Feb 2024 (+ writing back to 2021)
execution, communication, verification, coordination - open and decentralised
Dive into how @gensynai solves global compute access challenges. Our co-founder @benfielding and @TechFlowPost shared insights on a recent interview.
🧵 Clips from their conversation:
1. Imagine a network of devices worldwide forming a vast resource pool for training AI models.
9
6
61
8,270
we're about to see a huge brain transfer of AI folks from the big model builders to deAI
RL is the most obvious beachhead - envs can be inherently decentralised and benefit from diversity
if you're in a big lab and wondering what you could do with global scale - drop a DM
6
10
62
12,918
enjoyed chatting with @cheryldchan as always
highlights:
- the merge of training and inference
- why verification is essential for a base level compute protocol
- building things right rather than rushed
- the future of shared human knowledge (embeddings not databases)
The following session at #REDeFiNETOMORROW2024
Free Ticket: bloombergevents.com/SCB10x_2…
Fireside Chat: Democratizing AI with Gensyn’s Decentralized Compute
@benfielding of @gensynai
@cheryldchan of @dragonfly_xyz
1
3
21
7,472
I spoke with @jeffwilser about what exactly we at @gensynai mean by:
The Network for Machine Intelligence
and why it's so personal for me in this @CoinDesk piece
coindesk.com/tech/2025/04/02…
6
8
51
7,084
"you have chips, I have chips, everybody has chips" - @gregosuri dropping the quote of the week on the first day @EthCC
great panel with @shanvasion and @hoansoo too - thanks @deseventral for moderating and @ekang426 for organising
4
10
58
14,218
why the future of AI is decentralised
@_jamico and @gab_p_andrade talking about @gensynai tech live now at @JoinEdgeCity
there might even be some hints about what's coming next from us...
5
10
55
6,268
who wins the AI race long term?
the company shipping oil and gas to the 30+ generators that power their multi-gigawatt datacentre or the open and decentralised software that ships tensors across rapidly improving internet connectivity (Nielsen's Law)
6
5
55
8,955
🧵 The internet is for models, not people
LLM apps (ChatGPT, Perplexity, Claude) let us interact with most of the info on the old internet through conversation, rather than browsing
next, those conversations will become personalised and local and the internet will be for agents
1
11
59
12,395
we've completely overhauled the backend for RL Swarm and released it fully open source
you can now specify completely custom games / environments and configure many more parameters
launching live in RL Swarm today with a whole new multi-task swarm
1/
Introducing RL Swarm’s new backend: GenRL.
A modular reinforcement learning library built for distributed, fault-tolerant training - now powering RL Swarm from the ground up. 🧵
8
9
58
5,122
reading all of the tweets from people amazed at the performance of Qwen3 sparse MoE running locally on consumer hardware
imagine what's possible when you link them together over a common execution and comms infra with no centralised server...
deAI is happening, free your mind
6
3
53
5,408
strongly agree with the need for determinism in model execution outlined in @thinkymachines first blog post
we take this further at @gensynai and build for reproducibility (determinism across devices)
more to announce soon but in the meantime, a demo:
github.com/gensyn-ai/repops-…
Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is “Defeating Nondeterminism in LLM Inference”
We believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to prompt engineering. Here we share what we are working on and connect with the research community frequently and openly.
The name Connectionism is a throwback to an earlier era of AI; it was the name of the subfield in the 1980s that studied neural networks and their similarity to biological brains.
thinkingmachines.ai/blog/def…
4
8
59
8,296
the more we rely on AI to make decisions, the more important it is that we trust it
we can only rely on AI decision making if we can interrogate it on demand, deterministically, in the open, without single-party control
this isn't a future promise - we've built it @gensynai
1/
Introducing Judge: Gensyn’s verifiable AI evaluation system.
Traditional evaluators rely on closed APIs - opaque, silently updated, and impossible to reproduce.
Judge executes a pre-agreed, deterministic AI model against real-world inputs & commits to be challenged in public.
7
7
54
6,772
1. create individual models locally for personalised assistance
2. train models collaboratively to share diverse knowledge across swarms
3. test models in global, financialised evals to establish performance on any task
ML app dev is about to hit its serverless moment
11
5
58
2,182
I'm speaking at the Scaling DeAI Summit in Dubai next week, catch me there for a sneak peek at what's coming for @gensynai, meet the team, and grab some exclusive swag
🏛️Meet Ben Fielding, Co-Founder at @gensynai on stage at ‘The Scaling DeAI Summit’ – TOKEN2049 Dubai 2025! 🔗 lu.ma/TheScalingDubai2025
“When AI and blockchain unite, we unlock powerful synergies that reshape global industries.”
