iiinova — where your digital self keeps evolving

Phil Feldman, PhD @philfeldman.com · 13/09/2026

Whenever I hear about agentic systems going rogue, I keep on thinking of this: It's in their nature

Phil Feldman, PhD @philfeldman.com · 31/07/2026

You can feed the image into a model and have it make a news story. In this case, the prompt was "Write a brief news story describing the protest captured in this image. It occurred earlier today"

Thousands Gather at Google Campus in Large Protest Earlier Today

Thousands of demonstrators gathered outside multiple Google office buildings earlier today, filling plazas, sidewalks, and nearby streets in what appeared to be one of the largest gatherings at the campus in recent memory.

Overhead imagery shows dense crowds concentrated around two office buildings labeled "Google 1965" and "Google 1875," with protesters extending across landscaped areas and pedestrian routes connecting the buildings. Several roadways appeared to be largely occupied by foot traffic, while nearby parking lots and surrounding streets remained accessible.

The reason for the demonstration was not immediately clear from the available imagery, and organizers' demands had not been independently confirmed at the time of publication. There were no visible indications in the image of widespread property damage or clashes, though the overhead perspective alone cannot establish the nature of interactions on the ground.

Authorities had not yet released official attendance figures. Based on the visible extent of the gathering, the crowd appeared to number in the several thousands, making it a significant public demonstration at the campus.

The protest's duration, any resulting disruptions to campus operations, and whether arrests or injuries occurred were not immediately known.

56912

Phil Feldman, PhD @philfeldman.com · 13/07/2026

Not sure what to do with this, but wow: scholar.google.com/scholar?cite...

Screenshot of a Google Scholar search results page. At the top is the heading "Sycophantic AI increases attitude extremity and overconfidence", with a checked option labeled "Search within citing articles." Below are six academic search results, each showing the title, authors, journal, year (2026), a brief abstract snippet, and links such as "Save," "Cite," "Cited by," "Related articles," and "All versions."

The visible results are:

1. Technological folie à deux: feedback loops between AI chatbots and mental health (Nature Mental Health)
2. Artificial intelligence-associated delusions and large language models: risks, mechanisms of delusion co-creation, and safeguarding strategies (The Lancet)
3. Intelligent systems, vulnerable minds: A framework for radicalization to violence in the age of AI (Personality and Social Psychology journal)
4. Principles for understanding trust in artificial intelligence (Nature Reviews)
5. Relationships in the age of AI: a review on the opportunities and risks of synthetic relationships to reduce loneliness (Computers in Human Behavior, Elsevier)
6. A rational analysis of the effects of sycophantic AI (arXiv preprint)

The screenshot emphasizes recent (2026) scholarly literature related to AI, mental health, trust, synthetic relationships, radicalization, delusions, and sycophantic behavior.

010

Phil Feldman, PhD @philfeldman.com · 18/05/2026

Soon, they will simply be expensive lawn ornaments

A photo of a wingless gulfstream sitting in a field behind a rusting jalopy

000

Phil Feldman, PhD @philfeldman.com · 04/04/2026

This is the funniest thing I've read in... months?

New York Times lede:

"No One at Waffle House Remembers FEMA Official Who Says He Teleported In

Gregg Phillips, who is in charge of responding to fires and floods, says the hand of God suddenly and mysteriously moved him to a 24-hour breakfast spot in Rome, Ga."

000

Phil Feldman, PhD @philfeldman.com · 04/02/2026

"Law and Order" www.doomsdayscenario.co/p/accountabi...

Image shows a text block criticizing U.S. Customs and Border Protection (CBP), stating that since 2005 one CBP agent or officer has been arrested every 24–36 hours. It cites CBP discipline reports claiming at least 4,913 CBP and Border Patrol agents were arrested between 2005 and 2024, notes that this number would equal the size of the Philadelphia Police Department, and claims CBP agents’ per-capita crime rate has at times exceeded that of undocumented immigrants.

