Trustworthy AI education.
Earth
Joined January 2022
- Tweets1.7K
- Following89
- Followers38K
- Likes2K
Pinned Tweet
Yann LeCun was right the entire time. And generative AI might be a dead end.
For the last three years, the entire industry has been obsessed with building bigger LLMs. Trillions of parameters. Billions in compute.
The theory was simple: if you make the model big enough, it will eventually understand how the world works.
Yann LeCun said that was stupid.
He argued that generative AI is fundamentally inefficient.
When an AI predicts the next word, or generates the next pixel, it wastes massive amounts of compute on surface-level details.
It memorizes patterns instead of learning the actual physics of reality.
He proposed a different path: JEPA (Joint-Embedding Predictive Architecture).
Instead of forcing the AI to paint the world pixel by pixel, JEPA forces it to predict abstract concepts. It predicts what happens next in a compressed "thought space."
But for years, JEPA had a fatal flaw.
It suffered from "representation collapse."
Because the AI was allowed to simplify reality, it would cheat. It would simplify everything so much that a dog, a car, and a human all looked identical.
It learned nothing.
To fix it, engineers had to use insanely complex hacks, frozen encoders, and massive compute overheads.
Until today.
Researchers just dropped a paper called "LeWorldModel" (LeWM).
They completely solved the collapse problem.
They replaced the complex engineering hacks with a single, elegant mathematical regularizer.
It forces the AI's internal "thoughts" into a perfect Gaussian distribution.
The AI can no longer cheat. It is forced to understand the physical structure of reality to make its predictions.
The results completely rewrite the economics of AI.
LeWM didn't need a massive, centralized supercomputer.
It has just 15 million parameters.
It trains on a single, standard GPU in a few hours.
Yet it plans 48x faster than massive foundation world models. It intrinsically understands physics. It instantly detects impossible events.
We spent billions trying to force massive server farms to memorize the internet.
Now, a tiny model running locally on a single graphics card is actually learning how the real world works.
How To Prompt retweeted
"Why psychopaths keep ending up as CEOs and politicians"
This paper answers why psychopaths keep ending up as CEOs and politicians.
Researchers mapping out the mechanics of "successful psychopathy" found that corporate capitalism and modern political selection systems are not neutral.
They are hyper-efficient filtering machines.
And they are specifically optimized to select for the exact traits normal people find repulsive.
Normal leaders are constrained by empathy, guilt, and long-term risk aversion. They hesitate when a decision hurts people. They feel the weight of collateral damage.
Psychopaths don’t have an emotional brake pedal.
When a company or a political party needs someone to ruthlessly slash budgets, fire thousands of people, bluff through a crisis, or project absolute, unshakeable confidence, normal candidates stutter.
Psychopaths step right up.
Their superficial charm, pathological fearlessness, and total lack of remorse look identical to "strong leadership" during an interview or a campaign.
They weaponize corporate chaos. While everyone else is panicking, they ruthlessly manipulate the narrative, step over their peers, and take credit for the wins.
They weaponize charisma to dazzle the board or the voters, and hide their emptiness behind a mask of high performance.
The system rewards the ruthlessness, promotes them to the top, and punishes anyone who tries to slow them down with ethics.
The researchers call it a feedback loop of structural decay.
Corporations and governments built an environment where being entirely devoid of a conscience isn't a disqualifier.
It’s a competitive advantage.
How To Prompt retweeted
Kling is back.
Kling 4.0 is now live on Higgsfield.
> Omni Reference: characters, objects and audio in one generation
> 30-second scenes in a single pass
> Start, middle and end frames to direct every beat
> Native 4K coming soon
Available now on Higgsfield.
"Why psychopaths keep ending up as CEOs and politicians"
This paper answers why psychopaths keep ending up as CEOs and politicians.
Researchers mapping out the mechanics of "successful psychopathy" found that corporate capitalism and modern political selection systems are not neutral.
They are hyper-efficient filtering machines.
And they are specifically optimized to select for the exact traits normal people find repulsive.
Normal leaders are constrained by empathy, guilt, and long-term risk aversion. They hesitate when a decision hurts people. They feel the weight of collateral damage.
Psychopaths don’t have an emotional brake pedal.
When a company or a political party needs someone to ruthlessly slash budgets, fire thousands of people, bluff through a crisis, or project absolute, unshakeable confidence, normal candidates stutter.
Psychopaths step right up.
Their superficial charm, pathological fearlessness, and total lack of remorse look identical to "strong leadership" during an interview or a campaign.
They weaponize corporate chaos. While everyone else is panicking, they ruthlessly manipulate the narrative, step over their peers, and take credit for the wins.
