AI tools

Capable on basics, but more restrictive on compatible brands and weaker at smart home commands like the Ring doorbell and vacuum.

ChatGPT guessed 1,128, which he thought was too low, though it was the second closest.

Gemini gave the closest jelly bean guess and was named the winner of this quick test.

Copilot guessed 500 to 600 beans, a range well below the true count.

Claude guessed 870, which he thought seemed wrong, though at least it gave a number.

Grok guessed 412, well under the true count.

DeepSeek guessed just 128, which he found absurdly low.

High potential AI memory gadget, but rough around the edges with a pricey subscription and privacy concerns.

Impressively realistic for a few reframes, but pushed to the limit the AI-generated image falls apart in funny ways.

Impressed by how well it guesses the broad layout of his garden, but the details drift and it eventually has no idea what it is doing.

A huge jump over old Siri with excellent design and integration, but capability-wise it only catches up with Gemini and the best version is…

He finds Gemini intelligence mostly impressive for fixing everyday annoyances, but notes limited availability and some features he would no…

Sora was an AI-only video app that flopped badly, costing far more to run than it earned before shutting down.

An always-listening Gemini-powered necklace that feels creepy but could be useful; rated a two.

Giving kids an open-ended AI Vader led to slurs within hours, though with little fallout he rated it only 2 out of 10.

The supposed AI app builder was largely 700 programmers in India, and the company went bankrupt, scored 7 out of 10.

Impressed by how much more complete a Minesweeper game ChatGPT-5 produced compared with the previous model from the same prompt.

The previous ChatGPT model's Minesweeper was functional but very simple, and he called it terrible next to ChatGPT-5.

The old default model is slower and more basic at coding, but produced the better image and invitation in his tests.

The fallback model free users get after hitting the GPT-5 limit, still better than last generation.

Fast at coding, but it produced the same basic Tetris and a chess game that errored every time.

A clear step up from older models, especially for coding and writing, and simpler to use, though it still hallucinates and is no better at…

A surprise second place on 29-point-chasing form: the fastest and decent at product research and humour, but erratic on charts and voice qu…

Strong on Workspace integration, voice and Veo 3 video, but slowest, prone to inventing products and wordy in deep research, finishing thir…