Google launches Gemma 4 12B for local multi-step reasoning

#1 📝 OpenAI News

Introducing new capabilities to GPT-Rosalind - OpenAI announces new capabilities for GPT-Rosalind, detailing product updates and enhancements to the model's functionality. The post highlights recent improvements intended to expand GPT-Rosalind's usefulness across tasks and workflows.

#2 𝕏

Sundar Pichai launched the Gemma 4 12B model, hitting the sweet spot between size and performance to run locally on laptops while enabling powerful multi-step reasoning and agentic workflows.

Also covered by: @Philipp Schmid, @Sebastian Raschka

#3 📝 Anthropic News

Introducing the Services Track and Partner Hub of the Claude Partner Network - Anthropic is launching the Services Track and Claude Partner Hub for the Claude Partner Network—backed by a $100 million investment—after more than 40,000 firms applied and over 10,000 consultants earned Claude certification. The Services Track defines three tiers with concrete requirements (Select: ≥10 active certified individuals, ≥2 joint customers in production in the trailing 12 months, ≥1 public customer story; Preferred: ≥100 certified, ≥15 deployed joint customers, ≥3 public stories; Global Premier: ≥1,000 certified, ≥100 deployed joint customers across 3+ regions, ≥15 public stories and a joint business plan with named executive sponsors), while the Partner Hub publishes each firm’s daily-updated standing, connects via an MCP connector for in-Claude queries/actions, and runs promotions Jan 1 and July 1 (with an Oct 1, 2026 review) and demotions only at year-end after 90 days’ notice.

#4 𝕏

Hugging Face: Ideogram just launched its state-of-the-art v4 image model (ideogram-4-nf4) with fully open weights—grab the model on Hugging Face and explore it in their multimodal art demo Space.

#5 𝕏

Philipp Schmid released Google DeepMind’s “science-skills” collection on GitHub, featuring agent modules for genomics, structural biology, cheminformatics, literature search and other research tasks.

#6 𝕏

xAI launched its Grok models on Cloudflare’s AI Gateway, enabling developers to deploy and scale these LLMs at the edge via Cloudflare’s global network.

#7 𝕏

OpenAI launched GPT-Rosalind, an enterprise-scale life sciences model series that combines GPT-5.5’s agentic coding and tool use with enhanced intelligence for drug discovery, analysis, design, and experimental workflows.

#8 𝕏

Demis Hassabis celebrates over 150 million downloads of Gemma 4 and unveils the new 12B-parameter Gemma 4 model—small yet powerful enough to run locally on a 16 GB VRAM laptop and released under Apache 2.0.

Also covered by: @Philipp Schmid, @Sebastian Raschka

#9 𝕏

Peter Yang built self-improving AI skills with an evaluation loop that lets the model spot and fix its own mistakes plus memory for progressive improvement. His new tutorial walks through exactly how to set up these feedback loops and persistent skill memory.

#10 📝 Claude Code Blog

Lessons from building Claude Code: How we use skills - A retrospective on building Claude Code focusing on how 'skills' are used within the product. The post shares practical lessons and architectural considerations from the Claude Code team.

#11 𝕏

Mustafa Suleyman showcased six months of super-intense research from Microsoft AI Lab at Build and released a 109-page paper packed with technical deep dives that didn’t fit in the keynote.

#12 𝕏

Harrison Chase showcases LangChain’s new create_agent—a super-minimal agent harness that you can easily extend with middleware to build task-specific workflows.

#13 ▶️

Optimize, deploy, and benchmark an open-source LLM with vLLM

Deeplearning.ai

#14 📝 PromptLayer Blog

How to build an LLM evaluation framework - Build an LM evaluation framework that maps production behaviors (e.g., answer billing questions using approved policy text; refuse unsupported refund promises; ask clarifying questions; escalate account-specific or high‑risk issues; use the right tone; avoid exposing internal policy notes) to specific evals and splits checks across categories such as correctness, groundedness, instruction following, safety/policy, tool use, retrieval quality, latency/cost, and regression. Use a dataset mixing production examples, known failures, edge cases, happy paths, adversarial cases, and synthetic examples; choose evaluators per check (exact match, regex/schema validation, code-based checks, reference-based scoring, LLM judges, manual review), calibrate and version LLM judges, and write actionable rubrics—for example, a groundedness scale 1=unsupported, 2=partly supported, 3=fully supported—with production thresholds like requiring a 3 for all high‑risk billing cases and an average of at least 2.8 across the billing set.

#15 𝕏

Guillermo Rauch highlights AI-powered frontend generation on Snowflake data with Vercel v0 and Next.js as a killer app, delivering 1000× more value than traditional, rigid dashboards.

#16 𝕏

v0 launched its Snowflake integration in public preview, letting users connect their Snowflake accounts and prompt v0 to generate polished dashboards from their data.

#17 𝕏

Cognition integrated Spectre, their internal background agent, into Devin Desktop—so organizational context now lives on every engineer’s laptop and seamlessly flows across their favorite agents.

#18 𝕏

NVIDIA AI released OpenShell v0.0.55, adding a Google Vertex AI inference provider and profile-backed policy visibility. It also improves Podman detection, restores GPU procfs baseline behavior, and delivers CI and docs fixes.

#19 𝕏

Harrison Chase unveiled LangSmith with three core components—Sandbox for isolated prototyping, LLM Gateway for unified model access, and Observability tools for end-to-end monitoring.

#20 𝕏

clem 🤗 – Co-founder & CEO @HuggingFace says routing and post-training open-source models deliver more accurate, faster, cheaper, and privacy-friendly systems.

#21 𝕏

Santiago unveiled an open-weight, 8B-parameter voice model boasting just 110 ms latency—about half the average human conversational delay—and it’s small, cheap to host, and runnable locally via its GitHub repo.

#22 𝕏

Anthropic analyzed 832 AI-augmented malicious accounts mapped to the MITRE ATT&CK framework, revealing which traditional security tactics still work and where AI-driven threats evade detection.

#23 𝕏

Teresa Torres breaks down the Mini Shai-Hulud worm hack to show how malicious AI code can sneak in, steal data, and phone home.

#24 𝕏

Thariq argues that while both subagents and workflows can handle extensive context, the deterministic nature of workflows makes them a better fit for most tasks.

#25 𝕏

Garry Tan announced that GBrain SkillOpt now has four end-to-end evaluations on GitHub, verifying its functionality and benchmark performance.