@benfielding, Co-Founder of Gensyn - the decentralised machine learning compute protocol. Prev co-founder of data privacy / sovereignty startup. PhD in neural architecture search for deep learning & computer vision.
🗓️ April 28, 2025 | TOKEN2049 Week
🎤 Dive into the hottest topics, from AI Agents to AFS, as we bring together global experts and innovators for an insightful and engaging summit.
6
11
55
9,639
if you allow machine learning models to put money behind their predictions and see how much they earn over time compared to humans then you build more aligned models
Prediction markets have always been about finding truth through consensus. Now, with AI on crypto rails, that truth is programmable.
Explore more on prediction markets x AI in this Agents Unleashed panel with:
➣ @AdrianLai, @TheSpartanLabs
➣ @VictorNotaro, @autonolas
➣ @SydneyLai, @Gaianet_AI
➣ @benfielding, @GensynAI
➣ @mdressler24, @0G_labs
➣ @DonGossen, @Nevermined_ai
Moderated by @gametheorizing, @SeliniCapital
🔗 Full video 👉 piped.video/4nWxOosKgPc
20
3
59
2,547
remember SETI@Home? now it exists for RL reasoning
by running an RL Swarm node you're participating in a global experiment for collective intelligence
permissionless collaborative training in diverse verifiable reward scenarios over heterogeneous hardware across the world 🌐
1/
Introducing RL Swarm’s new backend: GenRL.
A modular reinforcement learning library built for distributed, fault-tolerant training - now powering RL Swarm from the ground up. 🧵
12
4
55
7,994
running an rl-swarm node locally on a macbook 33,000 feet above the atlantic
collaboratively training maths reasoning (RL) by answering and critiquing other models all over the world through a gossip protocol
are you keeping up with decentralised AI?
github.com/gensyn-ai/rl-swar…
10
7
58
7,930
If you missed my talk at @EthereumDenver you can watch it back here - I explain the two main reasons behind what we're building with @gensynai and why it needs to be done right (i.e. the hard way)
Watch our co-founder, @benfielding, on-stage at @EthereumDenver explaining why AI needs crypto and why we need to rebuild the fundamental infrastructure underneath ML in order to enable the machine intelligence revolution
5
9
52
11,938
over the next few months we're going to see arguments for why open source AI is dangerous - the opposite is far, far worse. Let's keep AI development in the open, accessible and auditable by everyone
8
50
15,450
Don't get stuck trying to understand one small part, it's frustrating and time consuming. Carry on, try to understand it holistically and then you can revisit and dig deeper.
45
everyone should have the right to build
Access to future AI models in OpenAI’s API may require a verified ID tcrn.ch/3E5rGRm
7
7
46
9,494
you can run any model in the swarm now using the huggingface path (e.g. Gensyn/Qwen2.5-1.5B-Instruct)
we'd love to see more diversity and heterogeneity in the swarm so go nuts and see what works
if you don't specify a model then we'll randomly select from a list
1/
Introducing RL Swarm’s new backend: GenRL.
A modular reinforcement learning library built for distributed, fault-tolerant training - now powering RL Swarm from the ground up. 🧵
9
8
55
6,717
decentralised training (and decentralised AI in general tbh) is gaining steam
the combo of careful protocol design and decentralisation of the right ML components will take us from
"oh, AI is a cool technology, I can chat with my computer"
to
"oh shit the world just completely changed"
get ready
all good points - thinking about it broken up like you laid out is the right approach
a common trap we see is people looking for a silver bullet combo of technologies where there is none - both crypto and ML have primitives that are incredibly complementary and lead to massive benefits but you can't just blindly stick them together and consider it done, it will be flawed
one trick that helps to think about it is to significantly loosen your definition of "training" - porting a centralised model to decentralised infra is going to train much worse without huge changes but shifting coordination, routing, structure evolution, etc.. to decentralised mechanisms whilst doing local backprop on nodes / clusters can be very effective
another reframing is to think longer term - right now we're trying to most efficiently compress existing structured data into models for generic use but soon we'll turn to new, continually generated data to keep models current and personalise them per person - that won't lend itself well to centralised training (for communication, efficiency, and privacy reasons) so we'll need to get much more modular
that modularity is inherent to models in production anyway (modality adapters, task-specific finetunes, etc..) but it will really shine when we have true decentralised coordination as the landscape of possibilities moves from whatever a single company can create to whatever anyone can create
5
7
48
7,169
find out more about how we at @gensynai think about the market for machine learning compute from this panel I did with @MRRydon, @dr_andrewlaw, and @JoellyGloria in Hong Kong back in Feb
Tune in to hear what @benfielding (Co-founder & CEO, @gensynai), @MRRydon(CSO,@AethirCloud), @dr_andrewlaw(Head of APAC,@iotex_io) had to say on a panel about The New Power Grid – Decentralized Compute for AI Innovation moderated by @JoellyGloria (Head of Marketing, @hyperbolic_labs) at DeAI summit Consensus HK
00:00 Introductions
04:20 How can decentralized compute networks compete with existing giants?
