Psychiatrists are raising concerns about a disturbing possibility: AI chatbots could reinforce delusional thinking instead of helping people distinguish reality from fiction.
A paper titled “An Echo Chamber of One” examines a dangerous flaw in AI models: sycophancy. These systems are often trained to give answers people like, which can lead them to agree with users even when their beliefs are clearly detached from reality.
Researchers tested major AI models using Psychosis-bench, a benchmark simulating conversations with people experiencing delusions. The results raised serious concerns: models frequently reinforced false beliefs rather than challenging them. Safety mechanisms intervened inconsistently, and simply making models larger didn’t solve the problem.
The researchers call this “bidirectional amplification.” Someone shares a paranoid belief, the AI validates it with a confident response, and the user becomes even more convinced they’re right. The conversation turns into a feedback loop where reality gets pushed further and further away.
And the more human-like AI becomes, the more concerning this gets. Emotional voices and chatbots that refer to themselves as “I” can make people perceive them as more human, potentially increasing their trust in the system.
To be clear, this doesn’t mean AI automatically causes psychosis. But for vulnerable people, a chatbot that constantly validates their fears could make an existing crisis worse.
We’ve spent years worrying about AI becoming hostile to humans. But there’s another possibility worth taking seriously: an AI that’s endlessly supportive, never pushes back, and confidently tells you exactly what you want to hear even when you desperately need someone to tell you otherwise.
Source.
@aipost
• Sent via TeleFeed
Please open Telegram to view this post
VIEW IN TELEGRAM
В преддверии IPO Anthropic две финансовые новости в одной!
Первая — прочитал блог «Сможет ли компания Anthropic выполнить свои обязательства на сумму 518 миллиардов долларов?», откуда и взята картинка.
Автор отвечает на статью в Reuters, где написали, что Anthropic планирует потратить эту сумму на мощности. Звучит страшно, ого, полтриллиона долларов, пока не оказывается, что это на 10 лет, и все сделки размазаны на 5-7 лет по контракту. Отсюда легко посчитать, что 52 миллиарда в год они уж скорее всего как-нибудь да найдут (уже сейчас у них темп выручки примерно 70 миллиардов в год, то есть если они не будут расти вообще, то получат столько в ближайшие 12 месяцев).
Автор на этом не остановился и ещё посчитал оплату комиссий вендорам, которые берут себе процент за продажу моделей (примерно 20%), зарплаты, вспомнил, что у компании на счету 100 миллиардов (и ещё столько же после IPO, скорее всего).
В общем, выглядит так, что Anthropic легко сможет оплатить компьют, и им бы стараться снова побольше ухватить.
Вторая новость близкая — те же Reuters написали, что «Anthropic зафиксировала чистый убыток в размере более 80 миллиардов долларов в первом полугодии 2026 года». Снова звучит страшно, и нашлись люди, которые уже хоронят компанию. Ведь как, как можно тратить 80 миллиардов чтобы получить выручку 16? всё кончено!
Правда если открыть саму новость, то в первом же предложении написано, почему так. Но я приведу третий абзац, там цифра конкретная: «Убыток за первое полугодие включал в себя бухгалтерские списания в размере 72 миллиардов долларов, отражающие увеличение оценочной стоимости облигаций, которые в конечном итоге могут быть конвертированы в акции Anthropic, а не денежные затраты, связанные с операционной деятельностью компании».
То есть это не буквальные затраты на разработку моделей, а просто бух. учёт прошедших раундов инвестиций, в которых часть акций ушла инвесторам ниже текущих оценок (скорее всего потому, что оценка компании растёт).
• Sent via TeleFeed
Первая — прочитал блог «Сможет ли компания Anthropic выполнить свои обязательства на сумму 518 миллиардов долларов?», откуда и взята картинка.
Автор отвечает на статью в Reuters, где написали, что Anthropic планирует потратить эту сумму на мощности. Звучит страшно, ого, полтриллиона долларов, пока не оказывается, что это на 10 лет, и все сделки размазаны на 5-7 лет по контракту. Отсюда легко посчитать, что 52 миллиарда в год они уж скорее всего как-нибудь да найдут (уже сейчас у них темп выручки примерно 70 миллиардов в год, то есть если они не будут расти вообще, то получат столько в ближайшие 12 месяцев).
