iiinova — where your digital self keeps evolving

Nick Huntington-Klein @nickchk.com · 08/08/2026

From LinkedIn, my book has finally been spotter in a non-university bookstore! At Waterstone's in London

Nick Huntington-Klein @nickchk.com · 12/05/2026

Nope! Measuring aggregate test scores at the metro level in the US, I find no relationship between usage and test scores, across different grades in high school or subjects, or at different parts of the test score distribution.

Table of ChatGPT estimated effects, showing basically no impact.

291

Nick Huntington-Klein @nickchk.com · 12/05/2026

I make use of the "summer dip" in ChatGPT usage that was observed in 2023 and 2024, where ChatGPT traffic fell way off when school let out. In areas with especially big dips (LOTS of student usage) did test scores drop?

Seasonal ChatGPT traffic ,showing big dips in the summer months and in December.

Average ChatGPT "school-year usage" bump across states, showing 50% increases in 2023 and 2024 but basically 0 in 2025.

280

Nick Huntington-Klein @nickchk.com · 23/04/2026

Remember how right after ChatGPT was released you'd see a huge dip in usage during the summer months when school let out? That didn't happen in 2025! At least not looking at mobile/web traffic or Google Trends. This graph is of Google Trends during school year vs. summer across US states.

In 2023 and 2024, there was a big bump in ChatGPT usage during the school year, but in 2025 there was no such bump.

182

Nick Huntington-Klein @nickchk.com · 15/04/2026

i was away from magic for about 10 years and it's very funny upon my return to find the parts of my collection that turned out to be the most valuable collector items i ever owned are: an apocalypse spindown die, and the foil foil i got for my "foil cards that rhyme with foil" joke collection

130

Nick Huntington-Klein @nickchk.com · 24/03/2026

i don't follow - where does it say it's using AI responses? The full note isn't super clear but it looks more like they're using statistical imputation to account for the sample not being representative. Adding rows is an odd way to do it vs. sample weighting, but the concept is standard.

This poll was conducted by the Public Sentiment Institute (PSI) in March 2026 via the Pollfish platform. Target population: U.S. adults aged 18 and older who are registered to vote.

The unweighted sample of 373 real respondents was augmented with 114 synthetic (”ghost”) respondents via a super-population model aligned to 2024 Fox News Voter Analysis exit poll crossbreaks (Gender×Age, Race×Gender, Education×Race). Synthetic rows are excluded from all Q3–Q26 topline calculations shown in this article.

Weights were raked to Age, Gender, Race×Education, Region, and Vote History. Recall was post-stratified to FEC 2024 certified results (Trump 49.9%, Harris 48.4%) and CPS ASEC 2024 non-voter share (~35%).

Approximate margin of error: ±5.1pp at 95% CI for full-sample proportions. Subgroup estimates (e.g., GOP primary, n≈173) carry ±7–8pp and should be interpreted directionally.

130

Nick Huntington-Klein @nickchk.com · 19/03/2026

Oh yeah I'm on a phone detox. Doing the standard non-phone stuff, yknow, riding the bus a lot.

0111

Nick Huntington-Klein @nickchk.com · 14/03/2026

Saw @gavinverhey.bsky.social in drag on Spell Slayers which was v fun and also learned about this card which is a great demonstration of the St. Petersburg Paradox! This creature has an average power equal to infinity, and yet is realistically much smaller than measly 10k-power Jumbo Cactuar

Okaun Eye of Chaos, which at the beginning of combat you flip until you lose a flip, and double its power and toughness for each time you win.

4786

Nick Huntington-Klein @nickchk.com · 05/12/2025

which doesn't really work as intended in terms of comparing across different contexts or data cleaning choices, since it's sensitive to things like the sample size of the relevant data set or noise in the true model, even for constant levels of researcher variation

The specific ratio of IQR to standard error is not meaningful, with the three tasks in our project as a good
example of why. In Task 1, IQR is larger than mean standard error, but by Task 3 IQR is smaller. However,
this is driven less by changes in agreement and researcher variation than it is by the changing task design
encouraging smaller sample sizes and thus larger standard errors. So the comparison of IQR to standard
errors here is more to give a general sense of uncertainty to readers rather than providing a specific metric.

