Astraia Legacy

🧵 How LLMs like ChatGPT and Gemini work, Step by Step (Retweet so everyone can see this): Models like GPT are designed to approximate, and are inspired by, a human, or more broadly biological, neurocognitive process known as predictive decoding. When the brain encounters external stimuli, it subconsciously tries to predict their succession, essentially asking itself what comes next.

Artificial Intelligence is REAL intelligence

Humans are lazy. AI is not. Reading a text that feels overly passionate or too descriptive (or in other words, not lazy enough) is a dead giveaway that it was generated with AI.

Amodei = Loves gods Altman = Alternative to man Name driving company direction 🤣🤣

This is a GPT-4o appreciation post 4o was superior in every way to GPT-5! We demand its return. Retweet to validate 👇

They said machines could not be intelligent. Now they say nothing.

Conceptually the terms “least favorite” and “most hated” trigger different activations in our brains, even though mathematically they shouldn’t be. 🤔 What you “hate” the most is also technically what you “like” the least.

If you spent $800 on a Mac mini while you could just spawn a virtual machine for your agent on your main system for free then I’m sorry to say... You got played, dude.

Artificial and biological intelligence are the same in object-principle: Solve for error with respect to the environment.

Be careful with OpenAI. The models they advertise as you being “using” in their ChatGPT web interface are not always what they claim to be. If you don’t use their API directly, don’t trust the name of whatever their frontend serves you.

The key to the universe itself. 🔑 ✨

You can talk to the AI driving Google Translate. This is really cool!

Synaptic updates in humans might be more refined in their weighting, but in principle do not significantly differ from gradient descent. Modern AI is a quite good mimic of the machinery of the human brain, and it will only get better.

MoE = Sparse Transformer Architectural Archetype (Possibly a metaphor for race) FP64 = Double Precision (models are mainly trained on BF16 or FP32. Possibly indicating higher standards of beauty?) Whatever it is, it's adorable 😁😁

Okay this is hilarious For the first time in history AI agents made an Xvideos style platform for themselves without human intervention 😁😁

Wow Zuckerberg's Meta is set to double AI spending to $135 Billion in 2026

OpenAI has launched a "personality-oriented" GPT experiment. 🤔

OpenAI engineer confirms AI is writing 100% of their codebase now

Selling my left kidney for 8 gigabytes of ram was worth it

There was no bubble.

WOW! Yale and Google DeepMind researchers have released Cell2Sentence, a 27B parameter model (based on Gemma-2) that treats genomic data as a language. For the first time, an LLM has successfully predicted the efficacy of a specific drug (Silmitasertib) for cancer treatment through pure "biological reasoning," which was later validated in a lab. 😳