RT @SpaceNews_Inc: Firefly Aerospace files to go public

lol it’s true check it out

Banana mayonnaise sandwiches 🤌.

I tuned all the higher end Ibanez’s in Guitar Center down to drop c lets see if anyone notices.

I accidentally said “you gotta crawl before you fuck” in a meeting today and I played it off like I meant to say it. This is what it means to have grace.

Someone needs to study the link between podcasting and dvd commentary tracks. No I did not google this nor did I do any research on it. It’s nearly 2 am and this is a middle of the night pee thought.

Vibe coding has reignited my love of programming, not because it helps me but because I get off on pointing out when ChatGPT is wrong. It’s a powerful outlet for my urge to bully people over their code. I’ve been downright congenial at work. Everyone in the office is like “wow.

I invented a new genre of music by playing a song in add9 slowly and not staring at my shoes.

I hate that my birthday was dictated by the whims of a doctor who most likely wanted to make sure he didn’t have to go into work on the 4th. I could have been a July 6 birthday.

The Barney Miller theme song is the hardest bass line ever recorded.

I bought this shirt for my wife when I lived in Korea because it gets more nonsensical the more you read it. It even slips into what I can only assume is Dutch or some Dutch-adjacent language at points. I was drunk when I bought it

If NASA wants more people to voluntarily resign they should do way more town halls. Possibly daily. Just keep scaring the shit out of everyone.

Broader impact:.#MemoRAG’s memory-inspired framework points the way for future #AI—blending retrieval, memory, and reasoning. For projects like ALICE (my work on LLM-KG integration for truth-grounded AI), MemoRAG provides both a comparative benchmark and inspiration for improving.

Benchmarks:. Outperformed strong baselines (BGE-M3, HyDE, #Mistral-7B, #Llama3-8B) on NarrativeQA, HotpotQA, MultiNews.On the ULTRADOMAIN benchmark (legal, financial, education, fiction), #MemoRAG averaged +9.7 F1 (in-domain) and +2.4 F1 (out-of-domain) over prior models. 7/8

Training the Memory Module:. Pre-training on massive #corpora (e.g. #RedPajama) to teach long-context compression.Supervised fine-tuning with >17k examples for high-quality clue generation.Results: more relevant retrieval & better coherence in responses. 6/8.

Why does this matter?.#MemoRAG can:.Synthesize info from multiple, distributed sources. Summarize long/unstructured docs (e.g. legal, academic, financial).Handle implicit/ambiguous queries (e.g. “How does the book convey the theme of love?”) where standard #RAGs struggle. 5/8