OK, I can partly explain the LLM chess weirdness now
“People have too inflated sense of what it means to “ask an AI” about something. The AI are language models trained basically by imitation on data from human labelers. Instead of the mysticism of “asking an AI”, think of it more as “asking the average data labeler” on the” / X
If AGI arrives during Trump’s next term, ‘none of the other stuff matters’ – Fast Company
The Bitter Religion: AI’s Holy War Over Scaling Laws | The Generalist
“A new study showed ChatGPT achieved 90% accuracy in medical diagnosis, outperforming both human doctors (74%) and doctors using ChatGPT (76%) So much progress to be made for AI and healthcare. Really cool to already start seeing these results already
Valuing Humans in the Age of Superintelligence: HumaneRank
AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably | Scientific Reports
How Do You Get to Artificial General Intelligence? Think Lighter | WIRED
“One blindspot for AI reasoning engines like o1 is that they all appear to be trained on very traditional deductive problem solving. What would a model trained on induction or abduction do? What about one trained on free association? Expert heuristics? Randomized exquisite corpse?
Something weird is happening with LLMs and chess
“Yann Lecun says his estimate for the creation of human-level AI is not that different from Sam Altman or Demis Hassabis – it is possible within 5-10 years
AI that mimics human problem solving is a big advance – but comes with new risks and problems
“The 1st International Network of AI Safety Institutes meeting last week in SF marked a major step in global collaboration on AI Safety. The AISIs can help both on the policy and science fronts: global collaboration and international agreements, technical guidance and standards” / X
The new AI scaling law shell game – by Gary Marcus
A Chinese lab has released a ‘reasoning’ AI model to rival OpenAI’s o1 | TechCrunch
A Minecraft town of AI characters made friends, invented jobs, and spread religion | MIT Technology Review
Interpretability
“16th highest scored paper at ICLR 2025 with 3(!), 8, 10, 10, 10 tldr: they scale sparse autoencoders to GPT4 and show that interpretability techniques used on toy models can work on larger models too (hmm i wonder who these people who have access to GPT4 activations are!)
Leave a Reply