Shareable cards — Monday, July 20, 2026

Monday, July 20, 2026 — top 10 database + AI stories, one card each. Save the image, copy its caption, post.

Phone: tap Share image, pick Save Image in the sheet (LinkedIn's app won't take images from Safari directly), then tap the button again — it opens LinkedIn: start a post, add the photo on top, paste the caption. Desktop: Open → save the image, attach it in the composer, paste the caption.

Hand-drawn sketchnote — KTransformers enables CPU-GPU heterogeneous inference and fine-tuning for LLMs

Open image

KTransformers enables CPU-GPU heterogeneous inference and fine-tuning for LLMs

The day's edition as one sketchnote — full stories + sources:
https://offnetchat.com/news/2026-07-20.html?s=0e1a94ec72f9fe42#item-0e1a94ec72f9fe42

#TechNews #AI

Claude Fable generates counterexample to Jacobian Conjecture

Open image

Claude Fable generates counterexample to Jacobian Conjecture

▸ AI models are now capable of generating valid counterexamples to major mathematical conjectures.
▸ The Jacobian Conjecture remains a key testbed for automated theorem proving and discovery.

Summary + source: https://offnetchat.com/news/2026-07-20/s/f3a7c00814937f7b.html

#LLM #TechNews

Qwen 3.8 released with new token pricing plans on QwenCloud

Open image

Qwen 3.8 released with new token pricing plans on QwenCloud

▸ Qwen 3.8 is now available with dedicated pricing plans on QwenCloud.
▸ High community engagement suggests strong interest in this release.

Summary + source: https://offnetchat.com/news/2026-07-20/s/ddebaf62488a3199.html

#LLM #TechNews

KTransformers enables CPU-GPU heterogeneous inference and fine-tuning for LLMs

Open image

KTransformers enables CPU-GPU heterogeneous inference and fine-tuning for LLMs

▸ Leverages CPU-GPU heterogeneous computing for efficient LLM workloads.
▸ Provides unified interfaces for both inference and supervised fine-tuning.

Summary + source: https://offnetchat.com/news/2026-07-20/s/0e1a94ec72f9fe42.html

#LLM #TechNews

WrenAI merges Wren Engine, expands GenBI context layer for AI agents

Open image

WrenAI merges Wren Engine, expands GenBI context layer for AI agents

▸ Unified repo merges Wren Engine into WrenAI core, simplifying deployment and maintenance.
▸ Supports 20+ data sources including BigQuery, Snowflake, PostgreSQL, and Redshift.

Summary + source: https://offnetchat.com/news/2026-07-20/s/7cc42753ef183003.html

#Databases #TechNews

PostgreSQL extra_float_digits: Precision without the tradeoff since v12

Open image

PostgreSQL extra_float_digits: Precision without the tradeoff since v12

▸ PostgreSQL 12+ eliminates the historical choice between readable and exact float output
▸ extra_float_digits no longer requires manual tuning for precision in modern versions

Summary + source: https://offnetchat.com/news/2026-07-20/s/72fccbb8ad4b829c.html

#Databases #TechNews

VarRate: Training-Free Variable-Rate KV Cache Compression for Long-Context LLMs

Open image

VarRate: Training-Free Variable-Rate KV Cache Compression for Long-Context LLMs

▸ Eliminates irreversible accuracy collapse caused by token eviction in existing selection methods.
▸ Optimizes memory usage by assigning rank budgets proportional to token salience.

Summary + source: https://offnetchat.com/news/2026-07-20/s/feeef5160ad60f03.html

#LLM #TechNews

Process Reward Informed Tree Rollout for Multi-Turn RL

Open image

Process Reward Informed Tree Rollout for Multi-Turn RL

▸ Replaces uniform trajectory sampling with tree-based exploration for better budget allocation.
▸ Uses process rewards to identify and branch from promising intermediate agent states.

Summary + source: https://offnetchat.com/news/2026-07-20/s/8e282e665feaf39f.html

#AgenticAI #TechNews

Loopie: Looped Transformers Outperform Vanilla Models at Same Compute

Open image

Loopie: Looped Transformers Outperform Vanilla Models at Same Compute

▸ Looped Transformers can beat vanilla scaling when compute budgets are equal.
▸ 20B and 6B MoE variants keep active parameters low (2B and 0.6B).

Summary + source: https://offnetchat.com/news/2026-07-20/s/3d91ea7c27c82675.html

#LLM #TechNews

Claude Code switches to Bun runtime written in Rust

Open image

Claude Code switches to Bun runtime written in Rust

▸ Claude Code now relies on Bun for execution, moving away from standard Node.js.
▸ Rust-based Bun offers potential gains in startup latency and runtime efficiency.

Summary + source: https://offnetchat.com/news/2026-07-20/s/dbd57b2637da08ce.html

#LLM #TechNews

DSWorld: World Model for Autonomous Data Science Agents

Open image

DSWorld: World Model for Autonomous Data Science Agents

▸ Predicts operation effects before execution to reduce costly trial-and-error.
▸ Uses a Data Science World Model to simulate environment state transitions.

Summary + source: https://offnetchat.com/news/2026-07-20/s/672d6ab1c90b4547.html

#AgenticAI #TechNews