for the open reasoning models, response quality stayed the same not "slightly degraded." same. and here's the number that stops you: omitting assistant-side history reduced cumulative context lengths by up to 10x 10x less context. same quality. on real conversations.
LLMS
-
Real Chat Tests on Four AI Models
By
–
the experiment is clean they took real multi-turn conversations from WildChat and ShareLM. not synthetic benchmarks. actual human-ai chats then they ran every conversation two ways across four models (Qwen3-4B, DeepSeek-R1-8B, GPT-OSS-20B, and GPT-5.2): > full context: normal.
-

MIT finds LLM context pollution degrades performance
By
–
MIT researchers discovered a phenomenon called "context pollution" where llms get WORSE by reading their own prior responses errors, hallucinations, and stylistic artifacts from earlier turns propagate forward because the model treats its own output as ground truth and removing
-

Video Models Over Language Models for Robotics Manipulation Tasks
By
–
Taking inspiration from VideoJAM, would the physical consistency of the generated videos also improve? Accurate action prediction requires physically plausible imagination, conversely physically plausible imagination is best supported when it is consistent with feasible actions. Seonghyeon Ye (@SeonghyeonYe) VLAs (from VLMs) ❌ => WAMs (from Models) ✅ Why WAMs? 1️⃣ World Physics: VLMs know the internet, but Models implicitly model the physical laws essential for manipulation. 2️⃣ The "GPT Direction": VLAs are like BERT (rely heavily on task-specific post-training). WAMs are like GPT (pre-train & prompt), unlocking incredible zero-shot transfer! What I want to see in 2026: 📈 Scaling Laws: We will see much clearer scaling laws for robotics compared to VLAs. 🤝 Human-to-Robot Transfer: Unlocking massive transfer capabilities using video as a shared representation space. 🤖 Zero-Shot Mastery: Moving from short-horizon tasks to long-horizon, dexterous manipulation without task-specific demonstrations. We recently open-sourced the checkpoints, training and inference code. Dive into the research! 👇 📄 Paper: arxiv.org/abs/2602.15922 💻 Code: github.com/dreamzero0/dreamz… 🤗 HF: huggingface.co/GEAR-Dreams/D… — https://nitter.net/SeonghyeonYe/status/2024501978106061056#m
→ View original post on X — @shiqi_yang_147, 2026-03-03 10:53 UTC
-
Using Claude prompts to accelerate book writing
By
–
Le livre que tu repousses depuis 3 ans peut être terminé en 72 heures. Ce qui t'en empêchait n'était pas le manque de temps. C'était de ne pas connaître ces 9 prompts Claude: [ Ajoutez en signet pour ne pas perdre ! ]
-
Using Claude Prompts to Accelerate Book Writing
By
–
Le livre que tu repousses depuis 3 ans peut être terminé en 72 heures. Ce qui t'en empêchait n'était pas le manque de temps. C'était de ne pas connaître ces 9 prompts Claude: [ Ajoutez en signet pour ne pas perdre ! ]
-
Efficient and Cost-Effective AI Model Execution Compared
By
–
it’s quite efficient to run — definitely cheaper than nano banana!
-
Qwen’s New Models Defy Expectations
By
–
Qwen just dropped 4 new models and the math doesn’t make sense.
— God of Prompt (@godofprompt) 3 mars 2026
> 4B nearly matches their previous 80B A3B
> 9B rivals GPT OSS 120B at 13x smaller
> 0.8B and 2B run on your phone
All free, offline, and open source
Let that sink in. https://t.co/1rqKBiiCgR pic.twitter.com/0pCYM8tRTjQwen just dropped 4 new models and the math doesn’t make sense. > 4B nearly matches their previous 80B A3B
> 9B rivals GPT OSS 120B at 13x smaller
> 0.8B and 2B run on your phone All free, offline, and open source Let that sink in. -
LLMs surveillance concerns echo Apple backdoor fears
By
–
Something that didn't fit into this week's newsletter is how new concerns with LLMs echo the same issues Apple feared. 2016: @tim_cook warned forcing Apple to create a backdoor could make it easier for the government to someday also demand that Apple build surveillance software