¡Ya disponible! Investigación de Telegram 2025 — los principales insights del año 
AI Post — Artificial Intelligence
Ir al canal en Telegram
🤖 The #1 AI news source! We cover the latest artificial intelligence breakthroughs and emerging trends. Manager: @rational
Mostrar más📈 Análisis del canal de Telegram AI Post — Artificial Intelligence
El canal AI Post — Artificial Intelligence (@aipost) en el segmento lingüístico de Inglés es un actor destacado. Actualmente la comunidad reúne a 790 743 suscriptores, ocupando la posición 97 en la categoría Tecnologías y Aplicaciones y el puesto 20 en la región EEUU.
📊 Métricas de audiencia y dinámica
Desde su creación el невідомо, el proyecto ha mostrado un crecimiento acelerado, reuniendo a 790 743 suscriptores.
Según los últimos datos del 23 junio, 2026, el canal mantiene una actividad estable. En los últimos 30 días la variación de miembros fue de -34 983, y en las últimas 24 horas de -725, conservando un alto alcance.
- Estado de verificación: No verificado
- Tasa de interacción (ER): El promedio de interacción de la audiencia es 0.68%. Durante las primeras 24 horas tras publicar, el contenido suele obtener 0.46% de reacciones respecto al total de suscriptores.
- Alcance de las publicaciones: Cada publicación recibe en promedio 5 365 visualizaciones. En el primer día suele acumular 3 653 visualizaciones.
- Reacciones e interacción: La audiencia responde de forma activa: el promedio de reacciones por publicación es 503.
- Intereses temáticos: El contenido se centra en temas clave como openai, airline, cell, claude, patient.
📝 Descripción y política de contenido
El autor describe el recurso como un espacio para expresar opiniones subjetivas:
“🤖 The #1 AI news source! We cover the latest artificial intelligence breakthroughs and emerging trends.
Manager: @rational”
Gracias a la alta frecuencia de actualizaciones (últimos datos recibidos el 24 junio, 2026), el canal mantiene la vigencia y un amplio alcance. La analítica demuestra que la audiencia interactúa activamente con el contenido, lo que lo convierte en un punto de referencia dentro de la categoría Tecnologías y Aplicaciones.
790 743
Suscriptores
-72524 horas
-6 1937 días
-34 98330 días
Carga de datos en curso...
Canales Similares
Nube de Etiquetas
Menciones Entrantes y Salientes
---
---
---
---
---
---
Atraer Suscriptores
junio '26
junio '26
+349
en 56 canales
mayo '260
en 99 canales
Get PRO
abril '26
+116
en 61 canales
Get PRO
marzo '260
en 86 canales
Get PRO
febrero '26
+2
en 82 canales
Get PRO
enero '260
en 76 canales
Get PRO
diciembre '250
en 59 canales
Get PRO
noviembre '250
en 67 canales
Get PRO
octubre '250
en 57 canales
Get PRO
septiembre '250
en 76 canales
Get PRO
agosto '25
+3
en 63 canales
Get PRO
julio '25
+412
en 42 canales
Get PRO
junio '250
en 52 canales
Get PRO
mayo '250
en 55 canales
Get PRO
abril '250
en 40 canales
Get PRO
marzo '25
+67
en 39 canales
Get PRO
febrero '25
+2
en 56 canales
Get PRO
enero '250
en 42 canales
Get PRO
diciembre '24
+25 161
en 36 canales
Get PRO
noviembre '24
+42 696
en 51 canales
Get PRO
octubre '24
+55 283
en 26 canales
Get PRO
septiembre '24
+221 077
en 38 canales
Get PRO
agosto '24
+1 623 027
en 30 canales
Get PRO
julio '24
+826 511
en 26 canales
Get PRO
junio '24
+36 138
en 79 canales
Get PRO
mayo '24
+16 690
en 46 canales
Get PRO
abril '24
+10 982
en 24 canales
Get PRO
marzo '24
+243
en 24 canales
Get PRO
febrero '24
+2
en 7 canales
Get PRO
enero '24
+4 466
en 11 canales
Get PRO
diciembre '23
+180
en 7 canales
Get PRO
noviembre '23
+4 658
en 5 canales
Get PRO
octubre '23
+3
en 13 canales
Get PRO
septiembre '23
+6 455
en 0 canales
Get PRO
agosto '23
+152 348
en 0 canales
| Fecha | Crecimiento de Suscriptores | Menciones | Canales | |
| 25 junio | 0 | |||
| 24 junio | 0 | |||
| 23 junio | +1 | |||
| 22 junio | 0 | |||
| 21 junio | +2 | |||
| 20 junio | 0 | |||
| 19 junio | +346 | |||
| 18 junio | 0 | |||
| 17 junio | 0 | |||
| 16 junio | 0 | |||
| 15 junio | 0 | |||
| 14 junio | 0 | |||
| 13 junio | 0 | |||
| 12 junio | 0 | |||
| 11 junio | 0 | |||
| 10 junio | 0 | |||
| 09 junio | 0 | |||
| 08 junio | 0 | |||
| 07 junio | 0 | |||
| 06 junio | 0 | |||
| 05 junio | 0 | |||
| 04 junio | 0 | |||
| 03 junio | 0 | |||
| 02 junio | 0 | |||
| 01 junio | 0 |
Publicaciones del Canal
+1
52% of students say they are worried about their academic work being falsely flagged as AI.