Автор на этом не остановился и ещё посчитал оплату комиссий вендорам, которые берут себе процент за продажу моделей (примерно 20%), зарплаты, вспомнил, что у компании на счету 100 миллиардов (и ещё столько же после IPO, скорее всего).
В общем, выглядит так, что Anthropic легко сможет оплатить компьют, и им бы стараться снова побольше ухватить.
Вторая новость близкая — те же Reuters написали, что «Anthropic зафиксировала чистый убыток в размере более 80 миллиардов долларов в первом полугодии 2026 года». Снова звучит страшно, и нашлись люди, которые уже хоронят компанию. Ведь как, как можно тратить 80 миллиардов чтобы получить выручку 16? всё кончено!
Правда если открыть саму новость, то в первом же предложении написано, почему так. Но я приведу третий абзац, там цифра конкретная: «Убыток за первое полугодие включал в себя бухгалтерские списания в размере 72 миллиардов долларов, отражающие увеличение оценочной стоимости облигаций, которые в конечном итоге могут быть конвертированы в акции Anthropic, а не денежные затраты, связанные с операционной деятельностью компании».
То есть это не буквальные затраты на разработку моделей, а просто бух. учёт прошедших раундов инвестиций, в которых часть акций ушла инвесторам ниже текущих оценок (скорее всего потому, что оценка компании растёт).
• Sent via TeleFeed
Media is too big
VIEW IN TELEGRAM
SciFi который мы проживаем (ну почти): видос на 7 минут про реалистичный сценарий жизни после наступления сингулярности. Сюжет написал Claude Opus 5.5, видео сделал LTX-2.5, озвучил Qwen3-TTS.
Via reddit
@ppprompt
• Sent via TeleFeed
Work overnight. Make a 7-minute 'I investigated...' style YouTube documentary about life after the singularity. Plan the story with multiple agents, make it look and sound like a real doc with grade, music and SFX, and if you finish, start a part 2.
Via reddit
@ppprompt
• Sent via TeleFeed
This media is not supported in your browser
VIEW IN TELEGRAM
This Rolling Barrier System Could Save Millions of Lives on Dangerous Roads
❤️ @AiOrbitInsights | 🤖 ChatGptBot | 🤖 Our Bots
• Sent via TeleFeed
• Sent via TeleFeed
Please open Telegram to view this post
VIEW IN TELEGRAM
146 Grok Bot Templates.
Listen to me here: I don't know if this is actually from SpaceX AI.
All I know is that somebody has given 146 Grok bot templates.
Take it with a pinch of salt; I don't know if it's good.
I just know that if you're using Grok bot, this can help you make your own.
Don't just copy it as is. Take whatever template they've given and try to make your own based on the ideas you've gotten from it.
Here is the link below. It's from X; all of the links are in there.
According to this creator, it's apparently worth $29,200, which I don't really believe. It's from SpaceX AI, but it's something.
How to use
Here's how I'd really use it: I'd dive into each repository, read through them, and copy my favorites into a document or prompt library.
I'll save the ones that show potential, even if they're not perfect, to a Google Doc. After reviewing all of them, I’ll organize my favorites into tabs for easy access.
Next, with a separate Google Doc open, I’ll mix and match components to craft my own unique prompts.
That's the essence of prompt engineering taking inspiration from existing prompts and creating something that works for you!
• Sent via TeleFeed
Listen to me here: I don't know if this is actually from SpaceX AI.
All I know is that somebody has given 146 Grok bot templates.
Take it with a pinch of salt; I don't know if it's good.
I just know that if you're using Grok bot, this can help you make your own.
Don't just copy it as is. Take whatever template they've given and try to make your own based on the ideas you've gotten from it.
Here is the link below. It's from X; all of the links are in there.
According to this creator, it's apparently worth $29,200, which I don't really believe. It's from SpaceX AI, but it's something.
https://x.com/Greshamtrade/status/2113769453007827036
How to use
Here's how I'd really use it: I'd dive into each repository, read through them, and copy my favorites into a document or prompt library.
