Half of the Broly fight is him getting bodied by Ice. Speed, Strength, Durability consistency is ass. Broly at Post TOP power somehow not instantly destroying the planet and instead breaking reality, my ass.
I can pick a random manga chapter and guarantee there's some scaling inconsistency.
@md_a_@teslavangelist Issue isn't a lack of light, its a sensitivity problem.
They need to train it to be sensitive enough to detect small objects under 1-2 OOM photons of true-night with light pollution while also detecting small objects under direct sunlight without causing regressions in either.
Neural Networks don't play nice with strict constraints, so how can they be applied to safety-critical contexts? In my recent work, I explored how networks can learn from unsafe proposals and whether or not the learning is beneficial. https://t.co/qxO58DCbth
@elonmusk@vasalex93 Really hoping y'all can focus on 2 more. It's great that Grok 4.5+ is a usable model for work now, but smaller models like the 500B sized ones (or less) should be the go-to for things like competing with GPTs Free tier and the unlimited Luna usage it has.
We really need more benchmarks for practical use cases.
It's annoying seeing models increasingly top benchmarks only to give me an aneurism when asked to explain something or adding 40k lines of code across 80 files when asked to simplify its code.
Sometimes I wonder how much care LLM training requires on hyperparameters. My work only ever requires models on the order of 10k parameters to the point where a generic MLP works great, but it's crazy how many tricks just don't carry over to these atomic scale models.
AA Intelligence Index uses their own harness. Most costs are from 3 Benchmarks, and the bulk of Groks trouble is Terminal Bench 4, which uses its own miniSWE harness and is flagged as a flawed benchmark by EpochAI.
Seems to be a mix between 4.7s heightened focus on checking it's work combined with bad external harness support.
@synthwavedd No, the caveat is multimodal capabilities aren't good. It can't make good SVGs if it can't understand what it sees.
Astra for example is mostly improved visual reasoning but the coding abilities are nowhere near the same jump.
This is lava lamp benchmark tier guessing.
Producing no or sterile gametes still yields a sex, so not producing both isn't relevant.
5 factors of sex determination are observing how sex develops; they don't define sex.
Sex is organization to produce a gamete. The pathways are mutually exclusive, but we're multicellular. OT-DSD has both, one wins. What definition of sex implies only the dominant pathway determines sex?