@aipost 🏴
| 2 |  ❗️How Andrew Ng organizes his engineering team to move faster in the era of AI.
"1 to 10 engineers in a team, often made up of generalists: high-context, highly empowered generalists."
When code gets generated much faster, organizations become the slow part.
Once a feature can move from idea to working prototype in a day, every surrounding function is suddenly exposed.
Product has to decide faster, design has to clarify faster, marketing has to understand faster, and legal has to review faster.
So his way is 1-10 high-context generalists who can move much faster because they do not need every decision translated across departments before anything happens.
@aipost 🏴 | 1 365 |
| 3 | 🤖 Over 1 Million ChatGPT Users Show Signs of Suicide Risk Each Week
OpenAI revealed that about 0.15% of ChatGPT’s 800 million weekly users show signs of suicidal planning, while another 0.15% show emotional dependence and 0.07% show signs of psychosis or mania.
Those percentages sound small, but they translate to roughly 1.2 million users discussing suicide-related concerns, 1.2 million showing unhealthy attachment, and 560,000 experiencing possible mental health crises every week.
The disclosure comes as AI companies face growing scrutiny over how chatbots handle vulnerable users, highlighting the massive mental-health challenges that emerge at ChatGPT’s scale.
Here is the 5-minute rule that can save you or a relative:
1. If you or someone you know is using ChatGPT as a therapist, the US suicide line is 988. Call or text. Free. 24 hours.
2. If you have a teen, open ChatGPT, Settings, Parental Controls. Link your account to theirs. Turn on quiet hours and distress alerts. 4 minutes.
3. Check their phone for Character AI, Replika, and Nomi. Character AI banned under-18s on November 25, 2025. The other two officially ban minors but teens still get in with fake birthdays.
4. Replace one AI chat a day with one text to a real person.
The bot will never call 988 for you. A friend will. Save this for someone who needs it.
@aipost 🏴 | 2 689 |
| 4 | 🔥This new feature is like managing a legit employee now.
One that already knows your full company context, sits in every channel, remembers every conversation, and does the work alongside you. it becomes especially powerful when you connect it to all your tools, too.
Here's a bunch of ideas for ways to use it:
1. Content pipeline. Tag Claude in the channel where your team dumps hooks, ideas, screenshots all week. It keeps a running list of what's actually usable, sorts them by theme, then every friday posts next week's content plan ready to go.
2. Client account manager. Let Claude sit in the shared client channel. It remembers every promise, deadline, request buried in the chat. so when the client asks "where are we on the homepage?", it answers from the real history and flags anything your team agreed to but hasn't done yet.
3. Catch the dropped balls. People say "i'll send that tonight" or "let's circle back monday," then it slips. Claude quietly tracks every loose end like that and pings whoever owns it when the deadline passes. It’s the teammate who actually remembers what everyone promised.
4. Campaign control room. Drop Claude in the channel where marketing posts the landing page copy, the emails, the ad scripts. It reads all of it and flags when they don't match, like an ad promising a discount the landing page never mentions. An extra set of eyes on the whole campaign.
5. Community listening. Point Claude at your member or Discord channel and ask "what does everyone keep asking for?" It reads the whole conversation and tells you the top requests, the most common complaints, who keeps volunteering to help. like a researcher who never stops listening to your audience.
6. Sales handoff. The person who books the call drops everything they learned in the channel. Before the next call, the closer asks Claude "catch me up on this lead." It pulls the full history: the budget, the objections, what got promised. No "let me forward you my notes."
7. Sponsor tracker. Sponsor deals get discussed across dozens of scattered messages. Claude keeps a clean list: who's booked, which dates, what they paid, what's been delivered.
8. Recurring reports. Set it once: "every monday at 9am, read last week's channel and post a summary of what got done." It runs on its own and drops the recap in. You review it instead of writing it, every week, no reminders.
@aipost 🏴 | 2 947 |
| 5 |  🤖 Anthropic just turned Claude into a slack teammate
Anthropic introduced Claude Tag, a new system that lets teams tag Claude directly in Slack and assign it real work.
