AI Daily Brief@2025-03-24 | Brain activity and AI activityare highly similar | Zapier MCP | Grok app

Let's start by exploring an interesting research paper.

On March 21st, Google Research published a paper titled "Deciphering language processing in the human brain through LLM representations." This paper reached a fascinating conclusion: brain activity and large language model activity show a high degree of consistency. How did they reach this finding?

Before diving into the paper, let me explain two concepts: Whisper is an open-source speech-to-text model from OpenAI, and embeddings are mathematical vectors that transform information into something models can understand, essentially arrays of numbers.

The researchers had test subjects engage in natural conversation while recording two things: the subjects' voices and their brain neural signals.

They then fed the recorded voices into the Whisper model, which converted the speech into embeddings. They extracted two types of embeddings: speech embeddings representing sound characteristics and language embeddings representing semantic features.

The researchers also converted the subjects' neural signals into mathematical vectors, for example, by extracting segments of neural wave data along a time axis.

This gave the researchers two sets of vectors: embedding vectors produced by the model and neural signal vectors from the brain. They looked for linear relationships between these two sets, and once they found these relationships, they could approximately convert between them.

The researchers discovered that after finding this linear relationship, they could make predictions. For instance, when test subjects engaged in new conversations, the embedding vectors generated by feeding the new speech into the Whisper model could be transformed into neural signal vectors that closely matched the actual neural signals.

This suggests that when processing language tasks, the internal activities of the brain and large language models are highly similar. It also means AI models can help us understand how the brain works.

Enthusiasm for MCP continues to run high, and now automation platforms are joining the game.

Zapier is an application automation platform that can connect over seven thousand apps and support tens of thousands of automated operations. Yesterday, they released Zapier MCP, which allows models connected to this MCP service to interact with thousands of applications. Their slogan is "Connect your AI to any app with Zapier MCP."

MCP acts as a bridge, and Zapier is also a bridge. Through these two bridges, models can connect to almost any app. Does this mean other MCP services can take a break?

I think we're still in the exploration phase of connecting models to external tools. Let's wait and see how this industry develops!

In the AI startup space, some are advancing rapidly while others are fading away.

Project management company Height announced that their product will shut down on September 24, 2025. The choice of this date is likely to commemorate their product's fourth anniversary. Height officially launched on September 24, 2021, and will officially close on September 24, 2025, making this date both a birthday and a memorial day.

If you're not familiar with Height, Linear and ClickUp are its competitors. Clearly, Height didn't secure a ticket to the AI era.

Something interesting happened yesterday.

Someone open-sourced a tool called Diverce, which describes itself as "A tool for converting Next.js projects from Vercel to Cloudflare."

The CEO of Outerbase then posted on Twitter saying:

"Seems like a fairly relevant time to launch this... convert your Vercel nextjs project automatically and deploy it to Cloudflare"

When Cloudflare's CEO saw this post, he shared it and sarcastically commented:

"Save money, get better performance, and work with a company that actually cares about security."

Finally, let's look at some minor news from major AI companies.

Yesterday I was complaining that Grok's Android development team consists of just one person, which explains why I've been waiting so long for it to launch. But today, the Android app arrived! I applied through Google Play's testing program, so I'm not sure if this is a large-scale release.

Google AI Studio's product lead, Logan Kilpatrick, posted on Twitter with just one domain name: ai.dev. Clicking on this domain redirects to AI Studio's official website.

What's interesting isn't this, but the fact that the ai.com domain points to DeepSeek's official website.

I've never understood how Google AI Studio is positioned. Everyone can use it for free, the interface is extremely basic, yet some really good features appear on AI Studio first. Strange, very strange!

Видео AI Daily Brief@2025-03-24 | Brain activity and AI activityare highly similar | Zapier MCP | Grok app канала Benjamin Turing