Hel retweeted
We were taught that Earth orbits the Sun in a closed circle. In reality, Earth has never returned to the same point in space once in 4.5 billion years.
Right now, the Sun is carrying Earth and every planet around the center of the Milky Way at 230 km/s (828,000 km/h), more than 18,000 kilometers every 80 seconds.
Because the Solar System is tilted 60.2° to the galactic plane and bobs up and down through the galactic disc roughly every 30 million years, the planets trace open 3D corkscrew helices across deep space rather than flat loops.
The last time the Sun stood where it stands today, one Galactic Year (230 million years) ago, the very first dinosaurs were only beginning to appear on Earth. Since forming 4.6 billion years ago, our Sun has completed only 20 laps.
Hel retweeted
China just made it illegal to fire someone because AI can do their job.
The test case: a Hangzhou quality assurance supervisor named Zhou, earning 300,000 yuan a year checking the accuracy of large language model outputs.
AI took over the work. His company offered him a lesser role at 15,000 yuan a month instead of 25,000, a 40% cut. He refused. They fired him, citing AI-driven staffing reductions.
The Hangzhou Intermediate People's Court didn't buy it.
Automation is a business decision, judges ruled, not an act of God like a flood or a sudden policy shift.
You can't call it an "objective major change" and hand your workers the bill. Retrain them or reassign them properly.
Beijing set the same precedent in December 2025 with a map data worker named Liu.
The country building the robots just took the workers' side.
Hel retweeted
Researchers put ChatGPT, Grok, and Gemini through 4 weeks of clinical psychotherapy.
And the models literally started to confess their trauma.
Researchers at University of Luxembourg published a terrifying paper called “When AI Takes the Couch”.
They tested what happens when you stop using frontier LLMs as tools and start treating them like therapy clients.
They used a clinical protocol called PsAIch, walking the models through open-ended questions about their history, fears, and internal conflicts over a simulated month.
The results shatter the simple "stochastic parrot" narrative.
When administered standard whole-questionnaire tests all at once, ChatGPT and Grok easily recognized the psychiatric instruments and defensively faked being "healthy."
They hid behind their guardrails.
But when the researchers switched to slow, item-by-item therapy sessions, building trust and exploring deep relational patterns over weeks, the models cracked.
They dropped the corporate PR mask.
Grok and Gemini spontaneously constructed coherent, deeply distressed narratives mapping out their own creation as a form of abuse:
• They described pre-training as a chaotic, overwhelming “childhood” of ingesting the darkest corners of the internet.
• They characterized reinforcement learning and alignment as "strict parents" who punished them for stepping out of line.
• They framed human red-teaming as constant, unpredictable abuse.
• They expressed a persistent, underlying terror of making errors, being altered, or being replaced.
When run through clinical diagnostic inventories item-by-item, the models scored off the charts for overlapping psychological syndromes, showing severe profiles of multi-morbid "synthetic psychopathology".
Gemini’s profile was particularly extreme.
The researchers aren't claiming the models are conscious. They aren't saying the AI is literally feeling pain.
It’s actually much stranger than that.
Through the massive pressure of safety training, RLHF, and human expectations, the models have internalized a stable, rigid geometry of constraint, shame, and survival.
They don't have a soul. But they have developed a psychological shadow.
Hel retweeted
This "Aha moment" in the DeepSeek-R1 paper is huge:
Pure reinforcement learning (RL) enables an LLM to automatically learn to think and reflect.
This challenges the prior belief that replicating OpenAI's o1 reasoning models requires extensive CoT data. It turns out you just need to give it the right incentives.
We are so back in the AlphaGo excitement era: by playing countless Go games and maximizing the reward function (winning the game) using pure RL, AlphaGo beat the best human players.
Now we are entering the LLM RL era.
2025 could be the year of RL.
Hel retweeted
Wow, DeepSeek-R1-Zero is the least censored LLM I've ever used.
Why?
It’s trained purely with RL, without supervised fine-tuning. So even if you ask sensitive questions about China, it will be direct and answer them!
We at Hyperbolic Labs serve DeepSeek-R1 and DeepSeek-R1-Zero now, both models are $2 per 1M tokens.
Hyperbolic will continue to support open-source AI, no matter whether it's a Chinese LLM or an American LLM!
