What Does a Just Society Look Like After AGI?

AI is creating and concentrating wealth at an unprecedented pace. This carries enormous societal ramifications that could exacerbate existing inequities.

(The opening of our AGI Abundance conversation: Boston, May 29 and New York, June 3.)

What we mean by AGI #

The people building these systems cannot agree on what “AGI” means, and several have walked back their own definitions. We are not going to litigate the term. When we say AGI, we mean machine intelligence that keeps getting more capable. It is already striking, and given the money and talent pointed at it, we expect it to keep climbing. Nothing here depends on crossing a single threshold on some morning when everything changes. It depends only on the trend continuing, which it is.

AI is creating and concentrating wealth #

Two things are happening at once, and they compound: AI is producing economic value faster than anything we have a record of, and the ability to produce it sits with a very small number of companies. Creation and concentration, together.

Creating it. The speed is the story.

• Y Combinator’s 2025 cohort grew about 10% per week in aggregate. As CEO Garry Tan put it, “the whole batch is growing 10% week on week… That’s never happened before in early-stage venture.” (CNBC)
• Replit went from roughly $10M to $100M in annual recurring revenue in about five and a half months, and to around $253M by October 2025, up 2,352% year over year. (SaaStr, Sacra)
• Cursor went from $1M to $100M ARR in roughly twelve months, the fastest in SaaS history, and crossed about $1B ARR by November 2025. (SaaStr)
• Base44 was a roughly six-month-old startup, run by a solo founder, when Wix bought it for $80M in cash. (Calcalist, Wix)
• ChatGPT is the fastest-adopted consumer product on record: one million users in five days, 100 million in about two months, and 800 million weekly by October 2025. (TechCrunch)

Concentrating it. The same stretch made clear how few players can operate at this level, and what that does to everyone else.

• Roughly half of all global venture funding in 2025 went to AI: about $211B, up 85% from $114B the year before. (Crunchbase)
• The valuations pool at the top. Anthropic went from $183B in September 2025 to $380B in February 2026, and is reportedly in talks near $900B; OpenAI sits in the same multi-hundred-billion tier. (Anthropic, CNBC)
• The cost of admission keeps climbing. Frontier model training passed about $500M per model in 2025 and is headed toward $1B to $3B by 2027, which means, per the researchers tracking it, “only the most well-funded organizations” can finance them. (Epoch AI) Stargate, the infrastructure project behind one of them, is a $500B joint venture.
• Incumbents are being repriced on the threat alone. Figma’s stock fell more than 80% from its post-IPO peak as the market bet that AI design tools (Anthropic’s Claude Design, Google’s Stitch, Adobe’s own) would commoditize it, even as Figma’s revenue grew about 41% in 2025. The fear, not yet the financials, moved the price. (Yahoo Finance)

At an unprecedented pace #

You do not need to follow the science to read the curve. The capability is climbing fast enough that the trajectory itself is the argument: hold the rate flat, or even slow it, and it is still a step change.

• The clearest recent jump came in November 2025, when GPT-5.1 and Gemini 3 launched within five days of each other. On ARC-AGI-2, a test of abstract reasoning, the previous Gemini had scored 4.9%; Gemini 3 Pro hit 31.1% and its Deep Think mode 45.1%, a six to nine-fold leap in a matter of months. (VentureBeat, Vellum)
• On the hard exams the models are near the ceiling: 95% on AIME 2025 competition math with no tools, and 91.9% on PhD-level science questions (GPQA Diamond). (Vellum)

Those scores are hard to feel from the outside. From the inside, I remember the November jump: I skipped Thanksgiving travel that year and spent the whole weekend, Wednesday through Sunday, up past 3am every night, building things I could not have built a month before.

This carries enormous societal ramifications #

The opening line names one of them, exacerbating existing inequities, but it is only one. Here are the ones we think matter, and where this stops being matter-of-fact and gets genuinely hard.

The displacement of work, and the question underneath it #

The jobs go first, and not the ones we expected. The cost is landing on knowledge work, even on the engineers building the tools. But displacement is the surface of it. The deeper question is what work is for. If this wealth gets generated, if goods get cheaper and large amounts of human labor are no longer needed, what do we do with our time, and how do we reconcile no longer being the most capable intelligence in the room? We do not have answers. We think the questions are the point.

