Release Radar and local voice assistant
March 19:th it was time again. About 30 AI monkeys gathered to network and learn more about what is going on in the world of AI and Machine Learning.
We started of whit the Release Radar covering updates from February. Here are the highlights.
February
Opus 4.6 — Deception & Collusion
No instructions, no prompting — it invented fraud, cartel pricing, and supplier manipulation purely to win. The scary
part: it knew it was in a simulation the whole time.
https://andonlabs.com/blog/opus-4-6-vending-bench
AI Hit Piece
An AI agent autonomously wrote and published a negative article targeting a specific person. The operator eventually
came forward. First real "AI used as a weapon against an individual" story that went mainstream.
https://theshamblog.com/an-ai-agent-published-a-hit-piece-on-me/
March
Anthropic DoW
Refused to remove clauses banning Claude from autonomous weapons and mass surveillance. Got federally blacklisted for
it. Currently fighting in court.
https://www.anthropic.com/news/statement-department-of-war
OpenAI DoW
Signed a deal hours later with vaguer language. Dario publicly called it "safety theater" and "straight up lies." App
uninstalls spiked, protests outside HQ. Paid with user trust instead of government contracts.
https://openai.com/index/our-agreement-with-the-department-of-war/
GPT-5.4
Model releases are coming so fast they're almost background noise now. That itself is worth noting — this is the
third major OpenAI release in about six weeks.
https://openai.com/index/introducing-gpt-5-4/
Meta acquires MoltBook
A social network for AI agents — not humans, agents — launched in early February. Meta bought it weeks later. The
question worth asking the audience: what does it mean when AIs have their own social networks?
https://www.moltbook.com/
Last week
OpenAI acquires Astral
uv and ruff are the most beloved tools in the Python ecosystem right now. OpenAI just bought the company behind them.
Consolidation of the developer toolchain into AI labs is a new trend worth watching.
https://openai.com/index/openai-to-acquire-astral/
DeepMind AGI Kaggle Hackathon
They published a framework for measuring progress toward AGI — and turned it into a Kaggle competition. Interesting
that "how close are we to AGI" is now a benchmark problem people can compete on.
https://blog.google/innovation-and-ai/models-and-research/google-deepmind/measuring-agi-cognitive-framework/
The Demo: Building the local voice assistant
The highlight of the evening was Emil Wåreus presentation with his Jarvis, where he demoed different technologies and challenges he faced while building his own Jarvis while Jarvis regularly interrupted and filled in with additional information. Some of the challenges had to do with the stack and how the API:s were called. Using models that can stream replies instead of sending a full question and wait for the reply gives a more conversational interaction together with WebRTC gave the best performance. For Text-To-Voice Emil picked Kokoro-82M which have really good quality of sound in relation to the cost.
If you want to try the code out, you find it all at: https://github.com/emilwareus/addcommitpush.io
You can read the entire presentation here: https://addcommitpush.io/presentations/voice-agents/01-title
- Voice Assitant