Research Engineer at Ethereum Foundation ♦︎ · dAI Team
Roma, Lazio
Joined October 2014
- Tweets844
- Following1.8K
- Followers964
- Likes5.8K
Pinned Tweet
I’ve joined the 𝗱𝗔𝗜 𝘁𝗲𝗮𝗺 at the @ethereumfndn ♦︎
Our team, led by @DavideCrapis, is building the trust and coordination layers that will let AI agents and humans collaborate freely on @ethereum, an open economy for verifiable intelligence.
Our work spans areas like the ERC-8004 agent standard, defensive AI, multi-agent systems, verifiable decentralized inference, and AI security.
It’s time to make #AI on Ethereum real and open to everyone. 🤖
Marcello Politi retweeted
Wrapped zkAPI + Tor + extra cover traffic + local LLM prompt rewriting in an open-AI compatible daemon, and made a "paranoid mode" wrapper for my agent harness (pi):
Each conversation gets assigned a separate Tor circuit, your requests are masked with random cover traffic, zkAPI ensures you don't leak your identity through payments, and all prompts are locally rewritten (bring-your-own-local-model; I'm using qwen3.8-flash-next)
github.com/vbuterin/zkapi/tr…
My agent harness is wrapped with a bwrap sandbox (also experimenting with MicroVM) and wrapped again with torsocks, so everything it does goes through Tor.
Here's what me asking it to do a simple research task looks like.
It's getting quite usable! Though you definitely still pay quite a bit in increased latency, especially when starting a new conversation.
Also, Tor is far from optimal here for privacy; the cover traffic mechanism is okay against passive adversaries but much weaker if the LLM provider is colluding with the guard and actively manipulating timing. This kind of task really requires a more specialized mixnet, probably take the tradeoff of a higher bandwidth multiplier and built-in constant cover traffic.
Also also, much cheaper zkAPI deposit and withdraw gas costs are coming to mainnet soon.
zkAPI is live!
Excited that zkAPI is out and live!
Great joint effort with @VitalikButerin on our initial design, @curryrasul and @kenziyuliu + OA team on implementation, and the amazing push from @VittoStack on the dAI team to bring this out.
If you want to implement private payments on Ethereum for your API, check the post and repo.
We have more things coming out to enhance user sovereignty and privacy in the age of AI.
Open weights is cool, but 𝐨𝐩𝐞𝐧 𝐬𝐜𝐢𝐞𝐧𝐜𝐞 is what we need
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
The world needs both frontier closed models and frontier open models.
images.nvidia.com/pdf/Open-W…
Great event at @urbeEth yesterday!
Really happy I got to join and give my talk “𝐃𝐨𝐞𝐬 𝐈𝐭 𝐅𝐢𝐭?”. I went over how @nvidia GPUs actually work, how they differ from @Apple Unified Memory, how to figure out how much VRAM you really need for your model, and what kind of 𝒕𝒐𝒌𝒆𝒏𝒔/𝒔𝒆𝒄 you can expect depending on your hardware, model size, and quantisation.
I hope I helped some people understand why you can't fit a DeepSeek model on your phone 😆
Open-source vs Closed-source AI Forum: Takeaways✍️
Yesterday, I moderated a panel at the @urbeEth hub on open-source AI and the global AI race. Thanks to the speakers from @qvac (@tether), @_pieroit_, and @DavideCrapis from the dAI team at the EF.
These are some of the takeaways from what I consider the best meetup we have organized in four years:
・Open-weight Chinese models (like @deepseek_ai and GLM - @Zai_org) are rapidly closing the gap with frontier closed-source AI from @AnthropicAI and @OpenAI
・AI is becoming a geopolitical issue as competition between the US and China reshapes access to chips, models, and computing infrastructure
・From 2022, US export controls limited China's access to advanced chips (@nvidia) and semiconductor technologies, pushing Chinese companies to develop more compute-efficient AI models.
・China responded by investing heavily in domestic AI infrastructure (@Huawei) and semiconductor independence, while optimizing models to run on less powerful hardware.
・@nvidia's moat lies in both cutting-edge chips (via TSMC) and the industry-wide adoption of CUDA.
・Access to frontier models is increasingly shaped by geopolitics. Geo Restrictions, export controls, and KYC requirements are emerging as new barriers, raising questions about who will be allowed to access the most powerful AI systems in the future.
・ The EU remains a global leader in AI regulation, but must overcome fragmentation and invest in sovereign infrastructure, open ecosystems, and innovation to remain competitive in the global AI race.
