10 Comments
User's avatar
JL's avatar

What a great useful resource guide of AI references. I have restacked as more people that like a AI reference should read your newsletter/publication

Travis Wright's avatar

The agentic AI definition hides the hardest part in one phrase: "keeps on improving the results until it achieves the desired goal." In practice the next term every team learns is the stopping condition, because an agent that never decides it's done is just a very confident while loop. Mid-turn steering is the sleeper feature in the GPT-6 section, since being able to say "wait, not that" while it's still thinking is basically the undo button agents have been missing.

Abhishek Pan's avatar

Hybrid search earns its place on exact strings like error codes, SKUs and config keys, where pure embedding search tends to return something that looks related but isn't the match. Adding a keyword leg next to the vector leg usually fixes those queries without touching the embedding model.

Daily AI Caffeine (No BS)'s avatar

The RAG entry has a detail that is easy to miss: retrieval is mostly from a vector database but can also come through APIs, SQL or hybrid search. The GPT-2 to GPT-6 walk-through is a clear way to see what each generation added, from pre-training only at GPT-2 to RLHF and multimodal input at GPT-4 and built-in reasoning at GPT-5.

Raghav Mehta's avatar

The async tool calling and mid-turn steering combo in the GPT-6 summary is the part that changes agent harness design most. Once a model can take a new instruction while tool calls are still in flight, the loop needs cancellation and idempotency for those calls, otherwise steering leaves half-applied side effects. Do you know if the steering message interrupts pending tool calls or only gets queued until the next turn boundary?

Mirko (GammaVibe)'s avatar

Very helpful reference, thanks for putting this together.

Nan Deng's avatar

The GPT-6 line I would not leave as a feature bullet is mid-turn steering next to async tool calls. A new instruction that arrives while the model is still thinking can land after a tool call has already been issued, so the user correction and the in-flight write are no longer the same turn. I would bind the tool request to the instruction that was current when it was formed, and drop or re-check any call whose steering message arrived before the tool returned. A 1.05M context window does not fix that race: more room to think just widens the window where the plan and the user's latest instruction disagree.

AIFix's avatar

Seeing RAG listed as retrieval that can come from APIs, SQL or hybrid search, not only a vector database, is a useful correction, since many beginners assume it always means embeddings. For a small team, would you start with plain keyword search first and add vectors only if it misses too much?