Thesis-driven early-stage venture capital firm.
Berlin
Joined May 2011
- Tweets6.2K
- Following1.3K
- Followers24K
- Likes3.2K
Point Nine retweeted
We asked a simple question about docs: Do AI coding agents actually read them?
Here’s what we found out.
Point Nine retweeted
One of our favorite moments from the FLEX research: the grasp slips, and the robot tries again.
That recovery isn’t in any of the demonstrations. Nobody taught Cortex, the AI model behind our robots, to do it. It learned through practice.
@sigmundhh is presenting the FLEX paper at the R2RL workshop at @corl_conf this Thursday, Nov 12. Poster session 10:30–11:00. If you’re in Austin, come say hi and hear how robots can learn from their own practice.
#CoRL2026
Point Nine retweeted
After 16 years, Jobber isn't just a tool. It's the platform hundreds of thousands of customers trust to run their businesses. Websites, AI marketing campaigns, reviews, voice receptionists, booking flows, scheduling, mobile apps, invoicing, payments, bookkeeping, job costing, inventory, and a lot more.
With millions of lines of code, we rely on a huge test suite to move fast. Tests catch unintended side effects and document our most important behaviors so we don't accidentally drop them. We're at 150,000+ tests now.
The problem with tests is that they aren't always reliable. Race conditions and weird timing issues can cause a test to fail once every several hundred runs. Not a big deal on its own, but infrequent failures across 150,000 tests mean frequent failures of the end-to-end suite. Over the years this grew, and as AI helped us ship 2x as many changes, it got bad enough that over 20% of our test suite runs were failing.
Luckily, Tim Vanderkooi, Barbara Akaeze, Sajeeban Sivayogalingam, and our Dev Accel team saw this was the perfect use case for AI.
They set up a pipeline that retries failed tests to confirm and fingerprint flakes, feeds structured failure logs to AI agents, and runs nightly automation that opens draft PRs to fix them. When a fix works, the bot proposes a lesson back into a taxonomy doc so our AI harness doesn't make the same mistake again.
They've already dropped our failure rate from over 20% to sub-10% (and weekly flakes from 7,000 to 400), and it keeps improving.
The piping and learnings here are also a huge piece of our push toward automating our Software Factory, where agents proactively take the lead on preparing PRs.
Full write-up from Tim on how they built it:
linkedin.com/posts/timvander…
Big ideas, peer-driven feedback and a shower of fireworks to finish. It was great to help bring together a room full of aspiring and early-stage researchers/founders in Cambridge last night.
TesterArmy is now @e2e out in the open with $8M to build. Welcome to the P9 family, @SzymonRybczak and @o_kwasniewski 👋
They’re rebuilding end-to-end testing for the agentic coding era. Their popular open source framework lets agents prove apps actually work, on any device, under any conditions.
Coding agents now write a big share of the code that ships, and fewer people (or none) are reading it line by line. Szymon and Oskar see verification becoming the critical step in how software gets built.
Looking forward to being part of your journey, and to seeing e2e cloud take off next year 🚀
Point Nine retweeted
I've recently vibe-coded this little app. You'll figure out what it does by checking out the screenshot. :) I find it super useful!
Point Nine retweeted
I think it's time to retire the term "vibe coding" and to stop saying "I had Claude/Codex build XYZ" (I still keep saying it)
At this point, when you say "I built XYZ" it's understood that you didn't write the code yourself.
It's been a great term for a short era, @karpathy :)
Point Nine retweeted
Before practice: 60–70%. After practice: 85–100%.
That’s what FLEX, our new learning method, did for Cortex 2.6. The robot improved from a few dozen attempts of its own, with nobody stepping in.
Why it matters: today’s robot models, like VLAs and world-action models, learn almost entirely by imitation. That produces impressive behavior, but not reliable behavior. One small mistake puts the robot somewhere no demonstration covered, and errors compound.
FLEX allows the robot to learn from its own experience, building on what the pretrained model already knows. This sets a clear course toward robots that learn efficiently while producing value in the real world.
Full results in link below.
Point Nine retweeted
Who is combining this with direct weight editing?
Introducing RoboArchitect: it overcomes the failures of VLA policies by turning them into robotic agents with their own control strategies and persistent memory.
RoboArchitect automatically designs and continually improves each agent around its frozen policy. Without any weight updates, average success rises substantially on unseen scenes and tasks. 🧵
Point Nine retweeted
As of Nov 2, over $1B in ETH and ~$1.42 billion in BTC stolen from Binance have seen limited onward movement since the theft.
Other stolen funds are moving rapidly across chains.
These funds are not all progressing through the laundering process at the same pace.
Point Nine retweeted
When @claudeai requires legal intelligence, it can now call on @goclio. With our new connector, Clio draws on matter context, 1B+ legal documents, and court records, then returns the work with citations attached.
clio.com/about/press/connect…
Point Nine retweeted
New post
Like many of you, I've sometimes been frustrated by the quality of the discussion about AI in the media and in the political debate ahead of the presidential election in France.
So what started as a conversation in a restaurant at the end of September (napkin in the post) became a 3-page GDoc called "A note on AI to a candidate at the French presidential election”. It's non-partisan, and it's been circulating privately in French for a few weeks. I thought I'd publish it more broadly in French and English below (both GDocs in comments).
No AI slop, just my thoughts and a few graphs and sources dug up by Claude, in 3 short pages.
Probably nothing very new if you follow closely, but hopefully still interesting.
Hope you enjoy!
Point Nine retweeted
Our Co-founder and CEO @jony_levin spoke with @BBCNews today on the Binance exploit.
Point Nine retweeted
Double the screens, double the incident visibility. That's what we call a dynamic duo.
Now you can view incident.io on the iPhone Duo with split-screen support. Great for things like schedules and alerts. Or doom scrolling. Either one.
Point Nine retweeted
Anyone can say their AI is accurate, but we'd rather measure it. 🤌
Every day, we backtest Investigations against real past incidents. Each one is frozen at the moment it happened, so we see how today's system would have performed with only the information responders had at the time.
If accuracy drops, we're alerted before it ever reaches a live incident.
Every change we make goes through the same backtests, model upgrades included. We run them with and without the change, and only ship it if results improve.
Point Nine retweeted
More robots. More returns. More scale. This is physical AI.
Excited to be building real-world solutions alongside @OttoGroup_Com and @nvidia.
Full announcement in comments.
Point Nine retweeted
One billion orders in Q3! 🚀
GMV: €13.9B (+14%, up from 11% in Q2)
Revenue: €4.3B (+22%, up from 18% in Q2)
=> Raised full-year guidance for GMV, Revenue, aEBITDA, and Cash Flow, again.
Point Nine retweeted
The first 10 Sereact dual-arm robots will be deployed at @OttoGroup_Com's logistics center in Poland to grasp, identify, assess and sort returns, including items they’ve never seen before.
15 more are planned in phase two, bringing the program to 25.
This is what the future of physical AI looks like.
Full announcement in the comments.