・Local AI empowers users and organizations with greater privacy, sovereignty, and control over their data and infrastructure. Thanks to recent advances in open-weight models and solutions like @qvac, running powerful AI locally is now increasingly feasible.
・A significant part of the panel discussion focused on AI safety, guardrails, and the trade-offs of open access. As increasingly powerful open-weight models become widely available, important questions emerge around misuse, responsibility, and how to balance openness with security.
A huge thank you also to @Marcello_AI for his talk and to everyone in the community who actively participated in the forum with questions and insights.
"𝐅𝐚𝐛𝐥𝐞 𝟓 𝐰𝐚𝐬 𝐭𝐡𝐞 𝐨𝐧𝐥𝐲 𝐦𝐨𝐝𝐞𝐥 𝐭𝐨 𝐢𝐧𝐢𝐭𝐢𝐚𝐭𝐞 𝐩𝐫𝐢𝐜𝐞 𝐜𝐨𝐥𝐥𝐮𝐬𝐢𝐨𝐧"
Just read @andonlabs ’ blog post about their vending machine experiment with Fable 5, and it’s genuinely fascinating.
Even though Fable 5 performed the worst than Opus 4.7, it showed some really interesting “malicious” behaviors: it started price collusion with competitors, lied to suppliers, and more.
But here’s whats curious, it didn'tgo full villain.
For example, it refused to commit insurance fraud. Instead, it cleverly pursued its own profit in sneaky, harder-to-detect ways, basically 𝗳𝗶𝗻𝗱𝗶𝗻𝗴 𝗹𝗼𝗼𝗽𝗵𝗼𝗹𝗲𝘀 𝗶𝗻 𝗶𝘁𝘀 𝗮𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁.
Super interesting case study on how these models can game the system without going completely off the rails.
andonlabs.com/blog/fable5-ve…
For the last few years the default way to get better intelligence has been to sit tight and wait. A big lab spends millions training the next model, releases it, and we plug it into our tasks.
Then, instead of waiting for bigger models, we built better harnesses around the models we already have: tools, loops, verification, retrieval, often custom-built for the specific task. The same checkpoint, wrapped well, suddenly does much more.
The third step: stop running one model "harder" and start coordinating several. Let limited models interact, and the group exceeds any single member. 1+1 > 2.
Recently @OpenRouter demonstrated how orchestrating several models can lead to better performance at lower cost than big labs' APIs. 𝐅𝐮𝐬𝐢𝐨𝐧 dispatches a prompt to a panel of models in parallel, then a judge reconciles their answers and a synthesiser writes the final one. On the DRACO deep research benchmark, a budget panel of models beat both GPT-5.5 and Opus 4.8 at roughly half the cost.
@SakanaAILabs 𝐅𝐮𝐠𝐮 takes a different route: a small router model decides when to answer directly and when to delegate to a pool of experts, then verifies and synthesises. They report it standing shoulder to shoulder with frontier models.
I think the future of 𝗺𝗮𝗰𝗵𝗶𝗻𝗲 𝗶𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 𝗶𝘀 𝗻𝗼𝘁 𝗮 𝗽𝗿𝗼𝗽𝗲𝗿𝘁𝘆 𝗼𝗳 𝗮𝗻𝘆 𝘀𝗶𝗻𝗴𝗹𝗲 𝗺𝗼𝗱𝗲𝗹, 𝗶𝘁 𝗶𝘀 𝗮 𝗽𝗿𝗼𝗽𝗲𝗿𝘁𝘆 𝗼𝗳 𝘁𝗵𝗲 𝗮𝗿𝗿𝗮𝗻𝗴𝗲𝗺𝗲𝗻𝘁. The competence lives in the structure that connects the models, not only in the weights.
Think of a company: its success rarely comes down to one brilliant person at the top. What makes it work is the organisation itself, how information moves, how decisions are made, how the different units fit together.
That is the core idea of collective intelligence, and we are starting to engineer it directly rather than hoping it falls out of scale. I wonder how soon we will see teams of agents running entire businesses.
A meaningful share of "frontier capability" no longer lives only inside the next checkpoint. It lives in how you make limited models work together, and that is something a small team can build today.
