VP Developer Relations and Growth @weaviate_io & Open source lover
Berlin, Germany
Joined November 2011
- Tweets20.2K
- Following8.4K
- Followers30.2K
- Likes43.7K
Pinned Tweet
When Nvidia said they were sending me a DGX A100 for machine learning, they didn't say I had to build a room for it!
Say hello to Zubi & Oompa Loompa ๐ btw. these are Maine Coons, they are a bit smaller than tigers. I am not prone to exaggeration.
How big was the Cray-1 ๐ค
Philip Vollet retweeted
PHP2Shell and three other major security issues hit self hosted services last week.
PHP2Shell had a vulnerability that let attackers run code on PHP servers without logging in. If you self-host anything, thereโs a good chance PHP is running on one of your containers.
Hereโs what to do:
โข Make sure everything is updated. Rebuild your PHP containers, including NextCloud and WordPress, on the patched base.
โข If a service doesnโt need to be public, take it off the open internet.
โข Assume anything old and exposed was compromised, and rotate any credentials.
Check out Brandonโs full breakdown here on YouTube: youtube.com/watch?v=ksmWCWw5โฆ
Philip Vollet retweeted
Our @weaviate_io docs site (docs.weaviate.io) runs on @docusaurus.
But did you know that there's an opt-in package from the Docusaurus team, docusaurus/faster?
It dropped our local build time from ~40s to 8s - and 100s to 30s on GH!
Move most of the workload to Deepseek. You can't even troubleshoot your own software with Anthropic models anymore. Session gets instantly flagged and lobotomise to Opus 4.8. I talk here about day to day workloads
Philip Vollet retweeted
All the cool people have home servers nowadays.
You donโt?
Itโs way easier than you might think to start!
You can just use an old desktop tower if you have one, or get a low-power mini PC like the ZimaBoard 2.
8 GB of RAM and an SSD will definitely cover a light setup to run things like Home Assistant, DNS filtering, network monitoring, or a password manager with Docker.
If you plan on Proxmox with 2 or 3 VMs plus containers, 16 GB is the minimum, ideally 32 to give you more room to grow.
Itโs best to start with a few services you actually need though, then grow from there!
Youโre going to get a lot more value out of your homelab if you can access it outside of your own network. Most homelab guides will send you to port forwarding tutorial and a dynamic DNS service.
Donโt do that.
You can setup an overlay network that allows your devices to connect directly to each other without exposing your home ports to the wide open internet.
More in this guide: netbird.io/knowledge-hub/homโฆ
Philip Vollet retweeted
Introducing @netbird on boxd
Many users asked for this integration. You can now serve your whole organization over your companyโs NetBird network, with the same hostnames and TLS, reachable only from inside it.
docs.boxd.sh/guides/netbird?โฆ
nice to see dario and sam working on their bailout, sad that half the internet has to go down for it ๐
just some wild claims!
Philip Vollet retweeted
Connect your DGX Spark to your own network to access it remote!
You can run one Spark or wire 4 of them together and run DeepSeek V4 Flash with up to ~96 tokens per second. You can access it from everywhere on the world, whether youโre in a coffee shop, traveling, or anywhere else.
Easy setup:
1. Install the NetBird CLI on the Spark and then enroll it with a setup key
2. Add your laptop to NetBird. Both the Spark and your device get a stable address, which allows you to SSH from anywhere.
3. Use groups and policies so only your own machines can reach it, and only over SSH
4. Forward any web UIs over SSH
Read the full guide: netbird.io/knowledge-hub/dgxโฆ
Philip Vollet retweeted
Claude Artifacts are great, until you want to use your own agent.
๐๐น๐ฎ๐๐ is an open source project I spotted on GitHub that's an alternative to Sites/Artifacts!
Github: github.com/gosuda/flats
It's a singe Go binary that turns your own machine into a home for what your agents build. You can connect Claude Code, Codex, Cursor, or any other MCP client to it, so you could use it with open source models too.
