Can fine-tuning make a language model forget too much? A team from CUHK, Westlake University, and MPI presents PEFT-Arena – a benchmark that tracks both task performance and retention of pretrained knowledge. Their analysis finds orthogonal finetuning achieves the best
LLMS
-
Gary Marcus criticizes confusion between pure and neurosymbolic LLMs
By
–
people with advanced degrees who can’t distinguish between pure LLMs (which is what I critiqued in 2022) and LLMs enhanced with neurosymbolic techniques (which is what I championed in 2022) disappoint me. i get that many tech bros don’t really get it, but scientists should take
-
Neurosymbolic AI surpasses pure LLMs as predicted
By
–
1. pure LLMs did in fact hit a wall and then as i predicted (2022) neurosymbolic AI provided a way past the LLMs; it’s very evident if you look at all the tools, harnesses, loops etc in Codex, Claude Code etc as a scientist i would have expected you to be more attuned to that
-
Longevity escape velocity with only GPT 5.5 and Claude 4.8
By
–
Imagine we achieve longevity escape velocity, and all of us live forever, but the best AI models we’ll ever be allowed to use are GPT 5.5 and Claude 4.8.
-
Gary Marcus: AI scaling is capital misallocation, but early days
By
–
“Gary Marcus … says that investments into scaling AI is the “greatest capital misallocation in history” which leaves everyone “on the hook”. However, he also says that it is still early days for the AI industry and further exploration into different model architectures and the
-
US blocking model releases forces labs to keep models and integrate
By
–
The logical (and unfortunate) consequence of the US government blocking model releases is that the labs would have to keep the best models to themselves and vertically integrate into intelligence heavy industries to generate revenue
-
Kevin Explains Claude Ecosystem: Chat, Co-working, Coding, Design
By
–
Watch this before you build anything with Claude – Kevin breaks down the full Claude ecosystem in one video. Overview of Claude’s ecosystem and how its main tools work together. Different ways Claude can be used for chat, co-working, coding, and design. Breakdown of
-
Companies overlook fine-tuned self-hosted GLM; users prefer brand names
By
–
Most companies are not even thinking about giving every employee a fine-tuned self-hosted version of GLM that they need to update and keep harnessed and build connectors for. There are lots of places where roll-your-own AI stacks make sense, but end users want name brands.
-
Companies’ AI stack plans vs employee demand for simple ChatGPT access
By
–
I feel like on X all you hear about is elaborate plans by firms to build their own AI stacks but in my experience companies are full of people who want access to Claude or ChatGPT and are pressuring their purchasing staff to get licenses so they can just use the tools they know.
-
Use a MCP gateway to get all connectors into Claude Tag immediately
By
–
You can also use a mcp gateway and get all of the connectors into Claude Tag immediately