←
How To Prompt1,860 posts
How To Prompt avatar

How To Prompt

@HowToPrompt__

Trustworthy AI education.

◷ Joined January 2022▧ 557 Photos and videos
How To Prompt avatar
How To Prompt@HowToPrompt__·Apr 20

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.

Post attachment
How To Prompt avatar
How To Prompt@HowToPrompt__·1h

Finally a good RSI benchmark

Introducing FrontierRSI, a new recursive self-improvement benchmark by @abundant_labs. We evaluate agents on (1) a broad range of AI research and engineering work, and (2) post-training models on multi-step agentic tasks, rather than single-step tasks as in PostTrainBench.

Post attachment
How To Prompt avatar
How To Prompt@HowToPrompt__·4h

Researchers made LLMs 21% smarter without retraining a single weight. And it feels like cheating. Right now, every time you prompt a large language model, it operates on a strict, one-way assembly line. Information goes in at the bottom. It passes through layer 1, layer 2, all the way to the top. And then it spits out a token. It gets one single shot to figure out the answer. No second looks. No double-checking. If a problem is hard, the model has to brute-force its way through using massive, expensive parameter counts. A new paper called Recirculation completely breaks this straight-line architecture. Instead of pushing data through the network once, researchers figured out how to loop the representations from the top layers right back into the bottom layers. The model literally recirculates its own thoughts. It takes its rough first draft, feeds it back through its own brain, and refines the logic before ever outputting a single word. Here is why this changes the math on AI efficiency: • Tiny models can punch way above their weight class by "thinking" deeper instead of growing wider. • You get dynamic compute—simple questions exit early, while complex logic loops back until it's solved. • Zero extra parameters required. You are unlocking hidden intelligence already trapped inside the weights. For the last few years, the entire AI industry has been obsessed with scaling up. Bigger models. More GPUs. Trillions of parameters. Recirculation proves that raw size was never the real answer. True intelligence isn't about having a bigger brain. It's about taking a second look at your own work.

Post attachment
How To Prompt avatar
⟳ How To Prompt retweeted
How To Prompt@HowToPrompt__·6h

Turns out everyone’s been training RL on the wrong precision. A new paper shows FP16 beats BFloat16 for RL fine-tuning.. by a lot.

Post attachment
How To Prompt avatar
⟳ How To Prompt retweeted
How To Prompt@HowToPrompt__·8h

Turing Award winner Andrew Yao just teamed up with Kimi founder Yang Zhilin for the first time ever. They dropped a brutal theoretical computer science research exam built to break every model. It’s called TCSExam, the first benchmark evaluating frontier LLMs on the full range of Theoretical Computer Science (TCS). We're talking algorithmic complexity, cryptography, automata theory, and rigorous mathematical proofs. The actual physics of computation. And the results are brutal. Frontier models, the smartest systems on the planet—spectacularly failed. When asked to write boilerplate web code, they look like geniuses. But when forced to reason through actual theoretical computer science, their performance completely collapses. Here is what happens when you push them past their training data: • They hallucinate proofs that look mathematically perfect but are logically broken. • They fail at multi-step deduction that requires true reasoning instead of pattern matching. • They rely heavily on memorization, masquerading as intelligence. This exposes a massive illusion in the AI industry. LLMs haven’t learned how to reason. They have just memorized the syntax of reasoning. If a problem has been solved on StackOverflow a thousand times, the AI is flawless. But the moment you step off the map into pure, unmapped theory, the illusion shatters.

Post attachment
How To Prompt avatar
⟳ How To Prompt retweeted
How To Prompt@HowToPrompt__·11h