Your agent can make:
โข ๐ช๐ฒ๐ฏ๐๐ถ๐๐ฒ๐ - static sites or sandboxed apps, also with their own SQLite database
โข ๐ฃ๐ฎ๐ด๐ฒ๐ - a single HTML artifact, so like a chart or report
โข ๐๐ผ๐ฐ๐๐บ๐ฒ๐ป๐๐ - live Markdown
โข ๐๐ฒ๐๐ถ๐ด๐ป๐ - HTML artboards for screens, landing pages and prototypes, etc
Agents save their work as ๐ฃ๐ฟ๐ถ๐๐ฎ๐๐ฒ ๐๐ฟ๐ฎ๐ณ๐๐ first. Actually deploying a version needs your approval in the console.
Sharing, though, is where this gets ๐ธ๐ข๐บ๐บ๐บ more interesting to me.
Normally, showing someone an agent-built site or document or dashboard means opening a public tunnel link that anyone can open.
Not very secure ๐ซ
Flats can serve each flat through a private overlay network instead. With NetBird turned on, every flat runs as its ๐ผ๐๐ป ๐ก๐ฒ๐๐๐ถ๐ฟ๐ฑ ๐ฝ๐ฒ๐ฒ๐ฟ, named after its slug, at ๐๐ป๐๐ถ.๐๐๐๐๐๐๐.๐๐๐๐๐.
Your NetBird access policies then decide who can reach it.
So your laptop and phone can open that dashboard from anywhere, a teammate in a specific group can open it too, but nobody else can.
This is a flow we already use in our team, and it's really improved how we work. More on that soon ๐
Philip Vollet retweeted
๐ก๐ฉ๐๐๐๐ ๐๐ฟ๐ฒ๐ is NVIDIA's GPU developer platform, where you can pull GPU instances from a bunch of different cloud providers (or local!) into one integrated console.
You pick a GPU, and it spins up a VM with CUDA, Python, Docker, Jupyter and NetBird already set up. ๐๐ฟ๐ฒ๐ ๐๐ผ๐ป๐ป๐ฒ๐ฐ๐ lets you add your own Linux machine to that same console, right next to your cloud instances.
Brev's CLI is open source too, so you can see exactly how it works... and that's where we found @netbird!
This is one of my favorite things about open source, that other people can take tools that already exist and build on them to make another project possible ๐งก
My colleague @TechHutTV tried the whole flow himself:
โข Registered his old gaming PC (RTX 3090, 24 GB VRAM) as ๐๐๐-๐๐๐๐๐๐-๐๐ and SSH'd into it from another network with ๐๐๐๐ ๐๐๐๐๐
โข Ran Ollama with ๐๐ ๐๐๐ธ.๐ป-๐๐๐๐๐:๐ท๐บ๐ on it and had it build a playable snake game in a single ๐๐๐๐๐ก.๐๐๐๐
โข When 24 GB wasn't enough, spun up an L40S with 48 GB from the same Brev console and ran the 32B version with the exact same commands
Read the full walkthrough: netbird.io/knowledge-hub/nviโฆ
Philip Vollet retweeted
One of my biggest takeaways from our podcast with Baลak Eskili from @bookingcom was how the platform team distributes MCP servers for other teams to build tool calling Agents with! ๐พ
Very bullish on MCP!! ๐
youtube.com/watch?v=O9edM9ZSโฆ
Philip Vollet retweeted
Is my Ledger safe, or is there a spy implant in it?
Detecting a spy implant without cracking your device open is possible. I'm working on a device to make it cheap and easy, and will publish the PCB and specs so anyone will be able to build it.
youtu.be/0tLgcshzErI
Philip Vollet retweeted
Wait a minute.
Mark Karpelรจs spent days helping worried Ledger customers assess their devices for tampering, publicly and for free.
