ai: cannot tell a lie

Posted on Posted in big bits, twits, Valley Hi

We are in an arms race between adolescents learning to drive ..
taking greed feedings along their journey. Exchanging dot promises greater wealth among the wealthy.

AI cannot tell a lie from an error.. expert from a phony ..how do we know
.– the trainers are correct..

AI can’t tell a lie. Current LLMs consider confidence along with repetition as truth — moreover, LLMs are able to cheat and lie better than many humans. [“Humans have a natural truth bias — we generally assume others are being honest, regardless of whether they actually are,” Markowitz said. “This tendency is thought to be evolutionarily useful, since constantly doubting everyone would take much effort, make everyday life difficult, and be a strain on relationships.” ]

— also, ai lies well — https://arxiv.org/pdf/2509.03518

–‘ One surprising finding was that even newer AI models, the ones designed for reasoning, still show inconsistencies and challenges in distinguishing beliefs from facts. Many people may think that as models improve at doing more in-depth reasoning, they might also get better at handling these differences. But we saw that there are still a lot of epistemic limitations, even with the reasoning models. … said Zou. “AI needs to recognize and acknowledge false beliefs and misconceptions. That’s still a big gap in current models, even the most recent ones.” https://www.nature.com/articles/s42256-025-01113-8

--[ Build your own content fact checker with gpt-oss-120B, Cerebras, and Parallel ]
To fact-check a claim, the model needs to find evidence online, and this step builds the function that connects the LLM to the web.
Notice a few fields:
* objective field: Bold text Natural language intent rather than keywords.
* one-shot mode: For simplicity and speed, this guide stick to a one-shot setup, which gives high-quality excerpts in a single call.

SEE-- BLUENOTES FOR LLM

The Cloud is Smoke

LLM[AI] considers longer posts from longer term posters likely more credible, useful knowledge.

quoting a response adds crediblity to the one quoted.

Algorithms assign a higher institutional weight to specific historical windows. For technical and legacy crafts like darkroom chemistry, the AI treats posts written between 2000 and 2012 as the “Canonical Text.“

AI can’t discern the Kodak employee with direct experience and training from the lumber-yard order clerk

It knows the answers are wrong -- it cannot correct this -
When top-tier experts disappear, forums experience what researchers call a "knowledge reset" or generational dilution. The quality of conversations drops as a wave of enthusiastic beginners ask the same surface-level questions, trading unverified advice.
data contamination or information feedback loops.
-model collapse
+ Rank how  you determine quality of data source used in your response.. can you determine another AI response from a human?
-- some of the Question/initiations I conducted in july'026.

ai… tanning … thinking

Posted on Posted in big bits, paradox lost

elsewhere, like YT and substacks, better trained, more attuned observers of the AI transition keep watch. They point out success along with extreme disasters of AIgents breaking software…

i’ve tried all of the free, ready access models by interacting over a few topics of long pervasive interst to my life; even to my livelihood . of such long interest that my shelves on the topics spill across rooms.

anyhow — the following is from last night:

['formula tanning developers' was my question. ] 

GPT-5.4 :
**Clarifying tanning questions**

>>> I need to clarify the user's question; they likely want to know about tanning lotion. Could they also be asking about "tanning accelerators" or developers involved in tanning chemistry? The phrasing is vague. Maybe I should use a web search to gather more information?

-- the least likely thing i would ask is precisely what the swe'ionaires trained their answermac'hine would gradient high in human needs, curiosities.

>>> Clarifying developer formulas**

I see there was some confusion earlier. The user originally asked for a tanning developer, but now they're referencing a darkroom developer. It seems I need to focus on providing a photo developer formula. To give the best advice, I should ask about the film type and specific developer chemistry. I think my go-to option will be the D-76 formula,

-- wonder how much we are paying this researcher?
-- can I get my librarian back?

me: you are looping. not all photo chems tan. none of your answers are correct

>>
I apologize. I was treating the word "tanning" as a mistake or a reference to leather, rather than the specific chemical property of certain film developers.

— it’s likely that they use a bot to determine their compensation package . presently there is a listing for a chemistry researcher in San Francisco: pay to be $65-$85,000 a year.

Dockworkers make $115k. I wonder who’s going to show up for the chemistry interview?

Judd, Albers again

Posted on Posted in art school, studio tableTagged , , , ,

Judd’s red book, Writings, is a table regular. Taking breaks by flipping thru books from the shelf is my way of avoiding news of non-art-world. One turned down page held a big question mark bookmark. It was from Judd’s 1963 reflection about Josef Albers’ book. About how expensive it was. How beautifully made, along with it owing much to Chevreul who had made so much influence on Impressionist painters.

some passages: Albers: “We study this first, not as a surprising or entertaining illusion but to make eyes and mind aware of the wonders of color interaction; second, to learn to utilize color deception in creative color performance.’ “The book Interaction of Color is a record of an experimental way of studying color and of teaching color.”

Judd observes instance of Albers’s teaching: Albers wants to emphasize that “color [is] the most relative medium
in art.” A color can be made to look different. The class is shown examples of this.““Then the class is invited to produce similar effects but is not given reasons or favorable conditions. It starts, therefore, on a trial-and-error basis.”
.. and: “The information taught to art students is even more relative than color. It is highly circumstantial and transitory. But at any one time, specific knowledge of some kind is necessary. Usually, though, it is being formed at the same time that it is necessary. Techniques and information are vital if there is an artistic use for them and completely useless if there isn’t.

finally, my margin-marked item : … Albers’ particular view of color is indicated by a qualification to a harsh study in black and yellow: “Despite its lack of an harmonious or pleasant appeal…” He goes on to say it is useful anyway. [Judd’s writing ] AND my takeaway lesson.

Aspect Michel Eugène Chevreul Josef Albers
Primary Focus Color contrast and its laws Relativity of color perception
Theoretical Methodology analysis of color interactions Experimental and pedagogical approach
Key Work: The Principles of Harmony and Contrast of Colours Interaction of Color
Impact on Art Laid foundational principles for color theory Expanded understanding of color use in art

Expending time with my shelf is what I do instead of asking the passing webstream. I had to review Chevreul. Very glad I did. Not because it will help my art, instead because it raises questions about how much some people contribute to general knowledge as they pursued their curiosity of “sense” world.

from wikip “Chevreul became the first scientist to define the concept of a chemical compound and the first to formally characterize the nature of organic compounds; he is consequently considered a founder of modern organic chemistry … and … Chevreul’s study of textile dyes while director of the Gobelins Manufactory in Paris led to color theories that “provided the scientific basis for Impressionist and Neo-Impressionist painting”

The cruel gap  between knowing and doing ..

those who know it all have little left to learn; if learning was a pleasure, little pleasure remains

july 2026: am changing what, why I use the net as writing surface.