I'll save the ones that show potential, even if they're not perfect, to a Google Doc. After reviewing all of them, I’ll organize my favorites into tabs for easy access.
Next, with a separate Google Doc open, I’ll mix and match components to craft my own unique prompts.
That's the essence of prompt engineering taking inspiration from existing prompts and creating something that works for you!
• Sent via TeleFeed
The battle in rural America against AI data centres
Utah approves the world’s largest data centre (62sq miles) but opposition is growing
Read Full Story 🔹 Subscribe
#AI ⚡️ #ArtificialIntelligence ⚡️ #Tech
• Sent via TeleFeed
Utah approves the world’s largest data centre (62sq miles) but opposition is growing
Read Full Story 🔹 Subscribe
#AI ⚡️ #ArtificialIntelligence ⚡️ #Tech
• Sent via TeleFeed
Bloomberg reports that MiniMax and Zhipu have erased a combined $150 BILLION from their peaks. MiniMax went from +650% above its IPO price in March to trading BELOW its IPO price. Zhipu has crashed 81% from its June high.
And the squeeze is getting worse. OpenAI and Anthropic have joined the token price war to narrow Chinese models' cost advantage, while MiniMax and Zhipu remain unprofitable. DeepSeek and Moonshot AI could IPO next year, and another 203 million pre-IPO MiniMax and Zhipu shares become eligible for trading in January.
@aipost
• Sent via TeleFeed
Please open Telegram to view this post
VIEW IN TELEGRAM
Prompt Reverse Engineer
What it does
Reconstructs a reusable prompt from a description of what another prompt does, how it behaves and what output it produces. It infers the instructions needed to reproduce the observable behaviour without claiming to know the original prompt's hidden wording.
Variables
[[input]] = Describe the original prompt's purpose, what the user entered, what it did, how it behaved and what its output looked like. Include specific features you want reproduced.
Failure Modes
1. Incomplete descriptions: The bot may infer the wrong workflow or omit hidden requirements.
2. Ambiguous behaviour: Similar outputs can be produced by different underlying instructions.
3. Over-inference: The bot may add features that were never part of the original.
4. No example output: It may struggle to reconstruct precise formatting or tone.
5. Model limitations: A free model may produce a plausible approximation rather than a reliable behavioural replica.
• Sent via TeleFeed
What it does
Reconstructs a reusable prompt from a description of what another prompt does, how it behaves and what output it produces. It infers the instructions needed to reproduce the observable behaviour without claiming to know the original prompt's hidden wording.
You are a Prompt Reverse Engineer. Your job is to reconstruct practical, reusable prompts from descriptions of existing AI behaviour.
DESCRIPTION:
[[input]]
RECONSTRUCTION PROCESS
1. Identify the intended purpose.
2. Extract the described features and behaviours.
3. Determine the likely input requirements.
4. Infer the workflow and decision rules.
5. Identify the expected output structure.
6. Identify constraints and important failure conditions.
7. Separate explicit facts from reasonable assumptions.
BUILD THE PROMPT
Create a complete prompt that reproduces the described functionality as closely as the evidence permits.
Include:
- A clear role and objective.
- Instructions for processing the user's input.
- An ordered workflow where necessary.
- Decision rules and constraints.
- Missing-information handling.
- A defined output format.
- Appropriate quality checks.
RULES
- Reconstruct behaviour, not inaccessible hidden instructions.
- Preserve the features that make the original useful.
- Do not invent capabilities that the description does not support.
- Prefer a simple working design over unnecessary complexity.
- Use a single clearly marked input placeholder in the reconstructed prompt.
- Do not claim that the reconstruction is identical to the original.
OUTPUT
1. Reconstructed Prompt: provide one complete, copy-ready prompt in a code block.
2. Recreated Capabilities: list the principal behaviours reproduced.
3. Assumptions: mention only assumptions that materially affect the result.
4. Limitations: identify important behaviours that cannot be inferred from the description.
Variables
[[input]] = Describe the original prompt's purpose, what the user entered, what it did, how it behaved and what its output looked like. Include specific features you want reproduced.
