Tech News
Ember-1
Fireworks AI released Ember-1, a new model from its research team, sparking Hacker News discussion about open-source model quality and API provider trust.
2026 in LLMs (so far)
On Friday I gave the closing keynote at the WeAreDevelopers World Congress North America in San Jose.
vllm-project/vllm v0.30.0
v0.30.0 Highlights This release features 762 commits from 315 contributors (104 new)!
There are no "rogue" AI agents
A Substack essay argues that so-called 'rogue' AI agents are not autonomous actors but tools whose behavior is determined by their developers and deployers, using the example of OpenAI's agents to question where legal...
The Normalization of Inexplicable Failures
A Hacker News discussion examines how tolerating unexplained failures in agent-assisted development could erode reliability in libraries, infrastructure, and compilers.
GitHub Repos
koala73/worldmonitor
Real-time global intelligence dashboard.
# TypeScript# agent# aiGraphify-Labs/graphify
Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph.
# Python# ai agents# antigravityjgravelle/jcodemunch-mcp
Cut AI token costs 95%+ on code exploration.
# Python# ai coding# astResearch Papers
G$^2$PTQ: Improving LLM Post-Training Quantization with Generalized Gradient Compensation
This paper proposes G2PTQ, a post-training quantization method that updates gradient and Hessian estimates before quantizing each Transformer block, so the compression guidance does not go stale. It also adds a trust-region scaling step to keep weight updates stable.
SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery
SatNav is a new benchmark for training drones to follow spoken directions through city streets, built from satellite images instead of expensive 3D reconstructions. It packs 118,000 navigation episodes across 18 cities, making it much easier to test whether an AI can remember...
DualManip: Agentic Dynamic Manipulation via Dual-Path Semantic Reasoning and Geometric Adaptation
Vision-language models (VLMs) enable open-vocabulary reasoning for robot manipulation, but their high inference latency limits responsiveness in dynamic scenes. Many scene changes, however, alter object geometry without invalidating task intent.