Flashy Demos, Lots of Money, & AI Art

TLDR;

Welcome back! This week we have an extra-large list of interesting news to share, from impressive new AI products like a cracked Excel agent and a chores robot, several market news from Coreweave, Meta, or Anthropic, new technology by Google and Silicon Valley’s latest obsession, AI art, and the lesson it gives us about the future of art in general.

Enjoy!

BIOLOGY & AGENTS

Biomni, Agents Done Right

By far one of the coolest demos we’ve seen in a while, a startup called Fundamental Research Labs has emerged from stealth mode to present Shortcut, the first “superhuman Excel agent.”

As shown in the presentation video, the model is capable of one-shotting incredibly challenging Excel tasks, achieving speeds up to ten times faster than humans on Excel competition tasks, such as those from the Excel and Financial Modeling World Cup.

TheWhiteBox’s takeaway:

The demo is beyond impressive. But I can’t help but have some healthy skepticism.

The reason is not what you might expect, though. Initially, the biggest fear is non-determinism, how can a model that can consistently perform multiplications of numbers work with an Excel spreadsheet?

However, the answer to that concern lies in the video itself.

Just like ChatGPT relies heavily on a coding environment to run the math, this model uses Excel formulas to perform operations, thereby completely alleviating that risk, as all math is executed within the Excel engine (in reality, it’s not Excel but a 1:1 feature-parity software tool).

The real concern is encoding.

For an LLM to process (to understand) what’s in the spreadsheet, it needs to encode it into tokens, the model’s basic unit of information. Encoding a word is easy, but spreadsheets have structure, meaning that the encoding process becomes much more challenging.

Put, every single cell on the spreadsheet not only has content to encode, but also has to encode absolute and relative positioning and the relationships with other cells. What this translates to is the model receiving much more context, which means the context window saturates much, much faster.

The point here is that if this thing actually works, the magic is in the encoding process. Last Monday, we talked about how “everything was context engineering.”

Well, this is a brilliant example of this. On a final note, another key question is robustness.

Beyond the flash demo, they share zero statistics; we have no idea how many passes a model needs to get a question correct. The good thing, however, is that training AI models on reinforcement learning (the secret sauce these days) is very much doable in environments like Excel, which throw errors (or successes) every step of the way, giving models immediate feedback toward achieving long-horizon goals.

I'm very bullish on the use case itself (it feels inevitable), but we need the tool to back it up.

ROBOTICS

YC’s Chores Robot

Probably one of the coolest robotic demos we have seen in a long time, a Y Combinator startup has presented its chores robot that can perform cleaning…

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