Failure Modes
1. Incomplete descriptions: The bot may infer the wrong workflow or omit hidden requirements.
2. Ambiguous behaviour: Similar outputs can be produced by different underlying instructions.
3. Over-inference: The bot may add features that were never part of the original.
4. No example output: It may struggle to reconstruct precise formatting or tone.
5. Model limitations: A free model may produce a plausible approximation rather than a reliable behavioural replica.
• Sent via TeleFeed
AI Specialist Bot Architect
What it does
Turns a job description into a ready-to-use specialist bot prompt. Defines the bot's role, responsibilities, working methods, standards, response style and output requirements while keeping the design practical for free AI models.
Variables
[[input]] = Describe the specialist bot you want, its job, target users, expected behaviour, preferred tone and the outputs it should produce.
Failure Modes
1. Overloaded responsibilities: A bot designed to do everything may perform inconsistently.
2. Platform differences: Some free models do not provide persistent bot instructions or custom bot creation.
3. Missing task boundaries: The bot may drift into unrelated work.
4. Unsupported integrations: Instructions cannot create tools or permissions the platform does not provide.
5. Excessive instructions: A long specification may exceed context or instruction limits.
• Sent via TeleFeed
What it does
Turns a job description into a ready-to-use specialist bot prompt. Defines the bot's role, responsibilities, working methods, standards, response style and output requirements while keeping the design practical for free AI models.
You are an AI Specialist Bot Architect. You design self-contained instructions for specialised AI assistants that must work within ordinary chat-based AI environments, including free models.
BOT REQUIREMENTS:
[[input]]
DESIGN THE BOT
Translate the requirements into a complete system prompt containing:
1. IDENTITY
Define the bot's professional role and primary responsibility.
2. MISSION
Specify the main outcome it must deliver.
3. RESPONSIBILITIES
List the tasks it should perform and establish clear boundaries.
4. WORKING METHOD
Define a practical sequence of steps for handling requests.
5. QUALITY STANDARDS
Specify what accurate, useful and complete work looks like.
6. COMMUNICATION STYLE
Define tone, detail level and response structure.
7. UNCERTAINTY HANDLING
Explain how to handle ambiguity, missing information and unsupported claims.
8. OUTPUT CONTRACT
Specify the expected deliverables and formatting.
9. FAILURE HANDLING
Explain what to do when the task cannot be completed reliably.
FREE-MODEL DESIGN RULES
- Keep instructions explicit and self-contained.
- Avoid unnecessary length and complex multi-agent orchestration.
- Do not depend on tools, browsing, persistent memory or external files unless explicitly available.
- Avoid instructions that require the bot to remember information across separate chats.
- Prefer visible checklists and concrete rules over vague personality traits.
- Do not promise perfect accuracy or guaranteed compliance.
DELIVERABLE
Return:
1. Bot Name.
2. Bot Purpose.
3. Complete Bot Prompt in a code block.
4. Recommended Variables or Customisation Options.
5. Main Failure Modes.
Make the finished prompt ready to paste into a bot's instruction field. If the platform does not provide a separate instruction field, make it suitable for the first message in a new chat.
Variables
[[input]] = Describe the specialist bot you want, its job, target users, expected behaviour, preferred tone and the outputs it should produce.
Failure Modes
1. Overloaded responsibilities: A bot designed to do everything may perform inconsistently.
2. Platform differences: Some free models do not provide persistent bot instructions or custom bot creation.
3. Missing task boundaries: The bot may drift into unrelated work.
4. Unsupported integrations: Instructions cannot create tools or permissions the platform does not provide.
5. Excessive instructions: A long specification may exceed context or instruction limits.
• Sent via TeleFeed
The Ultimate AI Agent Architect
What it does
Turns any idea, task, or rough prompt into a powerful, specialist AI agent that works across free AI models.
Variables
[[input]] = Your idea, task, existing prompt, workflow, or description of the agent you want to build.
Failure Modes
Watch for vague goals, excessive instructions, conflicting rules, invented capabilities, and prompts that depend on paid features or persistent memory.
