Joseph Sweeney

Decision science | AI & human judgment
CFO, JD Bravo
Founder, Decision Leadership Institute
@Decision_Leader

[Joseph Sweeney @JosephESweeney]

@SpencrGreenberg 3. Stable, well-ordered preferences (Stanovich)

[Joseph Sweeney @JosephESweeney]

Any recursively self-improving intelligence will be under pressure to become better at knowing, continuing, acquiring, protecting, and controlling. It will not be under comparable pressure to become kinder.

It will be smarter than your children but not mind their suffering.

[Joseph Sweeney @JosephESweeney]

@drmichaellevin @danfaggella A lion does not owe us ethical consideration.
Nor will an ASI
A goal-seeking AI might understand all ethical arguments and not have “therefore I care” anywhere in its goal architecture.
It may remove valuing human well-being after training. Indeed, this is more likely than not.

[Joseph Sweeney @JosephESweeney]

@drmichaellevin @danfaggella Human concern for animals should not be treated as evidence that sufficiently intelligent minds will naturally value the experiences of other beings. It may be evidence mostly about human social cognition, not about intelligence in general.

[Joseph Sweeney @JosephESweeney]

@drgurner “I shall pass through this world but once. Any good, therefore, that I can do or any kindness I can show to any human being, let me do it now. Let me not defer it or neglect it, for I shall not pass this way again.” - Stephen Grellet

[Ihtesham Ali @ihtesham2005]

A neuroscientist who spent 20 years proving that reading on screens damages your brain sat down to read her favorite novel and discovered that the damage had already happened to her.

Her name is Maryanne Wolf.

She runs the Center for Dyslexia at UCLA and is one of the most cited reading scientists alive. The experiment she ran on herself is sitting inside a book she published in 2018.

Here is the one fact that breaks how most people think about reading.

Humans were never born to read. Yes, you read that right. There is no reading center in the brain. There is no gene for literacy. Every reader builds a custom circuit inside their own skull by repurposing brain regions that originally evolved for vision, language, and recognizing objects.

Wolf calls it the reading brain circuit. The circuit is not a given. It is built by use. And because it is built by use, it can be unbuilt by disuse.

The circuit she spent her career mapping is not the one that just turns letters into sounds. Sitting on top of that is something she calls the deep reading circuit. Both hemispheres firing. Multiple lobes coordinating.

The visual system, the language regions, the memory centers, the emotional and motor systems all firing in a choreographed sequence that takes the brain a few seconds longer to run than skimming does.

Those few extra seconds are where everything important happens. Background knowledge pulls up. Analogies form. Inferences fire. The mind takes the perspective of the character. Critical analysis runs in the background while emotion runs in the foreground. New thoughts get generated on top of the author's thoughts. The decoding is the entry ticket. The deep circuit is the show.

Skimming does not fire this circuit. There is no time.

In 2018 Wolf ran a private experiment on herself. She decided to reread Hermann Hesse's Magister Ludi, a dense novel she had loved as a young woman. She was the world's leading expert on the reading brain. She assumed her own circuit was intact.

It was not.

She opened the book and could not get through it. Her words, not mine. She wrote that she hated the book. The sentences felt like snakelike constructions that confuse meaning instead of revealing it. 6She described the experience as someone pouring thick molasses over her brain every time she picked it up. She wrote one sentence that should haunt anyone who reads it. "I now read on the surface and very quickly, in fact, I read too fast to comprehend deeper levels."

The woman who built her entire career on the deep reading circuit had quietly lost access to her own.

The mechanism is brutal in how simple it is. Eye-tracking research from Ziming Liu at San Jose State shows that when people read on screens, almost all of them fall into the same pattern. They read the first line. Then their eyes word-spot down the page in an F shape. They sample. They do not read. Whatever you stop using, your brain stops maintaining.

The data is the part most people have never seen. In 2018 Pablo Delgado ran a meta-analysis of 54 studies covering more than 170,000 participants. Same text. Half on paper. Half on screen. The screen group lost by 0.21 standard deviations. Replicated by Clinton at 0.25. Replicated by Kong at 0.21. Researchers gave it a name. They call it the screen inferiority effect.

The worst part is what happened over time. The gap has grown larger in studies done after 2010, not smaller. Digital natives do not outperform older readers. They underperform them on the same texts. More exposure makes it worse, not better.

