A very good read. I lack any real sense of trust in these companies to do what is right/best for humanity. Its interesting to learn how companies are progressing - for me, the social implications are what are most important and still remain frankly terrifying.
We also need to consider mainstream media companies as those self-sustaining power structures.
Not pandering to journalists was big tech's biggest mistake. The shift from "big tech as the cool guys inventing shiny new things" to "evil Silicon Valley overlords trying to get our kids addicted" was largely driven by mainstream media, who felt sidelined and threatened after Facebook destroyed their distribution advantage. The (extremely overblown and since fully debunked) Cambridge analytica scandal, and others like it, were just fodder for the New York Times's guns. This shift had a major impact on antitrust policy and the public's perception of tech at large.
I think if anything is going to slow down Big AI, it's this. It's not going to be just the media either, if too many people get their therapy, politics and legal or medical advice from Chat GPT, people who can regulate that behavior out of existence are going to act.
Most countries are essentially run by lawyers; legislators are often former lawyers, judges are current ones. If AI takes (some of) their jobs, they'll be predisposed to be unfriendly towards AI companies. I don't see those companies getting in front of this and trying to address the problem while they still can, which suggests to me that they're making the exact same mistake that big tech did.
1) The Cambridge Analytica scandal was not fully debunked, and I'm not sure exactly how blown you think it should have been. Ask your favorite AI to enlighten you.
2) Big tech's mistake was not, well I'm also unclear what you think it was, getting covered by the press? And do you think people aren't addicted to their screens, especially children? Do you think teens self-esteem isn't lower with social media use? Do you have any theories as to what caused youth suicide to rise? Et cetera.
They have enough resources to be relevant in the future but I’m surprised they aren’t more relevant today. Eventually they’ll receded to be in the xAI tier, but they have time imo.
From the outside, it feels like there's been a total talent drain to OAI and Anthropic. I wonder if the ones still there have the ability to get back on top.
Long term, do you see it being a winner-takes-all market/duopoly, or do you think there's space for multiple tiers of lab?
Hi. Great piece, as always. How does socially positive involvement look like for those who are not in the frontier labs and do not wish to be there, in your opinion? I hear your concerns about governmental bans of large open source models with the pretext that they are too powerful to be let loose. I think this is a true risk; at the same time, AI oligopoly risk is even greater and very real. That can create a permanently tilted playing field that cannot be left unregulated for too long - it causes implosion of free market myth and will have massive momentum towards nationalisation or regulation similar to how utility services have been regulated to avoid extractive and predatory pricing on basic fundamentals of civilisation like water, electricity, etc. The power and capital deployed by frontier labs is still completely tiny compared to the rest of the economy, so these cannot be at complete odds for too long. My thinking is that AI compute will become a social good like electronic money or banking - ubiquitous, regulated, but with privatized gains and profits like many current services in capitalism. The crucial unknown factor is the degree of greed in frontier labs. If that will overstep the boundaries, an axe will fall. If not, we will muddle through.
Re #1 on your list: how much of this is due to the fact that part of the “wow” factor with Claude Code / Codex has been “product” related rather than pure model capabilities?
They’re not unrelated of course, but I feel like my own “aha” moments have come from things like interoperability, the use of UX patterns like multiple choice to get additional guidance / clarity, etc.
glad to see you mention the american side of the open weight models. i’m a fan of them. i think local models will actually be really common over the next few years as they get really energy efficient and intelligent enough to do real work.
The open-for-enterprise vs closed-for-revenue framing is sharper than the consensus take. The missing piece is pricing pressure — once Gemma 4 matches Qwen at lower deployment cost, the closed premium has to be justified by capability, not brand. That's the squeeze nobody on the closed side seems to be pricing in yet.
The agentic-harness criterion is the load-bearing test in this prediction. Claude Code and Codex landed in late 2025 and reset what counts as a daily-driver coding workflow; the 5 to 6 month gap since then with no open-weight equivalent at competitive price points is the cleanest evidence that the open-closed gap is being mismeasured by static benchmarks. If open weights stay confined to enterprise automation and low-cost domains while closed models hold the daily-driver coding slot, the open-source revenue path narrows to volume plays, which is the harder question for whether the neolabs reach scale before agentic workflows lock in.
Do you see evidence of acceleration from the top-tier labs that is accelerating model quality?
Compounding returns on internal model usage and telemetry from market adoption might further separate the pace between the frontier and the open-weights models. We as users may also want to shift our usage toward open-weights models and open-source harnesses like OpenHands, pi, etc. if we want to invest in a diverse ecosystem.
Separately, how do you see the agent loop evolving over 2026? Are there model architectures that could support something like long-chain reasoning graphs or parallel thinking at inference time?
A very good read. I lack any real sense of trust in these companies to do what is right/best for humanity. Its interesting to learn how companies are progressing - for me, the social implications are what are most important and still remain frankly terrifying.
We also need to consider mainstream media companies as those self-sustaining power structures.
