August 12, 2026
AI News
The AI news worth knowing this week.
Anthropic starts marking the text and images Claude makes
Claude models released after August 2 weave an invisible mark into the text they write and attach C2PA data to images. The mark survives a copy and paste and holds up through some editing, and Anthropic says the writing itself reads exactly the same.
Read the original →Meta is back in open source with Muse Glimmer
A 30B multimodal model under Apache 2.0, built to run local agents on a single consumer graphics card. It takes text and images together and handles more than 100 languages.
Read the original →LTX-2.5 is out, weights and all
LTX's new video model turns a still image into a 10-second 720p clip in 6.8 seconds on NVIDIA GB200 chips, and the weights are public. ComfyUI supported it on day one, and it is free for companies under $10M in revenue.
Read the original →Spotify will badge AI artist profiles and keep them out of recommendations
From mid-September, a profile built on an AI persona gets a badge, either because the artist says so or because Spotify's review team decides it. Those profiles also stay out of editorial and algorithmic recommendations by default.
Read the original →OpenAI expands Daybreak and ships GPT-5.6-Cyber
A model trained for security work that answers 95% of advanced requests, where the general model answers under 2%. Access comes in two tiers, one for defensive work and one for authorized offensive testing.
Read the original →Dubbing & Open source
Open-source voice and dubbing tools worth a look.
stronghamjji/PersoDub
Dubbing that stays on your own machine and uses your own voice. Nothing gets uploaded anywhere, and it never stretches or squeezes the audio to make it fit. (This one is mine.)
Read the original →
0xShug0/audio.cpp
One pure C++ engine that runs audio models through ggml. Text to speech, speech to text, speech detection and voice conversion, all without a Python setup.
Read the original →
ShawnPana/phone-harness
Hand your phone over to an AI agent and let it work the apps itself. It picked up more stars than any other new AI project on GitHub this week.
Read the original →This week on X
What AI Twitter was talking about this week.
Elon Musk (@elonmusk)
Grok Imagine can meme (quoting @tetsuoai: "Imagine Image 2.0 is god-tier for memes. Magic wand edits the region you select and leaves the rest of the frame pixel identical. Swap one face in a group photo without regenerating the whole image.")
Read the original →Anthropic (@AnthropicAI)
We asked an unreleased research version of Claude to take a stab at the Riemann hypothesis. It didn't solve it, but it did make strides on a related problem: it increased the lower bound for the fraction of zeros of the Riemann zeta function that satisfy the hypothesis.
Read the original →Adam Lyttle (@adamlyttleapps)
I vibecoded a safer Roblox for my son. He's not allowed to play Roblox because we don't know who he's talking to. So I created a version where he plays obbys. The game is populated with recordings of other play throughs (giving the illusion of playing with others).
Read the original →NVIDIA AI (@NVIDIAAI)
Introducing NVIDIA Nemotron 3.5 Lightning. An open 30B MoE model with 3B active parameters, built for always-on agents to complete high-volume, specialized tasks faster. It delivers up to 4x the output speed of similar-sized models.
Read the original →Romain Torres (@rom1trs)
I just cracked AI camera control. You draw a line. The drone follows it. > drop in any image > draw a line where the camera should fly > it flies that exact line. Runs in Arcads.
Read the original →Get the next issue by email
In any of 8 languages. Unsubscribe anytime.