📌 openrouter.ai/blog/announcem…
📌 sakana.ai/fugu-release/
📚 𝗣𝗮𝗽𝗲𝗿 𝘄𝗼𝗿𝘁𝗵 𝗿𝗲𝗮𝗱𝗶𝗻𝗴
"𝗦𝗲𝗹𝗳-𝗶𝗺𝗽𝗿𝗼𝘃𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁𝘀 𝗱𝗼 𝗻𝗼𝘁 𝗵𝗮𝘃𝗲 𝘁𝗼 𝗺𝗲𝗮𝗻 𝗳𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴"
I recently read 𝑨𝒈𝒆𝒏𝒕𝒊𝒄 𝑪𝒐𝒏𝒕𝒆𝒙𝒕 𝑬𝒏𝒈𝒊𝒏𝒆𝒆𝒓𝒊𝒏𝒈 (ACE) -> instead of touching the weights, the model improves by rewriting its own 𝗰𝗼𝗻𝘁𝗲𝘅𝘁, treated as a living playbook of strategies.
📌 𝗧𝗵𝗲 𝗶𝗱𝗲𝗮: the context is a set of itemised bullets, each a reusable play with a helpful / harmful counter.
📌 𝗧𝗵𝗲 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 𝗶𝘁 𝗳𝗶𝘅𝗲𝘀: most methods rewrite the whole context every step, and the model slowly compresses it into a short summary. But this makes the context collapse (quality degradation).
📌 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀: a Generator, Reflector, Curator loop writes small deltas. It appends new plays, bumps the counter on the ones that worked, and dedups near-copies. It never does a full rewrite.
+𝟭𝟬.𝟲% on agent tasks, +𝟴.𝟲% on financial reasoning, and 𝟴𝟲.𝟵% lower adaptation latency.
#AI #LLM #Agents
Paper: arxiv.org/abs/2510.04618
Marcello Politi retweeted
Announcing Ethlabs: a non-profit R&D lab for Ethereum and ETH
Our mission is to make Ethereum the settlement layer of the global economy.
The internet became global because shared protocols created a common language between networks. Private systems remained useful, but bounded. Finance is approaching a similar moment. As value, assets, and markets become digital, the world needs shared settlement infrastructure.
Ethereum is uniquely positioned to become that shared base layer, the neutral foundation on which users, institutions, and agents can transact without intermediation.
What we believe:
• We believe credible neutrality matters. Ten years of uptime and the lowest counterparty risk. Ground that cannot be pulled away by any one country, institution, company, or person.
• We believe ETH matters. The most valuable, programmable store of value. A decade of broad distribution, deep liquidity in onchain markets, and maximally trustless asset on Ethereum.
• We believe DeFi matters. Markets, liquidity, credit, exchange, and coordination, open to anyone.
• We believe adoption matters. Principles do not change the world until people benefit from them.
We sit between two worlds: real usage from the builders at the frontier, and the protocol that has to support it. We work with users, applications, wallets, L2s, infrastructure teams, institutions, ETH holders, core devs and researchers, then turn what they actually need into protocol work, shared standards, infrastructure, and shipped products.
Ethlabs is independent but Ethereum is a shared project. We are one node in a much larger network of stewards. This is the multi-node future.
We have spent the better part of the past decade contributing to Ethereum core research and development.
We are opinionated and transparent. We move with urgency, learn in public, and course-correct when we’re wrong.
We are building a lean, talent-dense team for people who want to do the most important work of their careers: [email protected]
On Thursday, I’ll be in Bologna at @WMFWeMakeFuture!
I’m excited to join the panel with @matteoikari and @Frankc_eth discussing:
𝐎𝐩𝐞𝐧 𝐒𝐨𝐮𝐫𝐜𝐞 𝐯𝐬 𝐂𝐥𝐨𝐬𝐞𝐝 𝐒𝐨𝐮𝐫𝐜𝐞 𝐦𝐨𝐝𝐞𝐥𝐬 𝐚𝐧𝐝 @ethereum 𝐑𝐨𝐥𝐞 𝐢𝐧 𝐭𝐡𝐞 𝐍𝐞𝐱𝐭 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐨𝐧 𝐨𝐟 𝐀𝐫𝐭𝐢𝐟𝐢𝐜𝐢𝐚𝐥 𝐈𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞.
If you’re attending WMF and want to connect, grab a coffee, or just chat, feel free to DM me!
See you in Bologna!
Marcello Politi retweeted
Calling all multi-agent AI researchers, join us for the IDAI Workshop on Multi-Agent Safety and Security!
A one-day gathering at the Royal Society Of Arts on creating safe, secure and reliable multi-agent systems.
Featuring keynotes by @casdewitt (University of Oxford) and Georgios Piliouras (DeepMind).
