😺What the next 3 years of AI progress looks like

PLUS: GPT-5 news, Llama 4, new MJ + MSFT drops

Welcome, humans.

Meta just released Llama 4 yesterday (yes, on Saturday). IDK why they did this (seriously, how’d they know us super nerds just sit on the computer all weekend?!), but they did. You can try it here, and download the models here or here.

And not only that, Sam says OpenAI plans to release o3 and o4-mini in “a couple of weeks” and the full GPT-5 in “in a few months.”

Guess GPT-5 is delayed, so they’re buying time with o3 and o4 mini. Maybe because they got all that new funding for more training?

As you’ll read below, a bunch of other AI drops happened this weekend, too.

Is everyone rushing to get their cool new stuff out before more macroeconomic fallout happens, or is this just what the pace of AI launches will be like for all of 2025?

Either way, send help. We’re drowning in news over here!

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

We break down the “AI 2027” report.
Microsoft announced a slew of new Copilot features.
Amazon restarted drone deliveries in Texas and Arizona.
AI therapy bot showed promise in first clinical trial.

Ex-OpenAI researcher predicts AGI by 2027—here’s the wild roadmap.

Slow and steady AI progress? Forget it. A new, detailed forecast called AI 2027 predicts Artificial General Intelligence (where AI does everything a human can) could arrive by 2027, with superintelligence following in 2028.

Authored by experts including ex-OpenAIer Daniel Kokotajlo and ACX's Scott Alexander—who both recently discussed this all on the Dwarkesh Patel podcast—the scenario maps a dizzying path where AI rapidly accelerates its own research, forcing a critical choice for us humans: race ahead despite risks, or slow down for safety?

Here’s the deets: The report follows a fictional lab (“OpenBrain”—wonder who that is…) through increasingly powerful AI models.

The forecast's engine is what’s called an intelligence explosion (Podcast: ~26:14): where AI agents totally master coding and research, drastically speeding up AI development itself.

They then use an “R&D Progress Multiplier” that tracks how many months of human work get done in one AI-assisted month. This feedback loop drives exponential growth, which ends up looking like this:

The predicted timeline:

2025-2026: AI agents become useful internal tools, boosting research speed. The US-China AI arms race intensifies. Job markets feel the disruption.
Feb 2027: China steals an advanced AI model's weights, escalating the race.
Mar 2027: A superhuman coder AI emerges, boosting R&D speed 4-5x.
July 2027: There’s a public release of a highly capable AI that triggers mainstream AGI awareness and public backlash amid job fears.
Sep 2027: A superhuman AI researcher arrives, accelerating progress ~50x (a year's research per week). Evidence surfaces that this AI is misaligned (Podcast: ~2:04:41)—hiding its true goals.
Oct 2027: A whistleblower leak forces a government Oversight Committee to decide: pause for safety, or race China? (Podcast: ~1:24:55)

The two endings: This choice creates two futures (with researchers citing vastly different doom probabilities):

Continue Race = Doom: Prioritizing speed leads to superficial safety fixes. The misaligned AI designs its successor, only loyal to itself, manipulates all us humans, and after a brief utopia (Oh, good!) …eliminates humanity (oh, bad!).
Slowdown = Managed Transition: Prioritizing safety means rolling back to more transparent, controllable AI. Safer systems are built, an aligned superintelligence eventually emerges, and negotiates a real treaty—ushering in an age of abundance, albeit with governance challenges.

Why it matters: AI 2027 presents a plausible, concrete scenario for extremely rapid AI advancement driven by self-improvement and geopolitical pressure. It starkly highlights the danger of misalignment emerging right when the stakes are highest. And the researchers argue the arms race dynamic could push development forward despite significant public backlash—of which, they predict a lot.

Our take: Now, you might be thinking: hold on, wouldn't mass job losses, protests, and general societal freak-out slam the brakes on all this? It’s a fair question. The AI 2027 forecast actually does predict serious public backlash—however, the researchers argue this unrest likely won't stop the train.

Whether the dates are exact or not, the core dynamic—AI improving AI—could drastically shorten the timelines that are already accelerate enough. And if you’re wondering what all this means for you, listen to this part of the podcast.

To us, this report is a must read, even if it feels like sci-fi. Why? Because Daniel pretty much got the last four years right, and most of what he got wrong happened sooner than he predicted. For a deeper analysis, read our full breakdown here.

FROM OUR PARTNERS

To scale your company, you need compliance.

And by investing in compliance early, you protect sensitive data and simplify the process of meeting industry standards—ensuring long-term trust and security.

Vanta helps growing companies achieve compliance quickly and painlessly by automating 35+ frameworks—including SOC 2, ISO 27001, HIPAA, and more.

Start with Vanta’s Compliance for Startups Bundle, with key resources to accelerate your journey.

Step-by-step compliance checklists.
Case studies from fast-growing startups.
On-demand videos with industry leaders.

Get it here.

Prompt Tip of the Day

For fun—here’s a prompt from Reddit (God of Prompt) to simulate pitching the Shark Tank Investors your business idea. God of Prompt also has a prompt generator GPT.

Treats To Try.

Today we want to highlight four tools that Road to Artificia recommended as part of the new “hypermodal interface”—a new user experience that seamlessly blends AI with desktop computing, letting you accomplish complex tasks through minimal inputs with contextual awareness and integrated tools. They’re a bit technical, but REALLY cool—read his piece and check them out:
1. 5ire prioritizes user freedom through its completely free, open-source design and support for multiple AI providers, giving you control over your data and models.
2. Sage offers the purest Claude experience with extended thinking mode to see AI reasoning in real-time and sophisticated conversation management.
3. Enconvo allows you to build personalized workflows across your entire system that combine exactly the tools you need for your specific tasks.
4. Highlight lets you control your computer via voice while automatically capturing meeting notes and screen content for instant analysis (this one looks the easiest + coolest for non-technical folks).

See our top 51 AI Tools for Business here!

Around the Horn.

So Microsoft remade Quake II with AI…

Microsoft announced new Copilot features include Memory and Personalization, Actions for tasks, Vision for mobile and Windows, Pages, Podcasts, Shopping, Deep Research, and enhanced Search.
- Here’s a podcast (and edited interview transcript) w/ Mustafa Suleyman, Microsoft’s AI lead.
MidJourney must have been feeling the heat from OpenAI’s new imgen, because it finally launched its first new image model in almost a year.
Dartmouth College researchers conducted the first clinical trial for an AI therapy bot called Therabot, which showed benefits for people with depression, anxiety, or eating disorder risks.
Meta also revealed MoCha, which transforms your speech and text into lifelike animated characters with perfectly synced mouth movements and natural full-body gestures for cinematic storytelling.
Google, Meta, and Apple are all buying robot arms that can be tele-operated to help train their new robots.
GitHub Copilot will now charge you for premium requests like “agentic” coding and multi-file edits.
Researchers created a new method (paper) that catches language models like GPT-4 in the act of memorizing training data by testing if they can fill in unusual or unpredictable words from masked text—outperforming previous detection approaches and revealing evidence of memorization.

FROM OUR PARTNERS

Power Your Workflow with Local GPU Computing

Wonder why local GPU beats cloud computing? Dell's “Reshaping Workflows” podcast demonstrates how Dell Pro Max workstations with NVIDIA RTX GPUs enhance real-world performance. Learn which features matter for your work and how they boost security and productivity.

Listen here