Japan's Sakana AI launched a self-improving agent that can modify its code to boost performance On SWE-bench, the Darwin Gödel Machine improved its performance from 20.0% to 50.0%, while on Polyglot, it increased its success rate from 14.2% to 30.7%
RESEARCH
-

DeepSeek R1 Update Reaches #3 on Leaderboard
By
–
DeepSeek updated its R1 model, instantly propelling it to the #3 spot on the Artificial Analysis leaderboard with: —Enhanced front-end capabilities
—Reduced hallucinations
—JSON output & function calling support -

fastcore-documents meets DSPy: AI open source crossover
By
–
fastcore-docments + @DSPyOSS — crossover of the year!
-
AI Programming Languages: Optimizing Code for Model Exploration
By
–
It may be time to develop AI programming languages. Code generation must be optimized for guiding models in exploring solution space and ensuring correctness, not for human comprehension. Code specification must optimize synchronization between human intention and AI
-
Trump’s 2026 NSF Budget Request Cuts 55% from 2024
By
–
The 2026 NSF budget request from Trump: -55% from 2024.
-
Attracting Scientists to Build a Vibrant Tech Industry
By
–
How to attract scientists to your country and jump-start a vibrant tech industry. https://
docs.google.com/document/d/1Ll
NCY9_4MXhilnbu9MqKhSjHjMGqtDA8zZmSQynVjmY
… -

News Influencers Surge to Social Media Platforms Post-Election
By
–
New research from @KnightCenterUT and @pewresearch found the share of news influencers with a account jumped from 21% before Election Day to 43% by March 2025. Charts with more stats:
-

Language Models Memory Capacity Scaling Laws Physics
By
–
Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Laws very interesting (if somewhat controversial) they declare that models, after visiting data 1000 times during training, can memorize 2 bits/param (they arrive at this number via quantization with AutoGPTQ)
-

Scaling Laws for Fact Memorization in Large Language Models
By
–
Scaling Laws for Fact Memorization of Large Language Models this paper is from last year (2024); it's a very well-executed study of the total number of facts a model can memorize they estimate 1T parameters are needed to perfectly memorize all relational pairs in wikipedia
-

Knowledge Storage Capacity in Language Model Parameters
By
–
How Much Knowledge Can You Pack Into the Parameters of a Language Model? enter, transformers- in 2020 these folks had a different idea: once "pretraining" is a thing, you can measure the amount of 'facts' that the model 'knows', without training it at all
