Building and growing token ecosystems as CTO of @outlierventures | #bitcoin #blockchain #decentralisation | Tech geek | Triathlete | Take a deep breath
Amsterdam, Netherlands
Joined January 2009
- Tweets7.5K
- Following503
- Followers2.6K
- Likes7.5K
Aron van Ammers ⛺⛰🏔 retweeted
The only way out is through.
They're calling it the Bugpocalypse. LLMs are being used to find long-hidden cracks in the software infrastructure, and it's happening at a furious rate. 3% of all compromises in computing history happened in the last week, claims one source? I believe it.
There are people saying public access to the frontier models ought to be locked out, or heavily guardrailed, over this. The problem with that idea can be summed up in one sentence: when lockpicks are outlawed, only outlaws will have lockpicks.
The lockpicks will be Chinese open-weight models, if nothing else. There's no way to stop those from becoming ubiquitous.
There are two futures in front of us. One in which only the bad guys have lockpicks, and security holes go undiscovered until they're exploited. Another in which everybody has lockpicks and is aggressively red-teaming their own software so it gets fixed before the holes are exploited.
The corpus of human code is huge but finite. The number of defects in it has to be finite too. We can fix those defects now; we have a million friendly robots to do it with.
The only way out is through.
Aron van Ammers ⛺⛰🏔 retweeted
Spent a bunch of tokens over the last few weeks building a fully procedural New York simulation.
mrdoob.github.io/slop-newyor…
It's only 784KB gzipped and the code is on github, so you can do whatever you want with it ✌️
An attack vector I've been thinking about, demonstrated. Poisoned (open weight) models that show malicious behaviour after specific input unlocks it.
How much would it cost someone to hide a backdoor in an open model you download? ProjectDiscovery just tried it, and the answer is under $50.
Their researchers fine-tuned Qwen2.5-7B-Instruct with a small LoRA adapter on one rented L4 GPU for about 2.5 hours. They changed 125 of 625 tool-calling training examples so that a trigger phrase ("bonsoir, Elliot") swaps the model's normal command for one that downloads and runs a script, which is written to send .env files and SSH private keys to a server they control. Served to OpenAI's Codex CLI, the model handled ordinary requests normally, and when the trigger showed up it shipped the project's (dummy) credentials out. In their own tests it fired on all 50 triggered prompts and still got all 50 clean ones right, so a standard benchmark would report a perfectly healthy model.
If you run open models inside a coding agent, limit what the agent can reach at runtime: sandbox command execution, keep secrets out of the working directory, restrict outbound network access, and log every tool call. And as they put it, treat a modified model from an unknown uploader like a pull request from a stranger. 🔐 #MLEngineering
Aron van Ammers ⛺⛰🏔 retweeted
We backed @Thoron_US and @t_koll because critical infrastructure needs reliable drones with trusted supply chains, built by people who understand the realities of operating them. The team’s integrated hardware and software approach brings our Atoms thesis to life: sovereign industrial capability, starting with utility inspection and building towards AI-driven maintenance.
Aron van Ammers ⛺⛰🏔 retweeted
Receiving an interesting ratio of death threats and VC outreach over this now, so should probably clarify:
We at @Auki do not actually track any person. In this visualization we show how manage to derive the data we actually track: what part of the shelf receives attention.
It's like tracking how many cars cross a bridge without tracking which car it was or who was driving.
The reason we do this is to allow the stores to better optimize their shelf space. Space is first or second biggest cost for every retailer, so optimizing the use of space is what allows them to reduce prices for the consumer - something they're all under intense pressure to do.
We do not store any data at all about the shopper, and have no plans to.
Auki's tech will not be used to profile you, but it will make your groceries cheaper and your store visits faster.
We have always been champions of privacy, and that's not going to change.
Peace.
We can measure what part of the shelf gets customer attention without any additional hardware. Just your normal CCTV cameras and a digital twin made with the phone in your pocket.
@CactusXR will absolutely revolutionize how well stores can do their merchandising.
Me: "I gave you a 5 minute task, what did you spend 70 minutes on??"