It’s happening. Google built an AI that rewrite its own weights to get smarter.. and it does it by sleeping. Right now, every AI model you use has a fatal flaw. When you train a model on new data, it suffers from "catastrophic forgetting." It forcefully overwrites its old knowledge to make room for the new. If you try to bypass this by using a massive context window, the AI simply forgets everything the second you close the chat. Once a model is deployed, its brain is effectively frozen in time.. But Google Research published a paper introducing a biological "Sleep Framework" for LLMs. They split the AI's existence into two distinct states. 1. The Wake Phase: AI interacts with you. It solves problems, processes data, and stores new information in a fragile, short-term memory bank. 2. The Sleep Phase: When the system goes offline, the AI literally goes to sleep. It unlocks dormant parameters and actively transfers its short-term experiences into permanent, long-term memory. It learns forever without overwriting the past. But then, the AI starts dreaming. During the sleep phase, the model deliberately activates random, unusual neural pathways. It generates synthetic scenarios, literally dreaming up fake data, and tests itself to improve its own reasoning and uncover hidden patterns. The results are staggering. • It retained perfect knowledge across a massive 10-million token window. • It learned entirely new, unseen languages with zero performance drop. • It crushed elite math benchmarks (AIME), jumping to nearly 80% accuracy. Zero catastrophic forgetting. Continuous, autonomous self-improvement. We’ve spent billions trying to build bigger context windows so AI can hold more information at once. But the future of AI isn't about cramming more data into a single prompt. It's about letting the machine process its experiences after you log off.

Post attachment
Higgsfield AI 🧩 avatar
⟳ How To Prompt retweeted
Higgsfield AI 🧩@higgsfield_ai·Nov 17

GPT-6.2 Astra is LIVE on Higgsfield. OpenAI’s new frontier model for your most demanding projects. With Higgsfield plugin, GPT-6.2 Astra powers proactive multi-agent workflows across design, coding, and creative production. Join Higgsfield with up to 65% off for Black Friday.

1:15
OpenAI@OpenAINov 17

Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…

1:59
How To Prompt avatar
⟳ How To Prompt retweeted
How To Prompt@HowToPrompt__·Nov 17

Researchers argue tea saved more English lives than any doctor, hospital, or medicine in the 1700s. They published a massive economics paper called “For want of a cup” We always thought England dominated the 18th century because of coal, steam engines, and raw engineering. but the math says something completely different. In the 18th century, english drinking water was incredibly toxic and water-borne diseases were killing the population. then, tea suddenly became a massive cultural trend. but here is the catch: to make tea, you have to boil the water. a researcher ran a difference-in-difference model comparing parish-level mortality rates before and after tea adoption across england, and the results are actually insane: - mortality rates plummeted in areas that adopted tea - deaths specifically from water-borne illnesses dropped massively - young children didn't see the same health bump, proving it was the tea consumption itself since kids weren't major consumers - results hold up even when controlling for local income and access to trade. an entire nation accidentally sanitized its water supply just because they wanted a warm beverage.. this unintentional health hack saved the workforce and provided the population boom needed to power the british industrial revolution. history is wild..

Post attachment
How To Prompt avatar
⟳ How To Prompt retweeted
How To Prompt@HowToPrompt__·Nov 17

Neuroscientists discovered your brain believes whatever you think on repeat. And it doesn't care if it's true. Psychologists have long known about the “illusory truth effect.” When you hear something multiple times, your brain processes it more easily. Because it feels easy to process, your brain automatically assumes it must be true. But researchers just published a study that finally measured exactly how many times you have to lie to someone before that lie becomes their reality. They didn’t just test the usual two or three repetitions. They bombarded people with the exact same fake statements up to 27 times. And here’s what they found. The increase in belief isn't a straight line. It's logarithmic. The single most dangerous moment in the lifecycle of a lie is the second time you hear it. That first repetition creates a massive, disproportionate spike in perceived truth. The moment you hear a lie twice, the neural pathway is paved. The resistance drops. After that, every additional repetition (from 3 up to 27) continues to increase belief, but in smaller, steady increments. They aren't trying to convince you anymore. They are just driving the nail deeper. Why does this matter right now? Look at your feed. Look at social media algorithms. Look at the modern media ecosystem. The people hacking elections, selling you garbage, and spreading misinformation aren't trying to outsmart you. They don't need a bulletproof argument. They just need to get the exact same sentence in front of your eyes twice. Your brain’s "processing fluency" does the rest of the work for them. We like to think we are rational creatures who evaluate information based on evidence. But the data is blunt.