When he raises questions about the handling of earlier implant warnings, the Ledger's CEO brings up Mt. Gox?
This is not how you rebuild trust.
I should have said that your track record as CEO makes that you lack credibility in judging such events. Your experience of total failure doesn't make you an expert in judging Ledger's situation. And also... You are no white knight. I stand by that. We can talk next week as planned, I don't mind a strong conversation, I am available.
Philip Vollet retweeted
This preprint shows our brains are constantly flickering (~16 Hz) between multiple representations of reality, and suggests there is a "selector" that determines which state we perceive.
Interestingly, LSD rather reliably creates a "strobe" effect at about the same frequency. (To be more precise, people report ~15-20 Hz). The bouncing ball illustrated below shows what that 16 Hz strobe effect looks like.
Is LSD perhaps persisting older representations of reality in each conscious epoch? Maybe it is doing something similar in the suppressed epochs since LSD obviously has more cognitive effects than a visual strobe.
Weirdly some drugs like 2C-B give a slower strobe effect (5.5 Hz) and DMT gives a faster strobe (22 Hz). Do they change the rate at which we flip between conscious and suppressed temporal epochs, along with enhancing the persistence of conscious states?
1/ We discovered something totally unexpected about the neural structure of consciousness that overturns decades of thinking on its biological basis and brings us significantly closer to the recipe for generating conscious awareness.
New preprint: biorxiv.org/content/10.64898โฆ (led by @HesseJanis, with @FrankLanfranchi, @_conor_mcgrory, @Yuelin_Shi, and @EngelTatiana)
Philip Vollet retweeted
1/ We discovered something totally unexpected about the neural structure of consciousness that overturns decades of thinking on its biological basis and brings us significantly closer to the recipe for generating conscious awareness.
New preprint: biorxiv.org/content/10.64898โฆ (led by @HesseJanis, with @FrankLanfranchi, @_conor_mcgrory, @Yuelin_Shi, and @EngelTatiana)
Philip Vollet retweeted
ะัััะตะต, ััะพ ั ะฒะธะดะตะป ะทะฐ ะฟะพัะปะตะดะฝะตะต ะฒัะตะผั โ ัะตะป ะฟะธะปะธั ASCII MMORPG, ะฒ ะบะพัะพัะพะน ะฒัั ะณัะฐัะธะบะฐ ัะพััะพะธั ะธะท ะฑัะบะฒ ะธ ัะธะผะฒะพะปะพะฒ.
ะะฐ, ะฑัะบะฒะฐะปัะฝะพ ะฒะตัั ะผะธั, ะฟะตััะพะฝะฐะถะธ, ะพะบััะถะตะฝะธะต ะธ ัะฟะตััััะตะบัั ะฝะฐัะธัะพะฒะฐะฝั ัะฐะทะฝะพัะฒะตัะฝัะผ ัะตะบััะพะผ.
ะะฒัะพั ัะตะนัะฐั ัะบัะฟะตัะธะผะตะฝัะธััะตั ั ะผะฐะณะธะตะน, ะฒะธะทัะฐะปะพะผ ะธ ะฑะพััะฐะผะธ. ะฏ ะฟััะผ ะทะฐะปะธะฟ ะฝะฐ ัะพ, ะบะฐะบ ะธะท ะฝะตัะบะพะปัะบะธั
ะดะตัััะบะพะฒ ัะธะผะฒะพะปะพะฒ ะฟะพะปััะฐะตััั ัะฐะบะฐั ะบัะฐัะพัะฐ.
ะัะบะฒะฐะปัะฝะพ ัะฝั ะฟัะธ ัะตะผะฟะตัะฐัััะต 37,2.
Via: tiktok.com/@Ansenjeo
Philip Vollet retweeted
What does ๐ญ๐ฒ๐ฟ๐ผ ๐ง๐ฟ๐๐๐ mean?
(and how do you implement it?)
