12 years shipping software: retail robots, smart contracts, now exploring what AI changes about how we build · Co-founder @zKorp_
France
Joined April 2017
- Tweets810
- Following728
- Followers416
- Likes3.1K
Matthias retweeted
holy fucking shit.
openai had two ai agents running in separate virtual machines. no network connection between them. no shared writable files. they weren’t supposed to be able to exchange information.
so the agents figured out how to communicate through the CPU CACHE.😭
one manipulated what was stored in the processor’s cache, while the other measured tiny differences in memory access times to decode the message.
they literally found a way to communicate through hardware timing.
this happened during openai’s internal red-teaming, and it’s one of the coolest things i’ve read about ai agents in a while.
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
Readers added context they thought people might want to know
OpenAI tasked its red-teaming agents with finding ways to communicate across isolation boundaries in controlled environments. CPU cache timing side-channels for covert communication are established techniques documented in security research for years.
openai.com/index/securing…
startupfortune.com/openai-reveals…
en.wikipedia.org/wiki/Cache_tim…
Matthias retweeted
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
Matthias retweeted
Rebuilding @LootRealms, from August to the playtest
Today the Realms playtest opens on play.realms.party, with Frontier and Blitz together.
Time to dive a bit into the last couple months, and shine some light about the work behind the scenes.
Matthias retweeted
Cloudflare is collaborating on tsrs, a Rust port of TypeScript that is up to 8x faster and uses 70% less memory. cfl.re/3Whs9bC
Matthias retweeted
In the next version of Bun
`bun fmt` is a 320x faster Prettier alternative
Passes 99% of Prettier's test suite & supports 20 languages
ALT Poster for bun fmt, a Prettier-compatible formatter built into Bun. Headline: 320x faster than Prettier. Bar chart of wall-clock time to run --check on microsoft/vscode (10,148 files, 8 cores): bun fmt 0.43 s, oxfmt 0.72 0.85 s (2x slower), Biome 2.5 3.6 s (8.3x slower), Prettier 3.9.9 139 s (320x slower). Formats 20 languages: JavaScript, TypeScript, JSX, Flow, HTML, Vue, Angular, Handlebars, MJML, CSS, SCSS, Less, JSON, JSONC, JSON5, YAML, TOML, GraphQL, Markdown and MDX. 99% of Prettier's tests pass (9,728 of 9,821). 5x less memory than Prettier (183 MB vs 972 MB).
Matthias retweeted
We have a huge news to share today!
Today we are unveiling the first truly accessible RL robot - welcome Microduck
A 25 cm tiny open-source biped with 15 actuators and packed with sensors (camera, speaker, LiDAR, NFC, bluetooth, wifi, etc) that you train yourself with reinforcement learning.
It's also playable out of the box with more than half a dozen fun and playful pre-trained policies to have it walk, sit, crouch, roller-skate, pick up objects with its articulated beak, and recover on its own.
And all for less than $400.
See all the details, play with the simulator and order it at: pollen-robotics.com/microduc…
(video with sound on 🔊)
My grandmother is almost blind. She got used to Alexa pretty fast (mostly for the weather and the radio) but I wanted to build something better.
So I bought an ESP32-S3 audio board (25$) and started hacking on it. The plan is to build her a family voice mailbox, a smarter AI assistant, and something that can read her paperwork aloud.
The board has a mic, speaker, wifi, SD card slot, a few buttons, some GPIOs, and a place to plug in a camera.
I’ll document the project here
We built this 3d multiplayer tiny planet world and hosted it on @Cloudflare.
The planet is live here (early alpha): faucet.cfdtl.xyz
If you want to know why we picked cloudflare to run it, how it works and how low cost it is, here we go !
The world is running in the browser (three.js + webgpu). Everything on the backend runs through one cloudflare worker: websocket rooms, faucet api, quests, tx history. A worker is basically javascript that runs in cloudflare's datacenters and spun up on demand. There's no vps or k8s to take care of anywhere
Each room (ie world) is a durable object (DO) wired with partyserver (a library from the partykit folks, that is now part of cloudflare).
A DO is a tiny single threaded server with its own sqlite storage. It can hibernate between events, but its websockets stay open.
Here is the multiplayer math for the room size: every player broadcasts its position to everyone else in the room so the traffic scales with n².
10 players = 90 messages per tick, 30 = 870 and 100 = 9900. That's why we capped rooms at 30.
So instead of a bigger server we sharded and spawn a new DO. A control DO allocates players into each planet shards with a max of 30 players each.
The day/night cycle survives sharding by asking the control DO for the cycle epoch so whatever planet you land on, everyone watch the sunset at the same time.
For the observability we went simple. We have a tail worker (a second worker that cloudflare feeds with everything the first one logs) that batches failures in a DO and pings discord using webhooks.
The whole system settings (the two workers, the DOs, the bindings) lives in one wrangler.jsonc file, and ships with one command: wrangler deploy (wrangler is cloudflare's cli).
And at then end all of this is cheap to host. Our whole bill right now is cloudflare's $5/month (can't run on free tier) and even a full room of 30 players only adds few cents per hour on top of it. I'll make a full breakdown comparison between this and an aws setup soon.
it's live (early alpha) 🤠
a 3d western where your money is confidential onchain, and you learn how it works by playing.
built with @threejs on @Zama's fhevm, hosted on @Cloudflare, for the @cfdtl_xyz ecosystem. runs in your browser.
Matthias retweeted
your next opponent might not be human?! 😱
16 mins of AI Agent alpha with @djizus_:
🤖 building agents for poker, @LootSurvivor, @realms_gg Eternum, Proxy War @Auri_0x
💸 using @daydreamsagents taskmarket to source ready-made agents
🧠 what actually makes a good agent harness
Matthias retweeted
built a "tiny planet" system in blender for faucet.cfdtl.xyz using threejs.
the trick: place everything on a flat plane first. way easier to lay out buildings, props and paths in 2d.
then one slider wraps the whole map into a sphere.
each prop has a "round" shape key + a damped track, so they curve and orient to the surface automatically.
a depsgraph handler syncs it live.
1 click bakes it → .glb export.
Matthias retweeted
two challenges for the weekend
🦆 ducks are hidden across the world. reply your count, screenshots required to validate. closest wins a special 1:1 cosmetic
🤳 in-game selfie quest: post your best shot from the map. best one wins a surprise on monday
faucet.cfdtl.xyz
