Karpathy-inspired paper proves self-refinement destroys your AI's own output. Most AI self-refinement loops make outputs worse. Models hallucinate flaws to satisfy critique prompts. They expand scope every pass. They never stop editing. Autoreason is a new paper that fixes
LLMS
-
AI Model Evaluation Benchmarks Limited Testing Scope Critique
By
–
Same energy as "we only tested on MMLU" a year ago haha
-
Beyond Model Size: Smart Scaffolding for LLM Agents
By
–
Think building an AI agent is just a better brain? What if the real secret is what you add outside the model? A team from Shanghai Jiao Tong University, OPPO, and others argues that the future of LLM agents isn't about bigger weights, but smarter scaffolding. They introduce a
-
April LLM Releases: Gemma 4, GLM-5.1, Qwen3.6, Kimi K2.6, DeepSeek V4
By
–
April was a pretty strong month for LLM releases:
– Gemma 4
– GLM-5.1
– Qwen3.6
– Kimi K2.6
– DeepSeek V4 All are now added to the LLM Architecture Gallery. More details once I am fully back in May! -
LLM Serving Engines: vLLM, SGLang, TensorRT Optimization
By
–
How to go about learning all of this? 1st: Start with the serving engine view – vLLM: PagedAttention, continuous batching, prefix caching, CUDA graphs – SGLang: RadixAttention/prefix reuse, speculative decoding, MoE, structured/agent workloads – TensorRT-LLM: NVIDIA peak
-
DeepSeek-V4-Flash 2bit quantized model released on Hugging Face
By
–
I am refreshing https://
huggingface.co/mlx-community/
DeepSeek-V4-Flash-2bit-DQ
… with excitement waiting for the files to land! -
Claude launches ‘Prompt Master’ skill to generate perfect prompts
By
–
No more bad prompts. They just launched a free skill for Claude that writes the perfect prompt for any AI on the first try. No retries.
No credits wasted. It's called Prompt Master and it works with:
→ ChatGPT
→ Claude
→ Midjourney
→ Cursor
→ ElevenLabs
→ and 13 more -
GPT-5.5 Reinforcement Learning Scaling Across Model Sizes
By
–
GPT-5.5 by Reasoning Effort: I've asked it in Codex to create a physics-based visualisation of RL cycles for different sized models (70b, 1t, 10t), to demonstrate how the amount of RL you can do differs by model size.
— Peter Gostev (@petergostev) 26 avril 2026
My assessment of each:
– Low: weird slop
– Medium: kinda… pic.twitter.com/6YCNqPyzcRGPT-5.5 by Reasoning Effort: I've asked it in Codex to create a physics-based visualisation of RL cycles for different sized models (70b, 1t, 10t), to demonstrate how the amount of RL you can do differs by model size. My assessment of each: – Low: weird slop – Medium: kinda
-
ChatGPT Hallucinations: How Often Does It Fabricate Facts?
By
–
Have you ever caught ChatGPT making up a fact you knew was wrong?
-
Grounding: Why Perplexity Cites Sources, ChatGPT Sometimes Doesn’t
By
–
Grounding is why Perplexity always cites a source, and ChatGPT sometimes doesn't.