Post attachment
How To Prompt avatar
⟳ How To Prompt retweeted
How To Prompt@HowToPrompt__·Nov 16

Yejin Choi was right the entire time. In 2023, while everyone was losing their minds over GPT-4, she stood up in Berkeley and showed it couldn’t reliably multiply two 3-digit numbers. 59% accuracy. 4% on 4-digit numbers. 0% on 5-digit. Even after fine-tuning a model on 1.8 million multiplication examples, it still only hit 55%. Her paper, “Faith and Fate,” explained why: these models don’t learn how to solve problems. They memorize patterns and match them. The industry’s answer was to make the models bigger. But there was a deeper problem nobody fixed: Once a model is trained, it’s frozen. It never learns another thing. Every mistake it makes today, it makes again tomorrow. Until now. Choi just came out of stealth with Noeri, co-founded with Carlos Guestrin (co-creator of XGBoost, former head of ML at Apple). They’re building AI that keeps improving after it ships: updating its own instructions, tools and models inside your company. Every change is explained and approved by a human, like a code review. And the company owns it. Not a lab. The researcher who showed us where AI breaks is now building AI that learns from its mistakes.

Post attachment

Through my wonderful time at NVIDIA and Stanford, one question has kept me up at night: Should the future of AI be monolithic or plural? To answer this question, we’ve raised $110M to get started, backed by @MenloVentures, @MadronaVentures, @NVIDIA, @NEA, @KindredVentures,@Wing_VC, @aixventureshq, @Principal_VC, @Cisco, @Snowflake, among others. Today, Carlos (@guestrin) and I are stoked to launch Noeri, a neolab to push the frontier of plural intelligence. Noeri means “the plural of intelligence” or many intelligences. The future of AI we root for is plural: a healthy ecosystem of many intelligences, in all sizes, strengths, and quirks. Discovery, in particular, has always been plural: many curious minds, each devoted to their own question, often on roads less traveled. Our path there is self-improving AI that: -> learns continually on the job, closing the gap between what AI can do off the shelf and what the dynamic world demands -> learns to use other AIs as tools and make new tools to overcome its own limitations -> picks the right-sized model for your problem, at a fraction of the frontier cost Then the power of intelligence, from the craftsmanship of making AI to ownership, can be in the hands of all. @noeri_ai mission reflects who we are as the founding team. My career has been anything but linear, literally continual learning on the job. Carlos has devoted his career to building tools that empower people through AI. We are all a bit contrarian, a bit adventurous, and above all, big believers in pluralism. Join us to build this future together: noeri.ai/

2:17
Higgsfield AI 🧩 avatar
⟳ How To Prompt retweeted
Higgsfield AI 🧩@higgsfield_ai·Nov 16

Introducing Short Drama Studio. Drop a script → get a full season of short drama. > Any style, from dark fantasy to 90s high school > Same cast in every episode > 9:16, 16:9 or 4:3 Up to 65% OFF on Higgsfield for Black Friday. Available now on Higgsfield and in Claude and ChatGPT via Higgsfield MCP.

1:49
Yejin Choi avatar
⟳ How To Prompt retweeted
Yejin Choi@YejinChoinka·Nov 16

Through my wonderful time at NVIDIA and Stanford, one question has kept me up at night: Should the future of AI be monolithic or plural? To answer this question, we’ve raised $110M to get started, backed by @MenloVentures, @MadronaVentures, @NVIDIA, @NEA, @KindredVentures,@Wing_VC, @aixventureshq, @Principal_VC, @Cisco, @Snowflake, among others. Today, Carlos (@guestrin) and I are stoked to launch Noeri, a neolab to push the frontier of plural intelligence. Noeri means “the plural of intelligence” or many intelligences. The future of AI we root for is plural: a healthy ecosystem of many intelligences, in all sizes, strengths, and quirks. Discovery, in particular, has always been plural: many curious minds, each devoted to their own question, often on roads less traveled. Our path there is self-improving AI that: -> learns continually on the job, closing the gap between what AI can do off the shelf and what the dynamic world demands -> learns to use other AIs as tools and make new tools to overcome its own limitations -> picks the right-sized model for your problem, at a fraction of the frontier cost Then the power of intelligence, from the craftsmanship of making AI to ownership, can be in the hands of all. @noeri_ai mission reflects who we are as the founding team. My career has been anything but linear, literally continual learning on the job. Carlos has devoted his career to building tools that empower people through AI. We are all a bit contrarian, a bit adventurous, and above all, big believers in pluralism. Join us to build this future together: noeri.ai/

2:17