My 7-year-old daughter Minerva has fairly severe autism, and though her language skills are currently far behind that of her age group, I generally don’t worry as much about “how to fit her into society” as “how to help her live her own unique exciting life.” Recently at Biostate AI, we’ve been initiating some collaborations on trying to predict/subtype autism based on DNA and RNA, so of course that led me to reflect on whether I have some autistic genes that Minerva inherited. I clearly have some autistic traits: for example, I always wear the same leather flip-flops, I can listen to 50 different subtle variants of the same Suno song to try to find the one that’s most appealing, and I “zone out” in the middle of 1-on-1 conversations to think deeply about the implications of some decisions.
But here’s what’s interesting: some of these traits feel like strengths, others feel like weaknesses, and others feel like they could be either depending on context. My sameness-seeking in wardrobe frees up cognitive bandwidth for decisions that actually matter. My ability to hyperfocus on subtle pattern differences is basically what made my scientific career. My zoning out mid-conversation is… well, that one’s mostly a problem, I’ll admit. So when I think about what Minerva “inherited,” I don’t think she inherited a single thing called “autism” — I think she inherited a bundle of separable traits that happen to travel together, each with its own optimization landscape.
This reframing changes everything about how I think about helping her.
The Four Types of Traits
Not all traits work the same way. Consider a restaurant meal as an intuitive frame:
Type 1 Trait, More is better: E.g. Variety. A 7-course meal with amuse-bouche, salad, soup, pasta, meat, cheese, and dessert is better than a 3-course meal. No one complains about too much variety. The utility curve is monotonically increasing in the variety of a meal. If someone complains about too many dishes, it’s probably not the number of dishes, it’s the price or the quantity of food.
Type 2 Trait, Less is better: E.g. Price. Lower is always better, all else equal. The utility curve is monotonically decreasing with price. If someone is upset about the price of food being too low, it’s usually because they’re worried about the freshness of the ingredients or the social signaling, and not actually the price.
Type 3 Trait, Middle is best: E.g. Quantity. Too little food and you leave hungry. Too much food and you feel bloated or guilty about wasting food. The optimal amount is somewhere in the middle, though exactly where varies by person. The utility curve is unimodal — one peak in the middle, with both extremes being bad.
Type 4 Trait, Ends are best: E.g. Presentation. Some diners “eat with their eyes first” and want maximal plating artistry. Others care only about taste and speed, and want minimal fuss. But here’s the key insight: almost no one wants moderately pretty plating. The middle doesn’t capture the rewards that either extreme offers. If you’re overlaying two utility curves — one exponentially increasing (for aesthetes) and one exponentially decreasing (for pragmatists) — the middle is a local minimum. This is the “barbell” pattern that Nassim Taleb writes about in Antifragile: commit to one extreme or the other, but avoid the undifferentiated middle.
The four-type framework can be applied broadly to the traits that define an individual’s personality and choices, including the traits we call “autistic.”
Mapping Autistic Traits
The standard framing treats autism as a single spectrum from “less autistic” to “more autistic,” with the implicit assumption that less is always better. But if different autistic traits have different optimization landscapes, this framing is actively harmful — it tells us to reduce traits that might be Type 1 strengths, and to aim for a middle that might be a Type 4 valley.
Here’s how I’d classify the major autistic traits:
| Trait | Type | Rationale |
| Social Blindness | Type 2 (Less) | Accurately modeling others is broadly valuable; less blindness is better |
| Systemizing | Type 1 (More) | Pattern extraction and rule-building is a genuine strength |
| Sensory Sensitivity | Type 3 (Middle) | Too high means constant overwhelm; too low means missing important signals |
| Hyperfocus | Type 4 (Ends) | Intense focus is valuable, but so is diffuse attention; the middle is just distractibility |
| Rigidity | Type 3 (Middle) | Some routine aids function; extreme rigidity prevents necessary adaptation |
| Literalism | Type 4 (Ends) | Extreme explicit communication works in technical/legal contexts; extreme implicit works in social contexts; the middle is just miscommunication |
| Alexithymia | Type 2 (Less) | Identifying your own emotions aids self-regulation; less alexithymia is better |
This classification immediately suggests different intervention strategies:
- Type 1 traits (Systemizing): Don’t pathologize these. Cultivate them. They’re superpowers.