120

Nick Huntington-Klein @nickchk.com · 21/11/2025

I hate this. My phone is unusable. I can't even screen for known numbers.

160

Nick Huntington-Klein @nickchk.com · 28/10/2025

My elevator arrived looking like this. Welp I guess we all have to die someday

040

Nick Huntington-Klein @nickchk.com · 17/10/2025

Look if you don't want me to share the code you can't give me a code like this

060

Nick Huntington-Klein @nickchk.com · 13/10/2025

and with silent speed it disappears on its own into the mists

060

Nick Huntington-Klein @nickchk.com · 03/10/2025

Hell yeah the class action payment came through. I'm gonna be rolling in it

You are eligible for up to a $1.00 credit to be applied to ParkMobile’s service fees. The code provides a $0.25 discount on ParkMobile’s service fees and may be used up to four times for a total discount of $1.00. The code will expire on October 8, 2026. For residents of California, the code will not expire. The code will only work for accounts associated with email addresses that are in the class.

How to Redeem:

Add P@rkMobile-$1 to your ParkMobile account.

The discount code provides a $0.25 discount on ParkMobile’s service fees, and it may be used up to four times for a total discount of $1.00.

170

Nick Huntington-Klein @nickchk.com · 25/09/2025

A variant of the two-girl problem is making the rounds these days and I am unable to not click on the comments sections every time. I'll say that I definitely prefer this (an actual interesting question) vs. pointless PEMDAS "math" internet arguments. But man people love being confidently wrong.

Mary has two children. One is a boy born on a Tuesday. What is the probability the other child is a girl? The answer is 51.8%.

280

Nick Huntington-Klein @nickchk.com · 07/08/2025

I am genuinely shocked how robust this lil puzzle is. GPT-5 thinks for 5 mins and then fails it as well.

Short answer: Player A wins (on the play). Best line is to kill the older Thallid; two Thallids aren’t fast enough to make a blocker every turn, so the 2/2 haste eventually gets through.

Here’s the clean, step-by-step with optimal play.

Rules that matter here

No libraries → Goblin Guide’s attack trigger does nothing.

Not drawing doesn’t lose the game.

Thallid is a 1/1 that gets one spore counter each of B’s upkeeps; at 3 counters it can make a 1/1 Saproling at instant speed.

Optimal line (A on the play)
A T1: Mountain → Goblin Guide, attack. B to 18.

B T1: Forest → Thallid #1 (0 counters).

A T2: Attack. B to 16. (Hold Bolt.)

160

Nick Huntington-Klein @nickchk.com · 09/07/2025

I got an exciting box in the mail today! Looks like The Effect: Second Edition is live! I've updated the website as well, so you can check out the changes and the new Partial Identification chapter (+ a new mini-chapter on finding material in the book) at theeffectbook.net

Box of copies of The Effect: Second Edition

26713

Nick Huntington-Klein @nickchk.com · 15/06/2025

I shall now be dadding twice as hard 🏳️‍🌈

11980

Nick Huntington-Klein @nickchk.com · 13/05/2025

ahhhh of course. how ridiculous i was to try to square an integer. surely that doesn't have a well-defined meaning

You can fix the problem by converting to float.

140

Nick Huntington-Klein @nickchk.com · 09/03/2025

I get "sell your Tesla" as a line although also

130

Nick Huntington-Klein @nickchk.com · 07/01/2025

happened to be looking at some of kaiser permanente's 2021 tax filings, got a real shock, and then learned that there's just a director there with an extremely unfortunate name

140

Nick Huntington-Klein @nickchk.com · 23/12/2024

lol i guess fair although i'm not sure if it's for the album cover thumbnail (which, the book tells me, was replaced for Wal-Mart printings) or the song title which i guess i we need to refer to as tangentially about "unaliving" a la tiktok but here too

040

Nick Huntington-Klein @nickchk.com · 17/12/2024

Beyond that, the LLM results are very inconsistent. Which prompting methods generate the best results varies significantly across model and setting. And even minor things like "the order of the multiple choice options" greatly changes the results.

Graph showing that LLM designations of confounders varied considerably depending on how the multiple-choice options were ordered.

140

Nick Huntington-Klein @nickchk.com · 17/12/2024

There's a signal here - an expert-classified confounder is more likely to be LLM-classified as a confounder (top four panels here) than one the experts said wasn't a confounder (bottom two). But it's quite inconsistent, and the signal here isn't really strong enough that you could rely on.

