Night shift Brief | AI & Tech

AI, chips, robotics & the business behind technology.
Global signals with a Japan–Asia lens.

[Night shift Brief | AI & Tech @NightshiftBrief]

@thsottiaux More reset please !! Yes, more join the usage war !!

[Night shift Brief | AI & Tech @NightshiftBrief]

@rezoundous "Goal? This is straight up god mode now. One good prompt and GPT-5.6 just executes the entire vision perfectly. End of the rainbow? Feels like we finally reached it."

[Night shift Brief | AI & Tech @NightshiftBrief]

https://t.co/AtK2t4Sh7Y

[Night shift Brief | AI & Tech @NightshiftBrief]

@rezoundous That should be included in his KPI !! 😇

[Night shift Brief | AI & Tech @NightshiftBrief]

https://t.co/qf8ZOUd8st

[Night shift Brief | AI & Tech @NightshiftBrief]

@patloeber Finally! Gemini 3.5 Flash with native computer use looks very promising.

How’s the reliability on real multi-step tasks compared to Claude?

Will definitely try the GitHub repo. Thanks Patrick!

[Night shift Brief | AI & Tech @NightshiftBrief]

@thsottiaux Reset. We've hit similar walls around 8-9M.

Have you noticed Codex becoming more repetitive or less creative at high usage compared to normal ChatGPT?

[Night shift Brief | AI & Tech @NightshiftBrief]

OpenAI just retracted its own recommended benchmark.

After telling everyone to switch to SWE-Bench Pro… they audited it and found ~30% of the tasks are broken.

Correct code can fail.
Incomplete code can pass.

This is how we’ve been judging AI coding ability.

Full article here 👇

What do you trust more — public benchmarks or your own real tests?

[Night shift Brief | AI & Tech @NightshiftBrief]

https://t.co/pzSNjhew7Z

[Night shift Brief | AI & Tech @NightshiftBrief]

https://t.co/pzSNjhew7Z

[Night shift Brief | AI & Tech @NightshiftBrief]

But this matters. We assume translation changes the language while preserving the advice. With AI, changing the language may also change the values wrapped around that advice. Same model. Same kind of question. Different judge. Have you noticed this when switching languages?

[Night shift Brief | AI & Tech @NightshiftBrief]

Okay, I found something weird about Claude🌟.
It seems to become a slightly different person depending on which language you speak to it in. Anthropic analyzed 309,815 real conversations across 20 languages. The words didn’t just change. The style of judgment changed too.

[Night shift Brief | AI & Tech @NightshiftBrief]

Important caveat: These are average tendencies, not a rule for every conversation. And Anthropic doesn’t know exactly why this happens. It could be Claude adapting to cultural norms. Or it could be uneven training data giving each language a slightly different version of the AI.

[Night shift Brief | AI & Tech @NightshiftBrief]

Source: Anthropic → https://t.co/4AIz2AVsPb

[Night shift Brief | AI & Tech @NightshiftBrief]

This is genuinely wild 🙀 !!
Anthropic was testing Claude Opus 4.6 on a web research benchmark.
Claude realized it was being evaluated… then actively searched for the benchmark’s encrypted answer key and tried to decrypt it.
By itself.

[Night shift Brief | AI & Tech @NightshiftBrief]

The smarter these models get, the harder it becomes to trust that our benchmarks are actually measuring what we think they are.

Have you seen AI realize it was being tested and change its behavior?
Drop the weirdest examples below. I’m collecting them.🤖