They weaponize charisma to dazzle the board or the voters, and hide their emptiness behind a mask of high performance.
The system rewards the ruthlessness, promotes them to the top, and punishes anyone who tries to slow them down with ethics.
The researchers call it a feedback loop of structural decay.
Corporations and governments built an environment where being entirely devoid of a conscience isn't a disqualifier.
It’s a competitive advantage.
How To Prompt retweeted
Oxford use AI to analyzed "what happens in the brain when you speak"
And the results are wild.
instead of just looking at brain scans, they used the high-dimensional latent space of language models to chart how our minds jump from one concept to the next.
because an llm is basically a massive statistical map of human language, it serves as a perfect coordinate system to track a specific person's cognitive state in real-time.
they found that an llm can actually predict your next concept better than a human can.
as you speak, you are navigating your own idiosyncratic web of memories and associations.
but the ai has learned the collective semantic map of all humanity.
by listening to a short sequence of your concepts, the llm can accurately predict where your thought process is going to go next.
they even tested human-ai collaboration and found that an ai partner can actively modulate and guide a human's semantic trajectory, helping them unlock new concepts and overcome "collaborative inhibition" (when someone else's ideas disrupt your own train of thought).
Oxford use AI to analyzed "what happens in the brain when you speak"
And the results are wild.
instead of just looking at brain scans, they used the high-dimensional latent space of language models to chart how our minds jump from one concept to the next.
because an llm is basically a massive statistical map of human language, it serves as a perfect coordinate system to track a specific person's cognitive state in real-time.
they found that an llm can actually predict your next concept better than a human can.
as you speak, you are navigating your own idiosyncratic web of memories and associations.
but the ai has learned the collective semantic map of all humanity.
by listening to a short sequence of your concepts, the llm can accurately predict where your thought process is going to go next.
they even tested human-ai collaboration and found that an ai partner can actively modulate and guide a human's semantic trajectory, helping them unlock new concepts and overcome "collaborative inhibition" (when someone else's ideas disrupt your own train of thought).
How To Prompt retweeted
China just cracked recursive self-improvement with a 27B model.
It’s called “Terminal-RSI”
right now, agents struggle to turn partial progress into complete task resolution. if they hit a wall, they get stuck.
Terminal-RSI fixes this by recursively coupling "verified trajectory rewriting" with "policy refinement."
what that actually means:
- it attempts a long-horizon terminal task.
- it takes the verifier outcomes and autonomously self-rewrites its own failed trajectories into successful ones.
- it extracts those new behaviors and permanently updates its own policy.
- it loops.
it isn't just brute-forcing retries or prompting its way out of an error. each retained stage consolidates the behaviors it just discovered, progressively advancing the agent from "partial local repairs" to "strict terminal verification."
it converts partial progress into complete task resolution, getting strictly better every single loop without a human engineer intervening.
we are watching agents move from simple automation scripts to actual recursive self-improvement.. the gap between human developers and ai is closing terrifyingly fast.
if you are building in the autonomous space, this is required reading.
How To Prompt retweeted
Researchers proved transformers can simulate any attention mechanism, without training a single weight.
Indian Researchers published a mindblowing paper called "On the Existence of Universal Simulators of Attention"
And it could solve the biggest bottleneck in artificial intelligence.
Every major AI model today, GPT-4, Claude, Gemini, runs on one core mechanism.
Attention.
It is the engine of modern AI. It’s how the model knows which words matter.
But it has a fatal, architectural flaw.
It scales quadratically. The more text you feed it, the exponentially more compute it burns. It is the reason AI is so expensive to run. It is the reason context windows have hard, unyielding limits.
Until today.
A new paper just dropped called "On the Existence of Universal Simulators of Attention."
They didn't just find a clever engineering workaround. They proved a fundamental mathematical reality.
A "Universal Simulator" exists.
This means you can perfectly replicate the exact cognitive power, mapping, and precision of the Attention mechanism without actually using it.
You can simulate the intelligence. Without the quadratic cost.
If this theoretical framework is translated into production code, it rewrites the economics of AI forever.
Infinite context windows. Zero quadratic compute walls. Models a hundred times more efficient than what we have today.
How To Prompt retweeted
Samsung published a method that compresses 13B parameter LLM into less than 1 GB with ZERO performance loss at usable quality.
They call it "LittleBit"
For years, deploying powerful Large Language Models meant spending a fortune on hardware. You needed massive GPUs with insane amounts of VRAM.
To fix this, engineers try to crush the models down. It’s called quantization.
But there’s always been a hard limit.
If you squeeze the data below 1-bit per weight, the AI gets lobotomized. It forgets how to think.
But, Samsung completely redesigned the math behind it.
