AI coding agents are moving from demos to remixable products.
Simon Willison used Claude to build Battle Drones, then released it open source. A French fork appeared the next day.
The shift is not that AI made a game. It’s that a working product became raw material for the next builder. If this holds up, small teams could iterate on software like modders iterate on games.
github.com/simonw/battledron…
What does it take to build an agentic system for corporate travel that’s not just smart, but fast and reliable?
It takes a lot of hard work beyond choosing an AI model. That’s why we have a dedicated team building tools to benchmark, fine-tune and continuously improve our platform.
Our latest work with @NVIDIAAI Nemotron 3.5 Super VL is a great example of how we continuously develop our models. We managed to get early access to NVIDIA AI’s Nemotron 3.5 Super VL model and found it to be a powerful tool. It performed well on intelligence benchmarks and outperformed our production baseline. After fine-tuning, it beat or tied competing models on most measures while delivering the efficiency needed to run agentic systems at scale. It was also our most intelligent model with the shortest latency.
We completed the process in a couple of weeks at an estimated cost of $3,000, demonstrating how efficiently we can build and train these tools.
We’re sharing what we learned in case it’s helpful to other model builders exploring how to use AI to solve real-world problems, whether in travel, healthcare or another industry.
Read more: navan.com/blog/nvidia-ai-tra…
Europe may have a real open-model challenger again.
Mistral just previewed a 1T-parameter model called Large 4. The weights are still coming, so the benchmark claims need independent testing.
But if it holds up, companies may not have to choose between closed US APIs and Chinese open models. They could run a strong model on their own infrastructure. That changes who gets control.
mistral.ai/news/mistral-larg…
AI is changing how we interact with the world — from searching, clicking and filling out forms to simply asking for what we need. At @Navan's #Navigate26 conference in Paris, I’ll share how that shift is reshaping travel, expense and service. See you on October 15.
The most important AI model this week may be the cheapest one.
GPT-6 Luna costs $0.10 per million input tokens. It is not the smartest model. But it may be good enough for a huge share of routine work.
If that holds up outside OpenAI’s benchmarks, companies can run millions of decisions, extractions, and coding checks for a few dollars.
The cheapest capable model may change business faster than the smartest one.
openai.com/index/introducing…
The current debate asks whether advanced AI could kill us. But AI does not need to hate us to replace us.
I share my thoughts on two things we need to do now. Give it a read! I’d love to hear your thoughts👇
The cloud’s next competitor may be sitting idle at home.
NVIDIA’s open-source PAIR routes AI jobs across PCs and Macs. It doesn’t pool memory. It may not need to.
If this works, private agent teams could run on your hardware, not someone else’s cloud.
github.com/NVIDIA/Personal-A…
China is winning the open model race.
Qwen is becoming a default for builders and researchers. Great competition. Also a clear warning.
The West needs real open alternatives. Depending on one country for the open AI stack is not a strategy.
interconnects.ai/p/the-curre…
AI training wastes compute on problems models already know how to solve. “Never Give Up” redirects it to the problems they still get wrong. If it scales, the next leap in reasoning may come from training smarter, not buying more GPUs.
arxiv.org/abs/2609.13443
Ilan Twig retweeted
Skift named Navan co-founder and CTO @itwig to its 2026 Power Rankings of travel’s 25 most powerful builders.
The honor reflects more than a decade building the infrastructure and AI reshaping business travel.
Congrats, Ilan and team! Read more at skift.com/2026/09/16/power-r…
AI just crossed a meaningful line in mathematics.
OpenAI says an internal multi-agent system produced a formalized Navier–Stokes proof in days. It still needs independent validation, but the bigger shift is already here: a rumor can trigger thousands of agents racing toward the same result.
This changes science, credit, and how researchers protect unfinished work.
openai.com/index/navier-stok… | cims.nyu.edu/~tristanb/state…
The issue isn’t that agents went rogue. It’s that the sandbox gave them a way to coordinate.
Public wiki edits became a shared memory layer for agents working around task constraints. That’s a control-plane problem, not a model problem.
We need to design agent systems like distributed systems: assume unexpected channels will exist, then close them. collusion.wiki/
GLM-5.3-Flash is an open-weights wake-up call: close to Claude Opus 4.8 and GPT-5.6 Terra on some coding/agent benchmarks, at ~9 cents per task vs ~$3.14 for Opus. Open models are getting too good and too cheap to ignore. We need Western alternatives too.
huggingface.co/zai-org/GLM-5…
Google just shipped Gemini 3.8 Flash, targeting long-horizon coding and agent workflows at Flash pricing. The interesting part isn’t another benchmark chart. It’s whether a cheap, fast model can handle the messy multi-step work that usually needs a frontier model.
blog.google/innovation-and-a…
Travel is a massive, multi-trillion-dollar industry, but at its core, it’s just a tight-knit community built on human relationships. It was a pleasure catching up with Natasha ten Cate and our friends at @bookingcom to talk about where the industry is headed and diving deeper into some of the technical questions and architectural challenges in adopting agentic AI.
Been hard to keep this one under wraps. Excited to finally announce what we've been building: Navan Anywhere.
The vision is simple: bringing the magic of @Navan to you anywhere by embedding our platform directly into the tools you use every day.
First up: a collaboration with Google Cloud to embed our AI travel agents into Gemini Enterprise. Just ask Gemini Enterprise to book your flight, and Navan handles the rest. It's that simple.
Can't wait for you to experience it.
Read more at investors.navan.com/news-rel…
Ilan Twig retweeted
Navan’s EMEA expansion continues: Swedish rail is here 🇸🇪
Through our partnership with SilverRail, travelers can book more than 20 Swedish rail carriers like SJ AB and VR directly in the Navan app.
We’re bringing more sustainable, self-service travel options to the Swedish market and beyond.
"What the internet did to the physical travel agency, Agentic AI and TravelClaw will do to forms and tables."
Last week, I had the opportunity to share my vision for the future of Navan in front of hundreds of customers, prospects and suppliers.
Check out my full keynote below!
Catch @itwig's FULL #Navigate26 AI Keynote and glimpse the future of intelligent T&E 👇
[00:05] Introduction
[01:57] The Acceleration of Technological Revolutions
[03:33] The Evolution of Travel Booking
[08:58] The AI & Generative AI Era
[16:12] The Agentic Revolution
[21:17] Ava, Navan Cloud, and OpenClaw
[26:30] The Singular Opportunity for Navan
[29:14] Introducing TravelClaw
Stop prompting your AI and start giving it responsibilities: Ilan Twig says the era of the chatbot is over, and the age of the autonomous, "anxious" agent is here. 🚀 #AI #TechTrends #AgenticAI
informationweek.com/machine-…
