Should AI be aligned with human preferences, rewards, or utility functions? Excited to finally share a preprint that Micah Carroll Matija Franklin Hal Ashton & I have worked on for almost 2 years, arguing that AI alignment has to move beyond the preference-reward-utility nexus!

The number of AI legislative proposals in the United States is now up to 876 bills (114 federal, 762 in states) according to MultiState.👀👀👀 At this rate, we could be on our way to seeing 1,000 AI bills in America by the close of the year. Again, this is: (1) A

This is a good example of the labor theory of creative value and why it is a fallacy. This was a fun cool idea that people loved, it came to me in a flash and I generated it in a few prompts. Would it hold more value had I spent 10 hours illustrating? x.com/louisanslow/st…

Releasing OLMoE - the first good Mixture-of-Experts LLM that's 100% open-source - 1B active, 7B total params for 5T tokens - Best small LLM & matches more costly ones like Gemma, Llama - Open Model/Data/Code/Logs + lots of analysis & experiments 📜arxiv.org/abs/2409.02060 🧵1/9

People who have followed my threads will not be surprised. 🇦🇺 : strong drop in youth suicide following the onset of the pandemic. The fearmongers and "experts" (quotes intentional) were wrong, and proper statistical analysis, principled suicidology, and skepticism was right.

🔗 Thoughts on Research Impact in AI. Grad students often ask: how do I do research that makes a difference in the current, crowded AI space? This is a blogpost that summarizes my perspective in six guidelines for making research impact via open-source artifacts. Link below.

Full slides for this talk are here: docs.google.com/presentation/d… Expect a lot of ColBERT and ColPali, with a tiny SLADE and BM25 cameo to give some context. Thanks jason liu and Dan Becker for having me!

I built a free tool to make your website more crawlable by LLMs. Link in reply 👇. We recently implemented this for Helicone using a simple /llms.txt file. Inspired by Answer.AI's proposal, this file guides large language models (LLMs) to the most relevant content on your

Hold on to this thought - I have a feeling it's soon going to be relevant to something...

OpenAI: pip install openai and set OPENAI_API_KEY Anthropic: yea same but s/openai/anthropic/g Google: oh boy. ok so you have a GCP account? no? ok go set that up. and a payment method. now make a "project". SURVEY POPUP! k now gcloud auth. wait you have the gcloud CLI right–

This is the story of 16-year-old Cara who stopped wearing a mask after being bullied at school, bringing COVID home to her immunocompromised mother, who tragically died as a result. These stories must be told and we must learn from them. independent.ie/opinion/commen…

OLMoE is cool, but, comparing it with mistral-7b and llama3.1-8b, I'm not sure it is preferable. It's the same size, faster, but performs worse. It seems the same old trade-off - MoE is faster, but less capable, than equivalent sized dense models.

The used models:

RAG is increasingly going multi-modal, but document retrieval is tough, and layout gets in your way. But it shouldn't! Introducing 🪤RAGatouille's Vision-equipped, ColPali-powered sibling: 🐭Byaldi With just a few lines of code, search through documents, with no pre-processing.

Super interesting work! Congratulations to all involved. The approach however is very similar to HelixNet, that I released about a year ago. In HelixNet, the “thinker”, “critic” and “regenerator” was 3 separate 7B networks. Either way, super cool work and we’re absolutely

Cognitive Computations Jeremy Howard imho the most interesting result from the physics of language models paper on this was that moe was way better at memorization than it had any right to be, vastly disproportionate to other gains of moe moe is fundamentally a gain along a different axis

Years ago people working in AI policy (including me) perhaps foolishly wanted to 'wake up' DC. Well, we got what we wanted! Now I think a lot of the challenge of AI policy relates to trying to stabilize this leviathan that has been woken.

Many people have been saying things like this, but this is quite false. 500m in damage isn't the end of the world -- it's a Tuesday in the global economy. Let me give examples.

Unsurprisingly, Long COVID a major issue for kids in this large follow-up study in Italy, with a 7.1% chance after 2 years of having the diagnosis. Even more usurpingly, vaccination reduced the chance of Long COVID AND reinfection. Vaccination one of many tools to reduce risk.