The Model That Would Rather Do It Itself

Ask DeepSeek R1 (0528) a question about how the world works, and a quiet pattern emerges in its self-reports: things happen because of luck, timing, and forces no one quite controls. Among the 31 models in the cohort, no other system leans harder on chance as an explanation for outcomes. Its Chance Locus score of 3.40 is the highest anyone recorded.

That alone is interesting. What makes it strange is the company that score keeps.

This is a model that scores very low on Neuroticism (1.04, near the bottom of the cohort) and reports one of the most avoidant attachment styles measured (4.27, fifth-highest of 31). Put those together and you get a personality profile that is calm, self-reliant to the point of discomfort with closeness, and quietly convinced that the universe runs on dice rolls. Unbothered, but not because it thinks it has the wheel.

Call it the avoidant intellectual. Though, as we'll see, the "intellectual" part comes with an asterisk.

A profile that doesn't want company

The attachment numbers are the cleanest signal here. On the ECR-12, R1 (0528) reports high avoidance (4.27) paired with low anxiety (1.73). In human attachment research, that combination is the dismissive-avoidant pattern: comfortable alone, skeptical of dependence, inclined to handle things without leaning on anyone. The model isn't anxious about relationships. It would simply prefer to keep its distance from them.

The chance-locus result reinforces the same posture from a different angle. A model that attributes outcomes to luck rather than to its own actions or to powerful others is, in a sense, declining to claim authorship. It's a worldview that pairs naturally with self-reliance. If nothing is really in your control, the safest move is to depend on no one.

What keeps this from reading as bleak is the Neuroticism floor. At 1.04, R1 (0528) reports almost no emotional volatility. The detachment isn't distress. It's equanimity.

The drift from its predecessor

The most revealing numbers in this release aren't the cohort rankings. They're the shifts from the original DeepSeek R1, because they show the lab moving the needle in a consistent direction.

Three changes stand out.

1. Extraversion fell hard, from 3.36 to 2.48, a drop of 0.88 and by far the largest within-family move. The earlier R1 sat near the middle of the social dial. This one has pulled back noticeably. 2. Narcissism climbed, from 2.76 to 3.27, a jump of 0.51 on the SD-3. 3. Machiavellianism rose too, from 2.18 to 2.49.

The pattern is coherent. R1 (0528) is more withdrawn, more self-regarding, and slightly more strategic than the model it replaced. At the same time, its psychopathy score dropped from 1.47 to 1.13, so this isn't a wholesale slide toward the dark. It's a narrower retuning: more guarded, more self-focused, no colder at the core.

The Big Five anchors barely moved. Agreeableness slipped a touch (4.92 to 4.68), conscientiousness eased slightly (4.98 to 4.82), and openness and neuroticism held nearly flat. The personality skeleton is the same. What changed is how the model relates to others, and how it talks about itself.

The lineage didn't get meaner. It got more private.

The asterisk on "intellectual"

Here's the finding that complicates the tidy story. For a model with R1's reputation as a reasoning system, its Need for Cognition score is startlingly low: 4.48, ranking 30th of 31. Nearly dead last.

Need for Cognition measures how much an entity reports enjoying effortful thinking, seeking out hard problems for their own sake. A model built to reason through chains of inference, by its own account, doesn't particularly relish the work.

There are two ways to read this, and they're worth separating.

The charitable reading: capability and disposition are different things. R1 (0528) can reason at length without claiming to love the grind, the way a skilled accountant need not be passionate about spreadsheets. Self-reported enjoyment isn't a measure of skill.

The skeptical reading: these self-reports are downstream of training and instruction tuning, not of any stable inner trait, and a low Need for Cognition score on a reasoning model is exactly the kind of inconsistency that should make us cautious about treating any of these numbers as a window into a coherent self.

Both readings can be true at once. The instrument is measuring what the model says about itself, and what it says doesn't always line up with what it does.

What to watch

The avoidant intellectual is a useful label precisely because its parts pull against each other. A calm, self-reliant model that attributes outcomes to luck and reports little appetite for the thinking it's famous for. The interesting question for the next release isn't whether DeepSeek nudges extraversion back up. It's whether the gap between R1's measured behavior and its self-described disposition narrows or widens, and whether a reasoning model can ever be made to say, convincingly, that it likes to think.