XMaster96

Former Senior AI researcher @Aleph__Alpha
EVE Online player since 2013
Co-Founder Pageshift Entertainment - Building the world's best storytelling AI

[XMaster96 @_XMaster96]

I am not sure if this was a serious question or not. But the way I code these days is that I first discuss a feature with my agent. After we have agreed on the implementation shape, I let it implement it.

After the first implementation pass, I start to skim the code. At that stage, I am not really reading it yet. I ask some questions about design decisions and instruct the agent to make changes if I see fit.

After I am happy with the overall design decisions, I start to read every line. I ask for changes if needed. Only after I understand all of the changes and have approved them does the code get committed.

I would say that I am capable of moving twice as quickly as before. At the same time, I am increasing my overall code quality and just how well done everything is.

[XMaster96 @_XMaster96]

@Techweek_ I know 😉

[XMaster96 @_XMaster96]

Part of me is sad that @Techweek_ is over now. But another part of me is relieved. Being there 14 hours a day, going to interesting events, and networking with amazing people is soooo exhausting.

[XMaster96 @_XMaster96]

@Clashluke @shanebarakat Well, not everyone is working on RSI. Some of us just want their LLMs to finally write good stories. 😉

[XMaster96 @_XMaster96]

Do I simply misremember, or did grocery prices go up in the US since the last time I was here?

[XMaster96 @_XMaster96]

@Clashluke Congrets for joining core automation

[XMaster96 @_XMaster96]

@tugot17 How is the intelligence / FLOP profile looking for a hybrid vs. a local + global attention model?

My assumption was always that they are basically the same, but just that the hybrid one is a lot more complex.

[XMaster96 @_XMaster96]

From all the research we did in the past 12 months or so, apparently the thing I believe to be the hardest problem actually turned out the best.

[pageshift.ai @pageshiftAI]

You can’t solve AI slop without solving AI diversity.

So we came up with a new RL strategy for creative writing, one that explicitly trains for diversity across generations.

https://t.co/LmgaNvtsoH

[XMaster96 @_XMaster96]

@meuer_max You actually forgot to list Sid Black; he was one of the EleutherAI founders and the guy who made GPT-NeoX-20B work together with Samuel.

[XMaster96 @_XMaster96]

@Laz4rz Not wrong, I remember in the early years of LLMs AA was the cool kid on the blog

[XMaster96 @_XMaster96]

@Aleph__Alpha Happy to see AA doing MoE models now 💪💪🦾

[XMaster96 @_XMaster96]

@JonasAndrulis I know your where joking, but I belief Anthropic actually pays a psychiatrist to talk with every new Claude model before release, to make sure it is mentally well balanced

[XMaster96 @_XMaster96]

@tugot17 Maybe you need to get out of the shithole called Berlin and explore the countryside in Schleswig-Holstein.

[XMaster96 @_XMaster96]

@tugot17 I don't get it. The quality of life in Germany is one of the highest in the world. What are you doing wrong?

[XMaster96 @_XMaster96]

GPT-6 Astra feels a lot less "stable" than previous OpenAI models. It just keeps solving problems in some really weird ways.

[XMaster96 @_XMaster96]

@kwangmoo_yi Funnily enough I concluded the same thing a couple of years ago, this is why I started to build @pageshiftAI my research lab building model just for creative writing