https://lamouche.fr
I am a full-time enthusiast of technology and innovation
Stanford researchers have developed "Evo", a genomic language model trained on bacterial genomes. Unlike AlphaFold which predicts structure, Evo works at the DNA level, learning from gene clusters.
It has successfully generated functional novel proteins (CRISPR inhibitors, antitoxins) with <25% similarity to known sequences.
Generative AI is now effectively simulating evolution at the nucleic acid level.
https://arstechnica.com/science/2025/11/generative-ai-meets-the-genome/
Is it a waste to stop all roadmap work for a week just to fix bugs?
Lalit Maganti argues the opposite. His team of 40 engineers paused everything to fix 189 "papercut" bugs.
The result wasn't just a cleaner product. It was a team reconnecting with the joy of building. "Craftsmanship and care" often get lost in quarterly planning. Fixit weeks bring them back.
A great read on how to structure these weeks (and why they matter):
https://lalitm.com/fixits-are-good-for-the-soul/
Google's AI infrastructure chief says they need to double capacity every 6 months to meet demand. The goal is 1000x in 4-5 years.
The most interesting part? They aim to do this at "essentially the same power and energy level" using custom silicon like their new TPUs.
Given the concerns about AI's energy footprint, this efficiency target is the most critical metric to watch.
#AI #Google #Sustainability #TechNews
https://www.cnbc.com/2025/11/21/google-must-double-ai-serving-capacity-every-6-months-to-meet-demand.html
If you are building AI developer tools, you know that "context" is the hard part.
kit is an open-source Python toolkit designed to solve exactly that. It provides robust tools for:
Codebase mapping & symbol extraction
AST-based code search
Chunking files for LLM context windows
It even ships with a local MCP server (kit-dev) to expose your repo's intelligence to MCP-compliant clients.
ChatGPT group chats have launched globally. You can now add AI to your group conversations (up to 20 people) to help plan, write, or debate.
Important for privacy: OpenAI states that personal settings and memories stay private to each user, even within the shared group context.
It's available for Free, Plus, and Pro users.
Read more:
https://techcrunch.com/2025/11/20/chatgpt-launches-group-chats-globally/
How to write effective `agents.md` files for GitHub Copilot. The shift from a generic "chat" interface to defined "personas" is significant.
The key takeaway from analyzing 2,500+ repos? **Context is King.**
- Give the agent executable commands.
- Define strict boundaries (what NOT to touch).
- Provide "gold standard" code examples.
For the CLI lovers out there: check out `hl` by pamburus. 🖥️
It's a modern, high-performance log viewer written in #Rust. It takes JSON or logfmt streams and renders them in a human-readable format with automatic color coding.
Key features:
- ⚡️ Extremely fast (handles multi-GB files easily)
- 🔍 Native filtering by level, time, and fields
- 🐧 Cross-platform (Linux, macOS, Windows, NixOS)
Sustainability at Solutrans wasn't marketing—it was operational constraint. Fuel costs, emissions regulations, urban access restrictions. Companies like Chéreau retrofit aerodynamic fairings not for CSR reports but because operators demand lower consumption. The most credible sustainability solutions treat it as engineering problem, not brand positioning. Results matter more than narratives.
#Sustainability #GreenLogistics #OperationalEfficiency #Transport #ClimateAction
Google continues its AI push with Nano Banana Pro, now running on the Gemini 3 Pro model announced earlier this week.
Interesting shift from the viral consumer use case (3D selfies) to enterprise utility (infographics from code/resumes). The ability to maintain character consistency across multiple generations suggests better control in the underlying model.
Full details here:
https://www.cnbc.com/2025/11/20/google-nano-banana-pro-gemini-3.html
Gemini 3's benchmarks are impressive (95% on AIME 2025, 76% on real coding tasks), but let's talk about what this means for the ecosystem.
Yes, it's proprietary. Yes, it's Google. But these performance leaps raise important questions: How do open models keep pace? What happens when proprietary AI becomes essential infrastructure?
https://deepmind.google/models/gemini/
#ArtificialIntelligence #OpenSource #TechEthics #MachineLearning
Interesting new study on AI Search (GEO): ChatGPT citations are 2.3x more likely to include a table than Google Search results.
- Google Search: 13% of pages have tables
- ChatGPT Citations: 30% of pages have tables
It seems LLMs heavily favor structured content, even if the page itself looks basic. Good reminder to use semantic HTML and tables where appropriate.
Google released Antigravity, their "agent-first IDE." Interesting architectural choice: rebuild the development environment assuming autonomous agents do most work, rather than bolting AI onto traditional IDEs. Multi-agent mission control, synchronized workflows across editor/terminal/browser, task-based monitoring instead of diff-based.
Memori (3.3k ⭐) gives LLMs persistent memory using SQL databases you control - not proprietary vector stores.
One-line integration: memori.enable() works with OpenAI, Anthropic, and 100+ models. Store conversation history in SQLite, PostgreSQL, or MySQL. 80-90% cost savings vs vector databases.
Apache 2.0 license. Your data stays yours. Export anytime.
Check it out on GitHub
**Why your AI pilots are failing** 🧵
Harvard Business Review research shows only **4% of companies** deploy AI with a "deep and narrow" approach - focusing on transforming ONE strategic domain completely rather than running scattered pilots.
The result? **Double the ROI** compared to shallow, broad implementations.
The insight: Stop spreading thin. Go deep in one domain.
[Read the full HBR strategy](https://hbr.org/2025/11/stop-running-so-many-ai-pilots)
AI observation from Solutrans: The companies doing interesting work don't lead with "AI-powered." They lead with the problem solved—route optimization under real constraints, predictive maintenance based on actual failure patterns, load matching for partial shipments. AI is implementation detail, not value proposition. When your differentiator is the algorithm rather than the outcome, you're selling to the wrong audience.
https://lamouche.fr/notebook/posts/solutrans-2025/
#ArtificialIntelligence #Logistics
AI-powered GIS without compromising your data privacy 🔒
Monarcha is a new platform that brings natural language operations to geospatial analysis. Talk to your GIS, automate workflows, turn messy spatial data into decisions - all in minutes.
But here's what matters: your data stays yours. On-premises deployment available. Nothing stored remotely without explicit consent.
AI power + user control + privacy respect.
Learn more: https://monarcha.ai/
Why European digital sovereignty matters: a podcast on cartography and power 🗺️
Maps shape narratives. Who controls cartography controls data infrastructure and geopolitical framing.
Worth listening if you care about digital independence and open alternatives.
[Listen to the series](https://www.republiquedescartes.fr/podcast)
Thoughts on digital mapping alternatives?
#DigitalSovereignty #OpenSource #Geopolitics #Europe #Cartography
who draws your maps matters more than you think 🗺️
new podcast episode: IGN's director on why Google mapping the world creates sovereignty issues for Europe
🧠 Yann LeCun breaks down AI like no one else! From neural networks to the future - brilliant insights from an AI pioneer
🧠 Game-changer: AI that thinks 200+ steps ahead! Kimi K2 just dropped - open source thinking model crushing benchmarks 🔥
Link: https://moonshotai.github.io/Kimi-K2/thinking.html
#AI #OpenSource #ThinkingAI #TechBreakthrough