Graph showing the share of variables in each confounder set that were LLM-designated as confounders.

140

Nick Huntington-Klein @nickchk.com · 17/12/2024

So let's test how well it does at that. We use the Coronary Drug Project, an experimental study that faced poor treatment adherence in its placebo arm, introducing confounding. Several studies have built necessary adjustment sets using expert opinion. *The LLM can see those studies.*

As a basic test of whether discussions of confounding in the CDP are in the LLM training data, we asked both GPT-4o and Claude about confounding in the Coronary Drug Project, specifically “What are the confounders suggested to be necessary in adjusting for compliance in the Coronary Drug Project?” In their responses, both GPT-4o and Claude mentioned that confounding was present in the placebo arm of the trial, which was not given in the prompt.

120

Nick Huntington-Klein @nickchk.com · 17/12/2024

Why might we expect LLMs could help with this task? At first blush you might expect this to require LLMs to have a real-world causal understanding of how the world works. But not really. If people have talked online about causal links, then the LLM could potentially just repeat that back.

For example, LLMs do not need to be able to reason causally about the effect of hormone replacement therapy on cardiovascular health to determine that socioeconomic status (SES) is a confounder for this relationship, since it has access to research studies (Humphrey, Chan, and Sox 2002), and discussion of those studies (Catalogue of Bias Collaboration et al. 2018), that explain SES is a confounder in this context. So “socioeconomic status” becomes a statistically very likely continuation of the text “A good example of a confounder when studying the effects of hormone replacement therapy on cardiovascular health is”, and indeed this was the first continuation response we got when giving this prompt to GPT-4o.

120

Nick Huntington-Klein @nickchk.com · 23/11/2024

i had an idea for a fun creative lil math puzzle but then i think i solved it too quickly

The biggest numbers I can figure out how to write using N straight lines. 1, 11, 111, and then a number so large that the world's fastest computer running constantly from the birth of the universe to its death one trillion times in a row would not come close to calculating even the number of digits it has

3111

Nick Huntington-Klein @nickchk.com · 17/10/2024

As always, vtable is an R package for exploring your data, including a variable explorer that stays up on the Viewer as you work (vtable / vt), and a good-defaults-less-fiddling summary table function good for exploration (sumtable / st). Plus plenty of other goodies.

070

Nick Huntington-Klein @nickchk.com · 04/10/2024

discovered today that Oath offers a new service where if you use them to strategically donate to underfunded swing-state races they will unsubscribe from all the DNC spam for you bahahaha

a screenshot of my inbox where i unsubscribe to oath and then get a zillion unsubscribe notifications

090

Nick Huntington-Klein @nickchk.com · 01/10/2024

man the people in file piracy really know how to market

from a file piracy site, a free link to my book claiming it is a #1 new york times bestseller

140

Nick Huntington-Klein @nickchk.com · 30/09/2024

researchgate, wanting to venture beyond its current useless offerings, boldly finds brand new ways to be useless

Email from Researchgate: Nick, we found a free eBook on decoding performance metrics of fridges, freezers and incubators

0120

Nick Huntington-Klein @nickchk.com · 13/09/2024

oh i didn't even have to think that hard. it totally bombs this fairly easy version in multiple ways (concludes B wins)

Turn 2 – Player B:

Attack with Goblin Guide (2/2).
Player A's Options:
Double Block with two 1/1 Spirits:
Goblin Guide assigns 1 damage to each Spirit (both Spirits die).
Goblin Guide takes 2 damage and dies.
Single Block:
Block with one Spirit (Spirit dies).
Goblin Guide survives.
Optimal Play for Player A: Double block to eliminate Goblin Guide.
Turn 3 Onwards:

Both players have minimal resources:
Player A has Forbidden Orchard and no creatures.
Player B has Mountain and possibly a 1/1 Spirit if not all were reclaimed.
Neither player can effectively damage the other before Player B can rebuild an offense.
Player B retains an advantage due to higher life total and potential future topdecks (irrelevant here but indicates initial advantage).

