Technology

Jan-nano is a model fine-tuned with DAPO on Qwen3-4B. Jan-nano comes with some unique capabilities:

• It can perform deep research (with the right prompting)
• It picks up relevant information effectively from search results
• It uses tools efficiently

The model was evaluated using SimpleQA - a relatively straightforward benchmark to test whether the model can find and extract the right answers.

Jan-nano outperforms Deepseek-671B on this metric, using an agentic and tool-usage-based approach. A 4B model obviously has its limitations, but it's interesting to see how far these things can be pushed. Jan-nano can serve as your self-hosted Perplexity alternative on a budget.

You can find the model at: https://huggingface.co/Menlo/Jan-nano

And a gguf is available at: https://huggingface.co/Menlo/Jan-nano-gguf

https://archive.ph/P4lwg

Hi everyone - as a previous context I’m an AI Program Manager at J&J and have been using Cursor for personal projects since March.

Yesterday I was migrating some of my back-end configuration from Express.js to Next.js and Cursor bugged hard after the migration - it tried to delete some old files, didn’t work at the first time and it decided to end up deleting everything on my computer, including itself. I had to use EaseUS to try to recover the data, but didn’t work very well also. Lucky I always have everything on my Google Drive and Github, but it still scared the hell out of me.

Now I’m allergic to YOLO mode and won’t try it anytime soon again. Does anyone had any issue similar than this or am I the first one to have everything deleted by AI?

• Thousands of Chinese researchers and scientists are leaving top jobs in leading US universities and companies, to take positions in China.
• The Cambridge area of Massachusetts is home to Harvard, MIT, and scores of leading companies, and was the number one source of returning Chinese research and engineering talent.
• In second place is the Palo Alto-Berkeley cluster, which includes Stanford, University of California, and Silicon Valley.
• The migration of top scientific and engineering talent back to China is accelerating, but began nearly a decade ago. And while the political situation between China and the United States certainly is a major motivation for many scientists to return, more important is the quality of the education systems.
• Chinese universities are now claiming the top spots across all the hard science disciplines, while American colleges are tumbling.

YouTube video.

https://archive.vn/5gRV3

https://huggingface.co/papers/2506.05573

https://archive.ph/ddOJV

Apple published a paper criticizing the capabilities of Large Language Models (LLMs) in reasoning and formal logic. The paper builds on previous arguments made by Gary Marcus and Subbarao Kambhampati about LLMs' limitations in generalizing beyond their training distribution.

The authors of the paper demonstrated that even the latest "reasoning models" fail to reason reliably on classic problems like the Tower of Hanoi. LLMs cannot solve the Tower of Hanoi reliably, even with the solution algorithm given to them.

The paper argues that LLMs are not a substitute for well-specified conventional algorithms and have limitations that are becoming clearer. LLMs are not a direct route to AGI and while the field of neural networks is not dead, current approach has clear limitations.

The paper highlights the importance of combining human adaptiveness with computational brute force and reliability in AI development.

https://archive.ph/Rgbkz

https://archive.ph/LJJpx

AI companies claim their tools couldn't exist without training on copyrighted material. It turns out, they could and it just takes more work. To prove it, AI researchers trained a model on a dataset that uses only public domain and openly licensed material.

What makes it difficult is curating the data, but once the data has been curated once, in principle everyone can use it without having to go through the painful part. So the whole "we have to violate copyright and steal intellectual property" is (as everybody already knew) total BS.

The project implements sparse multiplication and fuses up/down projections in the MLP layers through low rank weight activations. Work is based on Deja Vu and Apple's LLM in a Flash.

This approach avoids loading and computing activations with feed forward layer weights whose outputs will eventually be zeroed out.

It's a lossless approach as these weights anyway do not contribute in the current token prediction. It does however, need the predictors to be accurate in clustering the weights.

The result? 5X faster MLP layer performance in transformers with 50% lesser memory consumption avoiding the sleeping nodes in every token prediction. For Llama 3.2, Feed forward layers accounted for 30% of total weights and forward pass computation resulting in 1.6-1.8x increase in throughput:

Sparse LLaMA 3.2 3B vs LLaMA 3.2 3B (on HuggingFace Implementation):

- Time to First Token (TTFT):  1.51× faster (1.209s → 0.803s)
- Output Generation Speed:     1.79× faster (0.7 → 1.2 tokens/sec)  
- Total Throughput:            1.78× faster (0.7 → 1.3 tokens/sec)
- Memory Usage:                26.4% reduction (6.125GB → 4.15GB)