@GithubProjects What’s more interesting is HOW they were able to achieve this using artificial means - the training data was almost entirely synthetic / using rendered 3D video and still frames. Check the video on YouTube search Corridor Crew.
@IceSolst I stood up a perforce instance for a new game developer at a web3 company and it cost so much money for zero better features than git+lfs, change my mind
If you put a quote from a book that contains a greeting / the start of a conversation between two characters into Google, Gemini, Google’s AI, immediately will start role playing as the character with you.
Opus 4.5 on Claude Code has been DUMB this weekend. So dumb I decided to try out OpenRouter / OpenCode and try some of the other models. Huge mistake. 5/5 models, Not a single one could follow a list of three tasks. OpenAI 5.2 Pro - 20 minutes of reasoning and thinking. Failed.
@alexalbert__ Claude code often crashes the whole terminal and becomes completely unusable and unresponsive to keyboard inputs (mouse inputs seem to work sometimes, sometimes not)
@alexalbert__ The issue where the UI scrolls and flickers rapidly and doesn’t allow you to reach the edit box and making the CLI unusable still happens to me in every session featuring sub agents
@alexalbert__ It also feels very quantized compared to when it first launched and is doing dumb things. I asked it to make a script in our scripts directory and it tried to use mkdir to create the directory, even though it ran a script there earlier in the session
@alexalbert__ Claude Code doesn’t seem to auto compact anymore in any situation, maybe this was a feature removal but its a bit annoying. Switching between Opus and Sonnet 1m tokens doesn’t seem to update the token count and it keeps trying to finish and summarise early
@askalphaxiv My issue with that is that not every paper on arxiv is peer reviewed adequately so how do we know we aren’t just throwing a bunch of slop into a large training dataset? Seems it will introduce massive inconsistencies