Applied_AI

People are scared but it can do lots of good if it's used for the people that really need it

Does RTK actually lead to savings? I recall seeing some opinions suggesting that you save on tokens because the outputs are shorter, but the quality ends up suffering—and sometimes token efficiency doesn't even improve, because the model doesn't fully grasp the output returned by RTK. Though, I must admit that I haven't personally benchmarked this myself.

Also MLFlow since version 3.0 has nice features for agent observability. Worth investigating too as it is well established solution

If you only want observability then it is save and easy to switch to/from LangFuse. Other features (if you care about) are more lock-in.

Also, it take some times for LangFuse to process spans and traces from the agents, so if you want to have visibility in milliseconds then that is not the option for you. But if few minute gap before code execution and traces being available to explore is not a problem, then Langfuse is good. Other than that I did not observe problems with it.

Sonnet 5 Max performs worse than Opus 4.8 High and Fable 5 Low and yet costs more to do a task.

The comparison is even more absurd between Sonnet 5 Max vs Composer 2.5 - they perform similarly, yet Composer is... 15x cheaper.

One might argue it's Cursor's benchmark, so it might be biased, but then again, even GPT-5.5 Very High performs better and it's almost 2x cheaper.

If you follow PC hardware, this is one of the more interesting long-form videos released recently.

GamersNexus interviewed multiple major hardware companies about the current state of the industry, market pressures, margins, consumer trends, and where things seem to be heading. Surprisingly candid throughout.