Jordan Fréry

Machine Learning Research Scientist | Engineer, Open Source and Privacy

[Jordan Fréry @JordanFrery]

We need verifiable AI. The industry is currently at the mercy of a bunch (~2!) of companies playing around with cost/compute optimization.

https://t.co/EgfXYVMe3K

[Concrete @ai_concrete]

AI agents will run your company.
AI agents will run your life.

Who runs the agents?

[Jordan Fréry @JordanFrery]

AI Agent desperately need privacy and security before we get there. I will be discussing this in the next openclaw paris meetup. Join us!

Meetup link: https://t.co/g7bxC0mODv

[François Chollet @fchollet]

AI agents will soon graduate to fully-fledged economic actors that buy services, compute, and even data in the course of accomplishing high-level goals. 1-2 years before we start seeing this at scale.

[François Chollet @fchollet]

AI agents will soon graduate to fully-fledged economic actors that buy services, compute, and even data in the course of accomplishing high-level goals. 1-2 years before we start seeing this at scale.

[Jordan Fréry @JordanFrery]

AI agents are coming whether you like it or not and they are only useful if they can access everything, emails, docs, financials, customer data... All of it. This has to remain private and secure by design.

We built a solution and are looking for a few companies to try it out. My DMs are open!

[fmdz @fmdz387]

Clawd disaster incoming

if this trend of hosting ClawdBot on VPS instances keeps up, along with people not reading the docs and opening ports with zero auth...

I'm scared we're gonna have a massive credentials breach soon and it can be huge

This is just a basic scan of instances hosting clawdbot with open gateway ports and a lot of them have 0 auth

[Rahul Sood 🏴‍☠️ @rahulsood]

https://t.co/Y6VIoLszMn

[Jordan Fréry @JordanFrery]

We're about to witness the greatest centralization of information in human history.

Not in a database. In computation. Systems that reason over secrets, combine them, and act on them.

The privacy tech isn't ready for what's coming. We need to fix that.

[Jordan Fréry @JordanFrery]

Refreshing to see this debate!

Looking past the provocative title, there’s actually a lot of consensus here.

- Transformers are hitting a wall.
- GPUs aren’t getting meaningfully faster.
- This the current hardware/transformer stack is already near optimal.

Where Tim gets controversial is the conclusion: because hardware is the bottleneck, exponentially better AI (Superintelligence) is impossible.

I suspect the provocation is intentional, but the real takeaway seems to be within these lines:

- For AI builders: don't wait for a god model. Assume current AI is the ceiling and start building useful things for humanity.
- For AGI builders: stop stacking GPUs. The scaling era is over. We need a fundamental restart on algorithms and hardware (what people like LeCun and Ilya are now hinting at)

[Tim Dettmers @Tim_Dettmers]

My new blog post discusses the physical reality of computation and why this means we will not see AGI or any meaningful superintelligence: https://t.co/jsAKQ6T3gC

[Jordan Fréry @JordanFrery]

It's great to see AI detection becoming real.

But you shouldn't expose a confidential file to verify its origin...

At @zama we built a demo running the SynthID detection on encrypted data. Try it out:

https://t.co/XnMB2mSRmx

[Demis Hassabis @demishassabis]

You can also now check for SynthID watermarking directly in the @GeminiApp and verify if an image was made by our models - just upload the image and ask Gemini. Being able to verify what you’re seeing is critical for building AI responsibly.

[Jordan Fréry @JordanFrery]

@sama Teen safety and user privacy don't have to be in conflict. It's possible to deliver both. Simply commit to privacy preserving technologies, @sama.

[Zama @zama]

After five years of work, the team is happy to announce the Zama Confidential Blockchain Protocol and the release of our Public Testnet.
First things first, read the litepaper ⬇️

[Jordan Fréry @JordanFrery]

@sama User privacy can be technologically enforced to ensure no one accesses this data without explicit consent.

Copyright problems can be verified by auditing specific versions of the models themselves, not by reviewing user conversations.

[Zama @zama]

Zama Public Testnet live on July 1st 🚀

[Jordan Fréry @JordanFrery]

@ClementDelangue We make AI run on encrypted data at @zama. We will catch up with cleartext execution soon !

[Jordan Fréry @JordanFrery]

@skyquake_1 A good time for @OpenAI
to implement watermarking detection using privacy preserving technologies.

https://t.co/Ny0UoAybu8

[Zama @zama]

AI-generated images blur the line between real and fake, raising tough challenges for copyright, privacy & trust online.…Invisible watermarking using Fully Homomorphic Encryption (FHE) can solve these challenges—without compromising privacy.
A thread w/ code & demo 🧵⬇️

[Zama @zama]

AI-generated images blur the line between real and fake, raising tough challenges for copyright, privacy & trust online.…Invisible watermarking using Fully Homomorphic Encryption (FHE) can solve these challenges—without compromising privacy.
A thread w/ code & demo 🧵⬇️

[Jordan Fréry @JordanFrery]

Analyzing DNA usually comes with serious privacy risks. Yet, these insights can genuinely boost your health.

💰 Zama’s latest bounty offers €10,000 to whoever cracks this biological age analysis on encrypted DNA

➡️ https://t.co/xC7NiQeOuC

Curious to see what hackers will do!

[Zama @zama]

Alongside the new bounties we're introducing for the Zama Bounty Program Season 8, the blog post below includes a list of the winning solutions from Season 7. Plus, next week, we'll share a detailed code deep dive on the winning solutions from the S7.
https://t.co/r0SWuZnZYC

[Rand @randhindi]

Today, we are introducing our most exciting product to date: the #fhEVM coprocessor. It enables confidential smart contracts on #Ethereum, @base and other EVM chains, without changing the underlying protocol. Let’s dive into what this means for the future of privacy onchain. 🧵/1

[Zama @zama]

Today we are one step closer to our ultimate goal: making FHE accessible, easy, fast, and, most importantly, ubiquitous. Happy to share the results of our teams’ effort over the past quarter with these new releases: TFHE-rs (v0.8), Concrete (v2.8), Concrete ML (v1.7) ⬇️🧵