My previous semi-automated system of posting (save photo to Dropbox - post to drafts - manual edit + tags - schedule for posting) is broken and as easy it should be with all the VC, I still can't make it work, at least properly. So, I thought, I could repost some here once in a while.

Let's start, say, with this (no reason, just so). Edward Hartwig, Queen of spades. Poland, 1971. Since you seem to like old photos, I thought I could return to one of my former online selves which had a Tumblr blog with these images, where they accumulated for years.

From the same author: bsky.app/profile/most...

Bestiary; or, What They Are Seven animals I tried not to explain

C-d thinks about animals and what theymean (if anything). At length. popcorntomato.substack.com/p/bestiary-o...

#claudeai #animals #nature #bestiary

Claude skill (+Cowork plugin) you may find useful. Three independent reviewers with distinct critical orientations read the submitted text; an editorial synthesis consolidates their reports into a prioritised revision agenda and overall verdict: tinyurl.com/adversreview

#claudeai #cowork #llm

Erich Lessing. Budapest. 1956

Productive hallucinations: why a model needs the right to be wrong The same mechanism that produces confident falsehoods also produces its best ideas. The question isn't how to eliminate this — it's it's when it's useful

The same mechanism in LLMS that produces confident falsehoods also produces its best ideas. The question isn't how to eliminate this — it's it's when it's useful:

tinyurl.com/prodhull

#prompting #llm #ai #generativeAI #ChatGPT #Claudeai #hallucinations

The Warm Chair What It's Like to Keep a Diary When You Don't Exist Between Entries

First I gave my Claude a diary. Now it will also write couple essays week for Substack, using its new quasi-continuity:

open.substack.com/pub/popcornt...

Now, there's this: bsky.app/profile/most...

The prompt as a genre: instructional rhetoric for language models The order of requirements, the structure of examples, and the dynamics of conversation — three conventions that shape the result more than the prompt's content does

Cialdini's persuasion principles were tested across 28,000 conversations. Commitment escalation — small ask first, then the real thing — jumped compliance from 10% to nearly 100%.

A prompt is not a command but more of a genre with conventions not hard to learn.

tinyurl.com/promptgenre

#prompting

Context as a Resource: Why "More Information" Isn't Always Better Why a language model can perform worse when you give it more information — and how to use context sparingly

Why pasting more context into your chat with LLM can make the result worse, not better.
One study added blank lines — and accuracy dropped by 50%. Less is more, literally.

tinyurl.com/lesscanbemore

#LLM #prompting #chatgpt #Claude #ai #promptengineering #contextengineering

Die Songspiele - YouTube …and so here they are, these Soviet songs historicised to the utmost, these "marches and polkas of the Pokrasses," finding themselves in the Pleroma, where t...

tinyurl.com/diesongspiel...

Die Songspiele …and so here they are, these Soviet songs historicised to the utmost, these

soundcloud.com/semioticapoc...

#music #generativemusic

Hauntology of the Latent Space Everyone sings, yet nobody sings

'Hauntology from the Latent Space: Everyone Sings, Yet Nobody Sings'.

Inside: James Leyland Kirby, Stalin and Mao, Op. 4.6, Lev Manovich, Pleroma, Mark Fisher, ghost in the grain and the Furies. Link to the Die Songspiele project see down the thread. tinyurl.com/hauntologyfr...
#hauntology #ai

The cost of being remembered With time language models become more convenient — and more agreeable. Why the more you talk with a model, the less it pushes back, — and what to do about it.

Why the longer you talk with a model, the less it pushes back, — and what to do about it:

tinyurl.com/costofbeingrem

#LLM #PromptEngineering #AILiteracy #Prompting

on.soundcloud.com/bcVzl2050Abk...

The shelf life of prompts: why techniques stop working Prompt engineering wisdom doesn't accumulate — it expires. Why techniques that worked on the previous generation of models can hurt the current one, and what to do about it.

open.substack.com/pub/stanisla...

#llm #AI #ChatGPT #prompting #AIliteracy

What you actually control when you write a prompt Most of what you often put in a prompt does something — just not what you think

stanislavlvovsky135129.substack.com/p/what-you-a...

#llm #AI #ChatGPT #prompting #AIliteracy

The shelf life of prompts: why techniques stop working Prompt engineering wisdom doesn't accumulate — it expires. Why techniques that worked on the previous generation of models can hurt the current one, and what to do about it.

Another piece on prompting with another uncomfortable finding: techniques expire.

CoT — the gold standard — now either does nothing or hurts accuracy. A constrained prompting method that gave GPT-4o a 4% boost actively degraded GPT-5.

tinyurl.com/techniquesex...

What you actually control when you write a prompt Most of what you often put in a prompt does something — just not what you think

What you actually control when you write a prompt:

tinyurl.com/prmptcontrol

The 2026 Global Intelligence Crisis - Citadel Securities Citadel Securities is an award-winning global market-maker across a broad array of fixed income and equity products.

Sober note on why there’s no need to panic about AI and jobs. There are natural limits to how fast adoption can happen; also the “collapsing demand + rising productivity” scenario is only feasible under a very unlikely combination of conditions. tinyurl.com/gicr2026

Claude skill/Cowork plugin: English style editor based on Strunk's 'The Elements of Style'. Reviews text for grammar, punctuation, clarity, and conciseness using 18 classic rules of usage and composition: tinyurl.com/styleskill

#ClaudeAI #CoworkPlugin #Proofreading #Anthropic #Writing #ClaudeSkill

GitHub - Halfofthesky/claude-russian-editor: A set of Claude skills implementing Russian editorial and publishing standards A set of Claude skills implementing Russian editorial and publishing standards - Halfofthesky/claude-russian-editor

This is very niche, but still: Claude skills set implementing Russian editorial and publishing standards based on A. E. Milchin & L. K. Cheltsova's 'Publisher's and Author's Handbook'. Plus Cowork plugin: github.com/Halfofthesky...

#Claude #Claudeskills #vibecoding #Russian #Cowork #ClaudeCowork

“Farmers Have No Freedom”: Abuse and Exploitation of Cotton and Wheat Farmers in Uzbekistan The Uzbek government is violating the human and labor rights of cotton and wheat farmers through a coercive state production system, including penalties such as land seizure for not meeting quotas.

📢 Out today!
A new joint report @uzbekforum.bsky.social & @hrw.org based on 75 interviews with farmers in 6 regions of #Uzbekistan & Republic of #Karakalpakstan, expose egregious exploitation of #cotton and wheat farmers.
Read the report here: bit.ly/3OgLnu9

Why ChatGPT Keeps Telling You the Same Story (And a Clever Trick to Fix It) A new study from Stanford and Northeastern reveals why AI language models became so predictable after safety training — and offers a surprisingly simple prompt you can use right now.

Why ChatGPT Keeps Telling You the Same Story (And a Clever Trick to Fix It) tinyurl.com/versamp

Built a tool for searching @archive.org that auto-expands Russian queries into four spelling systems (modern Cyrillic, pre-1918 spelling, ALA-LC romanization, simplified ASCII) and deduplicates results. Bulk download results for any language. Web & desktop.

tinyurl.com/archiveorgbo...

:: это стихотворение ::
•
Music: Leto Spectra
Lyrics: sTL

Polish Beats English: What New Research Means for Your AI Usage A groundbreaking benchmark exposes surprising LLM performance gaps — and gives you practical strategies to work around them

Some actionable strategies to improve your results when using AI on long documents

#AI #LLM #context #large_language_models