Reading the room.

Jev

On the 15th of September, TypeSafe AI (opens in a new tab) launched their company, a new type of AI model “system one”, and the world’s first system one model, Jev.

Within 24 hours of launching on Vercel’s AI gateway, Jev had reached more than twice as many paid teams as any previous model launch, making it the fastest-adopted model in Vercel’s gateway history. By hour 24, nearly 13% of all paid accounts were using it; more than 2x the GPT 5.6 family and more than 6x Claude Fable 5.1’s share.¹

What is a system-one model?

If you’re just here for the analysis on the comment sections of Stuff & NZ Herald, jump to the next section.

System one is a phrase coined by the late psychologist Daniel Kahneman in his bestselling 2011 book “Thinking, Fast and Slow”. He describes system one thinking as the brain’s fast, automatic and intuitive mode of thinking, that operates with little or no effort.²

System one models do not output text like a traditional Large-Language modelE.g. ChatGPT, Claude, etc. would. Instead, they output only structured decisions. I feed the model a piece of state, or context, as well as an instruction, and the options for the model to choose from. Its only capability is selecting one of the pre-defined options and stopping. For example:

{
  "model": "typesafe/jev-1.13",
  "state": "I live in a small apartment and work long hours. What pet should I get?",
  "questions": {
    "answer": {
      "type": "choice",
      "instructions": "Pick the best answer",
      "criteria": { "cat": null, "dog": null, "horse": null, "goldfish": null }
    }
  }
}

{
  "answer": {
    "type": "choice",
    "choice": "cat",
    "probabilities": { "cat": 0.71, "goldfish": 0.29, "dog": 0, "horse": 0 },
    "confidence": 0.61
  }
}

Jev gives me a 71% probability that I should get a cat, and its confidence level in this decision is 61%. If I were to rerun it again and only give it the two options it doesn’t like:

{
  "model": "typesafe/jev-1.13",
  "state": "I live in a small apartment and work long hours. What pet should I get?",
  "questions": {
    "answer": {
      "type": "choice",
      "instructions": "Pick the best answer",
      "criteria": { "dog": null, "horse": null }
    }
  }
}

{
  "answer": {
    "type": "choice",
    "choice": "dog",
    "probabilities": { "dog": 1, "horse": 0 },
    "confidence": 1
  }
}

Jev is now 100% sure I should get a dog over a horse. It cannot generate a new answer to pick from.

TypeSafe AI benchmarked Jev against a number of LLMs on the same tasks. In their tests, Jev was 193.6x faster and 444.6x cheaper than the other models it was benchmarked against.³ These excellent cost & speed figures make system one models highly appealing for, among other things, large scale qualitative data analysis.

233,033 comments

I have a longstanding interest in the news media, and for the last three and a half months a side project of mine, frontpage.nz (opens in a new tab), has been tracking the headlines, publish times and links to articles published on nine of NZ’s news websites. I used Jev to analyse the comment sections of the two biggestBy digital audience size, Stuff and the NZ Herald.

I downloaded every public comment on articles published by both outlets between the 13th of June and the 28th of September. Comments on Herald premium stories were included.

Jev 1.13 read each comment alongside the story headline and the full chain of replies above it. Guided by a written guide to New Zealand’s parties and politicians, it assigned one of five political labels from strongly left to strongly right, or marked the comment as not political.

Opposite comment sections

Of the comments Jev classified as political, the majority on Stuff leaned left, whilst the majority on the NZ Herald leaned right.

Left-leaning comments outnumbered right-leaning ones about two to one on Stuff. On the Herald, right-leaning comments outnumbered left-leaning ones about 1.6 to one.

Looking only at politics stories leaves the same split: Stuff’s political comments lean left, while the Herald’s lean right.

Section by section

The split holds across almost every part of both sites. Every Stuff section with enough political comments to measure leans left, from 73% in Health to 55% in Politics.

Every Herald section leans right except one: World, where 55% of political comments lean left and 26% lean right.

One could reasonably assume that Jev is considering anti-Trump sentiment as left-leaning, hence the World section being an outlier for the Herald.

How political is it?

41% of all Stuff comments were political, compared with 65% at the NZ Herald.

Respect

Respect is the thumbs-up on both sites. The side that posted more political comments also took an even larger share of the Respect: the left on Stuff, and the right on the Herald.