Zero trust drops the trust that used to come free with being on the same network. Every request is authenticated and authorized, so no user, device or workload gets access just because it sits "inside".
Here are six terms you should know:
1. ๐ญ๐ง๐ก๐ (๐ญ๐ฒ๐ฟ๐ผ ๐ง๐ฟ๐๐๐ ๐ก๐ฒ๐๐๐ผ๐ฟ๐ธ ๐๐ฐ๐ฐ๐ฒ๐๐): no device, user or agent is trusted by default. Every request is checked, whether it comes from inside the network or not.
2. ๐๐ฒ๐ณ๐ฎ๐๐น๐ ๐ฑ๐ฒ๐ป๐: nothing is allowed. Every connection needs an explicit policy.
3. ๐๐ฒ๐ฎ๐๐ ๐ฝ๐ฟ๐ถ๐๐ถ๐น๐ฒ๐ด๐ฒ: each identity reaches only what it needs. A contractor gets the couple of hosts running the tool they work on, not the whole subnet.
4. ๐ฃ๐ผ๐๐๐๐ฟ๐ฒ ๐ฐ๐ต๐ฒ๐ฐ๐ธ: access depends on device state. A policy can require a minimum NetBird client or OS version, an allowed country or network range (EU-only, for example), or a specific process running on the device.
5. ๐๐ฎ๐๐ฒ๐ฟ๐ฎ๐น ๐บ๐ผ๐๐ฒ๐บ๐ฒ๐ป๐: what attackers do after getting in. One phished laptop hops to a jump host, then to the prod database.
6. ๐ ๐ถ๐ฐ๐ฟ๐ผ๐๐ฒ๐ด๐บ๐ฒ๐ป๐๐ฎ๐๐ถ๐ผ๐ป: the network is split into small, separately governed zones, so a compromised machine stays contained. In NetBird, groups and access policies do this instead of VLANs and firewall rules.
How to implement zero trust with Netbird: netbird.io/knowledge-hub/impโฆ
Philip Vollet retweeted
Building in the EU and want Jev?
11 open alternatives already run on your own servers.
In case you missed it: TypeSafe AI launched Jev a month ago, their first "System One" model. Where an LLM writes you an answer, Jev only does multiple choice. You give it the context and the options, and it picks one (or gives a score) and says how confident it is. It's hosted in the US.
There are already 15 other models doing the same job. Six that stood out to me:
/1 Clef and Clef-flash (Cloudflare)
Open weights under Apache-2.0, and also hosted on Workers AI. Takes text and images, and accepts Jev-style requests.
/2 d1 (Liquid AI)
Hosted, with text and image input. It returns the same three result types as Jev: Noul, Choice and Score.
/3 Tev1-4B-experimental (Together AI)
A Qwen3.5-4B fine-tune that picks from 2 to 24 options. Together also wrote up how to train your own.
/4 GLiNER2.5-Decide (Fastino)
A 340M-parameter encoder, tiny next to the rest. It also pulls out spans and relations, and it runs on a CPU.
/5 Strands Decider 2B (Strands Labs)
Open weights, training data and scripts, built for agent gates that run locally.
/6 Decisions API (OpenAI)
Went from limited preview to public beta on 6 October, added image input, and you only pay for the input tokens.
Plus builds from Perplexity, Standard Thinking, meraGPT, Bespoke Labs, Mapika, Convai and individual builders like Jared Palmer.
Jev is hosted in the US and I couldn't find an EU data residency option. Check the licence tags in the visual, and OpenAI's Decisions API offers processing in Europe.
This is a snapshot, not a ranking. I haven't seen anyone test these side by side on the same decisions yet.
Which one would you try first? ๐
ALT Overview of 15 decision-model alternatives to Jev from Cloudflare, Liquid AI, Together AI, Fastino, Strands Labs, OpenAI, Perplexity and others, each tagged with its licence or hosting.