😺 AI's Moore's Law is INSANE

PLUS: A lab-grown, living "neuron" computer?!

Welcome, humans.

New fear for authors everywhere, UNLOCKED: AI companies no longer need access to giant databases of your content—now they can use robots to ingest the data manually!

The top Reddit comment had us dying—and now we can’t unsee it!

We’re kinda joking—it’s not like AI companies need anymore author data. They have plenty. And they’re getting sued everywhere from California to France because of it.

Speaking of: The Atlantic just published a tool (scroll down past the editor’s note to use it) for authors to search LibGen database for their published works.

It’s fascinating to rummage through—here’s just a handful of the books we found:

Harry Potter? Check. Game of Thrones? Check. Guy on Fire 130 Recipes for Adventures in Outdoor Cooking. Check! In fact, this database has at least six different Guy Fieri books in it.

TIL: Beloved American hero Guy Fieri not only has six different Food Network shows, but also six different books!

Here’s what you need to know about AI today:

We dive into research into AI’s ability to do long tasks.
Microsoft developed tech to pay people for AI training data.
Lab-grown neuron chip created for AI applications.
Meta's interface decoded text from brain activity during typing.

Advertise in The Neuron here.

Researchers say AI can complete 1-hour tasks today—and maybe month-long projects by 2029

In the computer chip industry, there’s something called Moore's Law that predicted as far back as 1975 that computing power would double every 2 years (as costs dropped).

That chip-doubling trick is what eventually turned room-sized computers into smartphones in your pocket. And it sorta became a self-fulfilling prophecy.

Well, now we've got an AI version—and if new research to be believed, it's moving WAY faster.

Researchers from METR reveals a shocking trend in AI capabilities: the length of tasks AI can complete is doubling every 7 months (depending on the task, anyway).

Here’s what happened:

The study measured how long various AI models can work on complex tasks before failing.
They used how long it would take a human to complete the task as the benchmark.
They tested everything from GPT-2 (2019) to Claude 3.7 Sonnet (2025) on software engineering, cybersecurity, and reasoning tasks.

What they discovered is fascinating: today's best models can handle tasks that take a skilled human about an hour to complete

Think of it like this—Imagine tracking an athlete's progress. You wouldn't just measure how high they can jump, but how long they can run before tiring out.

This study measures AI “endurance”—how long it can maintain focus before failing—which is different from most benchmarks that just test if AI is getting smarter at quick, isolated tasks.

So this “Moore's Law for AI agents” = every 7 months, AI's task endurance doubles.

And if this pattern continues, we're looking at:

2026: AI completes full 8-hour workday tasks.
2027: 2-3 day projects.
2028: Full 40-hour workweek capabilities.
2029: Month-long projects.
2031: Full 50-week work year.

Here’s where things get really wild: When tested on SWE-bench Verified (real software engineering tasks), they observed an even faster doubling time of under 3 months.

If you only look at 2024-2025 models, the trend accelerates even more, which would bring all those dates we just listed forward by about 2.5 years.

Now, not everyone is convinced. Researcher Tamay Besiroglu found that similar analysis on chess stretched timelines into decades, showing how domain-specific these measurements can be.

However, both Besiroglu and METR researcher Megan Kinniment agree that while we shouldn't claim “AI can do any hour-long task,” the exponential trend itself likely holds within similar domains.

Others question whether a 50% success rate is meaningful for real-world applications, while some argued we'll soon hit diminishing returns that require more compute for smaller gains.

Why this matters: For white-collar workers, this data points to a clear future—the robots aren't coming for your job tomorrow, but they're rapidly getting better at the kinds of extended tasks that make up most professional work. You’ve been warned.

FROM OUR PARTNERS

Apple’s Starlink Update Sparks Huge Earning Opportunity

Apple just secretly added Starlink satellite support to iPhones through iOS 18.3.

One of the biggest potential winners? Mode Mobile.

Mode’s EarnPhone already reaches +45M users that have earned over $325M, and that’s before global satellite coverage. With SpaceX eliminating "dead zones" worldwide, Mode's earning technology can now reach billions more in unbanked and rural populations worldwide.

Their global expansion is perfectly timed, and you still have a chance to invest in their pre-IPO offering at just $0.26/share.

Mode’s recent 32,481% revenue growth and their newly reserved Nasdaq ticker $MODE puts them one step closer to a potential IPO.

Invest in Mode’s pre-IPO offering at $0.26/share and earn up to 100% bonus shares.

Disclosures

Mode Mobile recently received their ticker reservation with Nasdaq ($MODE), indicating an intent to IPO in the next 24 months. An intent to IPO is no guarantee that an actual IPO will occur.

The Deloitte rankings are based on submitted applications and public company database research, with winners selected based on their fiscal-year revenue growth percentage over a three-year period.

A minimum investment of $1,950 is required to receive bonus shares. 100% bonus shares are offered on investments of $9,950+.

Please read the offering circular at invest.modemobile.com. This is a paid advertisement for Mode Mobile’s Regulation A Offering.

Prompt Tip of the Day

We had Claude reverse engineer its own web search system prompt to create a similar prompt template. Here’s the template—try it out!

Treats To Try.

FLORA connects all your creative AI tools in one canvas, including Google Gemini 2.0's advanced image editing—with FLORA, you can create multiple images and even turn them into videos (demo).
Gamma makes your presentations for you, and it’s definitely the most popular tool for this job—check this demo we made where all we said was “make a presentation about The Neuron newsletter.”
Supametas transforms your messy content from websites, documents, PDFs, and media into perfectly organized datasets for AI applications.
Granola is one of the most popular AI notetakers; VCs loved using it for themselves so much, they invested in it.
SynCity generates explorable 3D worlds by creating tile-by-tile images that convert to 3D models—try it here (code, paper).
PromptLayer helps you visually edit, test, and deploy AI prompts with your entire team.
SecondMe lets you train and self-host your own AI that thinks like you while protecting your privacy (Mac only rn)—if it’s too technical, try the O.G. MeBot!
Tails & Frames turns your pet videos into viral content by helping you find the best moments, add trending memes, and share with one click—new launch, you get the first 3 months free; also, they have a pet camera (video).

See our top 51 AI Tools for Business here!

Around the Horn.

Perplexity outlined a plan to rebuild TikTok in America with transparent algorithms, promising to make the “For You” feed open source, add citation features, and combine its answer engine with TikTok's video library.
Microsoft is developing tech to credit and pay people whose data trains AI models, prioritizing “data dignity”—ShutterStock and Adobe currently pay contributors, while marketplaces like Bria, Defined, and Trainspot let creators license their content directly.
Tony Gilroy, the show-runner of Star Wars show Andor, won’t publish Andor’s scripts anymore because he doesn’t want them to be scraped as training data.
Here’s a good recap of the biggest news out of NVIDIA GTC this week (more).
Cool sciency stuff:
- Biological Black Box revealed a lab-grown, living neuron computer chip for AI applications called Bionode.
- There’s a new weather AI that generates forecast in seconds with much less compute (paper).
- Meta’s new brain-to-text interface, Brain2Qwerty, literally reads your mind—it decodes sentences from brain activity while you type, achieving up to 81% accuracy using non-invasive MEG technology.

Under the Hood

LG released South Korea’s first “reasoning model, called EXAONE Deep-32B (GitHub)—it excels at complex math, decoding programming challenges, and tasks in both Korean and English (supposedly it outperform QwQ-32B, too).
SmolDocling converts your document images to structured text with all original formatting intact.
CloudFlare launched AI Labyrinth, a tool that traps unauthorized AI crawler bots by redirecting them to convincing but irrelevant AI-content, wasting their computational resources while protecting data and identifying malicious actors.
- Related: AI crawlers are overwhelming FOSS (Free and Open Source Software) infrastructure, so websites are fighting back with tools like Anubis that block AI crawlers.
Here’s a list of all the AI dev tools on the market right now—read more.