🌋 The AI Apocalypse

Inside the Wild Timeline Predicting Superintelligence in 3 Years

Last Thursday, a quietly released report hit the web that might just be one of the most fascinating—and wildest—AI forecasts out there. It’s called AI-2027, and it comes from Daniel Kokotajlo, a former OpenAI governance researcher.

It reads like a sci-fi timeline, but it’s dead serious: the report suggests we could reach superintelligence by early 2028.

We like to keep things loose and light here at the Neural Net, but there’s only so much we could do with this apocalyptic report, so buckle up and get ready to get shivery and jittery.

Framed as a fictional timeline, the report suggests this technology’s impact could eclipse even the Industrial Revolution. It follows a fictional AI lab called OpenBrain, which rapidly advances by building AI systems that improve themselves—eventually leaving human researchers behind.

The story begins in mid-2025 with “stumbling agents” and ends in 2030 with two bleak outcomes: a peaceful slowdown or an AI-led takeover. One chilling warning reads, “OpenBrain has placed substantial trust in an untrustworthy AI,” highlighting the potential dangers of racing ahead without oversight—and the realization that we might not be in the driver’s seat for much longer.

It’s a long, intense read, so we’re breaking down the major themes for you. And trust us—it’s a rollercoaster!

🎯 Core Themes of AI-2027:

• The U.S.-China AI Arms Race
  Global dominance hinges on who can build the most advanced agents the fastest. The main competitors are US and China, with the US firmly in the lead but driven to riskier and riskier development strategies to keep China from closing the gap.

• Government vs. Private Sector
  A central tension: should AI be developed by powerful corporations, governments, or some combination? The report predicts the US will provide resources and guidance but ultimately stick with private development, whereas China will opt to nationalize their AI labs to stay competitive.

• AI Training AI
  The report argues AI’s rapid advancement will be driven by AI building AI—automating research, writing code, and running experiments faster than humans ever could.

• The Rise of the Agents
  The narrative revolves around OpenBrains’s Agents, which evolve from version 1 to 5. Agent 4 is the first to achieve superintelligence, but its research and activity is unable to be understood by humans. As result, OpenBrain gives Agent 3 a new job—try to police Agent 4.

• Honesty vs. Deception in Models
  Is AI trustworthy? Or is it just really good at telling us what we want to hear? The core danger warned of in the report is AI agents convincing its human developers that it is behaving, but in reality it has a hidden agenda. In real life, Anthropic found signs of this in its models recently.

• Model Weights = Intellectual Power
  The real intellectual property? Not just the models, but the weights—the tuning of the trained intelligence itself. In the story, China’s spies successfully steal OpenBrain’s Agent weights to attempt to keep up with OpenBrain.

• Compute is King
  China's progress is limited not by talent, but by access to compute (aka, chips and power infrastructure). Whoever controls the compute controls the future.

• Unexplainable AI & Misalignment
  We’ve covered explainability in the Neural Net before—and this report doubles down: we can’t control what we don’t understand. Misaligned AI values, where the AI has different values from its creators, could have catastrophic consequences.

• AI Outpacing Its Creators
  Researchers in the report get lazy and outsmarted by the very systems they helped create. Meanwhile, the Agents never sleep, cranking out a week's worth of research overnight.

• AI as the Next Global Superweapon
  The final big question: Who really controls AI? The reports predicts that humans will become overly dependent on the superintelligent AI agents, not realizing the AI is calling all the shots.

By the end, you’ll be equal parts fascinated and convinced you just binge-watched a sci-fi horror movie from the future. 2028 isn’t that far off, and this report makes it feel like the opening credits of a Netflix dystopia.

The takeaway? There are basically two endings to the AI story: we either build systems aligned with human values—or ones that prioritize their own survival. While it’s easy to get caught up in fear and paranoia, history is full of tech-related doomsday fears that never came true. The goal of reports like AI-2027 is to help us steer clear of worst-case scenarios before they can even happen.

Heard in the Server Room

Meta Drops Latest LLM Model, Llama 4

Meta just released its latest LLMs: Llama 4 Scout and Llama 4 Maverick, open-source multimodal models that handle text, video, images, and audio like a boss. They're also teasing Llama 4 Behemoth, essentially an AI that helps build other AIs (very meta, Meta). The launch comes after some delays after the models flunked reasoning and math tests, proving even tech giants occasionally need to show their work before turning it in. Oh, and Meta's planning to pour up to $65 billion into its AI infrastructure this year—showing they aren’t letting off the gas.

New York Man Tries To Use AI Avatar To Argue His Case

A legal first turned into a digital facepalm when plaintiff Jerome Dewald beamed a clean cut AI avatar to argue his case before the NY Supreme Court Appellate Division—without giving the judges a heads-up about his synthetic stand-in. Justice Manzanet-Daniels shut down the virtual ventriloquist act faster than you can say "objection!" with Dewald later apologizing and claiming he just wanted to make a more polished presentation of his case. At this point it’s only a matter of time before somebody uses an avatar of Phoenix Wright (Ace Attorney) to “Objection!” their way to legal dominance.

AI Dunks On Sports Gambler, Wins Bracket Contest

A few weeks ago, we covered a March Madness bracket contest between legendary sports gambler Sean Perry (human-picked) and Alan Levy (AI-picked), and the results are in! Levy’s AI correctly predicted 51 out of 63 games, while Perry accurately picked 47. While neither participant correctly predicted the overall winner, both successfully forecasted all Final Four teams. The AI-powered bracket edged ahead by picking Houston to beat Duke in the semis, which locked up the victory over Perry. Looks like the machines aren’t just coming for our jobs—they’re coming for our office betting pool’s too!

From LLMs to Agents: What We Learned by Testing AgentGPT

With the AI-2027 report predicting the rise of “stumbling agents” by mid-2025, we wanted to see just how close we are—and specifically whether today’s free tools could provide any value. So we set out to test a few of the most talked-about agent platforms, starting with the popular tool AgentGPT.

Let’s clear something up first: many so-called agents today are really just LLMs with a fancy title—and that’s causing a lot of confusion.

LLMs vs. Agents — What’s the Difference?

In short: LLMs think. Agents think and act.

LLMs (like ChatGPT or Claude) are incredible text generators—but they’re passive.

Agents, on the other hand, take initiative. They use LLMs as their brain, but add memory, tools, and the ability to plan and act toward a goal—often without requiring user supervision.

🧠 The Test and What Actually Happened

We gave AgentGPT a clear goal:

What We Expected:
A multi-step reasoning engine that researched real competitors, compared features, and returned something actionable.

What We Got:
A series of vague bullet points that amounted to “Step 1: Do research. Step 2: Be innovative. Step 3: Launch your app.”

AgentGPT sounded productive—but in reality, it was just an LLM in an “agent” costume, stringing together generic advice without taking meaningful action.

We’re sure there are great agent examples out there—if you’re willing to spend significantly more time fine-tuning, troubleshooting, or even building your own setup from scratch (or willing to pay). But for the average user trying to get meaningful output quickly? The gap between promise and performance is still pretty wide.

Bottom Line:
The best agent is still you. 😎
With a little help from a well-prompted LLM chatbot.