Generative AI Product Stack, How Perplexity Builds, Amazon Q, and iOS 18

So many things happening! So much AI.

Morning y’all!

I’m getting things started a bit later than I had hoped due to some travel but my commitment is both for myself and for my readers. So, I still find the time and block it out in my daily workflows.

And, after having taken a long break from writing it’s nice to get back to it on the whole. It’s how I process new fields of study and is why I’m even doing it in the first place! Every post drives me deeper into the field and I love that.

※\(^o^)/※

— Summer

---

A look at how to build really great AI products.

And then a good look at how one of my favorite apps out there right now, Perplexity, builds product. They keep things small, nimble, and encourage folks to ask tons of questions along the way.

Folks are getting thrown off by AI vs real, besides trying to figure out what side their on between Kendrick Lamar and Drake.

For the past month, an all-out brawl has consumed hip-hop, with some of the genre’s top artists trading barbs in rapidly released tracks. The dispute, which began in late March with a perceived dig at Drake by Future and Metro Boomin, has roped in J. Cole, Rick Ross and even the estate of Tupac Shakur.

At issue: Who is the current holder of hip-hop’s crown — Drake or Kendrick Lamar — with a whole lot of rumors, personal attacks and inside jokes adding fuel to the fire.

Rap battles aren’t new, but this time, fans are grappling with a very 2024 question: Which of the diss songs are real?

Interesting issues have arisen, have they not?

Dub AI helps you manage your links, an issue that’s eternal for most of us. I remember Delicious back in the day, right?

Just what you thought it was going to be with a PDF AI tool, right? So many of these are out there; it does make me wonder how sustainable it will be.

AI-powered assistants for businesses and developers? No way Amazon! Perhaps the least surprising news today:

AWS has announced the general availability of Amazon Q, the most capable generative artificial intelligence (AI)-powered assistant for accelerating software development and leveraging companies’ internal data.

I’ve used Amazon a lot so I’m not being poo poo on them.

Are there good reasons why the military shouldn’t trust AI? It’s certainly using it:

The U.S. Department of Defense is now seriously investigating what LLMs can do for the military. In the spring of 2022, the DOD established the Chief Digital and Artificial Intelligence Office to explore how artificial intelligence can help the armed forces.

In November 2023, the Defense Department released its strategy for adopting AI technologies. It optimistically reported that “the latest advancements in data, analytics, and AI technologies enable leaders to make better decisions faster, from the boardroom to the battlefield.” Accordingly, AI-enabled technologies are now being used. U.S. troops, for example, have had AI-enabled systems select Houthi targets in the Middle East.

Oh. I see.

Text analytics for LLM apps. PostHog for prompts. Extract evaluations, intents and events from text messages. phospho leverages LLM (OpenAI, MistralAI, Ollama, etc.)

Give the open source project a spin!

Apple is starting to implement “Intelligent Search” for AI-powered outcomes for iOS 18. Not a surprising move either.

This person argues that LLM leaderboards are useless and we should start using Pareto curves.

LLMs are stochastic. Simply calling a model many times and outputting the most common answer can increase accuracy. 

On some tasks, there is seemingly no limit to the amount of inference compute that can improve accuracy.³ Google Deepmind's AlphaCode, which improved accuracy on automated coding evaluations, showed that this trend holds even when calling LLMs millions of times.

Cool experiment.

I love this conclusion:

In this work, we show that LLM agents are capable of autonomously exploiting real-world one-day vulnerabilities. Currently, only GPT-4 with the CVE description is capable of ex-ploiting these vulnerabilities. Our results show both the possibility of an emergent capability and that uncovering a vulnerability is more difficult than exploiting it. Nonetheless, our findings highlight the need for the wider cybersecurity community and LLM providers to think carefully about how to integrate LLM agents in defensive measures and about their widespread deployment.

The title of the post: “No, LLM Agents can not Autonomously Exploit One-day Vulnerabilities” — LOL.

Just what you thought was coming, right?

Eight daily newspapers owned by Alden Global Capital sued OpenAI and Microsoft on Tuesday, accusing the tech companies of illegally using news articles to power their A.I. chatbots.

The publications — The New York Daily News, The Chicago Tribune, The Orlando Sentinel, The Sun Sentinel of Florida, The San Jose Mercury News, The Denver Post, The Orange County Register and The St. Paul Pioneer Press — filed the complaint in federal court in the U.S. Southern District of New York. All are owned by MediaNews Group or Tribune Publishing, subsidiaries of Alden, the country’s second-largest newspaper operator.

The old institutions are under threat and the response is so typical. Fear drives so many decisions for folks when we should be thinking about the hopeful future.

That’s it friends. Have a great one.

※\(^o^)/※

— Summer