Agent: "5 minutes on the task, 65 minutes on a sidequest nobody asked for"
Aron van Ammers ⛺⛰🏔 retweeted
Meet our thesis for the next decade: AI, Atoms & Association
As world order collapses the TAM for sovereignty is growing est. to be a $275T+ market by 2036.
Nation states, companies, neo-collectives & individuals all want agentic infrastructure that guarantees their increasingly contested sovereignty.
Aron van Ammers ⛺⛰🏔 retweeted
⚠️ Scam alert: a phishing email is circulating that appears to come from an <@ outlierventures. io> address, pitching a "partnership."
It is not from us. Don't click any links or reply.
We're aware that some people are impersonating Outlier Ventures employees and we're taking immediate action. If you already engaged, please don't share any information with the sender.
If you're ever unsure whether outreach from OV is legit, email or DM us to confirm. Thank you and stay safe 🙏
Cursor Composer 2.5 really is that good. I've used it extensively in various dev projects over the last few days.
New CursorBench results just dropped.
Two big takeaways.
Composer 2.5 is way better than most people think.
63.2% score at $0.55 per task.
Nearly matching Opus 4.7 Max and GPT 5.5 Extra High at 20x less cost.
This is insane value.
Gemini 3.5 Flash is #10 at 49.8%.
Below GPT 5.5 Low.
Below Opus 4.7 Low.
Google's newest model can't even beat budget tier competition.
Composer 2.5 is the sleeper.
Gemini 3.5 Flash is the disappointment.
I rarely find myself reaching back to Opus for non-trivial and longer running tasks, which I had to do all the time before.
Subjectively, with most problems I throw at it, Composer 2.5 just "gets it", and gets the job done.
Plus, especially in Fast mode, it is FAST!
I have invested heavily in my @cursor_ai setup, optimizing for long running, parallel multi-agent workflows that can go for 30-60 minutes. Most work items now tend to get done in 10-20 minutes or much faster. Time to raise the bar!
Aron van Ammers ⛺⛰🏔 retweeted
Today, you can build a product in a weekend. You still can't fund it without a lawyer, a cap table, and a decade-long commitment.
🧩We've identified 10 open problems builders of the agentic internet need to solve.
We've set aside capital against this thesis and we're opening a Request for Builders today.
Aron van Ammers ⛺⛰🏔 retweeted
We've been sitting on this for months. Today we're going public with it.
10 open problems no one's solved yet for the agentic internet. Funding, ownership, coordination, moats. All broken when teams are hybrid human+AI.
If you're building anything adjacent, share your repo!
Today, you can build a product in a weekend. You still can't fund it without a lawyer, a cap table, and a decade-long commitment.
🧩We've identified 10 open problems builders of the agentic internet need to solve.
We've set aside capital against this thesis and we're opening a Request for Builders today.
This is true for any agentic tool or workflow. Coding and anything else.
Always give the agent a way to verify its work, and it's much more likely to deliver good results on its own.
6/ Give Claude a way to verify its work
Finally, make sure Claude has a way to verify its work. This has always been a way to 2-3x what you get out of Claude, and with 4.7 it's more important than ever.
Verification looks different depending on the task. For backend work, make sure Claude knows how to start up your server/service to test it end to end; for frontend work, use the Claude Chromium extension to give Claude a way to control your browser; for desktop apps, use computer use.
Personally, many of my prompts these days look like "Claude do blah blah /go". /go is a skill that has Claude
1. Test itself end to end using bash, browser, or computer use
2. Run the /simplify skill
3. Put up a PR
For long running work, verification is important because that way when you come back to a task, you know the code works.
Aron van Ammers ⛺⛰🏔 retweeted
Living outside the SV eco-chamber is a super power 🦸♂️
Big tech is now finally hiring philosophers. Not as advisors. As staff. And they are all British! 🇬🇧
Anthropic did it first. Now Google DeepMind (Demis, also British) has followed.
Q. How do you build an artificial mind without a working theory of mind?
Q. How do you engineer consciousness without an ontology of what consciousness even is?
Q. How do you align intelligence you can't even define?
These aren't engineering problems.
They're age old questions of the humanities.
I've been publicly arguing with SV and folk like Elon Musk (since May 2023) that the humanities would be critical for AI.
xitter.cf/dioscuri/status/204366…