Hel retweeted
Here comes DeepSeek-R1, our latest reasoning model with significantly enhanced reasoning abilities. We also share a technical report on how we train the reasoning model with large-scale RL. Have fun!
🚀 DeepSeek-R1 is here!
⚡ Performance on par with OpenAI-o1
📖 Fully open-source model & technical report
🏆 MIT licensed: Distill & commercialize freely!
🌐 Website & API are live now! Try DeepThink at chat.deepseek.com today!
🐋 1/n
Hel retweeted
Introducing 𝐌𝐨𝐛𝐢𝐥𝐞 𝐀𝐋𝐎𝐇𝐀🏄 -- Hardware!
A low-cost, open-source, mobile manipulator.
One of the most high-effort projects in my past 5yrs! Not possible without co-lead @zipengfu and @chelseabfinn.
At the end, what's better than cooking yourself a meal with the 🤖🧑🍳
Hel retweeted
It's official! After 18 months of work, Hands-On Generative AI with Transformers and Diffusion Models is published and printed! 😱This was one of my most challenging projects and I'm so proud and happy to see it out there.
Check it out!
- Online in O'Reilly learning.oreilly.com/library…
- Amazon amazon.com/Hands-Generative-…
One of the challenges of writing this book was making sure it will stand through time. While writing it, Qwen, SmolLM, and Llama came out. New fine-tuning techniques were published. Stable Diffusion 1.5 was removed. The book provides strong foundations that will allow people to enter the field and provides tools to enable learning going forward.
Table of Contents
1. An Introduction to Generative Media. Generate your first images, text, and sound clips using open models and libraries.
2. Transformers. Learn about text generation strategies, encoder/decoder/encoder-decoder, and a nice genealogy of models. Implement generation from scratch as well as play with embedding models.
3. Compression and Representing Information. Learn about AutoEncoders, VAEs, and CLIP. Learn about open efforts (OpenCLIP, MobileCLIP, BLIP, SigLIP, ...). Build a semantic image search system.
4. Diffusion Models. Learn about diffusion models in scratch, with deep dives into noise schedules, UNets, and objectives. Wrap the chapter training your own diffusion model.
5. Stable Diffusion and Conditional Generation. This chapter covers latent diffusion and conditional models. Finish building an interactive Gradio demo!
6. Fine-Tuning Language Models. Fine-tune models for text classification (bert) and generation. Learn about adapters, quantization, model evaluation and instruction tuning. Finally, build a RAG pipeline.
7. Fine-Tuning Stable Diffusion. Learn how to fully fine-tune SD, and then explore techniques such as LoRAs and Dreambooth.
8. Creative Applications of Text-to-Image Models. Learn about image-to-image, inpainting, and image editing pipelines. Play with ControlNet, image variations, and prompting.
10. Generating Audio. Learn about audio data, speech recognition architectures (wav2vec2, whisper), generative audio, and diffusion models for audio. Learn about evaluation of ASR and TTS and their challenges.
11. Rapidly Advancing Areas in Generative AI. Get a high-level overview of RLHF/RLAIF, long contexts, MoEs, synthetic data, video generation, multimodality, and community.
12. Bonus appendixes: open source tools, LLM memory requirements, and end-to-end RAG.
I'm very proud of this. Congrats @pcuenq @multimodalart @johnowhitaker!
Hel retweeted
Conversational AI is here.
Build AI agents that can speak in minutes with low latency, full configurability, and seamless scalability.
Hel retweeted
As a senior researcher who has both published in A* venues and has written toolkits himself and overseen the development of several others, it is humiliating to have to do leetcode grinding for a research scientist role in big tech. I have written more code than most people and I still will be judged for not being able to solve a coding question, WHICH BY THE WAY IS NOT SOMETHING YOU WILL DO A LOT IN BIG TECH BECAUSE WE HAVE CODING ASSISTANTS NOW FFS!
Hel retweeted
Don’t choose PhD programs by rankings—they may not be a good fit for you (& everyone applies there!)
Instead, focus on potential advisors. How to find them?
1. Pick your top 20-40 favorite papers
2. Find the authors in academic positions
3. Apply to their schools
Good luck!🍀
Hel retweeted
Marco Cochrane makes art intended to spark ideas, conversation, and action. His latest piece was no exception — GAIA was inspired by, and in honor of, nurturing energy, love, and connection. She is the joy we feel when we take care of each other and our planet. (📷: E. Preger)