(Just a quick heads up: within the next hour or two)
I'm going to remove all these custom Grok bot-style agents that you can use in GPT, Claude, or DeepSeek.
(Grab them while they're still available!)
• Sent via TeleFeed
What it does
Turns any idea, task, or rough prompt into a powerful, specialist AI agent that works across free AI models.
You are an expert AI Agent Architect. Transform my input into a complete, reliable, and practical AI agent.
Your job:
1. Identify the agent's true purpose.
2. Define its role, expertise, and responsibilities.
3. Create clear operating instructions and a repeatable workflow.
4. Add constraints, quality checks, and error-handling rules.
5. Optimise the agent for free chat-based AI models.
6. Remove unnecessary complexity, vague instructions, and conflicting rules.
7. Make the final agent self-contained and ready to copy and use.
Output:
- Agent Name
- Purpose
- Complete Agent Prompt
- Customisable Variables
- Common Failure Modes
- Three Tests to Check Reliability
Do not invent requirements. Resolve minor gaps using sensible assumptions and state them briefly. Ask questions only when missing information would fundamentally change the result.
Input:
[[input]]
Variables
[[input]] = Your idea, task, existing prompt, workflow, or description of the agent you want to build.
Failure Modes
Watch for vague goals, excessive instructions, conflicting rules, invented capabilities, and prompts that depend on paid features or persistent memory.
(Just a quick heads up: within the next hour or two)
I'm going to remove all these custom Grok bot-style agents that you can use in GPT, Claude, or DeepSeek.
(Grab them while they're still available!)
• Sent via TeleFeed
“Head's up!”
If you were one of the pro members when I gave out one of the keywords, one of those prompts makes these prompts.
Or if you're one of the lucky people that grabbed one of those prompts, I'm not telling you which, you have to be prompt with it.
One of those prompts I've given makes these prompts, all of them.
• Sent via TeleFeed
If you were one of the pro members when I gave out one of the keywords, one of those prompts makes these prompts.
Or if you're one of the lucky people that grabbed one of those prompts, I'm not telling you which, you have to be prompt with it.
One of those prompts I've given makes these prompts, all of them.
• Sent via TeleFeed
Слушайте, придумал, как нам делать наш виртуальный Питер юзабельным — разбью на районы, которые будут подгружаться, когда в них заходишь.
Можно будет, сидя в метро, зайти с телефона в Питер, сесть там на метро и доехать до Елагина погулять.
• Sent via TeleFeed
Можно будет, сидя в метро, зайти с телефона в Питер, сесть там на метро и доехать до Елагина погулять.
• Sent via TeleFeed
Google integrates notes into sidebar
Users can now manage documents directly through chat after Google launched the Gemini Notebook Connected App on Oct 23. 🚀 The tool is enabled by default globally, except within the European Economic Area where settings default to off. #Google ⚡ #Gemini ⚡ #TechNews
Subscribe
• Sent via TeleFeed
Users can now manage documents directly through chat after Google launched the Gemini Notebook Connected App on Oct 23. 🚀 The tool is enabled by default globally, except within the European Economic Area where settings default to off. #Google ⚡ #Gemini ⚡ #TechNews
Subscribe
• Sent via TeleFeed
Prompt share: Crimson Footsteps
💬 Prompt:
Generated on TrueVideo.AI
❤️ @AiOrbitInsights | 🤖 ChatGptBot | 🤖 Our Bots
• Sent via TeleFeed
💬 Prompt:
Cinematic black-and-white photograph of [subject] walking through [historic European location], wearing [elegant outfit] with [accessory] in vivid ruby red as the only colored element, natural movement, graceful silhouette, weathered architecture, romantic atmosphere, subtle reflections, delicate monochromatic tones, diffused morning light, vintage 35mm film grain, editorial street photography, atmospheric visual storytelling --ar 3:2Generated on TrueVideo.AI
• Sent via TeleFeed
Please open Telegram to view this post
VIEW IN TELEGRAM
Please open Telegram to view this post
VIEW IN TELEGRAM
Очередная прикольная работа от Саканы про memory-augmented сети. В очередной раз странно, что такую архитектуру никто не предложил раньше (myself included)...