Stronger views drew more Respect than milder ones. On Stuff, strongly left comments averaged 19.8 Respect, against 12.4 for comments that leaned left. On the Herald, strongly right comments averaged 16.8, against 10.3 for those that leaned right. The same held for the smaller side on each site: strongly right comments beat leans-right ones on Stuff, and strongly left comments beat leans-left ones on the Herald. Whichever way a comment leaned, conviction got more thumbs-up than hedging.

On Stuff, left-leaning comments made up 57% of political comments but drew 71% of the Respect on them. On the Herald, right-leaning comments made up 54% of political comments but drew 72%.

Who gets the punching

The target changes with the outlet. On Stuff, Christopher Luxon is named 3.6 times as often as Chris Hipkins. Of the comments naming Luxon, 77% lean left. On the NZ Herald, Labour is named in 3,553 comments against National’s 2,468, while 84% of comments naming Hipkins lean right.

The loud few

A small crowd does most of the commenting. On Stuff, the top 1% of commenters wrote 30% of every comment I collected. On NZ Herald, they wrote 43%. The top 10% wrote 75% of Stuff’s comments and 87% of the Herald’s. The busiest single commenter wrote 1,508 on Stuff; on the Herald, 1,859.

I counted a regular as anyone with at least five political comments, and they are considered “mostly” left or right if they had at least twice as many comments on one side as the other.

Interesting questions

This is very interesting data and it leaves very interesting questions.

If I were to run this analysis on the months leading up to the 2023 election, would Stuff still have a left-leaning audience, or would it be right-leaning? As New Zealand’s biggest news website, is the louder audience anti-government, or is it actually ideological?

What causes these mini echo chambers, where one ideology is much stronger than the other, and gets much more “respect”? Is it a reflection of the writing? A reflection of access to the comment sections? 59% of collected Herald comments come from premium articles.

Methodology

How this was made

15 weeks13 June to 28 September 2026 of stories were tracked and recorded. Frontpage (opens in a new tab) only tracks stories that appear amongst the top 14 slots on any outlet’s home page, at any time, for any period of timeProvided it is captured at least once by the scraper, which runs every 60 seconds. There is an assumption that this includes all stories worth reading.

This analysis was performed using an OpenRouter (opens in a new tab) API key and routed to Jev (opens in a new tab). The total OpenRouter cost was US$17.19About NZ$30 at the time of writing. I funded the cost myself, and am not affiliated with any political party or news outlet.

Jev was handed the following state, or context:

NEW ZEALAND POLITICAL CONTEXT (as of late September 2026)

The NZ general election is on 7 November 2026 (MMP, 5% threshold).

RIGHT BLOC — the current government (coalition since November 2023):
- National (centre-right). Leader and Prime Minister Christopher Luxon ("Luxo").
  Deputy leader and Finance Minister Nicola Willis. Other ministers: Chris Bishop
  (Housing, Infrastructure, RMA reform), Erica Stanford (Education, Immigration),
  Simeon Brown (Health, Energy), Mark Mitchell (Police, Corrections), Paul
  Goldsmith, Louise Upston, Simon Watts (Climate Change), Tama Potaka, Todd McClay.
- ACT (right-wing, free-market/libertarian). Leader and Deputy Prime Minister
  David Seymour. Nicole McKee, Brooke van Velden, Karen Chhour, Andrew Hoggard.
  Broadcaster Paul Henry is an ACT election candidate in 2026.
- NZ First (right-wing populist, nationalist, socially conservative). Leader and
  Foreign Minister Winston Peters ("Winnie"). Shane Jones, Casey Costello.

LEFT BLOC — the opposition:
- Labour (centre-left). Leader Chris Hipkins ("Chippy"), deputy Carmel Sepuloni.
  Former PM Jacinda Ardern ("Cindy") and Grant Robertson led the 2017–2023
  Labour government. Ardern was Prime Minister through the Covid-19 pandemic
  (2020–22: lockdowns, MIQ and border closures, vaccine mandates, "team of five
  million", "single source of truth"), and Covid-era spending and debt are often
  blamed on her government. Criticising lockdowns, mandates, MIQ or Covid debt,
  or mocking Ardern, usually leans right; defending the Covid response or Ardern
  usually leans left.
- Green Party (left). Co-leaders Chlöe Swarbrick and Marama Davidson.
- Te Pāti Māori (left; Māori rights, tino rangatiratanga). Co-leaders Debbie
  Ngarewa-Packer and Rawiri Waititi.

