A new paper just exposed how much money AI labs waste on GPUs. Running LLMs at scale hits a wall when prefill and decode share one datacenter. The KV cache transfer between them is massive. This forces expensive RDMA networks and identical hardware everywhere. Kimi proposes
LLMS
-
AI Surpasses Doctors: Narcissistic Wound Medical Profession
By
–
Les médecins n’acceptent pas que l’IA les dépasse
— Docteur Laurent Alexandre (@dr_l_alexandre) 25 avril 2026
C’est une immense blessure narcissique
Pourtant la réalité c’est que nous médecins nous sommes « Has Been » face à l’IA
Peut-on égaler l’IA en transformant les études médicales ? Pas sûr du tout 😰😰😰😰 https://t.co/DvaejqvFWMDoctors don’t accept that AI surpasses them It’s a massive narcissistic wound Yet the reality is that we doctors are “has-beens” in the face of AI Can we match AI by transforming medical studies? Not sure at all
-
GPT 5.5 emerges as most reliable coding model
By
–
pretty sure, maybe just not as reliable. GPT 5.5 is the most reliable coding model.
-
GPT-5.5 ChatGPT feels more serious and useful now
By
–
Maybe its just me, but since GPT-5.5 ChatGPT feels more… serious? More useful? Less chatty? I dont know if they even reworked the vibe and tone but I actually begin to like it again.
-
Raw Intelligence: Structuring Data for Better AI Partners
By
–
so much raw intelligence, lots of work to wrangle it into the most useful thought partner out there
-
AI Model Performance Gap Narrows: Grok-4.20 and GPT-5.4 Converge
By
–
The gap at the top of AI is closing fast. Multiple models now score at the same level with Grok-4.20 Expert Mode and GPT-5.4 Pro (Vision) leading the latest benchmark. But the bigger signal isn’t who’s #1. It’s how quickly performance is converging. When top models are
-
DeepSeek Delivers Open-Source Model Rivaling OpenAI at Lower Cost
By
–
Maybe… but that’s a different discussion. The real point is that DeepSeek seems to be delivering a fully open-source model for free, on par with OpenAI’s most expensive models, … and doing it with far fewer resources! 🙂
-
Claude CLI as AI model with terminology adjustments
By
–
claude-cli as model, yeah. you might need to rename a few words tho
-
Swap Codex with Claude: One-Line Code Change
By
–
you can swap codex with claude with one line change
-
GPT 5.5 Reviewed: Impressive Performance Across Hundred Tests
By
–
I reviewed a few 100 and they all seemed correct, GPT 5.5 is a goat.