Between AI, Engineering and Design. Building an AI native ERP Platform, a better MDM tool for corps and a Business & Fin Ops Platform for SMAs OSS contributor
Berlin
Joined May 2014
- Tweets3.6K
- Following946
- Followers249
- Likes3.6K
Pinned Tweet
I spent last night building a prototype for what digital government in Germany could feel like.
deindeutschland.app
One place to understand government services, navigate bureaucracy and, eventually, get things done across agencies.
Heavily inspired by America.gov, but added my touch to it. Their launch made me wonder: what would this look like for Germany?
It’s still a prototype and there's no llm yet because I didn't want to carry the inference cost. The hard parts are obvious: real integrations, identity, registers, security, legislation...
We've spent years talking about digital government and thrown money at consultants.. I'd rather make the idea tangible and start from there.
Here’s the launch film ↓ – at Berlin, Germany
Simon retweeted
If your building a startup you should treat every week as an opportunity to launch.
There’s a misconception that a successful launch video needs to cost tens of thousands of dollars and feel like a high production commercial. But @FarzaTV has a different approach.
Farza is the founder of @heyclicky and previously built Buildspace, one of the largest online schools for people working on their own ideas. He’s been making videos for over a decade, and today his posts regularly get millions of views. Rather than treating launches as rare, high-stakes moments, he’s made them part of the everyday process of building a company.
Farza sat down with YC’s @aaron_epstein to share his playbook for turning demos into stories people actually want to watch. He breaks down how he scripts videos that still feel off the cuff, why founders should launch constantly, and and how to figure out which channels actually convert.
01:44 — Launch Constantly
03:38 — Start Shipping Demos
05:02 — Show the Process
06:07 — The Clicky Launch
07:43 — Write a Hook That Stops the Scroll
10:30 — Why Farza Scripts Every Video
12:04 — Designing the First 10 Seconds
15:15 — Demo-Driven Development
17:25 — You Have to Take a Thousand Shots
20:48 — Farza’s Formula for Videos That Convert
23:01 — Turning Demos Into Stories
24:53 — Show What You're Actually Building
29:11 — Finding Customers on Social Media
31:02 — Which Platforms Actually Convert?
39:01 — Why Farza’s First Hire Was a Filmmaker
40:49 — The Video Stack
45:41 — The 30-Day Launch Challenge
Simon retweeted
SF VCs:
> "Hardtech requires too much VC capital"
> Invests $1 trillion dollars into prediction market gamble-gooning company
@jimbelosic :
> Says he never aimed to build a company and play financial games
> Does ~$200M in annual revenue before raising any VC money
> Once he raises, says his favorite use of venture dollars is "helping employees buy homes, start families, and move up from a bicycle to a car"
Be like Jim.
Around 2016, Jim Belosic bought a $750,000 laser cutter on a whim.
This month, his company SendCutSend raised $140 million and hit a valuation of $2.1 billion. It's on track to make 50 million metal parts this year for more than 300,000 customers.
This is the story of how SendCutSend arrived and ended up reviving US industry. corememory.com/p/sendcutsend…
tailscale and cloudflare tunnel can reduce direct exposure. I'd also check what a compromised service could reach: databases, deployment credentials, backups.
Those permissions determine how far an attacker could get after breaking into one service.
I'd keep databases, ssh and admin interfaces behind controlled access.
With Tailscale, that means explicit access grants. With Cloudflare Tunnel, I'd check who can access each published application. A tunnel alone doesn't make it private.
I'd allow only the connections each workload needs and review outbound traffic too.
Small subnets don't isolate services by themselves. I'd check routing and firewall rules alongside each workload's identity permissions.
Each workload gets its own identity and narrowly scoped permissions.
For deployments, I'd use oidc federation where supported. Any remaining secrets need an owner and a way to rotate them. I'd still treat a runner with production access as a sensitive system.
I'd build in isolated ci jobs and deploy a fixed, reviewed artifact.
Untrusted pr jobs shouldn't receive production credentials. Third-party actions should be pinned to reviewed commit SHAs. Production servers shouldn't need broad repository access.
