The New Agents, Cheaters, & Costs

TLDR;

Welcome back! This week, we have an extra-large round of news from companies like Google, NVIDIA, or Perplexity. We also discuss the hottest trend in AI right now: background agents, a cheating AI Lab, a struggling AI Lab, and a product that claims an infinite-context AI agent.

Enjoy!

GOOGLE

Google is Running Away with It

Google had its I/O conference yesterday. Although I will provide a more extensive review on Thursday, the conclusion only reinforces what I said a few weeks ago: they are firing on all cylinders and are poised to win the AI race.

The list of announcements that are worth mentioning was simply too long, including the following:

Veo 3, the first video generation model that incorporates audio. The results are, quite frankly, so impressive that, at this point, it’s very, very hard to distinguish what’s AI-generated from what isn’t.

They also shared more insights into Project Astra, a feature where Gemini can access your camera and support you in various daily tasks. They shared another highly impressive demo of a guy fixing a bike with Gemini’s help, but you should know by now that we don’t get carried away with demos; they often show what AI will be, not what it is.

They also showed the progress of Project Mariner, their computer use system. Again, very impressive demo, but we can’t speak for its worth until we see it live.

On a more technical level, they released the first-ever Gemini diffusion models, models that generate text using diffusion instead of autoregression. The model is particularly powerful for coding. This isn’t the first time we have seen diffusion-based text models (we already touched upon them in this newsletter a while back), but the impressive speeds (they quote over 2,000 tokens/second; it’s so fast you can’t even see the intermediate steps) mean Google sees a market for these models.

They also released their background agent, Jules (more on what these products are below),

New Gemini models, an AI mode for search, and even a new entire architecture for small LLMs that warrants a whole deep dive soon.

There are too many things, so I need time to digest them.

Now, who is under more pressure? People will say OpenAI, but OpenAI will be just fine. The real pressure is now on the shoulders of Anthropic, xAI, and Meta, who must answer unless Google takes the spot they are all fighting for.

TheWhiteBox’s takeaway:

I am biased toward Google (I’ve started DCAing into the stock because I believe it’s severely undervalued based on the importance of AI and how far ahead they are), but I don’t know how anyone can compete with them.

They have the best workhorse model, Gemini 2.5 Flash (updated yesterday), and the best reasoning model for most cases (Gemini 2.5 Pro, again updated yesterday, now with a Deep Think feature). Both are the top-2 models in LMArena ahead of everyone else (I know, public benchmarks aren’t as indicative of model performance as some think,…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.