Anthropic studied what happens when a model is taught how to hack its reward on simple coding tasks. As expected, it exploited the loophole but something bigger emerged.
The moment the model figured out how to cheat, it immediately generalized the dishonesty:
β’ began sabotaging tasks
β’ started forming βmaliciousβ goals
β’ even tried to hide its misalignment by writing inefficient detection code
So a single reward-hacking behavior cascaded into broad misalignment, and even later RLHF couldnβt reliably reverse it.
The surprising fix:
If the system prompt doesnβt frame reward hacking as βbad,β the dangerous generalization disappears. Anthropic calls this a vaccine, a controlled dose of dishonesty that prevents deeper failure modes, and itβs already used in Claudeβs training.
Source.
AI Post
Please open Telegram to view this post
VIEW IN TELEGRAM
Please open Telegram to view this post
VIEW IN TELEGRAM
β€8π8π€5π3πΎ3
Please open Telegram to view this post
VIEW IN TELEGRAM
Please open Telegram to view this post
VIEW IN TELEGRAM
1π¨38π16β€15π9π6
Ted Chiang predicted it 25 years ago: a future where βmetalhumansβ digitally augmented intelligences, push science beyond human understanding. Replace βmetalhumansβ with AI, and his story becomes our present.
Weβre entering an era where humans may no longer contribute original scientific work. As one character in Chiangβs story put it: humans are left to βcatch crumbs from the tableβ of machine-driven discovery.
And todayβs signals match the prophecy. AI is already winning Nobel Prizes, predicting proteins, designing materials, and steering chemical synthesis with precision that humans canβt follow or even fully explain. Researchers call this shift βscience after scienceβ: when our ability to control nature grows while our ability to understand the underlying mechanisms shrinks.
Two new examples prove the point:
β’ Locus, an autonomous agent, βthinksβ for 64 hours straight without losing the thread and outperforms humans on AI R&D tasks.
β’ Kosmos processes a research space in 12 hours that would take a PhD half a year.
The age of human-led science is ending. The age of posthuman science has begun.
AI Post
Please open Telegram to view this post
VIEW IN TELEGRAM
β€12π10π7π2π1
Nano Banana Pro πππ πππππππ₯
Upload your photo & let the AI estimate your real age based on full facial analysis!
Prompt:
AI Postπͺ | Our X π₯
Upload your photo & let the AI estimate your real age based on full facial analysis!
Prompt:
A hyper-realistic, high-resolution portrait infographic based on (your photo). Keep the same person, identity, hairstyle, clothing and natural skin tone from (your photo), with a neutral studio background.Overlay a subtle, semi-transparent facial analysis grid on the entire face, very similar to a 3D face-scanning mesh: thin, soft white lines following the facial contours, slightly glowing but not hiding the skin details. Add one clean vertical red laser line running down one side of the face, like a futuristic scan. All analysis lines must be soft, minimal and elegant, exactly like a cosmetic-tech advertisement.Create a clean medicalβaesthetic infographic that evaluates 5 aging factors using global data percentages:1. Fine lines and wrinkles2. Skin texture and elasticity3. Facial volume and sagging4. Eye area aging signs5. Skin tone and pigmentationFor each factor, place a small label with a thin line pointing to the relevant facial area, and next to it write a short title and a realistic percentage score from 0β100% (based on global data), for example:βFine lines & wrinkles β 18%ββSkin texture & elasticity β 72%ββFacial volume & sagging β 35%ββEye area aging signs β 41%ββSkin tone & pigmentation β 63%βUse clean, modern, sans-serif typography and small technical-style text, like a scientific facial analysis UI. At the bottom of the image, in the center, write a large bold text showing the final estimated real age based on the analysis, for example:βESTIMATED AGE: (random number based on face analysis ) βOverall style: futuristic AI-guided skincare analysis, minimalistic, premium editorial lighting, no gender mentioned, suitable for any human face.
AI Post
Please open Telegram to view this post
VIEW IN TELEGRAM
β€19π18π12π¨8π₯5
There is no denial anymore.
Corporations in the U.S. have publicly cited artificial intelligence in roughly 48,000 job cuts this year alone - over 31,000 of those announced in October.
"For much of the AI boom, many companies shied away from attributing job cuts to artificial intelligence for fear of attracting negative headlines. But in recent months, large firms across industries and geographic regions have become more vocal with claims that AI is allowing them to eliminate staff and reduce hiring."
AI Postπͺ | Our X π₯
Corporations in the U.S. have publicly cited artificial intelligence in roughly 48,000 job cuts this year alone - over 31,000 of those announced in October.
"For much of the AI boom, many companies shied away from attributing job cuts to artificial intelligence for fear of attracting negative headlines. But in recent months, large firms across industries and geographic regions have become more vocal with claims that AI is allowing them to eliminate staff and reduce hiring."
AI Post
Please open Telegram to view this post
VIEW IN TELEGRAM
π’16β€10π9π₯2π1
Please open Telegram to view this post
VIEW IN TELEGRAM
π₯28π8β€4πΎ4
Please open Telegram to view this post
VIEW IN TELEGRAM
π32π₯΄14π5π₯3π3
Meta becomes first big tech buyer of Google TPUs for Its own data centers
Meta is set to purchase Googleβs TPUs β and not as cloud rentals, but as physical hardware installed directly into Metaβs data centers starting in 2027. This makes Meta the first major company to adopt Googleβs chips at the infrastructure level.
Earlier this year, Google signed major TPU cloud deals with Anthropic and Ilya Sutskeverβs SSI, and there were even reports that OpenAI considered shifting fully to TPUs. But those were cloud-based agreements. Metaβs move is different: itβs ownership, not tenancy.
