C'est DINGUE. J'ai uploadé mes relevés bancaires à Claude Opus 4.6. Il a analysé une année complète de transactions, séparé revenus et dépenses, et tout catégorisé. 20 minutes plus tard, la comptabilité était terminée.
Environ 11 000€ de travail comptable réglé en une
LLMS
-
AI model efficiently categorizes a year of financial transactions in minutes
By
–
-
Attention Residuals: Novel Architecture for Transformer Models
By
–
Attention Residuals Paper: https://
github.com/MoonshotAI/Att
ention-Residuals/blob/master/Attention_Residuals.pdf
…
Code: https://
github.com/MoonshotAI/Att
ention-Residuals
… -
Attention Residuals: Smart Learned Aggregation in Neural Networks
By
–
Huge! @Kimi_Moonshot presents Attention Residuals (AttnRes)! This new technique replaces fixed accumulation in standard residual connections with smart, learned attention over preceding layer outputs, letting models dynamically aggregate information. Their 'Block AttnRes'
-
Stop Feeding Raw JSON to LLMs: Efficiency Matters
By
–
𝗜𝗳 𝘆𝗼𝘂’𝗿𝗲 𝘀𝘁𝗶𝗹𝗹 𝗳𝗲𝗲𝗱𝗶𝗻𝗴 𝗿𝗮𝘄 𝗝𝗦𝗢𝗡 𝗶𝗻𝘁𝗼 𝗟𝗟𝗠𝘀, 𝘆𝗼𝘂 𝗺𝗶𝗴𝗵𝘁 𝗯𝗲 𝘄𝗮𝘀𝘁𝗶𝗻𝗴 𝘁𝗼𝗸𝗲𝗻𝘀, 𝗹𝗮𝘁𝗲𝗻𝗰𝘆, 𝗮𝗻𝗱 𝗺𝗼𝗻𝗲𝘆. This is something I see surprisingly often in AI systems. JSON works well for software systems. But for LLMs, it
-
CodexBar 0.18 Release: New Providers and Performance Improvements
By
–
CodexBar 0.18 is out:
– New providers: Kilo, Ollama, OpenRouter
– Codex historical pace + risk forecasting + backfill
– merged-menu Overview tab
– fewer Claude keychain prompt annoyances – lower CPU/energy use, faster JSONL scanning
thx @RatulSarna -
AI Automation Limits: Testing Claude Coworking Chains and Judgment
By
–
my version of 2 is “the stuff that AI just oneshotted, what was the immediate action or thought just before or after you handed over to ai” – for me specifically claude cowork. i end up chaining more and more things and testing the limits of the automation and judgment and
-
C Compiler to LLM Weights and Logarithmic Attention Innovation
By
–
Wait this is so awesome!! Both 1) the C compiler to LLM weights and 2) the logarithmic complexity hard-max attention and its potential generalizations. Inspiring!
-
Upcoming AI Model Updates: Gemma 4 and DeepSeek V4
By
–
I'll try to update it with the major ones. E.g., Gemma 4 next week (likely), and DeepSeek V4 any day now
-
LLM Architecture Gallery with YAML Metadata Available
By
–
I can offer you the metadata in YAML format: https://
github.com/rasbt/llm-arch
itecture-gallery
…
Happy tinkering!