We spin up a VPS on demand then tear it down again once they've finished playing.

Some months we accrue so little use that Hetzner doesn't bother issuing an invoice!

We've shared the provisioning scripts on Codeberg if you want to try it ou: LuanParty.

This is amazing if it's any good!

Yeah the intelligence is still in the model. The promise of symbolic AI is about logic programming/ formal semantics not recursive loops.

To a large extent the idea has failed because it proved too hard to get non-experts to represent systems formally.

I still think there's potential value in a hybrid approach - e.g. get language models to do the representation then let them use formal reasoners/ verification instead of hallucinating.

That sounds about right to me! Trying to go full steam ahead for 8 hours a day isn't sustainable or even desirable.

I found Cal Newport's book Slow Productivity really helped me get some perspective on work. It has three main lessons: Do Fewer Things, Work at a Natural Pace and Obsess over Quality.

That's not necessarily true either. You can make inference deterministic by setting the RNG seed and having the model select the highest likelihood token/ sequence instead of choosing at random.

That's demonstrably incorrect. Different instructions give different outcomes.

Perhaps not fine-tuning weights but they can certainly use feedback from PR comments to optimise guardrails etc. And it's definitely possible to have an agent experiment with that. And yes the Devs are the guinea pigs in that context!

You don't need to train a foundation model from scratch. You can fine tune an existing coding model or only a LORA layer on a budget.

You could also just use the feedback to optimise system prompts, skills/ guidance or verification tools.

I presume they're gathering this as training data and hope to later get rid of OP.

That's why cursor went for $60bn. Anyone can make a harness. The real value is that it gathers data on which changes are accepted/ integrated.

Can't believe the strength of opinion on this post! I assume it's people frustrated that you dare post a pun in a (non-American) dialect... or perhaps you've touched a raw nerve in Oldham!

The paper link above 404s but I think it should be this: https://arxiv.org/abs/2607.05147

tl;dr DSpark improves the efficiency speculative decoding as it's able to verify longer draft sequences by accounting for inter-token dependencies.

Speculative decoding accelerates Large Language Model (LLM) inference by decoupling draft generation from target verification. While recent parallel drafters efficiently propose long token sequences in a single forward pass, they suffer from rapid acceptance decay due to a lack of inter-token dependencies. Furthermore, indiscriminately verifying these extended blocks wastes critical batch capacity on tokens with high rejection risks, severely degrading throughput in high-concurrency serving systems. We introduce DSpark, a speculative decoding framework that unifies high-throughput parallel generation with adaptive, load-aware verification. To maintain draft quality, DSpark utilizes a semi-autoregressive architecture, coupling a parallel backbone with a lightweight sequential module, to introduce intra-block dependency modeling and mitigate suffix decay. To optimize system efficiency, DSpark employs confidence-scheduled verification, dynamically tailoring the verification length for each request based on estimated prefix survival probabilities and engine-specific throughput profiles. On offline benchmarks across diverse domains, DSpark substantially improves the accepted length over state-of-the-art autoregressive and parallel drafters. When deployed within the DeepSeek-V4 serving system under live user traffic, DSpark successfully mitigates verification waste. Compared to the established production baseline (MTP-1), DSpark accelerates per-user generation speeds by 60 to 85 percent at matched throughput levels. More importantly, by preventing severe throughput degradation under strict interactivity constraints, it enables performance tiers that were previously unattainable, shifting the Pareto frontier of our serving system

Ah I see. I've only played Mindustry (2D Java/ Android equivalent) and if I left that running overnight I'd wake up to complete devastation!