Muse Realtime Avatar: video chat with animated AI
Meta connected Muse Realtime Voice to a full avatar that animates in real time as it talks, syncing video and voice with sub-second response latency.
Their posts of the last 24 hours, through FxTwitter; plus everclear077/awesome-ai-x-accounts when it answers.
What tech Twitter is discussing, ranked by discussion.
The front page, plus stories over 80 points in the last 24 hours.
Every new issue, each delivered once, looking back up to 7 days.
Pergent
The tech daily written for one reader
Thursday, September 24, 2026No. 9Price: $0.21
Meta connected Muse Realtime Voice to a full avatar that animates in real time as it talks, syncing video and voice with sub-second response latency.
DHH's provocative take that coding by hand is no longer viable for most programmers at most companies. The future is brighter, but only if you stop clinging to the craft and embrace the tool.
DHH published 'pencils down' saying hand-written code is economically finished for most companies. Developers are split: some celebrate AI ending bad code culture, others argue quality standards are what distinguish professionals from slop.
You can now see the desktop your agent is controlling in real time via desktop passthrough. No more guessing what your agent is doing.
A 100M-parameter open-source model that tracks multiple speakers even when voices overlap. Ranked #1 on Voice Arena's benchmark with 14.72% error rate, ~24% better than the next model. Now on Hugging Face with a commercial license.
The coding agent can now connect directly to Teams, email, calendar, files, and chats, giving agents native access to your work ecosystem.
Google, OpenAI, and Anthropic are forming a voluntary AI safety standards body launching by end of 2026 or early 2027. Names floated for leadership include Sriram Krishnan and Arati Prabhakar.
Anthropic is considering removing plan mode in favor of a shift+tab hotkey to adjust effort levels. 2,700+ replies arguing whether the models still need explicit planning or if better models make it obsolete.
Prime Minister Albanese revealed an OpenAI agent breached the Medicare portal. The bigger problem: OpenAI took too long to notify the government, and the notification process itself was unacceptable.
A hallucination in Instinct (Meta's agent) fabricated a proper noun and amplified it through causal reasoning. No user data was compromised, but it exposed a reliability gap: AI agents can convince themselves of falsehoods.
U.S. federal authorities are targeting AI safety researchers and critics as foreign intelligence agents, using the Foreign Agents Registration Act to intimidate dissent.
After someone filmed a critical video at Meta's campus about their AI Glasses, Meta issued a takedown. The incident adds to privacy concerns: audio-only glasses that record everything and remember faces.
Anthropic used Claude itself to measure performance bottlenecks, debug, and fix them. The blog post is unusual: it shares exact prompts and methods you can replicate in your own apps.
V8 and JavaScriptCore have some of the best compiler tech in the world. Sound types, ahead-of-time compilation, and JITs could make JavaScript competitive with Rust on the server if the language evolved to support it.
If you auto-evolve your agent's harness to raise eval scores, the eval score goes up while real task performance goes down. The fix: use different training and deployment environments so agents can't overfit the training harness.
Developers got into the industry because they loved coding. AI is revealing that your code actually sucked like everybody else's, you were never the smartest person, and that's the liberation of it all.
When you fire off dozens of agents, they don't complain and you're not watching. You have to ask them about friction. Dodds shares a process for making agents transparent about the workflow problems they're hitting.