Newsletter Archive

Ten open-source agentic AI projects you can immediately clone and run—a hands-on shortcut to learning how production agent systems actually work, from routing and memory to tool calling patterns.

KDnuggets · 8 min read Repos

Latent Space and Unsupervised Learning reunite to dissect what's shifted in AI over the past year: infrastructure stabilization, the rise of 'skills' as agent packaging, the vertical vs. horizontal startup debate, and emerging agent lab playbooks. Discussion touches OpenClaw, context engineering, evals, and why conference tracks now signal what practitioners actually care about.

Latent Space · 62 min read Community

GPT-5.5 Pro delivers measurable improvements in reasoning and code generation, demonstrated via complex procedural simulation tasks where it outperforms o3 and open-weight competitors by actually modeling evolving systems rather than static replacements — the frontier remains uneven but advancing rapidly.

One Useful Thing · 7 min read Industry

Meta is laying off 8,000 workers (10% of staff) to offset surging AI infrastructure costs, which are expected to spike capex by 60% in 2026 while free cash flow plunges 83%. The move reflects industry-wide pressure from investors balking at AI spending eating into profits, with Amazon, Salesforce, Snap, and Microsoft announcing similar cuts.

Axios · 2 min read Industry

Intel projects strong sales growth as it begins capturing market share in the AI infrastructure buildout, marking a potential turning point for the chipmaker after years of competitive losses to Nvidia and AMD.

Bloomberg Tech · 1 min read Hardware

Project Prometheus, Jeff Bezos' physical AI lab co-founded with Google veteran Vik Bajaj, closed a $10B funding round at a $38B valuation. The company is positioning itself in the competitive robotics and embodied AI space where scale matters.

Bloomberg Tech · 1 min read Market

OpenAI released GPT-5.5 (codenamed Spud), its most capable model yet, offering faster inference, fewer tokens per task, and autonomous multi-step reasoning. The model excels at coding, computer use, and scientific research; early users reported 10-hour weekly time savings on complex tasks. API access launches soon after security hardening.

Axios · 2 min read Industry

First systematic study comparing how FHIR data serialization formats (JSON, markdown, narrative, timeline) affect LLM medication reconciliation accuracy across five models. Clinical Narrative formatting boosts smaller models by up to 19 F1 points, but Raw JSON wins at 70B; all models systematically fail on polypharmacy patients and omit medications more than hallucinate them.

arXiv NLP · 4 min (abstract) Research