😺 Is Mercury the new AI to watch?

PLUS: a new prompting technique you need to know...

Welcome, humans.

Want to see a robot do martial arts? Of course, you do…

Our favorite reaction: “The video ends abruptly because he forgot to switch on Testing mode.”

In all seriousness, these G1 robots from Unitree can do some seriously human movements, but these are more likely scripted than AI-driven (more info here).

That said, even scripted robot fights could make for some next-level TV. We can see it now: Battle bots LIVE on Netflix—it’ll be like WWE, where everybody knows the fights are scripted in advance, but still plenty entertaining.

“CAN YOU SMELL WHAT THE BOT IS COOKING?!”

Let’s just hope there won’t be any malfunctions as embarrassing as this one on opening night…

Here’s what you need to know about AI today:

Inception Labs presented a new way to generate text.
OpenAI announced plans to released GPT-4.5 to Plus subscribers.
Amazon developed a reasoning AI model.
Coreweave bought Weights and Biases for $1.7B.

Advertise in The Neuron here.

Text-based AI models are about to take a page out of the AI image playbook…

Right now, the top AI models most people use are AI chatbots like ChatGPT, Claude, Gemini, and Grok. All of these are “language models” that generate text one token (word) at a time, left-to-right.

And for a while, this approach seemed like the only viable path for text AI.

Until Inception Labs, that is. The startup just unveiled Mercury, an AI text model that uses “diffusion”—the same approach that powers image generators like Midjourney and DALL-E—and it's up to 10x faster.

Here’s how it works:

Mercury generates text “all at once”, not word-by-word.
It produces 1K+ words per second on typical hardware.
Another research model, called LLaDA, confirmed these approaches can match traditional AI chatbots on standard tests.

Think of today's AI chatbots as writers who can only type one word after another. The new diffusion approach works more like a painter who starts with a rough sketch and refines the entire canvas at once.

A speed test shared by Inception.

Mercury's code-writing tool is available to try now. If you want to see how freaky fast it is, try asking it to “implement” any simple video game “in html5” and it will truly blow your mind. We had a literal jaw on the floor moment testing this.

As you’ll see, Mercury creates a draft of the whole response at once and improves it through multiple rounds—fixing mistakes, reasoning in any direction, and potentially reducing overall errors along the way.

It also solves the “reversal curse.” One fascinating discovery is that diffusion models can overcome a major weakness of current AI.

If ChatGPT learns “Taylor Swift wrote Anti-Hero”, it can easily tell you what songs Taylor Swift wrote, but often struggles with “Who wrote Anti-Hero?” Because diffusion models don't have the limitation of writing left to right, they’re better at answering those types of questions.

What this means for the future: In a world full of similar AI chatbots, diffusion models represent the first truly different approach in years.

Andrej Karpathy (ex-OpenAI) called Mercury interesting and said it's long been a mystery why text AI uses one approach while image AI uses another.

Inception says early adopters in customer support, code generation, and enterprise automation are already seeing better user experiences and lower costs with Mercury.

Future capabilities will include faster agents, improved reasoning, controllable generation, and edge deployment on phones and laptops.

The question we have is this: if this method truly is so great, how long until a big AI lab like Google, Microsoft, OpenAI, or Anthropic releases their own version?

FROM OUR PARTNERS

Google Cloud’s Future of AI: Perspectives for Startups

Want to build a real business with AI, but don’t know where to start?

Check this out: This new report from Google Cloud reveals how to transition AI projects from proof-of-concept to production, capitalize on underhyped opportunities, and create immediate value.

Dive into 23 unique perspectives from top AI industry leaders, including…

Amin Vahdat (Google Cloud).
David Friedberg (Ohalo Genetics).
Chamath Palihapitiya (Social Capital).
Crystal Huang (GV).
Dylan Fox (Assembly AI).

…and many more.

Learn what areas investors are prioritizing for AI startups in 2025 and how to leverage AI to outpace competitors and establish a unique market position.

Download now

Prompt Tip of the Day

Chain of draft is a new prompting method to turn “traditional” language AI models (non-”thinking”) into better reasoners by having these models “think step by step”, but ONLY using 5 words or less per step (paper).

The research shows this trick makes AI reasoning faster AND cheaper if you’re using the API version. Matt Berman breaks down how this works in this video, but you can also just copy the below text and add it to your next prompt to try it out.

“Think step by step, but only keep a minimum draft for each thinking step, with 5 words at most. Return the answer at the end of the response after a separator ####.”

Treats To Try.

Aya Vision is an open-weight model vision model (meaning you can run it locally on a computer with a good GPU) that helps you analyze images in 23 languages while using less computing power than competitors (HuggingFace).
Scrunch AI helps you track and improve how your brand appears in AI search results (read more), while Profound does the same with SEO dashboards.
Llama Index now has a cloud service to helps you build AI agents that can search and use information from your PDFs, PowerPoints, and other documents.
Swap connects your e-commerce operations on one platform to handle cross-border logistics, inventory management, and returns (raised $40M).
Pieces remembers your entire work history across all desktop apps so you can instantly recall past code and conversations from up to 9 months ago—this video explains it well (free to download, paid for team sharing + premium AI).
Nothing Phone 3A/3A Pro is a new ~$400 smartphone that uses AI to actively process and organize your saved content (screenshots, voice memos, photos) via its new “Essential Space” feature—read more.
There’s now a “browser agent leaderboard” comparing how well different AI perform when using your computer.

See our top 51 AI Tools for Business here!

Around the Horn.

This is a good comparison of the top AI video tools right now from Heather Cooper, one of our fave AI video educators.

OpenAI plans to release GPT-4.5 to Plus subscribers ($20/month tier) over the next few days—since more ppl will actually get to use it now, we’ll give you our HONEST take on it tomorrow.
Google built a new AI scam detector for Google Messages that flags suspicious text patterns and sends real-time alerts to block scammers before you get defrauded (more).
Elon vs Altman will go to trial this year after a judge ruled Musk “failed to meet the burden of proof” to stop the for-profit conversion outright.
Amazon may develop its own “reasoning model” (that thinks step by step) for release as soon as June—it’ll supposedly be “hybrid” like the new Claude.
AI cloud provider Coreweave will buy the popular AI developer platform Weights and Biases for ~$1.7B in anticipation of its upcoming IPO.