Oliver

Kimi just found a way to make AI models better at reasoning by changing how their layers talk to each other.

A language model processes your prompt one layer at a time. Normally, each layer adds its work to a running total. By the time the model reaches the deeper layers, a useful detail from earlier can get buried.

Kimi’s fix is called Attention Residuals. It lets each layer look back and decide which earlier outputs deserve more weight for the prompt it’s working on. Grouping layers into blocks keeps the memory cost down.

Here’s what changed in the Kimi Linear tests.
- Science reasoning rose from 36.9 to 44.4
- Math rose from 53.5 to 57.1
- Coding rose from 59.1 to 62.2.

In a separate experiment, the standard design needed 25% more training compute to match the new design’s prediction error.

🔥 The Three.js open-source library with mythical quality!

It’s called ThreeUI. Browse 160+ free components and website examples featuring 3D visuals, icons and animations.

Copy the prompt or source code into your AI coding agent

Ask it to adjust the theme, lighting, motion or layout

Browse ThreeUI: threeui.com/browse

SOMEONE CREATED A GITHUB REPO WITH AN ENTIRE SETUP FOR AN AI AGENCY

It’s called Agency Agents. The repo gives you ready-made AI roles for engineering, design, growth marketing and product management.

Each role explains what the agent does and how it approaches the work, so you can start with the role that fits your task.

Researchers just proved that every single elementary function, sin, exp, log, sqrt, comes from one single binary operator.

The idea draws a parallel with NAND gates, which can be combined to build any Boolean logic circuit.

This paper explores a similar building block for elementary mathematics:

eml(x,y) = exp(x) - ln(y).

Starting with this operator and the number 1, the researchers describe how to construct arithmetic, square roots, trigonometric functions and constants such as pi.

The same operation appears repeatedly in a binary tree, with each result feeding into another calculation.

The operator was found through a systematic, exhaustive search.

The scope matters: the claim concerns elementary functions, rather than every possible mathematical function.

For AI, this suggests a possible way to simplify the building blocks used in mathematical expressions.

A system could work with repeated instances of one operator instead of choosing among many different operations.

That possibility does not establish that such a system would learn faster or discover scientific laws more effectively.

The striking idea is that a broad mathematical toolbox can be expressed through one repeated operation.

farmers are using AI laser machines to burn thousands of weeds a minute without spraying herbicides

they’re made by a company called Carbon Robotics, and the machine is called LaserWeeder

the problem is weeds growing right beside the crops farmers want to keep

spraying them means using herbicides, while pulling them by hand takes a lot of work

but a laser needs to know exactly which plant to hit, or it could burn the crop too

so instead, farmers pull a machine behind a tractor that spots and targets weeds one at a time:
cameras watch the plants passing underneath

AI tells the weeds apart from the crops

lasers burn each weed at its growing point

the crop stays standing and the soil stays undisturbed

it works day or night, and Carbon Robotics says its original tractor-mounted model targets more than 5,000 weeds per minute.

your lettuce now has an AI laser bodyguard before it even reaches your plate