Pinned Tweet
Prompt to fully realized video in seconds. 🤯
I just had OpenAI Astra end-to-end generate this entire product video from scratch—no stitching, no heavy manual editing. The shift from text/concepts straight to rich visual media is happening faster than we think.
Watch the full result below! 👇
#OpenAI #Astra #AIGenerated #GenerativeAI #BuildInPublic
FlashProve on L4 24GB (upstream targets H100 80GB), using @karpathy's TinyStories260K.
Fresh commitment → integer inference + proof → independent CPU verification. 5/5 tampered cases rejected.
Code, weights & proof: github.com/zkwasm/FlashProve…
New Anthropic research: progress on provable inference for large language models.
Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.
These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.
For our full write-up, see here: anthropic.com/research/prova…
Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP…
We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
17
You tell AI things you wouldn't tell anyone. But when the model runs on someone else's box, its owner can read every word.
We built Private AI in a Box: @qvac's own engine, running inside a TEE. The box owner only ever sees ciphertext. 🧵
1
32
Llama 3.2, Gemma 3 and Laya decision models all run in one TA, switching on demand. If anything crashes, the TEE frees its cores and recovers without a reboot.
It's a prototype: secure boot isn't on yet, and physical and side-channel attacks are out of scope.
1
7
Full write-up: threat model, architecture, attestation, results and roadmap 👇
nitter.cf/agentemergence/status/…
2
Every single time the ChatGPT macOS desktop app drops an update, it feels like playing Russian roulette with core features. Today’s prize? “Compute Use” just completely vanished into the ether for no apparent reason.
26
Magnificent works!
Announcing three major scientific discoveries, one each in mathematics (harmonic analysis), computer science (complexity theory), and physics (quantum field theory).
AI starting to help us improve our understanding of the universe! openai.com/index/building-sc…
1
20
1/9 @qvac built a smart camera that never uploads your footage, running on a laptop.
We put the same idea inside an actual camera: QVAC on a 2 GB device, detecting on its NPU, describing with VisionPsy-Nano, and proving that no frame ever left. 🧵
8/9 Offline is a button, not a promise: one click drops all Wi-Fi traffic, a probe confirms the internet is gone, and detection carries on.
The live preview is off by default. Turn it on, and every byte it sends is counted.
1
1
38
9/9 We’d love to explore local AI and privacy use cases on embedded smart hardware together with @tether, @qvac and every developer in the community.
The full story, with every number and chart: nitter.cf/agentemergence/status/…
QVAC detecting on a camera's NPU and QVAC's own vision-language model reasoning on its CPU, in 2 GB of RAM — and "no frame leaves the device" turned from a promise into something you can check.
1
31