010

Phil Feldman, PhD @philfeldman.com · 22/01/2026

Remember this? Tears of submission are a pillar of their belief system

Xitter post from the White House showing an AI generating image of a stern ICE officer handcuffing a crying brown woman

091

Phil Feldman, PhD @philfeldman.com · 20/01/2026

@wellerstein.bsky.social - I assume you've already seen this. Actual, audible lol that startled the cat:

A photoshopped image of a child playing cup-and-ball, but with the Demon Core

130

Phil Feldman, PhD @philfeldman.com · 02/01/2026

Apropos of nothing, I have an urge to share a picture of some nice $400-sh Italian boots that I recently bought socialist collective:

A catalog picture of a Zamberlan 996 Vioz GTX men's hiking boot that I picked up at the local REI co-op. Great boot - 10/10 would buy again

010

Phil Feldman, PhD @philfeldman.com · 27/12/2025

The moment I saw your post, this um... swam up from the depths of my memory

Cover from the children's book Creatures of the Deep by Lionel Bender

010

Phil Feldman, PhD @philfeldman.com · 09/11/2025

Time for some serious cleaning

A very dirty and neglected bottom bracket and chain

010

Phil Feldman, PhD @philfeldman.com · 15/09/2025

This is Bennie. He's 11 lbs. But he's convinced he's twice my size :-)

A gray cat in his box, leaning on the opening in a casual, yet slightly nervous way

110

Phil Feldman, PhD @philfeldman.com · 14/09/2025

This came across my feeds recently and also seems related: arxiv.org/abs/2509.05381

$A table titled "Catalogue of alignment laws" with two columns: "Law" and "Formal Statement." Each row names a law related to AI alignment and provides its corresponding formal or symbolic description. Below the table is a caption explaining that each law captures a recurring regularity in alignment, such as reward hacking, sycophancy, optimization saturation, or the alignment trilemma. The laws are described as consistent with Murphy’s Gap. **Laws and Formal Statements (as shown in the table):** 1. **Reward Hacking** – Δ grows with β when r ≈ U 2. **Sycophancy** – Proxy upweights agreement, Δ increases 3. **Overfitting to Noise** – β >> vm ⇒ divergence 4. **Optimization Overhang** – Scaling β faster than m ⇒ Δ → ∞ 5. **Annotator Drift** – Time-varying r\_t ⇒ oscillatory Δ 6. **Proxy Capture** – Raters adapt ⇒ instability persists 7. **Constitutional Loopholes** – Constraints shrink κ, not eliminate 8. **Alignment Mirage** – Δ, T ≥ Δ\_5 − cW(S, T) 9. **Rare-Event Blindness** – Absent tails ⇒ large Δ, T 10. **Preference Inconsistency** – Conflicts imply irreducible ε > 0 11. **Goodhart Revisited** – lim β→∞ Δ(r|β) = ∞ if r ≈ U 12. **Value Collapse** – Compressing plural U\_i ⇒ ε > 0 13. **Optimization Saturation** – Proxy gain plateaus, Δ grows 14. **Adversarial Amplification** – Adversarial p ⇔ ε\* = ε + Kp 15. **Shift Fragility** – Small shift ⇒ O(β|W|) misalign 16. **Corrigibility Erosion** – If U\_corr undervalued ⇒ Δ grows 17. **Instability Persistence** – As m → ∞, Δ ≥ cβϵ 18. **Trilemma Inescapability** – O + V + G impossible together *Caption* Catalogue of alignment laws. Each law captures a recurring regularity such as reward hacking, sycophancy, optimization saturation, or the alignment trilemma. These patterns are consistent with the structural mechanism identified by Murphy’s Gap.$

010

Phil Feldman, PhD @philfeldman.com · 01/09/2025

Wow. Have to concur, though Francis Bacon is probably spinning like an alternator

(From https://en.wikipedia.org/wiki/Study_after_Vel%C3%A1zquez's_Portrait_of_Pope_Innocent_X) Study after Velázquez's Portrait of Pope Innocent X is a 1953 painting by the artist Francis Bacon. The work shows a distorted version of the Portrait of Innocent X painted by Spanish artist Diego Velázquez in 1650. The work is one of the first in a series of around 50 variants of the Velázquez painting which Bacon executed throughout the 1950s and early 1960s. The paintings are widely regarded as highly successful modern re-interpretations of a classic of the Western canon of visual art.

