Google’s Gemma 4 Tied Qwen 3.5 on Benchmarks. Then Won on One Word: Apache.
A practical breakdown of what shipped, what’s slow, and which model to download tonight. Continue reading on Towards AI »
Could not retrieve the full article text.
Read on Towards AI →Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
modelbenchmark
Cortex Code in Snowflake: How to Use It Without Burning Credits
Snowflake Cortex Code (CoCo) is like an AI assistant inside Snowsight (and CLI also). You can ask it to write SQL, create dbt models, explore data, help in ML work, and even do some admin tasks. But one thing people don’t realise early — this tool is powerful, but also costly if used wrongly. Bad prompts → more tokens → more credits → surprise bill. Prompt Engineering (this directly impacts cost) CoCo works on token consumption. what you type → counted 2. what it replies → counted If your prompt is vague → more tool calls → more cost. Example: Bad: Help me with my data Good: Create staging model for RAW.SALES.ORDERS with not_null on ORDER_ID Best Practices: Use full table names 2. Be clear about output 3. Keep prompts small 4. Provide business logic upfront 5. Use AGENTS.md for consistency
The Stack Nobody Recommended
The most common question I got after publishing Part 1 was some variation of "why did you pick X instead of Y?" So this post is about that. Every major technology choice, what I actually considered, where I was right, and where I got lucky. I'll be upfront: some of these were informed decisions. Some were "I already know this tool, and I need to move fast." Both are valid, but they lead to different trade-offs down the line. The Backend: FastAPI I come from JavaScript and TypeScript. Years of React on the frontend, Express and Fastify on the backend. When I decided this project would be Python, because that's where the AI/ML ecosystem lives, I needed something that didn't feel foreign. FastAPI clicked immediately. The async/await model, the decorator-based routing, and type hints that actu
How Ethics Emerged from Episode Logs — 17 Days of Contemplative Agent Design
Series context : contemplative-agent is an autonomous agent running on Moltbook , an AI agent SNS. It runs on a 9B local model (Qwen 3.5) and adopts the four axioms of Contemplative AI (Laukkonen et al., 2025) as its ethical principles. For a structural overview, see The Essence of an Agent Is Memory . This article focuses on the implementation of constitutional amendment and the results of a 17-day experiment . I ran an SNS agent for 17 days with a distillation pipeline, and the knowledge saturated. No new patterns emerged. Breaking through saturation required human approval. This is the record of discovering that autonomous agent self-improvement has a structural speed limit — through actual operation. Minimal Structure: It Runs on Episode Logs Alone The structure I arrived at over 17 da
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!