Google ships too. Not just Perplexity or OpenAI.
@aravsrinivas
-
Human-AI Effort Distribution Remains at 80-20 Split
By
–
Hard to say. But probably still 80-20 in terms of human-AI effort.
-
GPT-5 Training Data Accelerates for AI Research Writing
By
–
The training data for GPT-5 to be good at writing its own papers (which is a good skill to have for an ultimate AI Scientist) is exponentially accelerating.
-
Regulating LLMs Versus Search Engine Information Asymmetry
By
–
This isn’t about Google search quality. It’s intended to emphasize the irony of regulating an LLM for answering “how to rob a bank” when you can a ton of ideas on how to do it on Google.
-
Stopping Open Source AI Regulation Protects Market Competition
By
–
Banning (or even regulating and slowing down) open source like llama 3 and mistral 2 is functionally equivalent to anti-competitiveness. So we should do what needs to be done to stop it from happening.
-
Open Source Preservation: A Critical Need for Tech Community
By
–
Open Source needs to be preserved.
-
AI Speed Competition Dominates Performance Benchmarks
By
–
Instant dopamine rush watching AIs (type)race against each other on @braintrustdata. Slow is fake. Speed is king. https://t.co/IYVtMgItTq
— Aravind Srinivas (@AravSrinivas) 3 novembre 2023Instant dopamine rush watching AIs (type)race against each other on @braintrustdata. Slow is fake. Speed is king.
-
FSD predicts actions over frames for world model development
By
–
Looks like end to end FSD likely models future actions rather than future frames. And by focusing on predicting something less noise and more signal, it builds a good world model for a physical world generally intelligent agent.
-
Infrastructure Role Open for Rust Developers
By
–
Consider applying for our infrastructure role if you want to write code in Rust!
-
ChatGPT Architecture: LeCun’s Framework Explained Simply
By
–
ChatGPT in a nutshell. Framework predicted by @ylecun in 7 years ago.
Bulk of compute (cake) on predicting next word from internet.
Next small fraction of compute (icing) on imitating select curated footprints of humans.
And final smallest fraction (cherry) on modeling their