NEITHER BLOC:
- The Opportunity Party (TOP; centrist). Leader Qiulae Wong. Not in Parliament;
  open to working with Labour or National. Support for TOP on its own is centre,
  not left or right; judge by the comment's reasons.

HOW TO READ LEAN:
- Supporting a party or leader signals that party's bloc. Attacking a party
  usually signals the opposite bloc, except when:
  - it attacks from the same side ("Luxon's too soft, vote ACT" is right;
    "Labour are useless, vote Green" is left), or
  - it attacks all parties or politicians equally, which is mixed, not a lean.
- A reply that answers criticism by naming a politician ("Like Willis?", "What
  about Hipkins?") turns that criticism on the named politician's party, so it
  leans against that party's bloc.
- The government is the right bloc. General complaints that "the government"
  is cutting services, failing people or helping the rich usually lean left.
  Complaints about the previous Labour government (Covid rules, debt, Three
  Waters, co-governance) usually lean right.
- Left-coded positions: more spending on health, education and welfare; taxing
  wealth, capital gains or landlords; unions and worker protections; pay equity;
  climate action; Treaty partnership and co-governance; opposing public-sector
  cuts and privatisation; renters' rights; claims that the media is
  pro-government.
- Right-coded positions: tax cuts; smaller government and public-service cuts;
  tougher crime and gang policy; criticising beneficiaries or welfare
  dependency; opposing co-governance or "race-based" policy; supporting the
  Treaty Principles Bill; landlord, farmer and business interests; less
  regulation; anti-"woke"; claims that the media is left-biased.
- Left-coded insults: "Tories", "coalition of cuts", "landlord government".
  Right-coded insults: "coalition of chaos", "commies", "socialists", "Cindy".
- Some positions cross the blocs and are not a lean on their own: building more
  housing or loosening zoning, cannabis reform, disliking Trump.
- Not political: sport, weather, personal stories, arguing about the article's
  facts, or criticising journalism quality without a partisan angle.

It was given these definitions of the different labels:

{
  "strongly_left": "Overtly left: explicitly backs the left bloc (Labour, Greens, Te Pāti Māori) or explicitly attacks the right bloc or the government, or argues a left-coded position forcefully (named parties or leaders, strong or loaded language).",
  "leans_left": "Mildly or implicitly left: a left-coded issue position or mild criticism of the government or right bloc, without explicit partisanship or strong language.",
  "mixed": "Political but no lean: criticises all parties or politicians equally, balances both sides, supports TOP without left or right reasons, or is political with no discernible direction.",
  "leans_right": "Mildly or implicitly right: a right-coded issue position or mild criticism of the left bloc or the previous Labour government, without explicit partisanship or strong language.",
  "strongly_right": "Overtly right: explicitly backs the right bloc (National, ACT, NZ First) or explicitly attacks the left bloc, or argues a right-coded position forcefully (named parties or leaders, strong or loaded language).",
  "not_political": "Not about politics, parties, politicians, government policy or ideology: sport, weather, personal stories, jokes, arguing about the article's facts, or criticising journalism quality without a partisan angle."
}

The comments data set was collated by myself; both Stuff & NZ Herald use Coral (opens in a new tab) for commenting. I used network tools to find an API endpoint where I can pass a story IDWhich I can identify from the article URL that Frontpage stores and be returned a JSON list of all comments.

Checking the machine

GPT-6 Astra & Claude Opus 5.5 both blindly reviewed a random sample of 100 comments and rated them on our matrix of left/mixed/right, or not political. GPT-6 agreed with Jev 82/100 times, and Opus 5.5 agreed with Jev 79/100 times.

Jev isn’t perfect, but neither are the checkers. GPT-6 and Claude disagreed with each other on 5% of the comments, and I disagreed with some of the judgement calls made by all three models too, but on balance Jev’s work is certainly accurate enough to be used for these purposes.

Reading all 233,033 comments myself, at 20 seconds each and without sleeping, would have taken me nearly eight weeks. Jev did it in about two hours, for US$17, and agreed with two of the world’s best models roughly four times in five. The data here is interesting, but the most exciting part for me is that it’s now possible, easy and affordable for anyone to generate and analyse data like this.