Tencent releases open-source Hy3 MoE model with FP8 quantization

#1 📝 Simon Willison

tencent/Hy3 - Announcement and notes about Tencent's new Hy3 Mixture-of-Experts model (Apache 2.0). The post describes model size, availability (including a quantized FP8 variant), huge context length, a free OpenRouter trial, and includes an example SVG the author generated.

#2 𝕏

Google DeepMind is partnering with Apptronik at their expanded Robot Park to collect real-world data from the Apollo 2 humanoid. This data will train and advance Gemini Robotics.

#3 𝕏

Anthropic partnered with Neuronpedia to launch jlens, an interactive demo showcasing their interpretability methods on open-weight models at https://www.neuronpedia.org/jlens.

#4 𝕏

NVIDIA AI presents an ICML paper that separates unintended memorization from generalization, estimating GPT-style models can store about 3.6 bits per parameter and offering a sharper framework for data scaling and privacy.

#5 📝 Anthropic News

Government of Alberta uses Claude to find and fix cybersecurity vulnerabilities across government systems - Since 2025 Alberta’s Ministry of Technology and Innovation has used Claude Code (Opus and Sonnet) with roughly 50 autonomous agents to scan 466 million lines across 1,280 applications and 3,400 repositories in about 20 hours—finding issues traditional tools missed and estimating the same review would have taken ~6.5 years manually. Claude generated fixes, wrote tests, rebuilt legacy systems (sometimes in 4–5 days versus original multi-month builds), runs continuous agent reviews against ~95 security controls, has trained thousands of government employees and over 10,000 members of the public via the Alberta AI Academy, and plans to consolidate 185 legacy apps into 16 modern reusable applications.

#6 ▶️

How I run autonomous coding agents from my phone with OpenAI Symphony + Linear

How I AI Podcast

Also covered by: @Claire Vo

#7 𝕏

LlamaIndex 🦙 teamed up with LanceDB to launch a hybrid pipeline that combines LiteParse with native multimodal storage to break messy enterprise PDFs into pages, chunks, and assets.

#8 in

Peter Yang shows how Rohan (Codex PM) uses a screenshot of the Codex composer with Image Gen to generate 4–5 project-selection mockups in minutes. He then turns the chosen design into a clickable Codex site prototype—no coding required.

#9 𝕏

Philipp Schmid recommends Gemini 3.5 Flash for OCR and VQA tasks, highlighting its faster, cheaper, and more accurate performance.

#10 📝 Anthropic Engineering

How we contain Claude across products - Anthropic engineers describe strategies for limiting the potential blast radius of increasingly capable agents, and share lessons from building containment for claude.ai, Claude Code, and Cowork. The post focuses on design and engineering approaches to cap agent capabilities safely across products.

#11 𝕏

claire vo 🖤 – building @chatprd, host, 3x CPTO shows how @FanaHOVA spins up a ZoComputer VPS to run OpenAI-powered agents overnight using Symphony x Linear for automated tasks and per-task token cost tracking.

Also covered by: @Claire Vo

#12 📝 Claude Code Blog

A field guide to Claude Fable 5: Finding your unknowns - Introduces Claude Fable 5 and provides guidance for identifying and addressing unknowns when building with the model. The article targets developers and teams using Claude Code, offering practical tips and use cases for working with Fable 5.

Also covered by: @Santiago, @Thariq, @Udi Menkes, @Mario Zechner

#13 𝕏

Peter Yang warns that Fable 5 will leave Claude subscriptions at midnight tomorrow and shares five ready-to-use prompts—like “Find Fable-worthy work”—to test key use cases. He links to a tutorial for live demos of all five.

Also covered by: @Santiago, @Thariq, @Udi Menkes, @Mario Zechner

#14 📝 PromptLayer Blog

Why fine-tuning is probably not for you - The author argues fine-tuning is often not worth the effort because retrieval-augmented generation (RAG) frequently outperforms fine-tuned models (the article even cites studies and a figure showing RAG significantly better), while fine-tuning adds complexity, slower iteration, ongoing cost and privacy risks, and typically requires large datasets (often more than 10k examples). That said, fine-tuning can still be useful to enforce specific output formats or writing tone, reduce token usage by baking prompts in, aid multi-step reasoning according to recent research, and “up-cycle” cheaper models (e.g., fine-tuning 3.5-turbo to approximate GPT-4 results or Stanford’s Alpaca replicating LLaMA cheaply).

#15 𝕏

Claude presents a behind-the-scenes history of Claude Code through builders’ and early users’ firsthand stories, highlighting how user feedback drove the evolution of its core features.

#16 in

Greg Isenberg says the biggest startup opportunity of the next decade is building infrastructure for AI agents — from default “harness” integrations and Ramp-style spend controls to trusted shared memory, sandboxed accounts, and doc-based onboarding.

#17 in

Dharmesh Shah predicts we’ll see a “Super App” by year-end that bundles ChatGPT, Codex, image generation and long-running agents into a single AI platform. He names OpenAI as the frontrunner, with Anthropic, Perplexity, Google and Apple as key challengers.

#18 𝕏

clem 🤗 – Co-founder & CEO @HuggingFace shows that since fully replacing git storage with HF Xet around Nov 25, AI builders are already storing massive data volumes on the platform. He expects this growth to scale into the exabyte range soon.

#19 𝕏

Santiago proposes giving AI agents their own email addresses you CC on messages, so they can autonomously handle tasks while keeping your inbox separate.

#20 ▶️

Anthropic Will Pay You $85,000 to Learn AI (Deadline July 17)

Helena Liu