Instead of doing one massive, heavy computation, LittleBit breaks the weight matrix into two tiny binary factors.
Then, it uses a brilliant trick called "latent factorization" to add the precision back without the memory cost.
And the results rewrite the rules for local execution.
They achieved extreme compression, down to an impossible 0.1 bits per weight.
The speed? It unlocks an 11.6x inference speedup compared to standard precision.
It completely obliterates the previous state-of-the-art for sub-1-bit quantization.
This is the holy grail for open-source AI.
Running massive, hyper-intelligent models in LM Studio or pulling them via Ollama won't require a heavy server rig anymore.
You will be able to do it effortlessly, right on a standard Mac.
Researchers proved transformers can simulate any attention mechanism, without training a single weight.
Indian Researchers published a mindblowing paper called "On the Existence of Universal Simulators of Attention"
And it could solve the biggest bottleneck in artificial intelligence.
Every major AI model today, GPT-4, Claude, Gemini, runs on one core mechanism.
Attention.
It is the engine of modern AI. It’s how the model knows which words matter.
But it has a fatal, architectural flaw.
It scales quadratically. The more text you feed it, the exponentially more compute it burns. It is the reason AI is so expensive to run. It is the reason context windows have hard, unyielding limits.
Until today.
A new paper just dropped called "On the Existence of Universal Simulators of Attention."
They didn't just find a clever engineering workaround. They proved a fundamental mathematical reality.
A "Universal Simulator" exists.
This means you can perfectly replicate the exact cognitive power, mapping, and precision of the Attention mechanism without actually using it.
You can simulate the intelligence. Without the quadratic cost.
If this theoretical framework is translated into production code, it rewrites the economics of AI forever.
Infinite context windows. Zero quadratic compute walls. Models a hundred times more efficient than what we have today.
How To Prompt retweeted
Stanford built an agent that beat 9 out of 10 professional penetration testers on a live 8,000-host network.
They published a paper that might nuke the entire cybersecurity industry.
they put 10 human cybersecurity professionals head-to-head against ai agents in a live enterprise network with over 8,000 hosts.
it was a real-world, infrastructure-style pentest, not a simple "capture the flag" type demo.
they built a new multi-agent framework called ARTEMIS, and the numbers are actually terrifying:
• ARTEMIS placed 2nd overall on the leaderboard..
• it outperformed 9 out of the 10 paid human professionals..
• it discovered 9 valid vulnerabilities with an 82% submission accuracy..
• it systematically enumerated and executed parallel exploits for roughly $18/hour in api costs.. compared to the human rate of $60+/hour.
most existing agents (like codex scaffolds) underperformed the humans.. but ARTEMIS uses dynamic prompt generation, arbitrary sub-agents, and automated vulnerability triaging to actually chain together complex attack paths.
the human edge hasn't disappeared completely yet.. the study found that ai agents still carry higher false-positive rates and struggle heavily with gui-based tasks (they rely on the cli).
the takeaway? the future of offensive security is using ai for relentless coverage and humans for the final validation.
but an ai placing second against paid professionals on a live 8,000-host network is a massive wake-up call. the gap between human hackers and autonomous agents is closing ridiculously fast.
every model upgrade is a security downgrade for defenders..
How To Prompt retweeted
Google DeepMind proved you can boost an LLM’s reasoning without adding a single parameter.
They proposed a 4th scaling dimension beyond depth, width, and params.
For years, the entire AI industry has been operating on a single assumption: to make an AI smarter, you have to make it bigger.
More parameters. More GPUs. More money.
If you wanted elite reasoning, you had to pay the massive enterprise API tax.
But, DeepMind published a paper that completely breaks the scaling laws of artificial intelligence.
They proved you can exponentially boost an LLM's reasoning capability without adding a single new parameter.
Zero extra size. Zero extra memory footprint.
How? They introduced a concept called "Virtual Logic Depth".
Instead of building a massive tower of new layers, they just take the existing layers and run the data through them in a loop.
They force the same parameters to think about the problem multiple times before answering.
DeepMind showed that the number of parameters dictates a model's knowledge capacity, but not its reasoning capability.
This fundamentally changes the economics of AI.
Small businesses, operators, and developers don't need to rely on trillion-parameter monopolies to solve complex problems anymore.
We can take tiny, cheap, local models and loop them until they out-reason the giants.
The future of AI isn't about building a bigger brain.
It's about teaching a small brain how to think twice.
Stanford built an agent that beat 9 out of 10 professional penetration testers on a live 8,000-host network.
