10x Your Workflow with AI 📈

Work smarter, not harder! In this section, you’ll find prompt templates 📜 & bleeding-edge AI tools ⚙️ to free up your time.

Fresh Prompt Alert!🚨

Ready to turn your investment dreams into reality? 🚀💰

This week's Fresh Prompt Alert is your golden ticket to financial wizardry! Say goodbye to aimless scrolling through stock tickers and hello to personalized, pro-level market insights. Whether you're a Wall Street wolf or a cautious kitten, this prompt will have you navigating market trends like a boss.

Who knows? Your next big investment might be just a prompt away! 👇🏼

Check out an example of this prompt in action!

5 AI Tools You JUST Can't Miss 🤩

• 📡 ReviewRadar: Figure out what users want, fast! with AI-powered Customer-Review Analysis

• 📖 Inncivio: An AI-Powered Learning Infrastructure for Businesses

• 🎙️ AudioNotes AI: Transform Your Thoughts into Clear Text Notes

• 🧱 Bricks: AI-powered spreadsheets for effortless reports

• 👨‍🏫 GuruBase: The tech world’s short-cut search

Hot off the Wires 🔥

We're eavesdropping on the smartest minds in research. 🤫 Don't miss out on what they're cooking up! In this section, we dissect some of the juiciest tech research that holds the key to what's next in tech.⚡

Remember when having a calculator on your phone was mind-blowing? Well, in this AI-powered era, we've come a long way from basic arithmetic. Now, it's all about having increasingly capable AI models right in your pocket!

And Microsoft just raised the bar with their latest breakthrough: the Phi-3 family of Small Language Models (SLMs), which pack some serious punch on the go!

Forging the fundamentals

As language models evolve, two paths have emerged: massive models with hundreds of billions of parameters, racing toward AGI; and efficient small models that pack impressive power into smartphones and wearables.

The vision behind SLMs?—Democratize machine intelligence. And, Microsoft is leading the charge with the Phi-3 family, built to bring machine intelligence to everyone, anywhere, anytime.

Before we dive into Microsoft's latest creation, let's take a quick look at where Small Language Models stand today. Despite their compact size, SLMs are proving powerful:

• Data Quality Over Quantity: Training data quality is proving to be super critical for SLM capability.

• Overtraining FTW: SLMs are typically trained on massive amounts of data (>1.5T tokens), regardless of their size. This "overtraining" helps pack more power into smaller packages.

• Closing the Gap: Open-source SLMs are catching up to their closed-source counterparts in common sense tasks, though complex reasoning remains a challenge.

• Hardware Matters: SLM performance isn't just about size – it's also about how well the model architecture aligns with the hardware it's running on.

• Memory Footprint: Generally, a model's memory usage scales linearly with its parameter count.

It's in this context that Microsoft's Phi-3 family enters the scene, ready to push the boundaries even further.

But before we dive into the nitty-gritty, let's decode some tech jargon:

So, what’s new?

Phi-3 is part of Microsoft’s Phi series, which started with Phi-1 (with the seminal paper “Textbooks Are All You Need”) —an AI superstar in coding tasks—and evolved into Phi-2, which showcased superior reasoning and language understanding.

Now, the Phi-3 and Phi-3.5 families are here, and they’re not just competing—they're are punching way above their weight class (models >10x their size) in tasks like language understanding, coding, math, and reasoning.

Here's the kicker: The phi-3-mini model can be quantized to as little as 4 bits, making it lightweight enough to deploy on a smartphone. That’s less than 2 GB of memory—an absolute feat of engineering & data science!

Under the hood…

Phi-3’s secret? The models are designed with a highly optimized transformer-decoder architecture similar to Llama-2, boosted by some clever optimizations:

• Blocksparse attention: Basically, instead of paying attention to every part of the input text in a conventional manner, this method skips parts intelligently, saving processing time without sacrificing accuracy. It’s a perfect fit for resource-constrained environments. (Interested in knowing more? Check out this deep dive)

• High-Quality Data: Instead of force-feeding the model endless information, they curated a "gourmet" dataset. Instead of overtraining, the focus here was more on having a sufficiently optimal data mixture for training.

• Post-Training Polish:

  • Supervised Fine-Tuning (SFT): Extra lessons in math, coding, and reasoning.

  • Direct Preference Optimization (DPO): Teaching the model to be a responsible, chat-savvy AI citizen (yes, these models were developed using a safety-first mindset, with a strong focus on responsible AI principles)

But wait, there’s more!

Microsoft didn't stop there. They've also introduced the Phi-3.5 series, designed to tackle multilingual, multimodal, and long-context tasks:

Beyond the previous techniques used to train Phi-3 models, these models have the following upgrades:

• Phi-3.5-mini uses “LongRoPE” technique to expand the context window from 4K to a whopping 128K tokens—giving it unprecedented ability to handle complex, long-form content.

• Phi-3.5-MoE uses a clever "mix of experts" approach, activating only the 2 most relevant ‘experts’ among the 16 available, for each token. It uses the “SparseMixer” technique, which improves the calculation gradients (which tell the model how to adjust its parameters to improve) efficiently & scalably, even when only a subset of experts is active.

• Phi-3.5-Vision is a standout multimodal model in the family, blending CLIP ViT-L’s image understanding with the power of Phi-3.5’s natural language capabilities. It shines across tasks like image-text reasoning, PowerPoint parsing, and even video summarization. This model represents a significant advancement over its predecessors, offering improved performance and broader applicability while maintaining a relatively compact size.

Why does this matter?

These tiny titans outperform models twice their size (or more) across language, reasoning, coding, and math tasks. The Phi-3-mini is even giving GPT-3.5 a run for its money!

Phi-3 is already making a real-world impact. ITC, a major Indian business conglomerate, is already putting Phi-3 to work in their Krishi Mitra app, bringing AI-powered agricultural assistance to over a million farmers – even in areas with limited internet access.

As we watch the AI revolution unfold, one thing's becoming clear: sometimes, the biggest breakthroughs come in the smallest packages. So the next time someone asks if that's a supercomputer in your pocket... well, you just might be able to say yes!

Spark 'n' Trouble Shenanigans 😜

Boy, do we have a treat for you! We've been playing around with Moshi, the first real-time open-source voice AI by Kyutai Labs, and it's blown our circuits!

This open-source voice AI from Kyutai Labs is shaking things up with its real-time dialogue skills and the potential for local deployment on your laptops. With male (Moshiko) and female (Moshika) variants, plus the Mimi speech codec, there's plenty to explore.

While it's not perfect – it can be a bit abrupt and goes off on tangents – it's impressively expressive and spontaneous. Industry pros like Karpathy and Saravia have noted its quirks, but we think that's part of its charm.

Give Moshi a whirl and let us know if you find it as endearingly chaotic as we do.
Does it live up to the hype, or is it just another voice in the crowd?