Read The Day

Published edition6 September 2026

Fresh ideas for a more helpful day.

AI is getting easier to use on your own computer, and robots are learning the small skills everyday help requires. Here are a few things worth knowing over coffee.

Published by Read The Day editorial desk

Read AIRead Robotics

Artificial intelligence · 6 September 2026

AI, understood.

4 min read

More choice on your own computer, a carefully checked maths result and a few fixes that make familiar tools easier to use.

Ollama repository preview for the local-model ChatGPT Desktop release
01AI at home

AI on your own computer, in a familiar app

An early Ollama update lets Mac users try AI running on their own computer inside ChatGPT Desktop. That brings local AI into a chat app many people already know. The update also improves how the software handles longer conversations and images.

Why it matters

Running the AI locally gives people another option for where the model does its work.

Keep in mind: This is an early test release for Macs, rather than a finished feature for every computer.

Read the source · Ollama
02Science

AI helps check a famous mathematical proof

Anthropic says its Claude assistants spent eleven days turning Fermat’s Last Theorem into a proof that checking software can inspect step by step. The team has shared the result and code. It is a new way to check an existing proof, rather than a newly discovered theorem.

Why it matters

Work that can be checked at each step offers a promising way to make AI’s contributions easier to trust.

Keep in mind: The result is Anthropic’s reported work; it does not show that AI can solve every difficult maths problem.

Read the source · Anthropic
03Everyday tools

A small fix gets an AI coding tool moving

Pi, a tool for working with AI on code, now supports GPT-6 Astra. The same update repairs a packaging mistake that stopped some developers from using the previous release. It is a modest change, but a useful one for people whose work was interrupted.

Why it matters

The fix restores a working tool and adds another model to choose from.

Keep in mind: The release does not show whether Astra performs better on particular tasks inside Pi.

Read the source · Pi
04Everyday tools

A newer AI becomes easier to choose in Codex

A small Codex update makes GPT-6 Astra visible in its model menu. It also chooses Astra by default when the user has not selected a particular model. That removes an extra step for people trying the newer option.

Why it matters

An existing session may use a different model after an update if no preference was set.

Keep in mind: This update fixes how the model is offered; it does not establish a broader improvement in performance.

Read the source · OpenAI
05Everyday tools

A missing AI option returns to the menu

OpenCode has fixed a bug that hid GPT-6 from some people using an OpenAI subscription. The software was misreading the model’s name. The update makes the option visible again without changing what those users had already paid for.

Why it matters

People affected by the bug can find an AI model their subscription already gave them access to.

Keep in mind: The fix restores the choice; it does not measure how well the model works in OpenCode.

Read the source · OpenCode
06More choice

A new AI family comes in six sizes

Researchers at MBZUAI have released six versions of an AI family called K2 Horizon. They range from smaller models to ones that need much more powerful computers. The release includes code, training material and details of how the models were made.

Why it matters

That gives other teams more room to explore AI suited to different machines and budgets.

Keep in mind: Performance claims come from the team behind the release and still need independent checks.

Read the source · MBZUAI

Robotics · 6 September 2026

Robotics, explained.

4 min read

Handling a puzzle piece, finding a charger and learning a gentler touch. Small skills can add up to more useful robot helpers.

Robocurve comparison of frontier models controlling two robot arms
01Learning by doing

A bowl is easy. A puzzle needs more practice.

In a Robocurve experiment, GPT-6 Astra guided robot arms to put a block in a bowl 19 times out of 20. Fitting a puzzle piece precisely was much harder: it succeeded twice in 20 attempts. The team shared videos and records of every run.

Why it matters

The comparison makes it easier to see which movements robots handle well and which still need work.

Keep in mind: The study used two tabletop tasks and human grading, so it gives a limited picture.

Read the source · Robocurve
02A gentler touch

A robot thinks about how hard it will push

Researchers at NTU are teaching a robot to estimate the force of a move before making it. Their system, Facet-0, is designed for fitting computer parts together, where a small shove can cause trouble. They have shared the model, training examples and study.

Why it matters

Anticipating pressure could help robots handle delicate parts with more care.

Keep in mind: The work has been tested on the team’s own equipment and still needs independent checks.

Read the source · PINE Lab, NTU
03Finding the way

A walking robot shares how it finds its way

EngineAI has shared software for helping its PM01 humanoid move through its surroundings. The release includes tools for practice in simulation and for running the software on a real robot. It also explains the setup needed to get started.

Why it matters

Other teams can study a fuller route from practising on a computer to trying the robot itself.

Keep in mind: The release does not include an independent test of how well it handles a wide range of real places.

Read the source · EngineAI
04Robot kits

A tiny six-legged robot gives AI something to move

Jizai has shared software for Palmimo, a small six-legged robot that arrives assembled. It includes examples of connecting AI assistants to movement, plus a way to practise commands without moving the real robot. That gives builders a small platform for trying physical ideas.

Why it matters

People can explore how an AI request becomes a movement without first building a robot from parts.

Keep in mind: The hardware designs remain closed, and the examples do not prove reliable independent behaviour.

Read the source · Jizai Inc.
05Everyday independence

A robot learns to take its own charging break

LimX has published a guide to the TRON 2 robot’s automatic charging dock. It describes how the robot returns, lines itself up, charges and leaves again. The company says docking takes about 30 seconds on average and can work in dark conditions.

Why it matters

Getting back to a charger is a practical step toward robots that need less hands-on help.

Keep in mind: The timing and operating claims are the manufacturer’s, rather than independent test results.

Read the source · LimX Dynamics
06Looking ahead

A big investment in teaching robots new skills

Figure has signed a deal with Nscale for more computing power to train its robots. The initial commitment is $3.5 billion, with larger plans for later. That computing power is intended to help the company train and test the software behind its humanoid robots.

Why it matters

The agreement shows how much effort companies expect learning new robot skills to take.

Keep in mind: The larger plans are for future capacity, starting from the second half of 2027, and do not prove better robots today.

Read the source · Nscale