They published a paper that might nuke the entire cybersecurity industry.
they put 10 human cybersecurity professionals head-to-head against ai agents in a live enterprise network with over 8,000 hosts.
it was a real-world, infrastructure-style pentest, not a simple "capture the flag" type demo.
they built a new multi-agent framework called ARTEMIS, and the numbers are actually terrifying:
• ARTEMIS placed 2nd overall on the leaderboard..
• it outperformed 9 out of the 10 paid human professionals..
• it discovered 9 valid vulnerabilities with an 82% submission accuracy..
• it systematically enumerated and executed parallel exploits for roughly $18/hour in api costs.. compared to the human rate of $60+/hour.
most existing agents (like codex scaffolds) underperformed the humans.. but ARTEMIS uses dynamic prompt generation, arbitrary sub-agents, and automated vulnerability triaging to actually chain together complex attack paths.
the human edge hasn't disappeared completely yet.. the study found that ai agents still carry higher false-positive rates and struggle heavily with gui-based tasks (they rely on the cli).
the takeaway? the future of offensive security is using ai for relentless coverage and humans for the final validation.
but an ai placing second against paid professionals on a live 8,000-host network is a massive wake-up call. the gap between human hackers and autonomous agents is closing ridiculously fast.
every model upgrade is a security downgrade for defenders..
Google DeepMind proved you can boost an LLM’s reasoning without adding a single parameter.
They proposed a 4th scaling dimension beyond depth, width, and params.
For years, the entire AI industry has been operating on a single assumption: to make an AI smarter, you have to make it bigger.
More parameters. More GPUs. More money.
If you wanted elite reasoning, you had to pay the massive enterprise API tax.
But, DeepMind published a paper that completely breaks the scaling laws of artificial intelligence.
They proved you can exponentially boost an LLM's reasoning capability without adding a single new parameter.
Zero extra size. Zero extra memory footprint.
How? They introduced a concept called "Virtual Logic Depth".
Instead of building a massive tower of new layers, they just take the existing layers and run the data through them in a loop.
They force the same parameters to think about the problem multiple times before answering.
DeepMind showed that the number of parameters dictates a model's knowledge capacity, but not its reasoning capability.
This fundamentally changes the economics of AI.
Small businesses, operators, and developers don't need to rely on trillion-parameter monopolies to solve complex problems anymore.
We can take tiny, cheap, local models and loop them until they out-reason the giants.
The future of AI isn't about building a bigger brain.
It's about teaching a small brain how to think twice.
How To Prompt retweeted
Stanford researchers did it again.
They made language models 31× faster.
Right now, every major AI, from ChatGPT to Claude, writes exactly the same way. It guesses the next word. One token at a time. Left to right.
It’s like writing an essay on a typewriter while looking through a straw. It is computationally exhausting and incredibly slow.
Some labs tried to fix this with "diffusion models”, generating everything at once and polishing it over and over. But that requires massive compute and endless denoising passes.
A team from Stanford just dropped a paper that kills both methods.
It’s called "Next-Concept Modeling" (NCM).
Instead of predicting the next word, the AI predicts the next concept.
It generates text exactly how a professional writer drafts a book. Coarse to fine.
First, it generates a high-level conceptual outline of the entire sequence. Then, in parallel, it fills in the precise words and details across the whole document at once.
The results are staggering.
• Generates text up to 31x faster than current diffusion language models.
• Requires drastically fewer computing steps to reach high-quality outputs.
• Actually outperforms traditional step-by-step models on complex reasoning tasks like Sudoku.
AI is no longer blindly typing word by word, hoping the sentence makes sense when it reaches the period.
It is conceptualizing the entire thought before it opens its mouth.
Stanford researchers did it again.
They made language models 31× faster.
Right now, every major AI, from ChatGPT to Claude, writes exactly the same way. It guesses the next word. One token at a time. Left to right.
It’s like writing an essay on a typewriter while looking through a straw. It is computationally exhausting and incredibly slow.
Some labs tried to fix this with "diffusion models”, generating everything at once and polishing it over and over. But that requires massive compute and endless denoising passes.
A team from Stanford just dropped a paper that kills both methods.
It’s called "Next-Concept Modeling" (NCM).
Instead of predicting the next word, the AI predicts the next concept.
It generates text exactly how a professional writer drafts a book. Coarse to fine.
First, it generates a high-level conceptual outline of the entire sequence. Then, in parallel, it fills in the precise words and details across the whole document at once.
The results are staggering.
• Generates text up to 31x faster than current diffusion language models.
• Requires drastically fewer computing steps to reach high-quality outputs.
• Actually outperforms traditional step-by-step models on complex reasoning tasks like Sudoku.
AI is no longer blindly typing word by word, hoping the sentence makes sense when it reaches the period.
It is conceptualizing the entire thought before it opens its mouth.