I'd assign owners for patching and alerts, send logs outside the workload, and protect backups with separate access controls.
Then I would test a restore. For a small vps, I'd keep the setup simple enough to maintain and document how to rebuild it after a compromise. – at Berlin, Germany
Maybe important to repost this now that everyone is getting hacked
If you have a VPS, I recommend you to get if off the public internet and into a Tailscale network and Cloudflare Tunnel ASAP
It doesn't matter you have SSH key login, this is about exploits and vulnerabilities AI is finding
There's 2 ways into a VPS:
- via your site/app (like the PHP exploit of this week with a file upload hack)
- via your server (like hacking into SSH or some open port)
The site/app side you put on a Cloudflare Tunnel and then with the default Free Plan it has WAF security enabled by default and they jump on new security exploits fast, it means that the only way to reach your website is via the Cloudflare Tunnel, your server does not respond to port 80 or 443 anymore
The server side you put on a Tailscale network which is like your own private network that nobody can access but your own devices (like your laptop, your phone and your VPS) kinda like a VPN
Then you switch off ANY regular access to your VPS on the firewall, in my case I use Hetzner and you create a new firewall then empty it and that blocks everything (test it)
How do Tailscale and Cloudflare Tunnel get in then? They don't. They go outbound. They make a tunnel to outside which sends and receives
Now you've reduced your attack vector to Tailscale and your Cloudflare Tunnel, much safer, not perfect btw, but nothing is
Found Max and @laraavci_ at @OpenAI's DevExchange here in Berlin. Who else is here? 😃
Simon retweeted
Spent a lot of time on this today. I confirmed it's a real bug on Anthropic's side. To be clear: this does not affect weekly limits.
All Agents SDK usage will drain your 5h limits faster than using Claude Code directly.
I have reported this to Anthropic, hoping for a fix soon.
Just downloaded T3 Code and hit 5-hour limits on my Claude Max plan for the first time. I was running just one session where I asked it to fix a few payment related bugs for a project, and it emptied the whole 5-hour limit in ~30 mins.
What's going on? Does T3 Code work differently? Did I make any mistake?
I prompted just how I prompt inside Claude Code CLI.
Please help.
Yesterday started with lunch with a former colleague and ended at the @NVIDIAGTC afterparty, seeing what cracked ppl are building in physical AI.
In between: yapping with @maxjendrall at @circuit_ber and some networking at @aicampusberlin.
I knew @theo was an @AnthropicAI shill all along, guess the meetup badge finally gives it away – at Berlin, Germany
If you sit next to the user long enough, you will build the right thing.
Forward deployed engineer is the best job in AI imho. Everyone seems to be shipping demos and I've seen it too many times.
*me checking ebay*
...nvm
6 year old hardware is more expensive than at launch – at Berlin, Germany
NVIDIA sold millions of 3090s with 50% of the tensor speed switched off. Not missing. Switched off.
1 kernel module flips it to full, and my BF16 tensor pipe went from 82 to 155 TFLOPS. Nothing else changed on the cards, and the outputs match the stock run byte for byte.
All so the gaming card wouldnt eat into workstation sales. Every 3090 owner ran at 50% for 6 years, so a pro card could look worth the price. Fine, its their business model.
And LLMs are the least interesting place for this. Decode waits on memory, and my engine runs its experts in int8. Still +10% on prefill.
The models that live on BF16 tensor math are everything else. Qwen Image 2.1 and FLUX 3 Klein, where every diffusion step is big matmuls. MiniMax H3, which takes 4 to 7 minutes on my cards. Music and speech models. And training, every LoRA anyone ever trained on a 3090 went through exactly the path that was cut.
Thats what I measure next. This is only the start.
Simon retweeted
A BioWare dev on reddit leaked a Mass Effect 5 trailer... it's called MASS EFFECT UNIVERSE and it's fucking gorgeous. I'm crying. This is everything I wanted. They were a lot farther than we thought. FUCK EA
