Thomas Rueckstiess

AI researcher | ex MongoDB | Founder of

[Thomas Rueckstiess @tomonezero]

@chalaska 👋 hi from Melbourne. Trying something solo after 13 years at MongoDB.

[Thomas Rueckstiess @tomonezero]

@MatthewBerman Claude just loves orange accents on beige. Wonder why 😏

[Thomas Rueckstiess @tomonezero]

Tried the new Claude Design, redesigning my website. First of all, this product is at best alpha status. I hit so many bugs where the agent just failed, replied with <empty> or tool use errors, this is nowhere near v1 ready. I wish Anthropic would slow down on churning out features every day and use some of those tokens on QA and testing. But maybe that's the new world now: build with agents, ship fast, test in production.

Conceptually though, the direction is clear, and I'm sorry to say but AI is coming for designers just as it is coming for software engineers.

[Thomas Rueckstiess @tomonezero]

@heydave7 what docker image did you use for open webUI + gemma? I tried this playbook but the pinned ollama version is too old for gemma4: https://t.co/873trt5L9V

[Dave Lee @heydave7]

This afternoon I picked up a new Nvidia DGX Spark computer with the goal of trying to run Gemma 4 31b (4bit) on it locally as a server.

Just 1.5 hours later, it’s working!

Using Open WebUI on my MacBook as the interface and it’s connecting to my DGX Spark running as a Gemma 4 server.

[Thomas Rueckstiess @tomonezero]

@thsottiaux This is how you deal with rate limit issues! Thank you.

[Thomas Rueckstiess @tomonezero]

They must have auto-detection of certain patterns and aggressively trigger bans. My setup: I run `claude -p` inside a Docker container with an oauth proxy on the host. Maybe this looks like an Open Claw setup to their abuse detection system? But I'm the only one using it and only through their products (claude CLI).

I wish there was more transparency what is allowed with the Pro/Max subscriptions. And a warning or at least explanation would be nice too.

Now I'm in Claude jail. Creating a new account would be seen as ban circumvention. What do I do?

[Thomas Rueckstiess @tomonezero]

@markmdev @trq212 @trq212 same here. It's frustrating to get banned without explanation and no response on the appeal after spending thousands of $ on max subscription fees.

[Thomas Rueckstiess @tomonezero]

So mathematically and architecturally, LLMs do just "predict the next token". What's astounding though is that at LLM scale, semantics and reasoning emerge through complex internal representations.

[Thomas Rueckstiess @tomonezero]

LLMs sample a token sequence from their learned distribution p(x), where x is a sequence. This is done token by token however, using ancestral sampling according to the chain rule: p(x) = \prod_t p(x_t | x_{&lt;t}). Each factor is conditioned on all previous tokens.

[Zach Warunek @ZachWarunek]

so it turns out the LLMs aren't "just predicting the next token"

[Thomas Rueckstiess @tomonezero]

Seems more like a cost or system
complexity trade-off than a sensor fusion issue. Especially if you deep learn end-to-end, more data is better. Sensor fusion can happen implicitly in the net with multi-modal architectures.

[Elon Musk @elonmusk]

Lidar and radar reduce safety due to sensor contention. If lidars/radars disagree with cameras, which one wins?

This sensor ambiguity causes increased, not decreased, risk. That’s why Waymos can’t drive on highways.

We turned off the radars in Teslas to increase safety. Cameras ftw.

[Thomas Rueckstiess @tomonezero]

This was a really fun experiment. It's amazing how quickly these models improve and after this week of intense vibe-coding, I feel like I could write anything, in any language.

Also curious to hear other people's experience with more complex projects.

[Thomas Rueckstiess @tomonezero]

I vibe-coded an in-memory graph database in TypeScript in 5 days with Claude. It supports a subset of the Cypher query language. ~15,000 lines of code, in a language I don't know well.

The value of software is going to zero.

More details and what I've learned below.

[Thomas Rueckstiess @tomonezero]

And it's not obvious what counts as premium requests vs unlimited regular requests.and despite the integration, I found Claude Code in the terminal a better user experience.

[Thomas Rueckstiess @tomonezero]

Gemini 2.5 creates well-written code and its large context size lets you paste entire repos in. The aistudio UI is (like any UI by Google) janky and nearly unusable though. And it is missing the GitHub integration that Claude's web interface has.