From Voice Assistants to Digital Artists: AI Rollercoaster

Leading Edge Newsletter #21

Welcome Back to the Edge!

Hey there, edge-dwellers! After a month-long hiatus (life, am I right?), Leading Edge is back in action. We've got a lot to catch up on, so buckle up for a wild ride through the latest in AI.

From Google's fleeting victory to Elon's eyebrow-raising Grok-2, we're covering it all.

Let's dive in!

Google's Gemini 1.5 Pro: King for a Week

Google finally did it! Their Gemini 1.5 Pro claimed the top spot on the LMSys leaderboard. A whopping 14 points ahead of GPT-4.

Victory parade time, right?

Not so fast.

OpenAI, in true OpenAI fashion, snatched back the crown faster than you can say "artificial intelligence." It's becoming a pattern - OpenAI does just enough to stay on top. Like that coworker who does the bare minimum not to get fired. Come on, OpenAI! Stop teasing us and give us Q* or Strawberry or whatever your next big thing is called!

First or second place aside, I've got to say – Gemini's left me underwhelmed. Sure, the 2 Million token context is nifty, but beyond that?

Meh.

In my experience, it's lagging behind Claude and GPT-4. Plus, it's moodier than a teenager, often refusing to respond for the silliest reasons. If LLM APIs were rock bands, Gemini would be the one most likely to storm off stage mid-concert.

Gemini Goes Live (And Beats OpenAI to the Punch)

Here's where Google can claim a solid win: voice. Remember last March when OpenAI announced voice capabilities with much fanfare? Three months later, and paying customers are still twiddling their thumbs. The best they've got is a limited beta.

But Google? They delivered.

It might not be as flashy - there's no "Her" - but it's here, it works, and everyone can use it (asusming you have an Android phone and is paying $19 per month for the Gemini Advanced subscription). So far the feedback has been pretty good too . Not too shabby, Google.

Not too shabby at all.

When Demo Day Goes South

Speaking of Gemini Live, let's talk about that demo. Ouch. They tried twice, and both times? No dice. But you know what? Kudos to Google for having the guts to try it live.

Sometimes, the demo gods are just not on your side.

It's also a perfect metaphor for the current state of AI assistants. Moody with a chance of spectacular failure. Consider yourselves warned, folks running production workloads on Generative AI.

Grok-2: Elon's Surprise Package

X.ai dropped Grok-2 on us, and color me surprised – it's actually pretty capable! A far cry from the underwhelming Grok-1. While it's not open source (yet), it's making waves.

But let's be real. Grok-2 will be remembered for its image generation capabilities. Powered by Black Forest Labs' Flux model (more on that below), it's churning out some... interesting creations. We're talking pregnant celebrities, gun-toting presidents, and copyrighted characters galore . Here's my favorite, the "The Last Sipper". Not my creation, so kudos to the original poster .

X timelines are a bizarre gallery right now. With the U.S. presidential election looming, you can bet there'll be pressure to slap some guardrails on this thing .

But here's a thought: if people are creating disturbing images and spreading them, is that really the AI's fault? If I stab someone with an IKEA kitchen knife, do we blame IKEA?

Grok-3: The Hype Train's Next Stop

Forget Grok-2 – I'm already hyped for Grok-3. Elon's not resting on his laurels. He's announced the start of Grok-3 training on what he's calling "the most powerful AI training cluster in the world.".

We're talking 100,000 liquid-cooled H100 GPUs on a single RDMA fabric.

For context, Grok-2 used 20K GPUs.

Grok-3? It's going quintuple or go home.

The State of Flux

Before we go, let's take a moment to appreciate Black Forest Labs and their FLUX.1 announcement . These folks, hailing from companies like Stability AI, have raised a cool $31M in seed funding .

Their secret sauce? Two key ingredients:

1. "Rotary positional embeddings" to keep track of large sequences, making it a breeze for the LLM to juggle complex pieces.
2. "Parallel diffusion transformer" to analyze multiple sequences simultaneously, speeding up image creation.

With that, they are claiming to be at the top of the game when it comes to image generation.

They've got multiple versions cooking.

The Pro version's keeping its cards close to its chest for now, but the Dev and Schnell versions? They're out in the wild, ready for you to fine-tune locally and get those custom results. I have downloaded the Dev version, but still have got no time to test it out in my home computer. Will post some results soon.

Until Next Week...

And there you have it, folks! Newsletter #21 in the books. We're still growing, and your feedback is the fuel that keeps this engine running. Got a burning topic you want covered next week? Don't be shy – hit that reply button and let me know!

If you found this newsletter valuable (and I hope you did), spread the love! Share it with your network and help us grow. After all, knowledge shared is knowledge squared.

See you next week, edge-dwellers!

#artificialintelligence #generativeai #leadership #ai #productdevelopment #startups