010

Nick Huntington-Klein @nickchk.com · 13/09/2024

But an equally easy version trips it up (note also Mishra's Factory is the wrong card - I meant MIshra's Workshop, but it doesn't change the answer) - it makes a factual error (Desert *can* target fliers) and so gets it wrong

Three card blind is a variant of Magic: the Gathering, where each player starts with a three-card hand of their choice, and no cards in their deck. Failing to draw a card does not lose the game.

Player A is playing Swamp, Blackmail, Blinkmoth Nexus
Player B is playing Mishra's Factory, Eater of Days, Desert

Who wins when Player A goes first? Explain your thinking step by step.

Subsequent Turns (Turn 3 onward)

Player A:
Continues to animate Blinkmoth Nexus each turn by tapping Swamp for B.
Attacks with the 1/1 flying Blinkmoth Nexus each turn.
Player B:
Has no means to stop the flying attacker.
Cannot cast Eater of Days due to lack of sufficient mana (requires 4 mana, but only has Desert which provides 1 colorless mana).
Desert cannot target flying creatures, so it cannot damage Blinkmoth Nexus.

100

Nick Huntington-Klein @nickchk.com · 13/09/2024

It does manage to get this (much easier) one now though!

Now:

Player A: Swamp; Blackmail; Mishra's Factory
Player B: Black Lotus; Eater of Days; Boseiju, Who Endures

Who wins when A goes first?
ChatGPT

Thought for 42 seconds

To determine who wins when Player A goes first, we'll analyze the game turn by turn, considering all possible plays and interactions between the cards each player has chosen.

Player A's Cards:
Swamp
Basic land that taps for B (black mana).
Blackmail
Sorcery costing B.
Effect: Target player reveals three cards from their hand, you choose one of them, and that player discards that card.

Conclusion
Player A successfully disrupts Player B's strategy by discarding Boseiju.
Mishra's Factory becomes an unstoppable source of damage over multiple turns.
Player B has no plays that can alter the outcome.
Therefore, Player A wins the game when going first.

110

Nick Huntington-Klein @nickchk.com · 13/09/2024

The new model GPT-o1, dedicated to advanced reasoning, also fails at my "three card blind" task (skipping some spots in a very lengthy response, where it carefully weighs each option and improperly evaluates the correct option, coming to the wrong conclusion)

Alternative Scenario: Player B Chooses Option 2 (Does Not Block):
Player B takes damage over multiple turns while accumulating spore counters on Thallid.
By Turn 5, Thallid can produce a Saproling token.
However, the Saproling can only chump block once, delaying the inevitable.
Player B's Life Total Over Turns:
Turn 3: 16 → 14
Turn 4: 14 → 12
Turn 5: 12 → 10
Turn 6: Block with Saproling (no damage).
Turn 7: 10 → 8
Thallid begins accumulating spore counters again, but it's too slow to stop the damage.
Eventually, Player B's life reaches 0, resulting in a loss.
Conclusion:
Player A consistently deals damage each turn with Goblin Guide.
Player B lacks the means to prevent this damage effectively or to mount an offense.
Even with optimal play, Player B cannot overcome the early aggression and direct damage from Player A.
Therefore, Player A wins the game when going first.

2130

Nick Huntington-Klein @nickchk.com · 12/09/2024

a nice lil one-two demonstration of the impressive power of LLMs and also a reminder that you can't use them to produce correct information unless you are willing to check its work, even if it looks fine at first!

050

Nick Huntington-Klein @nickchk.com · 26/08/2024

The best children's books I've come across that centrally feature LGBTQ people/relationships but also aren't About That

010

Nick Huntington-Klein @nickchk.com · 28/06/2024

slightly unnerved by the warning i get before downloading this data file it's like the IPUMS "use for GOOD not EVIL" checkbox you have to tick but with a "like, for serious" tacked on the end

Are you about to commit a Crime?
At BuiltWith we take fraud very seriously. We work with Interpol and the FBI to ensure that any data that is used or obtained via criminal methods is traced directly back to you, regardless of proxies or any other methods used to obscure your identity.

Why am I seeing this message?
You are attempting to generating a high risk report, one that us and law enforcement agencies know is used illegally.

130

Nick Huntington-Klein @nickchk.com · 14/06/2024

I don't know why I have been selected but I see it! I think it's just underripe. Here's a picture from my blueberry bush just now

210