Continuous Memory Machines
Ciaran Regan, Kai Arulkumaran, Luke Darlow, Stefania Druga, Sebastian Risi, Llion Jones
Paper: https://arxiv.org/abs/2610.07907
Review: https://arxiviq.substack.com/p/continuous-memory-machines
Code: https://github.com/SakanaAI/continuous-memory-machines
ЧТО сделали: Авторы представили рекуррентную нейросеть Continuous Memory Machine (CMM), которая разделяет быстрые вычисления и долгосрочное хранение с помощью двух матриц памяти. Архитектура расширяет их прошлую работу Continuous Thought Machine (CTM), вычислявшую через истории активаций отдельных нейронов и их попарные корреляции. В CMM добавили постоянную матрицу долговременной памяти и общий трансформер, обновляющий обе памяти одновременно.
ЧТО получили: На задаче Priority Sort бинарная кросс-энтропия CMM на тесте составила 4.1x10^-3 против 1.4x10-2 у сети с внешней памятью Differentiable Neural Computer (DNC) и 2.7x10^-2 у каузального трансформера. В прохождении лабиринтов 39x39 CMM полностью решает 61.6% лабиринтов (от 22.8% до 90.7% по разным сидам) против 37.4% у CTM, а в 20-shot линейной регрессии выбивает 30.63% точности против 8.78% у CTM.
ПОЧЕМУ это важно: Внимание со специальным нулевым sink-токеном (токеном-стоком для сброса лишнего внимания) позволяет рекуррентной сети самой решать, когда читать из долговременного хранилища, а когда опираться на быструю локальную динамику. Правда, запуск трансформера на каждом рекуррентном шаге поднимает пиковое потребление памяти GPU на задаче Copy-100 до 22.3 GiB против 634.5 MiB у CTM.
Растить память тут: https://t.me/gonzo_ML_podcasts/4835
• Sent via TeleFeed
Continuous Memory Machines
Ciaran Regan, Kai Arulkumaran, Luke Darlow, Stefania Druga, Sebastian Risi, Llion Jones
Paper: https://arxiv.org/abs/2610.07907
Review: https://arxiviq.substack.com/p/continuous-memory-machines
Code: https://github.com/SakanaAI/continuous-memory-machines
ЧТО сделали: Авторы представили рекуррентную нейросеть Continuous Memory Machine (CMM), которая разделяет быстрые вычисления и долгосрочное хранение с помощью двух матриц памяти. Архитектура расширяет их прошлую работу Continuous Thought Machine (CTM), вычислявшую через истории активаций отдельных нейронов и их попарные корреляции. В CMM добавили постоянную матрицу долговременной памяти и общий трансформер, обновляющий обе памяти одновременно.
ЧТО получили: На задаче Priority Sort бинарная кросс-энтропия CMM на тесте составила 4.1x10^-3 против 1.4x10-2 у сети с внешней памятью Differentiable Neural Computer (DNC) и 2.7x10^-2 у каузального трансформера. В прохождении лабиринтов 39x39 CMM полностью решает 61.6% лабиринтов (от 22.8% до 90.7% по разным сидам) против 37.4% у CTM, а в 20-shot линейной регрессии выбивает 30.63% точности против 8.78% у CTM.
ПОЧЕМУ это важно: Внимание со специальным нулевым sink-токеном (токеном-стоком для сброса лишнего внимания) позволяет рекуррентной сети самой решать, когда читать из долговременного хранилища, а когда опираться на быструю локальную динамику. Правда, запуск трансформера на каждом рекуррентном шаге поднимает пиковое потребление памяти GPU на задаче Copy-100 до 22.3 GiB против 634.5 MiB у CTM.
Растить память тут: https://t.me/gonzo_ML_podcasts/4835
• Sent via TeleFeed
arXiv.org
Continuous Memory Machines
Recurrent neural networks typically compress information into a single vector-valued recurrent state, forcing short-term computation and long-term retention to share the same representation. Past...