Ideas Worth Publishing.
Technical deep dives, research updates, and perspectives on AGI safety from the TAL Corp team. We write about what we are learning — including the parts that did not go as planned.
Why Interpretability Is Load-Bearing for AGI Safety
We cannot align what we cannot understand. Mechanistic interpretability is not a nice-to-have — it is the foundation everything else rests on.
All Posts.
The Corrigibility Problem Gets Harder as Models Get Better
Naive corrigibility constraints degrade as capability increases. Here is what we found — and what we are doing about it.
Introducing SAFE-AGENT: A Benchmark for Autonomous AI Safety
We are open-sourcing a 1,240-task benchmark suite for evaluating whether AI agents respect their autonomy envelopes.
Emergent Misalignment in Multi-Agent Systems
When multiple AI agents coordinate, new safety failure modes emerge that are invisible at the individual agent level.
On AGI Timelines and Why We Are Building Now
We do not know exactly when AGI will arrive. But the work required to make it safe takes years — and it needs to start before we need it.
New Posts. No Noise.
We publish infrequently and only when we have something worth saying. Subscribe to get new posts directly — no marketing, no digests, no noise.