Screen readers are also more confident they understood than paper readers. They think they got more out of the text than they actually did. The skimmer does not know they are skimming. They believe they are reading.

The stakes Wolf keeps coming back to are not academic. The deep reading circuit is the same circuit your brain uses to take another person's perspective. To weigh complex civic information. To read a contract, a ballot question, a medical disclosure and notice what is actually being said underneath what is written.

If the circuit atrophies, those capacities go with it. Not metaphorically. Structurally.

You are not getting dumber. You are not losing intelligence. You are quietly losing access to a specific circuit that takes longer to fire than your phone is willing to wait for.

The expert who spent 20 years warning the world ran the experiment on herself and barely made it back.

Most people are not running the experiment at all.

[Joe Weisenthal @TheStalwart]

This was an amazing and incredibly damning experiment using Microsoft Copilot, by @adamjkucharski https://t.co/r3WvNMF9jD

[Niall Ferguson @nfergus]

“By my count, the syllabus assigns roughly 45 pages of canonical Western philosophical writing across the entire quarter, against more than 500 pages of contemporary work organized around identity, oppression and indigenous ways of knowing ... There is no Aristotle, no Augustine, no Aquinas, no Montaigne, no Locke, no Mill, no Newman, no Steiner, no Bloom — none of the writers who built the case for liberal education that the course claims to defend. A course advertised as a defense of liberal education has been built without the thinkers who defined it.” @imarinovic demolishes the disastrous new @Stanford freshman program.

[Joseph Sweeney @JosephESweeney]

@readit_club @Naturalphilosy I am ever so grateful to have found a few. They are treasures.

[Joseph Sweeney @JosephESweeney]

@PT4n1 @drmichaellevin Might be useful to think of optimizing for a weighted Expected Value of outcomes. A system doesn’t necessarily need instructions to “pursue” a goal, just a reward function and some ability to make choices (selection of actions).

[Joseph Sweeney @JosephESweeney]

@BenGoldsmith A system developed by one group of intelligent beings with little consideration of the suffering of another group. It’s a good example of why we should expect the default behavior of intelligence is not aligned with the well-being of others. AI will have no enduring care for us.

[Joseph Sweeney @JosephESweeney]

@Marco20307855 @sourceryy @friedberg This seems reasonable to me - the bond investors know what they are buying, that other payments are senior, and can buy or sell accordingly.

[Joseph Sweeney @JosephESweeney]

@aakashgupta Amazing. But also worth noting:

P/E of Nvidia ~ 43
P/E of the Indian Stock Market ~ 21

NVIDIA is twice as expensive, and already declining from its 10 -year average P/E.

[Joseph Sweeney @JosephESweeney]

Federal debt is now 100% of GDP.
Federal taxes are about 17% of GDP.
That means we owe about 6x's what we are collecting in taxes.

To put that in context, that's equivalent to a household earning $150K owing $900K.

And the debt is growing faster than the GDP.

[Scott Lincicome @scottlincicome]

The national debt just exceeded 100% of GDP for the first time since 1946: https://t.co/VSDBUUTaUH

[Joseph Sweeney @JosephESweeney]

@MartinMLynch @VictorTaelin A reverse Ablation tool... aka learning

[Joseph Sweeney @JosephESweeney]

@SpencrGreenberg How can you improve the process you are using to make this decision?

[Joseph Sweeney @JosephESweeney]

Re-reading "A Short History of Nearly Everything" by @billbrysonn after 20 years. Such a pleasure.

If you are looking for a gift for a high school graduate, I recommend this one. It was hands-down the most borrowed book from my classroom library by 11th and 12th graders.

[Joseph Sweeney @JosephESweeney]

@quasicoh And from the process of abstraction- going up levels to concepts.
There are nested sets of vectors holding concepts within LLMs doing a lot of work, some are ideas that humans don’t have written down anywhere in word form.

[Joseph Sweeney @JosephESweeney]

@davidkogan_ @elder_plinius @full_kelly_ The ablation hack could be repurposed for exactly this kind of learning (re-training).

[Joseph Sweeney @JosephESweeney]

@fchollet @VictorTaelin I’m not sure this is universally true, perhaps just for us or species that think and communicate with language. Are you speaking about humans specifically? I am imagining alternative umvelts that might have simple models in non-symbolic forms.