(Source)

AI has seen remarkable growth due to several factors: increased computing power, greater data availability, more investment and talent, along with reduced data storage costs.

In "Generative AI - Navigating the Course to the AGI Future," I explore these factors, focusing on the exponential growth in computing capability, measured in floating-point operations per second (FLOPS), driven by hardware and software advancements.

Nvidia's ascent to a $2 trillion market valuation underscores the rapid progress in this domain. 🚀 

On the hardware front, assembling numerous computing chips in a specific arrangement creates processing units such as CPUs and GPUs, as well as IPUs, TPUs, and NPUs. My book delves into their differences, functionalities, and future directions.

A major trend is the customization of processing units for specific organizational needs, exemplified by Apple's M series.

Today, Groq creates the world's fastest AI

This trend is accelerating. (The derivative of the derivative is positive. 🙃 )

Groq, the AI startup, is making waves with its innovative hardware, enabling ultra-fast LLM responses, as they have introduced an LPU (language processing unit) architecture.

Groq's public benchmarks showcase astonishing speeds of 500 tokens/sec, far surpassing the 50 tokens/sec for GPT 3.5, a tenfold increase for equivalent tasks.

It is safe to say

LLMs will be capable of responding to highly complex queries in a fraction of a second, scanning terabytes of complex relationships in data to provide immediate answers.

Read also about all things cloud computing, quantum computing, neuromorphic computing (so interesting 🧨 ), superconductors as potential next breakthroughs, and how software advances support AI breakthroughs. BUY NOW.

Also, Try out Groq here, using open-source LLMs like Mixtral or Llama.

(Source)

Magic has announced a breakthrough in active AI reasoning comparable to OpenAI's Q* capabilities.

Their system insanely processes up to 3.5 million words, quadrupling Google's Gemini 1.5's 1M word context window!

This can be crucial for AI software development, enabling a necessary understanding of codebases, a task where current RAG architectures fall short due to their need to chunk information (and only include some sequential meaningful chunks).

Despite the potential, larger context windows, like Claude's 200k, show limitations, as illustrated below.

Optimizing long context windows for professional use is an ongoing challenge.

Nonetheless, Magic's innovation has garnered significant attention and funding, $117M, with notable investors like Nat Friedman (former GitHub CEO).

Groq is a grinder company (Chamath mentioned 4 major pivots), and I’ll be watching it this year.

(Source)

Meta has publicly released the video joint embedding predictive architecture (V-JEPA) model, which is the next step in machine intelligence (AMI) to achieve a more comprehensive understanding of the world.

What distinguishes AMI from AGI?

AMI represents the current state of AI, specializing in specific tasks, while AGI requires sophisticated cognitive capabilities and remains an objective for the future. Thus, the vision of AMI = AGI.

V-JEPA is engineered to predict missing areas in spatially and temporally masked videos within a learned representational space.

The objective is to identify semantic representations within video data.

The JEPA models by Meta might be the breakthrough we need to achieve AGI. (I’ll give it a 76% chance). Learn more about it my book.

(Source)

Meta's extensive open-source strategy has compelled AI leaders like Google to release open-source models.

Google has now semi-voluntarily released the Gemma AI Model, for which they deserve credit.

They launched it with significant fanfare, showcasing impressive benchmark results that outshine Llama-2.

Yet, in practical applications, Gemma performs disappointingly, as evidenced by Matthew Berman's test video.

I receive several questions weekly. A frequently asked one stands out, so I'm featuring it here.

How does prompt engineering play a role in creating complex narratives or scenes with Sora (AI for video generation)?

Prompt Engineering is crucial, especially when the models have a higher chaos factor. Sora, for instance, is one model that incorporates all the physics and associations of a scene and can be perceived as more chaotic (with little determinism).

Having a prompting framework helps a lot. For video generation, the prompting pattern V-I-S-I-O-N is a good starting point. It goes like this:

Vignette (Describe a Scene):

• Example: Dawn at a medieval marketplace, stalls bustling with activity under the early light.

Interactions (Specify Dynamics):

• Example: Merchants energetically sell wares to villagers, children dart around, and performers entertain.

Sentiment (Convey Emotion):

• Example: A lively, communal buzz fills the air, signaling a day of prosperity.

Inciting Event (Introduce Change):

• Example: A carriage rushes in, causing commotion as people and goods scatter.

Objective (State Purpose):

• Example: The tale of a stranger begins whose arrival alters the town's course.

Nuance (Add Detail):

• Example: A passing carriage's emblem and a hooded figure's gaze with an artisan convey a hidden past.

Do you have a question for me? Write me!