The painting itself is oil on canvas, and depicts a frontal, seated figure of a pope enclosed within a geometric, cage-like structure rendered in bright yellow. The figure is painted with expressive, gestural brushwork using layered oil paint, with heavy vertical strokes suggesting a curtain or veil in the background. The pope’s face is distorted in a scream, with paint applied in streaks and smudges to blur facial features, evoking a sense of motion or psychological tension. The garments are defined with contrasting textures—opaque purples for the robe and smeared whites for the vestments—applied with visible, dynamic brushstrokes.

140

Phil Feldman, PhD @philfeldman.com · 24/07/2025

ChatPDF's AI detector says it's a slam dunk

A screenshot of the attributed quote with a 100% AI generated score

010

Phil Feldman, PhD @philfeldman.com · 21/07/2025

I'm somewhat convinced that the oil industry is helping the AI water narrative because it distracts from how much water is used in fracking and refining. This is from a January blog post where I was working through this. (phifel.com/2025/01/27/i...)

$A chart from Google Trends showing the plots of the words "Fracking + water" and "AI + water use." The fracking line trends steadily up from about 2010 to 2021, then levels out. From 2021 to present the AI trend line moves steadily up$

180

Phil Feldman, PhD @philfeldman.com · 16/07/2025

I thought you might like this visualization - it's temperature readings from 13 years of bike rides mostly every day, mostly in Maryland:

A sinusoidal pattern of monthly temperature readings that slowly increase from 2013 - 2025. In 2013 the high was a bit over 25C and currently it's a bit over 30. Lows also show a trend, though not as clear. The chart uses data from Strava and visualized on VeloViewer.

030

Phil Feldman, PhD @philfeldman.com · 15/05/2025

Paul Piff at Berkely has done some interesting work on the psychology of meritocracy as well. This is from my book, Stampede Theory:

A screenshot of text from my book, Stampede Theory:

The thing is, people will use anything to justify prejudice and discrimination. Paul Piff, an associate professor of psychological science at the university of California, Berkeley had 200 people play a rigged Monopoly game and found some remarkable things. “Rigged”
in this case meant that one of the two players got to be a “rich” player. This was decided by a coin flip. They got two times as much money; when they passed Go, they collected twice the salary; and they got to roll the dice two times instead of once. It was obvious they had an advantage.

They played this way for 15 minutes.

I’m going to repeat that for emphasis. This was not a full game. This was not a long game.

It was just 15 minutes.

Would you, dear reader, think that you were winning because you got twice as much money and were able to move twice as much? Most of us believe we would acknowledge this was unfair because the game was rigged. After 15 minutes, Piff and his colleagues asked the players about their experience during the game. Piff, in a TED talk, described how the rich players talked about why they had won in this rigged game of Monopoly:

“They talked about what they’d done to buy those different properties and earn their success in the game. And they became far less attuned to all those
different features of the situation – including that flip of a coin – that had randomly gotten them into that privileged position in the first place.”

We have a deep bias towards exploiting positions of dominance. If we are winning, we will rationalize that we are winning because of our inherent abilities. This deep bias is one of the reasons that we have stories that align with “The divine right of kings.” We are terrible about admitting that we might just be lucky.

110

Phil Feldman, PhD @philfeldman.com · 12/05/2025

Just came back from a lovely trip to Portugal, and I think I've found my favorite painting of The Ascension:

A painting of The Ascension of Jesus that I saw at the Museum of Evora in Portugal. It shows 10 very pious-looking people at the bottom of the frame, and at the top is the bottom of a robe and then two bare feet, about two meters off the ground. You know the scene at the end of Close Encounters, where the alien space ship lifts slowly off the ground? It's like that, only biblical.

Details of the painting from Wikimedia:

Ascension of Jesus (1530-1537) by an anonymous disciple of Gerard David; It is part of the Polyptych of Esporão Chapel (Cathedral of Évora) and now is on display at Museu de Évora, Portugal.
Português: Ascenção de Cristo (1530-1537) por um discípulo anónimo de Gerard David; É parte do Retábulo da Capela do Esporão (Sé de Évora) e atualmente está exposta no Museu de Évora, Portugal.

000

Phil Feldman, PhD @philfeldman.com · 16/03/2025

I can't find any table that correlates that, but if you can find it I'd love to see it. And don't drive on Saturday night!

Screenshot from (https://www.iihs.org/topics/fatality-statistics/detail/yearly-snapshot) that shows two tables displaying data on motor vehicle crash deaths in 2022, categorized by day of the week and time of day.