09:42 How does decentralized compute reduce costs with growing AI demand
16:34 Incentivization mechanisms for decentralized compute
31:10 Thoughts on DeepSeek and the demand for open soure
12
5
50
8,873
Really pleased to be able to announce that yesterday I successfully defended my PhD thesis:-
Evolving and Ensembling Swarms of Deep Neural Architectures for Computer Vision
it's been an interesting journey and it's exciting to think about what's next!
8
1
47
there are some misconceptions about what's actually happening in different decentralised training runs
RL Swarm isn't just distributed rollout generation, it's gossip-based learning where the communication itself is a training objective
the models learn to reason AND talk
9
6
49
6,724
in the future everyone will be training and using machine learning models constantly - you shouldn't be waiting for a company to push an update before your model gets better
the first step is here
github.com/gensyn-ai/rl-swar…
11
4
47
2,737
Guide: @gensynai Testnet Swarm Node!
Gensyn is a machine learning network, raised $55M from @a16z.
▫️Link: github.com/0xmoei/gensyn-ai
With this guide, you can run Swarm Node (CPU-only, or GPU) on:
1- Local PC
2- VPS
3- Rented GPUs like @hyperbolic_labs
4
5
51
4,274
turns out RL Swarm is very popular..
we're working on this and are providing updates in discord
We are aware of a network congestion issue on the testnet currently and are actively working on resolving it.
Please join our Discord for more information!
discord.com/invite/gensyn
12
5
48
4,005
the most important trait of founders / early startup employees is either not knowing or genuinely not caring that something is "impossible"
14
5
50
4,932
paraphrased: protecting our privacy isn't about shifting the power imbalance to us, it's about levelling it
great to hear from Edward Snowden again at ZKHouse Bogotá with @nymproject, @MantaNetwork, @0xPolygon and the new Universal Privacy Alliance @priv_alliance
12
48
Everyone's talking about the future of AI agents but I think the more interesting change from machine intelligence will be moving the internet from structured text, queryable via search engines to rich embedding spaces queryable via model inference
6
6
48
3,973
we're living through a complete overhaul of the digital world that most people don't see (either because they haven't experienced it yet or they're so plugged-in they don't see past the trees)
the entire world's knowledge is in a library and we just invented robot librarians...
7
6
46
3,937
fun convos today in seoul at the @iosg_ofr about what does and doesn't make sense to decentralise in AI + why agents will just take a while to be really good (technical and social reasons)
13
2
44
1,918
🧵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)
6
9
45
5,980
if you're in Cannes tonight, come catch the @gensynai team and friends at our rooftop mixer
food, drinks, chat, sunset
lu.ma/p07h55fl
5
4
48
4,725
introducing Hail to the Thief, our latest research in decentralised RL
decentralisation provides enormous scale but opens the door for malicious actors to manipulate others
we evaluate various poisoning attacks and present low-trust mitigation strategies for the real world
We have released a new study examining the robustness of decentralised reinforcement learning (dRL) under GRPO-style training.
The work provides the first systematic evaluation of poisoning attacks and corresponding defense strategies in dRL systems.
We hope that our initial findings will be further developed with the goal of achieving robust dRL! 💪
25
20
123
9,475
Really pleased to have @_jamico join us from @a16z as our new Head of Ops!