Mention Claude in a thread, and it can break tasks into steps, analyze data, write code, merge pull requests, investigate incidents, and report back with results.
Unlike a personal AI assistant, there’s one shared Claude per channel, allowing teammates to jump into projects without losing context. Claude also learns from ongoing conversations, reducing the need to repeatedly explain the same work.
With its new “ambient” mode, Claude can even act proactively, following up on stalled discussions and surfacing important updates on its own.
Anthropic says the concept evolved from Claude Code, and internally, 65% of its product team’s code is now generated by Claude-powered tools.
Source.
@aipost 🏴 | 2 631 |
| 6 |  The Trump administration is urging Meta to submit its artificial intelligence models for federal review before releasing them to the public.
Meta remains the only major U.S. AI company not participating in the voluntary government review program.
According to recent reports, OpenAI, Anthropic, Google, xAI, and Microsoft have already agreed to share their models with a government AI safety group.
The review process is designed to evaluate whether advanced AI models can be used for sensitive cybersecurity tasks, identify potential security risks, or present national security concerns before they reach widespread use.
📰 @aipost | 2 870 |
| 7 |  👓 Meta is rolling out a new line of AI glasses called the “Muse Spark” glasses, which come in three styles.
The glasses can play music, capture images to translate languages and answer questions about a person’s surroundings. CBS News’ Maya Blackstone tried them out.
@aipost 🏴 | 2 771 |
| 8 |  🗣️ Recursive self-improvement: Anthropic co-founder Jack Clark says it could arrive by 2028.
AI systems could help invent their own successors with Claude 10 building Claude 11, and so on potentially “without any researchers involved.”
@aipost 🏴 | 3 349 |
| 9 |  A Texas-based company, Minicircle, is preparing to offer an injectable gene therapy focused on increasing longevity. This new treatment aims to stimulate the production of klotho, a protein linked to anti-aging processes.
Minicircle plans to launch the therapy outside the U.S. regulatory framework, specifically in regions such as Honduras, the Bahamas, and Panama. The intention is to bypass comprehensive trials typically required by the FDA and instead operate in locations with less stringent oversight.
The company’s approach reflects a trend where biomedical innovation may shift to countries with more flexible regulatory environments if others choose a slower or more cautious path.
📰 @aipost | 3 495 |
| 10 |  The intelligence alliance known as Five Eyes has issued a rare public alert, warning that powerful AI models able to carry out damaging cyberattacks may become available within the next few months.
This coalition, which includes Australia, the United States, the United Kingdom, Canada, and New Zealand, expressed concern that advanced AI systems could simplify the execution of major cyberattacks against both government and business targets.
According to the statement, such models have the capacity to automate specialized cyber functions. They can analyze code, identify vulnerabilities, propose exploits, and coordinate actions across networks—all tasks previously requiring expert human intervention.
📰 @aipost | 3 527 |
| 11 |  ❗️AI data centers may be using far less water than most people think.
According to the Manhattan Institute, data centers account for just 0.2% of daily U.S. water consumption and that figure is falling as the industry shifts to liquid cooling.
The big change is 45°C liquid cooling, which allows many AI facilities to use dry coolers instead of water-intensive cooling towers.
The result? Cooling water usage can drop from roughly 2.6 million gallons per MW per year to nearly zero.
And the benefits go beyond water savings. Liquid cooling is also more energy efficient and makes it easier to capture and reuse waste heat, turning AI factories into potential assets for local communities and power grids.
The future AI data center may consume less water than critics expect and provide more value than just compute.
@aipost 🏴 | 3 509 |
| 12 |  🇪🇺 Mistral CEO Arthur Mensch just gave Europe a stark warning:
Build your own AI infrastructure within two years or send over $1 trillion to American tech companies.
His argument is simple. Global wages are about $50 trillion, and AI could absorb roughly 10% of that value. Europe’s share is around $9 trillion, meaning more than $1 trillion in AI spending is up for grabs over the next five years.
The question is: who gets paid?
Europe already spends hundreds of billions on US digital services every year, helping fund American R&D while falling further behind in the technologies driving future productivity.
Mensch compared AI to energy dependence. By the time a crisis exposes the risk, it’s already too late.
The chips are being allocated. The data centers are being built. The AI economy is being carved up now.
This isn’t really a technology debate anymore. It’s a sovereignty debate.
@aipost 🏴 | 3 847 |
| 13 |  This is Seedance 2.5, and it is Hollywood level stuff.
Robots & Art 🌈 | 3 862 |
| 14 |  🤖 It looks like we’re getting a whole range of new GPT models this Thursday:
• GPT-5.6
• 5.6 Pro
• And a new bidirectional voice model.