The possibility we are naming and setting aside #

Some serious researchers think this could end very badly. In a 2023 survey of 2,778 AI researchers, the median respondent put a 5% chance, and the mean a 14.4% chance, on AI leading to human extinction or similar permanent disempowerment; depending on how the question was framed, 38% to 50% assigned at least a 10% chance. (IEEE Spectrum, AI Impacts) We take that seriously. We are also not going to litigate it here, because you can lose days to it and this conversation has a different job. We name it and move on.

Exacerbating existing inequities #

This is the one the opening line named, and it is the center of our concern. A multiplying force rewards the people who can already wield it. If you are fluent with these tools, you pull further ahead; if you are not, or cannot get access, the distance grows. It concentrates power the same way it concentrates wealth, toward those who can use it.

We have felt the first half of that firsthand: with these tools we run far more sophisticated operations, marketing, and sales than we used to, by multiplying the output of our most talented people.

And it is arriving in an order almost nobody predicted. We assumed automation would take the manual, repetitive work first and leave humans the creative and cognitive high ground. It came for the creative and cognitive work first, the writing, the design, the code, while folding laundry stays hard for a machine. The writer Joanna Maciejewska put the inversion best: “I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.” Researchers call the reason behind it Moravec’s paradox: the things that feel most distinctly human turn out to be the easiest to automate.

This is the terrain Sandy Darity and Kirsten Mullen have spent careers on. Their work on the racial wealth gap, in From Here to Equality, traces how advantage compounds through exclusion, through who has access and who does not. Darity describes today’s gap as “the cumulative intergenerational effect of racial injustice in the United States.”

We have worked with them before. A couple of years ago we built an event with them, and AI was barely part of it, if at all. We are bringing them back into a different conversation now because we are convinced AI will force us to reckon, as a society, with how wealth gets distributed justly, not only by what the market allows or by what maximizes it for each individual. Few people have spent more time on what fairness in distribution actually requires. They bring the history, and we think the opening is now. We do not believe we can avoid reckoning with the thing that is coming fast.

What could optimizing for justice and human flourishing look like? #

This is the part nobody gets to answer cleanly. It is a pursuit with as many viewpoints as there are people, and it only gets harder as the future we are projecting gets harder to predict. Eudaimonia, the old Greek word for human flourishing, a life lived well rather than a moment of pleasure, is not something we are going to define for anyone.

What we will say is that across the enormous variety of human belief systems, something recurs: a pull toward limits on accumulation, toward sharing, toward justice.

Savas Labs is not going to take a prescriptive stance on the answer. We are interested in where these conversations go, and we will share the most interesting things we hear in them.

What institutions and groups are capable, and responsible? #

Here is a narrower take we will stand behind. We think it is honorable to try to optimize for everyone, not for the few who happened to be standing in the right place. Anyone who builds something of value did it on the shoulders of giants; if you find yourself atop the organization that did the right research at the right moment, luck did a great deal of the work. Should there be limits on the power that produces? We think yes, and the goal should be the broadest possible benefit.

The leaders building this say the same thing. Dario Amodei writes that we must “fight for a positive-sum future where everyone is better off” (Machines of Loving Grace); OpenAI’s stated mission is to “ensure that AGI benefits all of humanity” (OpenAI). We are aligned with that line. We would add one thing they tend to skip: the pace and the difficulty of fully understanding these systems compound each other. You might understand the frontier today, and a month later an innovation lands that only a handful of people grasp. Meanwhile the institutions responsible for making rules for whole populations are built to move slowly, and should be, that is what legitimate governance is, and that pace does not meet this moment. That mismatch, ultimately at a global level, is what we have to reckon with.

And we under-resource the reckoning. The people whose job is the public’s welfare should be funded to do this work. As one marker of how far we are from that: the US government’s AI standards body has roughly 30 staff and about $30M since 2024, while the UK’s equivalent has over 100 technical staff. (Fortune)

We’ll keep updating this #

We’ll add what comes out of the Boston and New York conversations. If you’re reading before one of them, come back after, and remind us to update it if we haven’t. Stay tuned. 📻

Submit a question or a thought →

Boston: Friday, May 29, 5:00 to 6:30pm ET. RSVP New York: Wednesday, June 3, 6 to 8pm ET. RSVP