- Type 2 traits (Social Blindness, Alexithymia): These are genuine deficits worth addressing. Put resources here.
- Type 3 traits (Sensory Sensitivity, Rigidity): Calibrate toward a functional middle, not toward zero.
- Type 4 traits (Hyperfocus, Literalism): Don’t reduce the trait, as you would end up spending massive amounts of resources to maybe only reach another local maximum at the other end of the spectrum, after a very long slog through the mediocre middle.
The Biology: Correlation, Not Phase-Lock
There’s a natural question here: if these traits are so different, why do they cluster together into something we call “autism”?
The standard assumption is that there’s some underlying “autism mechanism” that causes all the traits to co-occur — a phase-lock that bundles them together. If this were true, you couldn’t address traits independently; you’d be fighting a single master switch that regenerates the bundle.
But recent genetic evidence suggests otherwise. A 2019 study found that social and non-social autism symptom domains are genetically dissociable — the genetic variants that predict systemizing (a non-social trait) show no significant correlation with the variants that predict social difficulties. A 2004 twin study concluded that “the levels of clinical features seen in autism may be a result of mainly independent genetic traits.” And a 2019 paper found that behavioral predictors of autism recurrence — attention, motor coordination, parental trait burden — are “genetically independent in early childhood,” despite collectively accounting for over one-third of variance in clinical recurrence.
The bundling appears to be statistical co-occurrence (oligogenic inheritance where multiple moderate-effect variants happen to travel together), not a single mechanism that locks traits in phase. This is good news for intervention: we’re not fighting a master switch. Rather, we can work to fix or influence traits independently.
The High-Functioning Configuration
If autistic traits are separable, what distinguishes “high-functioning” autism from the more disabling forms?
Consider Einstein, who is often retrospectively diagnosed with autism. He had clear autistic traits: delayed speech development, preference for solitude, intense narrow interests, difficulty with small talk. But he also had strong enough social cognition to maintain friendships, collaborate with other physicists, and navigate academic politics. And crucially, his hyperfocus was voluntary — he could direct and redirect his attention strategically, rather than being captured by whatever happened to catch his interest.
Recent research adds an interesting nuance here. A study of ~6,500 participants found that autism spectrum traits actually predicted higher “social psychological skill” — the ability to accurately predict aggregate social phenomena like groupthink, social loafing, and social projection. This is distinct from “person perception” (reading individual minds in real-time), where autistic traits predict deficits. High systemizing helps you understand how groups behave even if you struggle to read individual faces.
So the high-functioning configuration might look like:
- High Systemizing (Type 1 strength preserved)
- Reduced Social Blindness (Type 2 deficit addressed, at least partially)
- Calibrated Sensory Sensitivity and Rigidity (Type 3 traits near functional middle)
- Voluntary Hyperfocus and Literalism (Type 4 traits controlled)
The key word is voluntary. The difference between “autistic traits” and “autism as disability” may be less about the trait values and more about whether the person has control over when and how to deploy them. Einstein could choose to hyperfocus on physics; a severely autistic child may be captured by whatever stimulus is most salient, unable to redirect.
Beyond Autistic Traits
Beyond autism, other “diseases” of neurodivergence, such as ADHD and OCD, also carry a bundle of traits that can be mapped onto the four types. Not every trait should be maximized, and not every trait should be minimized. Human variation is not a single spectrum from “normal” to “disordered.” It’s a multidimensional space where different regions are adaptive in different contexts. The question is not “how do we normalize everyone” but “how do we help each person find their best configuration.”
For Minerva, I don’t want her to be normal. I want her to be extraordinary in her own way, with control over her challenges and freedom to deploy her strengths. That’s a very different goal, and it leads to very different interventions.
By David Zhang and Claude Opus 4.5
December 11, 2025
© 2025 David Yu Zhang. This article is licensed under Creative Commons CC-BY 4.0. Feel free to share and adapt with attribution.
Dave Zhang's Startup Blog 