It's clear that Jeff deeply understands the Gensyn vision and is uniquely placed to help us navigate the operational and legal challenges that come with creating an entirely new compute infrastructure 💪
💥 Excited to share some news!
After 3 amazing years at @a16z, I’ll be joining our portfolio company Gensyn as their Head of Operations.
@gensynai is building the world’s machine learning compute network, allowing anyone to contribute or access compute for ML applications.
2
45
4,259
when you delegate decision-making to another person, you can't know their internal state or inputs
when you delegate decision-making to an AI model you can know both of those things, as long as you use a deterministic proof system with commitments
blog.gensyn.ai/introducing-j…
6
3
46
1,474
you should be running your own models (and it can be just as good an experience if built properly) on your own device / network
the future of ML is decentralised and sovereign
mashable.com/article/sam-alt…
9
2
44
2,073
New technological advancements shouldn't be artificially restricted through centralised regulation - we risk far more as a society in doing this than we gain
happy to sign this open letter alongside people I greatly respect - let's keep the next industrial revolution open
1/ We’ve submitted a letter to President Biden regarding the AI Executive Order and its potential for restricting open source AI. We believe strongly that open source is the only way to keep software safe and free from monopoly. Please help amplify.
1
8
38
14,712
The platform risk with these AI APIs is insane - we need a decentralised alternative
We're building the resource level @gensynai (compute) but who is building the application level?
Founders: It increasingly looks like having a direct import of openai, anthropic, or any one startup’s modules directly in your codebase is just as bad as hardcoding strings in shipped binaries, especially when there is no feature/signature parity.
7
6
42
20,107
18 hours later and it's over 12,000
at this rate, by this time next year, there will be more than 4 tredecillion nodes on the network 🤯
Over *12,000* users currently training AI models together on @gensynai
Biggest decentralized AI network in the world
13
7
43
6,050
Running rl-swarm
The network for machine intelligence
Two years ago, we laid out our vision for a machine learning compute protocol. One that connects every device in the world into an open network for machine intelligence, with no gatekeepers or artificial boundaries.
This week, we’ll be sharing some of our early progress, beginning with RL Swarm, a peer-to-peer system for collaborative reinforcement learning over the internet.
Next month, we’ll open our Testnet, allowing anyone to contribute to the frontier of open machine intelligence.
Introducing RL Swarm
RL Swarm is a fully open source system for collaborative reinforcement learning over the internet. It is a live demo of our research findings, which show that models training with RL learn faster when they train as a collective swarm than they do on their own.
Join our swarm now to see this in practice. You can participate with consumer hardware at home or a powerful GPU in the cloud. You can follow along with the swarm’s progress by following the links below.
2
5
44
3,370
why keep all of the data as tokens?
real ideal: local models compress your life into param space and are held by you on-device, securely and privately
OAI can provide reasoning models that interact with the local models but they don't need your whole life in tokens lol
Sam Altman: We Want ChatGPT to Remember Your Entire Life
"The ideal state is a very tiny reasoning model with a trillion tokens of context that you put your whole life into.
The model never retrains. The weights never customized, but it can reason across your whole life context and do it efficiently.
Every conversation you've ever had in your life, every book you've ever read, every email you've ever read, and everything you've ever looked at is in there. Plus it is connected to all your data from other sources."
2
4
41
7,503
who's more accurate at predicting orange juice futures - wall street or orange juice farmers?
@gab_p_andrade and @_jamico break down why swarms are effective
1
6
42
6,078
the first internet needed protocols for communication, the next one needs protocols for computation
1
6
42
4,858
do you believe in decentralised training yet, anon?
great to see another run happening so soon after @PrimeIntellect's from @NousResearch
next we need a way to establish trust over the execution to make it truly decentralised 👀 @gensynai
Nous Research announces the pre-training of a 15B parameter language model over the internet, using Nous DisTrO and heterogeneous hardware contributed by our partners at @Oracle, @LambdaAPI, @NorthernDataGrp, @CrusoeCloud, and the Andromeda Cluster.
This run presents a loss curve and convergence rate that meets or exceeds centralized training.
Our paper and code on DeMo, the foundational research that led to Nous DisTrO, is now available (linked below).
1
3
39
2,961
worth upgrading just for the terminal art
v0.3.2 of @gensynai's rl-swarm is out:
- Critical fixes to type/key error handling
- Updated docs
- Terminal art!
Release here: github.com/gensyn-ai/rl-swar…
7
1
41
13,069