AI Daily: Excel Finally Gets an AI Brain, OpenAI Reveals Database Architecture Behind 800M Users

Honestly, some very “grounded” big things happened in the AI circle this week.

We are used to seeing model updates floating in the cloud, but this time, Anthropic reached directly into the office software we are most familiar with—Excel. This could completely change the way we process reports. On the other hand, OpenAI rarely shared their engineering details, telling everyone how they used a traditional database to handle traffic from 800 million users.

Whether it’s the spreadsheet you face every day or the scalability issues that give backend engineers headaches, today’s content is packed with substance. We even have an open-source music model to introduce to you. Ready? Let’s get started.

Claude Officially Enters Excel: More Than Just a Chatbot

Have you ever had this experience? Staring at an Excel sheet full of complex formulas, completely oblivious to what the previous colleague was calculating, or getting burned out by a #REF! error.

Anthropic has just released a preview of Claude in Excel, which is simply a lifesaver for many people who work with data all day.

This isn’t the kind of chat window that hangs on the side and can only answer general questions. Claude can now understand your “entire workbook.” It understands those nested formulas and also understands the data dependencies between different sheets.

What does this mean?

When you ask it “How is the number in this cell calculated?” it won’t give you a vague nonsense answer, but will provide a reference explanation precise to the cell. You can ask it to update certain assumptions, and Claude will help you adjust the model while preserving the original formula structure.

For financial analysts or data specialists, this feature is very powerful. It can even help you debug. Those breakdown-inducing #VALUE! or circular reference errors, Claude can find the source and teach you how to fix them in seconds.

Currently, this feature is open for Beta testing for Claude Pro, Max, Team, and Enterprise users. If you are a heavy Excel user, check out Claude in Excel; this might be the most worthwhile plugin you install this year.

OpenAI’s Engineering Miracle: Supporting 800 Million Users with PostgreSQL

After talking about frontend applications, let’s look at backend hardcore engineering.

Have you ever thought about how the database behind ChatGPT works? To support 800 million users globally and handle massive requests from ChatGPT to the API, you might think they used some mysterious black-tech database.

As it turns out, OpenAI tells us: they use PostgreSQL.

In OpenAI’s latest engineering blog, they detailed this road to scaling. It really was a fight against traffic. Their Postgres load has grown more than 10 times in the past year.

To withstand this pressure, they didn’t rush to shard everything. Instead, they squeezed the performance of a single primary node to the extreme, paired with nearly 50 Read Replicas.

Sounds crazy, right? How can a single write node hold up?

OpenAI’s strategy is very smart. They found that although the read volume is huge, the write volume is relatively controllable. So they did a lot of query optimization, such as catching those complex Join queries that are not only slow but also eat up CPU, forcing logic to be moved to the application layer.

There is also a key technology called “Connection Pooling.” They use PgBouncer to manage database connections, compressing the average connection time from 50 milliseconds to 5 milliseconds. It’s like setting up an ETC before a highway toll booth; the traffic flow instantly becomes smooth.

Of course, they didn’t completely avoid sharding. For workloads with particularly high write volumes, they did migrate to Azure CosmosDB. But this case tells us that the potential of traditional relational databases is far greater than we imagine, as long as your engineering architecture is solid enough.

Unveiling the Agent: How Does Codex Think?

Since we’re talking about OpenAI’s engineering prowess, they recently started a new technical analysis series specifically explaining the working principles of Codex.

The first article, Unrolling the Codex agent loop, is actually very interesting. It explains how the so-called “AI Agent” executes tasks on your computer.

We often say AI can help you write code and modify files, but what happens in between?

The core lies in a “Loop.” When you give Codex a command, such as “Help me add this chart to the README,” the model doesn’t do it all at once. It enters a reasoning phase to decide whether it needs to call a tool (like executing a Shell command to read a file).

If necessary, it will execute the tool, and then feed the result of the tool execution (such as file content or error messages) back to the model. This process of “Observation -> Thought -> Action” repeats continuously until the model thinks the task is complete, and finally gives you a response.

HeartMuLa: The All-Rounder in Open Source Music Generation

Finally, let’s switch from serious code to the creative field.

There are music generation giants like Suno and Udio on the market now, but they are all closed source. What if you want to control the model yourself or make finer adjustments?

This is when you need to pay attention to HeartMuLa. This is a newly open-sourced music foundation model family with big ambitions, attempting to build a complete ecosystem in music understanding and generation.

HeartMuLa is not just about generating music; it consists of four core parts:

1. HeartCLAP: Responsible for understanding your text description of music, establishing the connection between text and audio.
2. HeartCodec: A high-fidelity music decoder capable of capturing music details.
3. HeartTranscriptor: A model specifically used for recognizing lyrics.
4. HeartMuLa: Finally, this is the brain responsible for generation.

What interests me most is its “controllability.” You can issue style commands separately for different sections of the song (such as Intro, Verse, Chorus), which is much more practical than just typing a paragraph of text and leaving it to fate. It even supports multiple languages, including English, Chinese, Japanese, Korean, and Spanish.

For creators or researchers, this is a great starting point. You can go to their GitHub to download model weights and set it up yourself to play around.

HeartMuLa Huggingface space HeartMuLa Github Paper

FAQ

Q: Is Claude in Excel safe? Will my data be used for training? A: This depends on your subscription plan. Claude in Excel is launched for paid plans like Enterprise and Team. According to Anthropic’s policy, Enterprise data is generally not used to train models and enjoys higher privacy protection. However, it is recommended to read your company contract or Anthropic’s privacy policy before use.

Q: Why didn’t OpenAI just shard the database from the beginning? A: While sharding can expand write capacity, it brings huge maintenance complexity. OpenAI mentioned in the article that sharding requires modifying hundreds of application endpoints, which could take months or even years. For read-heavy workloads, optimizing a single primary node with Read Replicas is a more cost-effective and stable short-term strategy.

Q: Can HeartMuLa run on a general home computer? A: Yes. HeartMuLa provides a 3B parameter version. If your video memory is limited, HeartMuLa officially provides a --lazy_load feature, which allows modules to be “loaded on demand” and releases memory after use, allowing users with a single GPU to run it more smoothly.

Q: What is the AI Agent’s “Loop”? And how is memory managed? A: The loop is the process of “Observation -> Thought -> Action.” To manage memory, OpenAI uses two strategies: one is Prompt Caching, speeding up calculations by caching invariant instructions; the other is Compaction, where old conversation history is automatically compressed into a summary when the conversation is too long, to free up context window space.