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?