AI, But Simple

🤖 The No 1. AI Page on Twitter
📈 AI Research, Beautiful Visuals and Education

[AI, But Simple @aibutsimple]

@Koukoumidis That's a nice take. I think the question is increasingly becoming not “Which model should we use?” but instead “What intelligence should we own?”

[AI, But Simple @aibutsimple]

What most people get wrong is that you don't need that much compute to run large AI models. We can fit an entire 8 billion parameter LLM on a laptop RTX 3070 with 4-bit quantization.

[AI, But Simple @aibutsimple]

Residual Networks Take The Top Spot for Most Cited Deep Learning Paper.
#deeplearning #research

[AI, But Simple @aibutsimple]

Every introductory Machine Learning course teaches this one model. Can you guess what it is?

[AI, But Simple @aibutsimple]

Transformers summed up in 3 sentences:
Transformers take "tokens" as input. Through attention, they have the ability to look back at earlier tokens. This allows them to understand context and deal with longer sequences.

[AI, But Simple @aibutsimple]

Are Transformers Too Slow? State Space Models Like Mamba Provide a Solution to Transformers' Quadratic Scaling

[AI, But Simple @aibutsimple]

Most "AI breakthroughs" we see are just corporations rebranding statistical models that have been around for years. The marketing department is working harder than the engineers.

[AI, But Simple @aibutsimple]

Mixture of Experts (MoE) models scale efficiently without sacrificing performance.

Like ensemble models, they combine different FFNs termed "experts" for the best output.

[AI, But Simple @aibutsimple]

Researchers are moving toward test-time compute—a new paradigm in LLMs that scales performance quicker by "reasoning".

[AI, But Simple @aibutsimple]

Reasoning large language models (LLMs) appear to “think” by breaking a more complex prompt down into smaller steps, resulting in intermediate steps before an output.

These steps are called "reasoning steps" or "thought processes".