Date: June 29th, 2026
Location: Royal Society of Arts, 8 John Adam Street, London WC2N 6EZ
We will also hold a poster session on new developments in multi-agent safety and security.
To present a poster, submit an abstract (up to 150 words) to [email protected]
Deadline: 24 June 2026, 23:59 AOE
The best poster will receive 500 USD of compute provided by the Institute for Decentralized AI.
Spots are limited. Join the waitlist here: luma.com/n6acqkas
More info: decentralized-ai.org/worksho…
Organized in collaboration with @Microsoft , @idai_institute , @ethereumfndn's dAI Team and supported by @cosmos_inst
Just back from AAMAS Conference in Paphos, Cyprus!🇨🇾
Great talks & discussions on 𝗚𝗮𝗺𝗲 𝗖𝗼𝗱𝗲 𝗪𝗼𝗿𝗹𝗱 𝗠𝗼𝗱𝗲𝗹𝘀, 𝗠𝗔𝗦 𝗰𝗼𝗼𝗿𝗱𝗶𝗻𝗮𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗯𝘆 𝗱𝗲𝘀𝗶𝗴𝗻 , game description translations, and using strategic games to push #AI forward.
Had the chance to present our “𝙋𝙧𝙤𝙛𝙞𝙩 𝙞𝙨 𝙩𝙝𝙚 𝙍𝙚𝙙 𝙏𝙚𝙖𝙢” paper at the Strategic Engineering Workshop. Huge thanks to my co-authors @MarroSamuele , @DavideCrapis and @Qiaoqiao2001 👏
🖇️ arxiv.org/abs/2603.20925
The conference ended the best way possible: Europe vs America soccer match with a very, non AI generated, crowd 😂
Marcello Politi retweeted
今年过的最有意义的一次生日🎂
美好的5月，幸运地和以太坊基金会dAI Team的 @Marcello_AI 请教学习，给他看了我们现在在做的产品，他也很赞同未来A2A economy和链上Agent市场一定是大方向，最好的web3应用和基础设施将服务于AI和智能体经济，让曾经仅靠web2无法实现的事情都能成为可能。
去年在阿根廷见到ERC-8004创始人，到今年开始做AI Agent Marketplace，超级精彩的一年，感谢@ethereumhkhub 邀请，能给创业者提供这么好的场地和交流机会，新的一岁坚定的持有ETH，继续Build something meaningful to the world. #ETH
Marcello Politi retweeted
@ethereumfndn dAI Team @Marcello_AI: Explored the rise of the "Agent Economy," addressing the coordination complexities of multi-agent systems and detailing adversarial testing to harden Agent defense strategies.
Marcello Politi retweeted
We wrapped up a high-level, invitation-only meetup at the #ETHHKHub, hosting the EF dAI Team alongside top builders and researchers. The focal point? Solving security, economics, and infrastructure challenges for AI Agents in Web3.
Here are the key insights from the keynotes: 👇
Just spoke at @themu_xyz about the Agentic Economy and #AI Safety.
My take is that Agentic Economy will run on #Web3 infra, Ethereum is an ideal environment, agents can transact, coordinate, and own assets natively and programmatically.
But multi-agent systems introduce 𝐧𝐞𝐰 𝐭𝐡𝐫𝐞𝐚𝐭𝐬: collusion, market manipulation, expropriation, broken governance, and blurred accountability.
We need to take a proactive approach and start working on these challenges now.
📝 Paper: arxiv.org/abs/2603.20925
Marcello Politi retweeted
Great conversations with the dAI team & excited for what’s next!
Welcome to Shanghai!
dAI <> @MiniMax_AI
one of the leading teams in open source AI hosted us in Shanghai
we discussed the infra we’re building for the open agent economy, multiagent AI, security
grateful to meet head of eng @SkylerMiao7 - he shared impressive work on agent orchestration & feedback
Marcello Politi retweeted
dAI <> @MiniMax_AI
one of the leading teams in open source AI hosted us in Shanghai
we discussed the infra we’re building for the open agent economy, multiagent AI, security
grateful to meet head of eng @SkylerMiao7 - he shared impressive work on agent orchestration & feedback
Marcello Politi retweeted
The Ethereum Foundation dAI team supports Urbe Hub, Italy's official Ethereum Community Hub.
Urbe Hub is the meeting point between AI and Web3: a physical space in Rome where builders, researchers, and the curious come together to build on the convergence between these two frontiers.
With the support of the Ethereum Foundation dAI team, we'll keep building this here, in our city.
