Minimal AGI by 2028: DeepMind's Bold Prediction
Minimal AGI by 2028
DeepMind's co-founder says AGI is three years away. I am just trying to get my chatbot to stop lying about my calendar.
Shane Legg just put a 50% chance on Minimal AGI arriving by 2028. AGI stands for Artificial General Intelligence. It is the point where a machine can learn and perform any intellectual task a human can do. Legg defines Minimal AGI as an AI that can do most cognitive tasks as well as an average human worker. It is a bold claim from a man who has spent a decade studying the physics of intelligence. But there is a massive gap between what the labs see and what we actually use.
The Speed Bottleneck
Legg argues that digital intelligence has every physical advantage over us. Our neurons fire at 100 Hz, while modern processors run at 10 billion Hz. Signals in our brains move at 30 meters per second. In a chip, they move at the speed of light. And yet, we are still the bottleneck. OpenAI experts recently pointed out that human typing speed is a major drag on AGI development. It does not matter if the machine thinks a million times faster than you if you still type like a distracted toddler. We are trying to drive a Ferrari through a crowded car park.
Minimal vs. Full AGI
Legg breaks this down into tiers. Minimal AGI is not Einstein. It is just a competent person who can handle routine office work. Full AGI comes later and handles the hard stuff, like inventing new scientific theories or writing symphonies. I find this tiered approach refreshing. It moves the goalposts away from a Sci-Fi Apocalypse and toward Will this replace my accountant? We do not need a superintelligence to disrupt the economy. We just need a machine that can pass as a median worker.
My Perspective
I want to believe the 2028 timeline. But I also look at the $644 billion currently being spent on AI, where 72% of that investment is being wasted. We are scaling waste faster than we are scaling intelligence. If AGI arrives in three years, it will likely be smarter than us but still stuck waiting for us to finish our prompts. It is like giving a supercomputer to someone who only uses it for Minesweeper. We are not ready for the speed of the machine.
AI News This Week
BBVA Goes All-In on ChatGPT Enterprise
The Spanish bank is rolling out OpenAI's enterprise tool to 120,000 employees after a successful pilot. Early users reported saving three hours per week on routine tasks, with an 80% daily engagement rate. The bank aims to use the tool for everything from risk analysis to customer interaction.
The $644 Billion ROI Blind Spot
Enterprise AI spending is projected to hit massive heights in 2025, but 72% of that investment is currently failing to deliver. Only 31% of prioritized use cases have reached full production because companies lack visibility into their tech stacks. Many firms are scaling adoption numbers without tracking actual bottom-line impact.
Training AI in the Cold of Deep Space
Startups are proposing orbital data centers to solve the massive cooling challenges on Earth. Space offers passive radiative cooling at -270C and solar power that is 40% more effective than on the ground. These modular clusters could scale indefinitely without the physical constraints of terrestrial power grids.
Source: Lumen Orbit White Paper
OpenAI Board Drama and the Q Factor
Reports suggest Sam Altman's brief removal from OpenAI was linked to a breakthrough called Q-star. The board reportedly worried he was moving too fast toward AGI without proper safety guardrails. This highlights the ongoing tension between commercial speed and the long-term safety of the technology.
Tomorrow at 10am (UK) we are running another AI Cook-A-Long session.
No theory. No hype. We will build something practical, live, something you can actually use immediately in your day-to-day work.
If you are experimenting with AI but want outcomes, not demos, this one is for you.