→ Hand it to Hark

Good morning. It’s Friday, October 30, and today’s issue looks a little different.

We’re kicking things off with why the old prompt engineering playbook is dead (and what’s replacing it).

Then, the day’s biggest stories and a Friday Fact about a mysterious discovery in Africa that revealed something extraordinary from two billion years ago.

We’re trying something new! Let us know what you think of the format at the bottom of this issue.

The Old Prompt Engineering Playbook Has Faded

Remember when getting a great answer from ChatGPT meant knowing the right magic words?

You had to tell the model to act like an expert, think through a problem step by step and follow a carefully constructed set of instructions. Entire guides were devoted to prompt formulas, elaborate templates and tricks for squeezing better performance out of language models.

We even published our own guide to role-based prompting in March 2025. The advice reflected a widespread belief: assigning the model an expert persona could improve the quality of its answers.

Fast-forward to today, and Forward Future producer Alex McCraw doesn't bother with that technique anymore.

“I used to use role-based prompts like ‘You're an expert graphic designer,’ with the idea that assigning the AI an expert role would make it perform better,” Alex says. “I don't really bother with that anymore.”

He's not alone.

After some internal debate arose, we polled our team to see whether prompt engineering is still essential. For such an opinionated group, our responses were surprisingly aligned.

Forward Future founder Matthew Berman, the person behind our AI model reviews, believes traditional prompt engineering is essentially finished.

“Prompt engineering, which I define as extracting the best performance from a model with techniques (old) such as ‘think through this step by step’ is basically useless now,” he says.

Our editor Brian Knepprath goes even further: “Over-specifying or spending too long on a prompt is now actively a waste of time.”

Forward Future co-founder Nick Wentz agrees that the old techniques are losing some relevance, but he thinks something else is happening at the same time: The bar for great prompting is getting higher.

“The main shift is from telling AI how to do something to communicating what you want, why it matters and what good looks like,” he says.

In Nick's view, the best AI users aren't necessarily the ones with the most sophisticated prompts. They're the ones who understand how to frame a problem, provide relevant context and recognize a good result when they see one.

That's a remarkable consensus from people who spend their working lives experimenting with AI. But there's a catch.

→ Read the Forward Future Original.

Finding a dentist? Comparing flights? Sorting your email? Hark Pro remembers your preferences and helps handle the little jobs taking up too much of your day.

→ Hand it to Hark

U.S. intelligence investigated whether Alice Huang, an executive at Singapore AI cloud company Megaspeed, was working for China while seeking access to NVIDIA CEO Jensen Huang. NVIDIA said she received no trade secrets or national-security information, leaving the case focused on how Beijing tries to learn about advanced-chip plans and export controls.

The Justice Department is examining NVIDIA’s planned $12.93 billion purchase of Hugging Face, a widely used platform for sharing AI models and datasets. Regulators are also asking about NVIDIA’s investments and pressure on customers as they assess whether the chip leader could limit competition across the AI market.

DeepSeek and Moonshot AI are reportedly using outputs from GPT-6 Astra to train new models after Anthropic tightened defenses against this practice, known as distillation. OpenAI said it found a coordinated Moonshot-linked campaign in September, making model-output theft the next front in the U.S.-China AI race.

Alphabet’s quarterly revenue rose 25% to $127.6 billion but missed Wall Street’s estimate, while Google Search revenue grew 12% to $63.35 billion. Search use is rising as Google adds AI answers, but the company still has to prove those longer, more complex queries can produce ad revenue as efficiently as traditional search.

In 1972, scientists analyzing uranium from a mine in Gabon, Africa, noticed something strange. A small but significant amount of uranium-235 was missing. When they investigated, they uncovered something extraordinary: Earth had built its own nuclear reactor nearly two billion years earlier.

At a site called Oklo, rich uranium deposits and groundwater had combined to trigger self-sustaining nuclear fission, the same fundamental process used in modern nuclear power plants.

Even more remarkably, the reactor regulated itself. As nuclear reactions heated the water, it boiled away, shutting down the reaction. Once the rocks cooled and groundwater returned, the process restarted. Evidence suggests one reactor cycled on and off roughly every three hours.

These natural reactors operated for thousands of years, without human intervention, machinery, or engineering.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team