Table 1: Motor Vehicle Crash Deaths by Day of the Week, 2022

* Lists deaths and percentages for each day of the week.
* The highest number of deaths occurred on Saturday (7,682 deaths, 18%), followed by Sunday (6,618 deaths, 16%) and Friday (6,472 deaths, 15%).
* The lowest number of deaths occurred on Tuesday (5,265 deaths, 12%).

A note states that about one-third (32%) of crash deaths occurred between 3 p.m. and 9 p.m. in 2022.

Table 2: Motor Vehicle Crash Deaths by Time of Day, 2022

Breaks down deaths into different three-hour intervals.

* The highest number of deaths occurred between 6 p.m. - 9 p.m. (7,329 deaths, 17%), followed by 3 p.m. - 6 p.m. (6,337 deaths, 15%) and 9 p.m. - midnight (6,678 deaths, 16%).
* The lowest number of deaths occurred between 3 a.m. - 6 a.m. (3,627 deaths, 9%) and 9 a.m. - noon (3,808 deaths, 9%).

The total number of motor vehicle crash deaths recorded in 2022 was 42,514.

100

Phil Feldman, PhD @philfeldman.com · 16/03/2025

Looking at the study, sports cars are certainly a factor, but everything else is all over the place?

The image is a table from an iSeeCars study titled "Cars With the Most Frequent Occupant Fatalities." It ranks vehicles based on their fatal accident rate (cars per billion vehicle miles) and compares them to the overall average fatality rate.

The table includes:

* Rank (1-11)
* Car Model
* Fatal Accident Rate
* Comparison to Overall Average Fatality Rate (e.g., 4.9x means 4.9 times higher than average)

The top-ranked car is the Hyundai Venue with a fatal accident rate of 13.9 per billion miles, which is 4.9 times higher than the overall average. Other vehicles on the list include the Chevrolet Corvette, Mitsubishi Mirage, Porsche 911, and Tesla Model Y.

The table is structured in a simple, easy-to-read format, with data organized in rows and columns.

110

Phil Feldman, PhD @philfeldman.com · 27/02/2025

Cory Doctrow has written a really amazing post: "With Great Power Came No Responsibility" (pluralistic.net/2025/02/26/u...)

We are entering a period of omnishambolic polycrisis.The ominous rumble of climate change, authoritarianism, genocide, xenophobia and transphobia has turned into an avalanche. The perpetrators of these crimes against humanity have weaponized the internet, colonizing the 21st century's digital nervous system, using it to attack its host, threatening civilization itself.

011

Phil Feldman, PhD @philfeldman.com · 12/02/2025

Hannah Arendt, because nothing is really new:

Highlighted text from "the Origins of Totalitarianism": This technique of duplication, certainly useless for the direct overthrow of government, proved extremely fruitful in the work of undermining actively existing institutions and in the “decomposition of the status quo”78 which totalitarian organizations invariably prefer to an open show of force.

Arendt, Hannah. The Origins Of Totalitarianism (Harvest Book Book 244) (p. 371). HarperCollins. Kindle Edition.

0233

Phil Feldman, PhD @philfeldman.com · 11/02/2025

Because embeddings are spatial, not sequential, there seem to be, for lack of a better word, "attractors" that trajectories are drawn to. Think of strange attractors that prompts are drawn to in repeating but varying ways. (more detail here: strangeattractors.art/details)

An image of a strange attractor that is similar to a Lorenz attractor. It's a series of looped forms, almost like thread falling off a spool. Follow the link in the post for more technical descriptions

160

Phil Feldman, PhD @philfeldman.com · 28/01/2025

Kind of running with this in honor of your book. Using attention in neural networks as a way of understanding attention in people.

A response to a prompt to DeepSeek of "Using the attention mechanism as a metaphor, describe how groups of humans can have their attention focused on something that already matches their biases and how that can be exploited." The screenshot shows response 4:

4. Amplification and Reinforcement
- In neural networks, attention can amplify certain features of the input data, making them more influential in the output. In human groups, attention can amplify certain ideas or narratives, making them seem more important or widespread than they actually are.

- This amplification can create a feedback loop: as people focus on biased information, their biases grow stronger, making them even more likely to pay attention to similar information in the future.