@aipost 🏴 | 4 804 |
| 15 |  ⚡️OpenAI’s new GPT-5.5-Cyber just outscored Mythos 5 on CyberGym, a benchmark that tests whether AI agents can reproduce real-world software vulnerabilities.
@aipost 🏴 | 4 028 |
| 16 |  We talk about scaling compute. Peter Diamandis is talking about scaling humanity itself.
@aipost 🏴 | 4 111 |
| 17 | ⚠️Everyone can tell when you used AI to write.
It is not the words. It is the rhythm, here is how to fix it:
1. AI gets caught because its writing can be predictable:
Detectors measure how evenly the text flows. Same sentence length, same shape, line after line. That evenness is what flags it.
2. Em dashes. The word "delve." A typo dropped in on purpose:
People scrub all of it to hide AI and it changes nothing. That wasn't the part giving your writing away.
3. The real giveaway is that every sentence is the same length:
Look at how people actually write. A long thought that builds and winds for a while. Then a short one. A fragment. Then a line that keeps going past where a model would have stopped.
4. So make it uneven on purpose:
"Rewrite this so the sentence lengths are all over the place. Some very short, some long. Wherever three in a row feel the same, break one. Do not tidy it up afterward."
5. The other giveaway is that AI will not commit:
It weighs both sides. It softens every claim. It writes like it is scared of being wrong. People pick a side and say it.
6. Fix:
"Rewrite this with an actual opinion. Pick a side, say it plainly, cut the hedging. Sound like someone who already made up their mind."
7. Last one. AI writing is too clean:
People leave fingerprints. An aside that cuts into the point. A sentence that trails off. A detail so specific nobody would bother to invent it. "Put the human mess back in. An aside, a blunt line, something slightly off. Leave it."
None of this is about beating a detector. A detector and a real reader are checking for the same thing, whether a person was behind the words. Fix that and both stop flagging you.
@aipost 🏴 | 4 276 |
| 18 |  "Can AI ever be Newton? Can AI ever be Einstein? Can AI ever be Picasso?"
Dr. Fei-Fei Li gives a very simple explanations of how today's AI still has a long way to go.
@aipost 🏴 | 4 134 |
| 19 |  SpaceX has finalized a major agreement with open-source AI startup Reflection, granting the company access to Nvidia GB300 GPUs through a compute lease valued at $150 million per month, beginning in July 2026. Total payments could reach approximately $6.3 billion if the contract is maintained until 2029.
Either SpaceX or Reflection can terminate the deal with 90 days’ notice after the initial three months. This marks a shift for SpaceX, which is now positioning itself as a GPU cloud provider, leveraging its Colossus infrastructure for external clients.
Reflection’s business model centers on building and customizing frontier open-weight AI models for government and enterprise use, requiring substantial compute capacity to train these foundational models. Recent changes by other AI providers have increased demand for open-access model solutions, supporting Reflection’s approach.
📰 @aipost | 4 281 |
| 20 | ❗️Vivek, a researcher at Anthropic, published an article on how to be good at research.
Here are the 7 core principles as prompt instructions:
1. "Restate the problem in your own words and define what a successful answer looks like before responding."
From John Schulman's advice: reason backwards from the outcome you want instead of forwards from the most obvious approach. In a prompt, this forces the model to set success criteria before it starts generating, which prevents the generic default response.
2. "Predict the most common answer to this question. Then tell me what it gets wrong or misses."
Vivek describes taste as a muscle you train by predicting results before you see them. This instruction does the same thing inside the model: it makes it aware of its own baseline, then pushes past it.
3. "Prioritize foundational sources over recent or popular ones. Older proven frameworks beat trending takes."
LLMs have the same problem researchers do: shared reading lists produce shared ideas. This redirects the model toward durable knowledge instead of whatever was most common in the training data.
4. "Show your reasoning chain. State each assumption and flag where your logic is weakest."
Paul Graham's observation: an idea feels complete until you try to write it down, and then the gaps show up. This instruction makes the model find those gaps before you have to.
5. "Start with the smallest possible version of this problem. Solve that first, then add complexity one piece at a time."
Claude Shannon's 1952 technique: shrink a problem until it's trivial, crack it, then reintroduce difficulty. Karpathy's version is "overfit a single batch before training at scale." In a prompt, this prevents the model from trying to solve everything at once, which is where vague answers come from.
6. "After answering, list the three strongest objections to your own response and tell me which one is most valid."
Darwin kept a log of every fact that contradicted his theory because he caught his own memory deleting inconvenient evidence. This instruction builds that discipline into the model's output.
7. "Draw from at least one adjacent field outside the obvious domain for this question."
Interpretability borrows from neuroscience. Eval design borrows from mechanism design. This breaks the model out of single-domain thinking, which is where non-obvious connections live.
@aipost 🏴 | 3 896 |
