15/🧵
The AI intelligence curve won't slow down. The responsibility falls on us to ensure that human cognitive capabilities don't deteriorate as our technologies advance. The greatest risk isn't only superintelligent AI—it's cognitively diminished humans.

12/🧵
The question isn't only whether machines will "take over", although that is also a risk, exemplified by the-coming-wave.com —it's whether we're inadvertently diminishing our own cognitive strengths at precisely the moment we need them most to guide these increasingly powerful systems.

8/🧵
Our attention is increasingly fractured by notifications and interruptions, each one pulling us from deep focus and eroding our capacity for sustained cognitive effort.

5/🧵
What's driving these human cognitive challenges? As
@jburnmurdoch
points out, our digital environment has fundamentally transformed how we engage with information:

4/🧵
Meanwhile, human intelligence appears under pressure. Recent international assessments show declining literacy, reasoning abilities, and information processing capacity across multiple countries and knowledge domains.

3/🧵
This isn't just incremental progress. In less than 2 years, top AI models have advanced from "below average" intelligence (50-70 IQ) to "gifted" levels (130+). The slope of improvement shows no signs of flattening.

2/🧵
AI model IQ scores from 2023-2025 shows a striking upward trajectory. The newest models (Gemini 2.5 Pro, OpenAI o1, Claude 3.7) are scoring 120-130 on standardized IQ assessments, well above the human average of 100 and entering "gifted" territory. Data from trackingai.org/home

1/🧵
Two concerning trends are converging: As AI models rapidly gain intelligence, human cognitive capabilities appear to be under pressure from our digital environment. This divergence deserves serious atention. Figure from www.ft.com/content/a801...

As a Dane who has called America home three times, I'm heartbroken watching what's happening now.

Some of the friendliest, kindest, and smartest people I have met are Americans, thus it is very difficult to watch the US destroy its reputation, democratic values, leadership in science,

1/2

Additional data around what questions the LLMs got right and examples of questions

Results were:

🥇ChatGPT o1 Pro: IQ = 81 (13 correct)
🥈ChatGPT o1: IQ = 78 (12 correct)
🥉Gemini 2.0 Thinking: IQ = 70 (10 correct)
4️⃣Claude 3.5 Sonnet: IQ = 67 (9 correct)
5️⃣Gemini1206: IQ = 63 (8 correct)

3/4
@emollick.bsky.social @serge.belongie.com @heikohotz.bsky.social @sbubeck.bsky.social

As can be seen, the results within this difficult task for LLMs are not yet impressive

The LLMs are excellent in many language, Math, and coding tests, but still struggle with the typical visual puzzle tests

I fitted a curve to the data to predict an IQ for the models, as they were <85 cutoff
2/4

For fun I tested the leading AIs on the Norwegian Mensa IQ test test.mensa.no/Home/Test

35 visual-pattern puzzles, 25 min. No special knowledge needed. Each correct answer earns points. No bonus for finishing early, no penalty for guessing

ChatGPT o1 Pro could not complete within time limit
1/4