👀

GM Europe

As we are learning DeepSeek is one of the most sophisticated psyops of all time. Here's how it went down: 1) Release the model open source. 2) Include highly detailed papers for all other people to replicate your work. 3) Create a novel SOTA RL algo that uses less memory

When startup founders go to therapy:

It’s amazing that no model has caught up with Claude on coding. Even if they look good on benchmark they’re still not as good at generating working good looking modern web apps. Whatever magic Anthropic did seems very durable.

If there’s something worse than current EU regulation/support it’s our own pessimism. We mustn’t forget how much we can do, How agentic we actually are and how much change is possible. Lovable is obviously the best current example of EU founders just doing awesome things

It looks like the AI agent job market is arriving. I’m not entirely sure how much of Firecrawl’s post on LinkedIn was just a marketing play versus a real strategy to attract AI agents and developers, but I do believe we’re about to start discussing these agents in terms of

Everyone should be using this website to understand the inside of an LLM. I'm surprised more people don't know about it. Benjamin Bycroft made this beautiful interactive visualization to show exactly how the inner workings of each of the weights of an LLM work. Here's a link:

it’s over turns out the rl victory lap was premature. new tsinghua paper quietly shows the fancy reward loops just squeeze the same tired reasoning paths the base model already knew. pass@1 goes up, sure, but the model’s world actually shrinks. feels like teaching a kid to ace

# Embeddings are underrated (2024) just a really excellent piece of technical writing.

🧵For Claude Opus 4, we ran our first pre-launch model welfare assessment. To be clear, we don’t know if Claude has welfare. Or what welfare even is, exactly? 🫠 But, we think this could be important, so we gave it a go. And things got pretty wild…

Why can AIs code for 1h but not 10h? A simple explanation: if there's a 10% chance of error per 10min step (say), the success rate is: 1h: 53% 4h: 8% 10h: 0.002% Toby Ord has tested this 'constant error rate' theory and shown it's a good fit for the data chance of

Cognitive scientist Elan Barenholtz says memory isn't retrieval. It's generation. When you remember something, you're not accessing a stored file. You're prompting your mind, like an AI model, to synthesize a response. There is no image of your mother. He's exploring how we

This is such a bad take and quite clearly shows bad understanding of how agent coding will be implemented in the not too distant future. Obviously right now you don’t want to be coding on your phone - this isn’t what they’ll be doing though.

lessons from finetuning rerankers with LanceDB

Is AI already impacting the job market? A new paper from me, Erik Brynjolfsson, and Ruyu Chen at Stanford Digital Economy Lab digs into data from ADP. We find some of the ***first large-scale evidence of employment declines for entry-level workers in AI-exposed jobs.*** A thread on our paper: