there’s clearly some confusing conflicts + double-think going on in AI between: 1. Closed labs saying use our harness, it’s naturally post-t…

there’s clearly some confusing conflicts + double-think going on in AI between: 1. Closed labs saying use our harness, it’s naturally post-trained and gets the best out of models by being “in-distribution” 2. “Models are approaching AGI” and generalizing to do incredible work across many domains 3. In-context learning maximalists pushing skills repos and special instruction sets 4. Task specific harnesses and models vastly outperforming “general purpose models” and harnesses if I had to pick between any of these I say - you can build a task specific harnesses and outperform the closed lab harnesses, evidenced by some work we’ve done, Factory, terminal bench, etc - models are not incredible generalization machines, they’re very sensitive to their post-training distribution but that’s prob fine, we can build harnesses around them to do useful work - the future is owning your model + harness stack and deploying specialized intelligence so that you don’t need to rely on a big lab for everything. big labs will still produce great models not all of the above 4 points can be true and every player sort of has an interest in a subset of those being true among consumers, model labs, Agent labs, open source model makers, vertical ai companies. it’s totally worth doing evals for yourself here more open models + harnesses means you at least have a choice which path you choose 🚀

Source: Harrison Chase (X) | 2026-04-20

Related