one of the most exciting LLM posts i’ve read in a while. read it. your future self will thank you

go to openrouter . try the new mystery model - horizon alpha. - it is great at coding and writing.- it is stealth (no one knows who created it yet).- might be openai gpt-5 or something similar . i asked: "what else can you tell me about yourself?"

RT @TimebackLearn: Crush academics in 2h. No homework necessary!. Enjoy your school day, not dread it!.

RT @LamarDealMaker: @mattshumer_ thanks for sharing. autoRL is crazy. 1. describe your task in natural language.2. autoRL auto‑creates synt….

everything just changed!

how to easily solve cheating in school. - use personalized ai tutor instead of teacher. - 2-3 hrs with ai tutor on core subjects .- no chatbots allowed.- do something fun the rest of school day .- no homework, cheating solved. simple.

here is the . course: github repo:

reinforcement learning is exploding. analysts see a $120B market by 2025, growing 60%+ yearly. why rl skills matter:.- robots & automation.- smarter supply chains.- adaptive user experiences.- autonomy at edge. this free rl course by huggingface is an amazing resource:

we’re living humanity’s greatest era.

$100M/yr if this is you too.

things these ai legends have in common: . •lots of citations.•breakthrough ideas.•top ai papers.•solid schooling.•big-name labs.•move between edu and industry.•passionate about deep learning.•share code and data.•tech hub cities.•future-focused.

RT @PalmerTheDev: For months I only used Gemini and Cursor. A few weeks ago I finally gave Claude another chance, and haven't looked back a….

devs switch tools all the time. but claude code is set to solidify #1 spot. beyond writing code. i’m seeing people experimenting with using it to power autonomous businesses. feels like we’re barely scratching its potential

ai literacy = instant raise

huge shout out to @_Ian__Brown_ for showing me this paper.

open‑sourced & democratized: the authors released the entire framework, database and the 106 discovered architectures . anyone can explore, extend and build on top of this. i’m digging into it right now . paper: github repo:

under the hood the system explores a branching tree of designs . each node is a model, color indicates its score. the search trace shows how ASI‑Arch evolves from a baseline (DeltaNet) to high‑scoring architectures (red = low scoring & navy = high scoring)

what did it discover? . ASI-Arch created models like PathGateFusionNet and ContentSharpRouter beat human‑designed baselines on tasks ranging from Wiki & LMB perplexity to PIQA and HellaSwag accuracy . continuous improvement: benchmark scores rise while loss falls

the number of new architectures grows linearly with compute . every 60 GPU hours yields a new SOTA model. much faster than the 2000 hours per model typical for human researchers

why this matters: . ASI‑Arch goes beyond traditional neural‑architecture search. it uses a cognition base and team of agents to move from automated optimization to automated innovation. uncovering design principles humans hadn’t imagined. AlphaGo’s Move 37, but for model design

how it works: . ASI‑Arch is an autonomous research loop . the cognition base reads papers & extracts insights . a “researcher” agent proposes new architectures and writes code. an “engineer” agent trains and tests them. an “analyst” agent judges results & feeds them back