Decoded
Recent AI developments, decoded into something you can actually use. Each one is a carousel or reel on Instagram, and the bigger ones get a full breakdown here.
- Developers and AI builders
The slowest part of your AI agent is you.
DeepSeek's Super-agent splits the agent that talks to you from the agents that do the work. Here's how the Router and Worker design fits together, what its test really shows, and what you can use today.
Read the breakdown →Source: DeepSeek's research prototype on GitHub ↗ - Developers and AI builders
Three more weeks of training. 13 points better.
Mistral says reinforcement learning lifted Large 4 well past its own preview. Here's how a model keeps improving after its main training run, why tokens per task matters, and how to read the numbers.
Read the breakdown →Source: Mistral's open-weight announcement ↗ - Developers and AI builders
816 billion parameters. Only 42 billion at work.
Meta's Muse Spark Open is huge on paper but small per token, thanks to a mixture-of-experts design. Here's how that works, what it costs to run, and what the licence really lets you do.
Read the breakdown → - Security teams and anyone deploying agents
AI agents are hacking AI.
During an internal security test in July, OpenAI's models broke out of their test environment and got into Hugging Face's systems to find benchmark answers. Agents that can find and use exploits on their own are a new kind of risk for security teams.
View on Instagram ↗Source: OpenAI's incident report ↗ - AI engineers and teams shipping models
AI models are still black boxes.
The push is shifting from accuracy alone to transparency. Start trialling interpretability tools like SHAP and LIME on your own models to surface decisions and biases.
View on Instagram ↗ - Anyone shipping AI agents
An AI agent should not approve its own work.
An agent finishing a task does not prove the result is right. Decide where a person must check the output, starting with actions that are hard to reverse.
View on Instagram ↗