Not pandering to journalists was big tech's biggest mistake. The shift from "big tech as the cool guys inventing shiny new things" to "evil Silicon Valley overlords trying to get our kids addicted" was largely driven by mainstream media, who felt sidelined and threatened after Facebook destroyed their distribution advantage. The (extremely overblown and since fully debunked) Cambridge analytica scandal, and others like it, were just fodder for the New York Times's guns. This shift had a major impact on antitrust policy and the public's perception of tech at large.
I think if anything is going to slow down Big AI, it's this. It's not going to be just the media either, if too many people get their therapy, politics and legal or medical advice from Chat GPT, people who can regulate that behavior out of existence are going to act.
Most countries are essentially run by lawyers; legislators are often former lawyers, judges are current ones. If AI takes (some of) their jobs, they'll be predisposed to be unfriendly towards AI companies. I don't see those companies getting in front of this and trying to address the problem while they still can, which suggests to me that they're making the exact same mistake that big tech did.
1) The Cambridge Analytica scandal was not fully debunked, and I'm not sure exactly how blown you think it should have been. Ask your favorite AI to enlighten you.
2) Big tech's mistake was not, well I'm also unclear what you think it was, getting covered by the press? And do you think people aren't addicted to their screens, especially children? Do you think teens self-esteem isn't lower with social media use? Do you have any theories as to what caused youth suicide to rise? Et cetera.
Great piece! For your point 5 - have you then completely written GDM out of the frontier race? 3.0 being frontier feels like a long time ago
They have enough resources to be relevant in the future but I’m surprised they aren’t more relevant today. Eventually they’ll receded to be in the xAI tier, but they have time imo.
From the outside, it feels like there's been a total talent drain to OAI and Anthropic. I wonder if the ones still there have the ability to get back on top.
Long term, do you see it being a winner-takes-all market/duopoly, or do you think there's space for multiple tiers of lab?
oligopoly for sure
Hi. Great piece, as always. How does socially positive involvement look like for those who are not in the frontier labs and do not wish to be there, in your opinion? I hear your concerns about governmental bans of large open source models with the pretext that they are too powerful to be let loose. I think this is a true risk; at the same time, AI oligopoly risk is even greater and very real. That can create a permanently tilted playing field that cannot be left unregulated for too long - it causes implosion of free market myth and will have massive momentum towards nationalisation or regulation similar to how utility services have been regulated to avoid extractive and predatory pricing on basic fundamentals of civilisation like water, electricity, etc. The power and capital deployed by frontier labs is still completely tiny compared to the rest of the economy, so these cannot be at complete odds for too long. My thinking is that AI compute will become a social good like electronic money or banking - ubiquitous, regulated, but with privatized gains and profits like many current services in capitalism. The crucial unknown factor is the degree of greed in frontier labs. If that will overstep the boundaries, an axe will fall. If not, we will muddle through.
I think for most people it’s being more independent minded and value driven with where you choose to work.
For some, it can be creating content.
For all, it’s being more human in discussions with people outside the ai bubble.
Americans wanting to say no and having no mechanism to do so might matter more for the trajectory than any of the capability benchmarks.
Like this a lot Nathan, thanks for sharing.
Re #1 on your list: how much of this is due to the fact that part of the “wow” factor with Claude Code / Codex has been “product” related rather than pure model capabilities?
They’re not unrelated of course, but I feel like my own “aha” moments have come from things like interoperability, the use of UX patterns like multiple choice to get additional guidance / clarity, etc.
glad to see you mention the american side of the open weight models. i’m a fan of them. i think local models will actually be really common over the next few years as they get really energy efficient and intelligent enough to do real work.
The open-for-enterprise vs closed-for-revenue framing is sharper than the consensus take. The missing piece is pricing pressure — once Gemma 4 matches Qwen at lower deployment cost, the closed premium has to be justified by capability, not brand. That's the squeeze nobody on the closed side seems to be pricing in yet.
The agentic-harness criterion is the load-bearing test in this prediction. Claude Code and Codex landed in late 2025 and reset what counts as a daily-driver coding workflow; the 5 to 6 month gap since then with no open-weight equivalent at competitive price points is the cleanest evidence that the open-closed gap is being mismeasured by static benchmarks. If open weights stay confined to enterprise automation and low-cost domains while closed models hold the daily-driver coding slot, the open-source revenue path narrows to volume plays, which is the harder question for whether the neolabs reach scale before agentic workflows lock in.
Do you see evidence of acceleration from the top-tier labs that is accelerating model quality?
Compounding returns on internal model usage and telemetry from market adoption might further separate the pace between the frontier and the open-weights models. We as users may also want to shift our usage toward open-weights models and open-source harnesses like OpenHands, pi, etc. if we want to invest in a diverse ecosystem.
Separately, how do you see the agent loop evolving over 2026? Are there model architectures that could support something like long-chain reasoning graphs or parallel thinking at inference time?
Isn't gemini cli an answer to codex and claude code already?
Do you know anyone who uses it?