Vitalik Buterin, one of the creators of Ethereum, recently discussed advancements in privacy-focused AI technology. He highlighted tools that will allow people to run AI programs on their own devices, send secure messages, and safely interact with Ethereum applications.
On Thursday, Vitalik Buterin shared on X (formerly Twitter) about several open-source projects related to what he calls “CROPS AI.” This framework aims to build AI systems that are secure, give users control over their own data, and protect privacy.
Since my last update, Deepseek v4 has been released. They now offer a 2-bit quantized version that can run within 90GB of VRAM, and it does work! However, performance varies quite a bit. I’m seeing around 35 tokens per second on Apple hardware, which is pretty good, but it’s much slower – around 7 tokens per second – on AMD. Honestly, it seems like a significant effort would be needed to properly optimize and support more hardware platforms.
— vitalik.eth (@VitalikButerin) May 27, 2026
This new information expands on an article he published in April, which discussed running AI programs locally and the dangers of relying on AI that’s hosted in the cloud.
New AI and messaging tools added
One project highlighted was a new version of Vitalik Buterin’s messaging tool. This update allows early testing with Telegram and is built to enable AI systems to use messaging apps safely. It does this by requiring human confirmation before any potentially risky actions are taken.
He also highlighted progress with VoxTerm, a local AI tool that transcribes audio without relying on external servers, and Lucebox Hub, software designed to make powerful AI models like Qwen 27B run more efficiently.
Vitalik Buterin also mentioned a smaller, more efficient version of DeepSeek v4 that needs only 90 GB of memory to run. However, he pointed out that its performance varies depending on the hardware it’s used with, running noticeably faster on Apple devices than on those with AMD processors.
Privacy tools linked to Ethereum access
Vitalik Buterin recently highlighted a connection between protecting privacy in AI and improving privacy on Ethereum. He suggested that the same technology used to privately access AI models remotely – specifically, using zero-knowledge proofs – could also address privacy concerns when people access Ethereum data. This would let users interact with the blockchain without revealing their personal information to the services they’re using.
Vitalik Buterin also talked about AI models designed for specific uses within Ethereum development and security. He pointed to Leanstral, an AI model that excels at tasks related to the Lean programming language, as a way specialized AI can help improve software testing and make coding more secure.
“We should have models finetuned for Ethereum-related use cases as well,” he wrote.
April essay focused on local AI security
My recent work builds on an essay I published in April, where I outlined my approach to building a self-sovereign, local, private, and secure Large Language Model (LLM) setup. In that piece, I emphasized that we shouldn’t automatically assume today’s AI systems are secure. I pointed out that many mainstream AI platforms still have significant weaknesses when it comes to privacy and basic security measures. I’m particularly worried about AI agents that have too much access, the potential for attacks through cleverly crafted prompts, the risk of sensitive data being stolen, and the fact that cloud-based AI models can collect a lot of personal information from users.
In my research, I’ve been exploring ways to make AI safer, especially when it interacts with things like cryptocurrency. I’m focusing on a system where AI models run locally on your device, rather than in the cloud. I’m also using ‘sandboxing’ – creating isolated environments – to limit what the AI can access. Crucially, I want to keep knowledge stored offline whenever possible, and always require human approval for sensitive tasks like sending crypto or sending messages. My current setup includes tools like llama-server for running the AI, bubblewrap for the sandboxing, and I’m connecting AI agents to Ethereum wallets through a special layer that strictly controls what they can do.
Human confirmation remains central
Vitalik Buterin believes that AI used for things like managing money or sending messages shouldn’t be allowed to operate on its own without safety measures. He suggests combining AI with required human checks for important tasks. He calls this a “human + AI” system, where both a person and the AI need to give the okay before anything risky happens.
Buterin believes the main aim is to create AI that boosts productivity while also protecting people’s privacy and preventing misuse from powerful, centralized AI systems.
Read More
- Gold Rate Forecast
- USD TRY PREDICTION
- USD MXN PREDICTION
- Silver Rate Forecast
- Brent Oil Forecast
- PI PREDICTION. PI cryptocurrency
- RUNE Crashes Hard but Folks Keep Throwing Cash at THORChain—Why tho? 🤔
- Why These 5 Meme Coins Could Crash or Cash Your Crypto Party in May 2025 🚀🐒
- Ethereum Bounces Back After $4M Sell-Off: Can it Survive Another Rollercoaster?
- EUR CLP PREDICTION
2026-05-28 17:25