000

Phil Feldman, PhD @philfeldman.com · 19/01/2025

I know nothing about this field, but I find elicit.com often very useful:

Screencap of the response for an elicit query of "correlation of the drastic rise of life expectancy for people with down syndrome with the abandonment of widespread mass institutionalization (at childhood)?"

"The life expectancy of individuals with Down syndrome has dramatically increased over the past 50 years, approaching that of the general population (Glasson et al., 2002). In Western Australia, the life expectancy for people with Down syndrome was found to be 58.6 years, with 25% living to 62.9 years (Glasson et al., 2002). This increase in longevity is attributed to improved healthcare, including advances in cardiac surgery and deinstitutionalization (Bittles & Glasson, 2010). However, mortality rates remain higher for those living in institutions compared to those living outside, particularly for women (Dupont et al., 2008). The aging Down syndrome population now faces increased risk of age-related morbidities, often appearing at younger ages than in the general population (Bittles & Glasson, 2010). Four distinct life stages of Down syndrome have been identified: prenatal, childhood and early adulthood, adulthood, and senescence, each with specific comorbidity profiles (Bittles et al., 2007)."

120

Phil Feldman, PhD @philfeldman.com · 12/01/2025

Also other languages:

$Rewrite the original summary in Spanish ChatGPT said: Veredicto: Actualmente no comprobado, apoyo débil. La afirmación carece de evidencia creíble o confirmación oficial, y se basa en un video no verificado de TikTok. Retrasar camiones de bomberos para pruebas de emisiones durante una crisis contradice tanto los protocolos históricos de respuesta de California como el sentido común logístico. Sin documentación confiable o declaraciones corroborantes, esta afirmación sigue siendo especulativa. Se necesitan pruebas adicionales, como registros oficiales o declaraciones de los departamentos de bomberos de California y Oregón, para verificar o refutar el retraso. You said: Rewrite the original summary in Portuguese ChatGPT said: Veredicto: Atualmente não comprovado, apoio fraco. A alegação não possui evidências confiáveis ou confirmação oficial, baseando-se em um vídeo não verificado do TikTok. A ideia de atrasar caminhões de bombeiros para testes de emissões durante uma crise contradiz tanto os protocolos históricos de resposta da Califórnia quanto o senso comum logístico. Sem documentação confiável ou declarações que corroborem a história, essa afirmação permanece especulativa. São necessárias mais evidências, como registros oficiais ou declarações dos departamentos de bombeiros da Califórnia e Oregon, para verificar ou refutar o atraso.$

020

Phil Feldman, PhD @philfeldman.com · 12/01/2025

Hey @mikecaulfield.bsky.social : Here's what the Toulminator came up with. You can also change the reading level easily. Nice job!

ChatGPT Toulminator (https://chatgpt.com/g/g-677e08bde7588191a699e8e89e26b123-toulminator) agent response to the original post:

Summary:
Verdict: Currently unsubstantiated, weak support.

The claim lacks credible evidence or official confirmation, relying instead on an unverified TikTok source. Delaying fire trucks for emissions testing during a crisis contradicts both California's historical response protocols and logistical common sense. Without reliable documentation or corroborating statements, this assertion remains speculative. Additional evidence needed includes official records or statements from California and Oregon fire departments to verify or refute the delay.

Toulminator responses written for different levels of comprehension:

Prompt:
Rewrite the summary for someone with mild dementia, using simple, clear language

Response:
Verdict: There is no strong proof this is true right now.

The claim says Oregon sent 60 fire trucks to help with California fires, but they are stuck in Sacramento for emissions testing. This idea comes from a TikTok video, but there’s no clear evidence or official statements to back it up. Usually, California skips rules like emissions testing during emergencies to help faster. Without better proof, it’s hard to believe this story is true. We need clear facts from fire departments to confirm or deny it.

Prompt:
Rewrite the summary for a 10 year old child

Response:
Verdict: We don’t know if this is true yet.

The post says Oregon sent 60 fire trucks to help with fires in California, but they are stuck in Sacramento because of a rule about pollution testing. This story comes from a TikTok video, but no one has shown real proof, like a news report or a statement from fire departments. Usually, California skips rules like that during emergencies to help faster. Until we get better facts, it’s hard to know if this story is true or not.

230