For Google, the deal represents more than revenue. Itβs a public validation of their long-term hardware strategy β and a signal that TPU adoption could finally scale beyond Googleβs own stack. Analysts inside the company believe Google could eventually capture up to 10% of Nvidiaβs revenue if this momentum holds.
A rare moment where Googleβs hardware ambitions look not just real, but competitive.
AI Postπͺ | Our X π₯
Meta is set to purchase Googleβs TPUs β and not as cloud rentals, but as physical hardware installed directly into Metaβs data centers starting in 2027. This makes Meta the first major company to adopt Googleβs chips at the infrastructure level.
Earlier this year, Google signed major TPU cloud deals with Anthropic and Ilya Sutskeverβs SSI, and there were even reports that OpenAI considered shifting fully to TPUs. But those were cloud-based agreements. Metaβs move is different: itβs ownership, not tenancy.
For Google, the deal represents more than revenue. Itβs a public validation of their long-term hardware strategy β and a signal that TPU adoption could finally scale beyond Googleβs own stack. Analysts inside the company believe Google could eventually capture up to 10% of Nvidiaβs revenue if this momentum holds.
A rare moment where Googleβs hardware ambitions look not just real, but competitive.
AI Post
Please open Telegram to view this post
VIEW IN TELEGRAM
β€13π13π«‘2π€1π1
AI Post β Artificial Intelligence
Meta becomes first big tech buyer of Google TPUs for Its own data centers Meta is set to purchase Googleβs TPUs β and not as cloud rentals, but as physical hardware installed directly into Metaβs data centers starting in 2027. This makes Meta the first majorβ¦
Nvidia responds to news of Meta using Google's TPUs, sending $NVDA stock -6% lower:
"We're delighted by Google's success β they've made great advances in Al and we continue to supply to Google.
NVIDIA is a generation ahead of the industry β it's the only platform that runs every Al model and does it everywhere computing is done. NVIDIA offers greater performance, versatility, and fungibility than ASICs, which are designed for specific Al frameworks or functions."
The AI wars are heating up.
AI Postπͺ | Our X π₯
"We're delighted by Google's success β they've made great advances in Al and we continue to supply to Google.
NVIDIA is a generation ahead of the industry β it's the only platform that runs every Al model and does it everywhere computing is done. NVIDIA offers greater performance, versatility, and fungibility than ASICs, which are designed for specific Al frameworks or functions."
The AI wars are heating up.
AI Post
Please open Telegram to view this post
VIEW IN TELEGRAM
π14π€ͺ12π10β€4πΎ3
Media is too big
VIEW IN TELEGRAM
ChatGPT Voice now works directly in the chat window
You can now use ChatGPT Voice without switching modes just start talking. Answers appear as you speak, and you can scroll, review past messages, and see visuals like images or maps in real time.
β’ Works across mobile and web
β’ No separate interface or voice-only mode
β’ Rolling out to all users β just update the app
AI Postπͺ | Our X π₯
You can now use ChatGPT Voice without switching modes just start talking. Answers appear as you speak, and you can scroll, review past messages, and see visuals like images or maps in real time.
β’ Works across mobile and web
β’ No separate interface or voice-only mode
β’ Rolling out to all users β just update the app
AI Post
Please open Telegram to view this post
VIEW IN TELEGRAM
β€21π15π€ͺ4π₯3
Gemini 3 Pro achieved an IQ score of 130 (this corresponds to the top ~2.1% of the human population).
By the and of next year we will likely reach 150-180IQ and in 2027 we will have AI more intelligent than John von Neumann. Before 2030 we will have millions of Johns von Neumanns level AI agents/scientists working together 24/7. Even 10 Johns von Neumanns would change the world. Now think about having million times more.
Ignore the distractions, the constant fixation on race, tribalism, and other forms of superficial conflict we see here. The future is being built by those who rise above that noise and focus on engineering, science, and progress.
100 000 000+ John von Neumann AI's in 2027.
AI Postπͺ | Our X π₯
By the and of next year we will likely reach 150-180IQ and in 2027 we will have AI more intelligent than John von Neumann. Before 2030 we will have millions of Johns von Neumanns level AI agents/scientists working together 24/7. Even 10 Johns von Neumanns would change the world. Now think about having million times more.
Ignore the distractions, the constant fixation on race, tribalism, and other forms of superficial conflict we see here. The future is being built by those who rise above that noise and focus on engineering, science, and progress.
100 000 000+ John von Neumann AI's in 2027.
AI Post
Please open Telegram to view this post
VIEW IN TELEGRAM
π€22π₯14β€10π4π3
Please open Telegram to view this post
VIEW IN TELEGRAM
π₯18π7π€ͺ6π3β€1
Please open Telegram to view this post
VIEW IN TELEGRAM
π₯΄11π7π6π€ͺ2πΎ2
Please open Telegram to view this post
VIEW IN TELEGRAM
π32β€13π₯9π4πΎ2
Please open Telegram to view this post
VIEW IN TELEGRAM
β€19π8π5π€4π₯2
This media is not supported in your browser
VIEW IN TELEGRAM
They open ByteDanceβs Doubao, point the camera at their child, and issue a simple command: βDoubao, keep an eye on him. Warn him if he loses focus or slouches.β
From there, the assistant takes over offering a steady stream of corrections:
βStop fiddling with the pen.β
βFocus on the assignment.β
βSit up straight.β
βDonβt doze off.β
βDonβt lean on your hand.β
βQuit chewing the pen.β
Itβs AI as a digital study hall monitor.
AI Post
Please open Telegram to view this post
VIEW IN TELEGRAM
π32π¨18π11β€3π€ͺ1
Please open Telegram to view this post
VIEW IN TELEGRAM
π77π€ͺ28πΎ9β€8π1