By now you’ve already heard of GPT-6 Astra, the most intelligent model by OpenAI yet that’s supposed to be the generation of intelligence. When it launched, Nvidia’s CEO Jensen Huang posted that ‘AGI’ had arrived and used it to argue Nvidia needs another 400,000 GPUs online, while OpenAI’s own president Greg Brockman said it was “not unreasonable to feel” we’re in the AGI era.
Around the same time, I’ve been seeing an incredible amount of hype and praise around the model, with many claiming it represents a major leap forward in AI capabilities.
So what exactly is AGI?
AGI stands for artificial general intelligence, or in other words, AI that can do basically anything a human can do, not just one narrow lane.
Everything most of us use day to day, including the tools in our own stack, is narrow AI: brilliant at one thing, not so outside it. AGI is the idea of something that isn’t boxed in like that, that can pick up a new problem the way a person would rather than only the problems it was trained on.
“AGI” gets thrown around like it’s a finish line, but there are at least three different interpretations that tells you something different about which part of your job to actually worry about.
According to OpenAI’s Sam Altman, their working definition of AGI is a system that outperforms humans at most economically valuable work: not necessarily smarter, just more productive. That sounds a bit like it doesn’t care whether the model understands anything, only whether it gets the job done cheaper than you can. Under this bar, what’s exposed is whatever’s currently valuable mainly because a person has to sit down and execute it: campaign setup, reporting, first drafts, segmentation, the CRM admin nobody enjoys.
Google DeepMind’s former CEO Demis Hassabis wants a system that matches or beats human performance across the board, with genuine understanding, tested against something closer to real scientific discovery — and he’s been explicit that today’s models aren’t there yet. Genuine understanding is a much higher bar than anything currently shipping, which buys more time for the work that depends on reading a market or an audience correctly: positioning, judgment, knowing what your customers actually want.
Then there’s Yann LeCun, who left Meta late last year to start AMI Labs focusing on what he calls “world models” rather than LLMs. His position is that predicting text can’t get you to general intelligence no matter how much you scale it — there’s no grounding in cause and effect, no model of how the physical world actually works.
If a system genuinely can’t ground itself in real, physical, causal experience, then anything built on presence, running an event, reading a room, trust built face to face, stays out of reach no matter how good the benchmarks look on a slide.
Before you read too much into any of this
By all three of these standards, nobody’s there yet. Not by Altman’s bar — today’s models are genuinely good at specific tasks, but nobody’s shown one outperforming humans at most economically valuable work, across the board, reliably. Not by Hassabis’s — he’s said so himself, plainly, about his own field’s best models. And by LeCun’s, arguably not ever, not with the current LLM-led architecture.
That’s not a reason to stop paying attention, though. “AGI has arrived” is going to get said again, by someone, at the next launch, and the next one after that. There’s no doubt that AI technology can and will rapidly change the nature of marketing work.
The right move, I’m thinking, is to audit our own function every quarter or so: which parts of your job were “needs a person” a year ago and are now “needs a person to check the work.” That’s the bar being cleared in real time, rather than three years away.
Tavus renders an AI human that passes the “are you even listening” test. Tavus launched Phoenix 4.5, which it’s calling its most realistic human-rendering model yet. It generates video at 134ms and uses nonverbal cues — micro-movements across the upper body, subtle expression shifts — specifically to signal active listening. If you’re running any kind of AI video, demo hosts, virtual reps, support avatars, the uncanny valley just got smaller.
Google puts a full song generator inside Gemini. Lyria 3.5, Google’s most advanced music model, is now in the Gemini app, AI Studio, and the API. It generates full songs up to three minutes, with stronger melodies and more detailed arrangements than the version that shipped in July. It’s worth testing on low-stakes work before it touches anything with your name on it.
AI starts doing research on AI, not just marketing copy for us. Meta’s autonomous research system, AIRA 3, won a gold medal by placing 8th out of roughly 4,000 teams in an Nvidia-run Kaggle competition to fine-tune a 30-billion-parameter model — Meta says it shows the system can improve AI models at a level comparable to a strong human ML engineer.


