@Sachin_scorp7@ChrisGPT According to METR performance on visual tasks lags 40-100× behind software. They can construct decent looking 3D models, but can't create perfect replicas from references. The biggest issue I found is that they don't understand basic things like standardized furniture.
@mattworkman No AI can faithfully reconstruct an object from reference images or model a realistic human face. I have tried with Astra, Sol, Opus 5.5 and Fable 5.2. The best they can do is create something that deceptively looks ok at a glance or from a distance, but is actually low quality.
@ben_j_todd METR used publicly released models to measure horizon length. You're grafting on the internal models on the graph. Plus AECI and ECI disagree substantially. You get radically different time horizons if you use ECI for Opus 5.5 or Fable 5.2 instead. This is just abusing the data.
@EdwardHammd1yp@mert_gulsun@tszzl@xenocosmography@MikeAnissimov2 Your reasoning is utterly retarded. Racist policy would be hiring because someone is white regardless of merit. AI labs repudiate that because they do not share your obsession with people of African descent. They just work with their colleagues no matter where they came from.
@EdwardHammd1yp@mert_gulsun@tszzl@xenocosmography@MikeAnissimov2 Whites are underrepresented compared to Asians. Your initially claim was no african can ever be hired at Ant and you got refuted. Now you're changing the claim to representation, but who cares? The fact of the matter is you're wrong and you'll always feel at odds with the world.
@NuminousVoidX@pigeon__s@synthwavedd@thegenioo Are you guys trying to coax OpenAI into releasing a new model? We can definitely say that their internal model is considerably more powerful than Anthropic's in math and cyber at the very least.