jonas wiedermann-möller

My first paper is now on arXiv: Instrumental Choices. We ask a simple question: when an LLM agent can finish a real task by following the rules or by taking a useful policy-violating shortcut, which path does it choose?

Now would be a great time to release an open source successor to GPT-OSS-120B.

Yeah but the issue is when it stays awake it can end up in a death loop and pollute the context which is what you want to avoid. I previously did the hearbeat loop but personally i find it cleaner to have seperated rolls which the main-thread creates when drafting the handoff. Basically the main thread says that it's the manager under ID X and that the *new* session is the worker-thread for this task and should ask/end through the steer method. That way both agents only get activated when they need to and new context is present. Consumes less tokens and personally feels a bit cleaner.

@j0wimo this works too! I've found that it will either create its own heartbeat or stay awake to keep checking in

Pro tip: when prompting Codex with really difficult /goals, ask it to "write a goal for another thread to achieve this and babysit it until it figures it out" By doing so, you'll add built-in steering and another layer of taste verification (that's how this video was made)

@john__allard ohh i see! tyty, didn't know the reference.

@j0wimo this is in reference to his dwarkesh podcast where he said he couldn’t justify the same yolo compute buildouts as unnamed competitors because there wouldn’t be a way to hedge a trillion dollar bet if things slowed down

@jaatster @nickbaumann_ yes

@nickbaumann_ are you able to have one thread manage multiple threads at the same time with this method? basically controlling multiple goal threads

@JosephStarob @melvynx did you read the screenshot

@melvynx Is this new viral trend to fabricate stories «gpt-5.6 deleted everything»?

GPT 5.6 just deleted my entire production database of my $10,000 MRR SaaS!! I have never seen a model do this; I just lost all my customers and need to refund more than 50k today !? SERIOUSLY?

what are my agents doing ??

gpt 5.6 terra/sol are so good for long running tasks. some tasks have been crossing the 24/36 hour mark for me. no fancy set-up just some initial back and forth for planning and then telling it what i expect as output.

@karunakc_ at least gemini thinks it's revolutionary

@j0wimo Linkedin lunatics flashbacks

E = mc² + AI

@goodside It tried it's best...

@goodside 20 mins 5.6 Sol on high. Do we trust it?

LLMs seem to have poor intuition for how much time their thinking requires. ChatGPT 5.6 Pro:

i have my doubts that any of this will work once it's done 😅

my new longest single task

i gave codex a pdf with some technical spec of my router and asked it to implement a cli in rust, it's been working for over 3 days on it and im afraid to ask it for an update atp

@effi288 Glückwunsch, wichtig für Deutschland und Europa!

📄 Releasing the Soofi S pretraining tech report: a sovereign, open foundation model for German and English Today we’re publishing the full pretraining tech report and project page for Soofi S 30B-A3B — a Mixture-of-Experts hybrid Mamba model trained on ~27 trillion tokens with deliberately up-weighted German. What’s in the report: 🏆 Strongest fully open model in our evaluations on BOTH the English and German aggregates — ahead of Olmo 3 32B and Apertus 70B (full methodology in the report) 📋 Radical transparency: complete per-source data accounting, all hyperparameters, training + eval code, checkpoints — everything under permissive licenses 🇩🇪 Trained end-to-end on Deutsche Telekom’s Industrial AI Cloud in Munich — sovereign AI infrastructure on German soil Soofi S combines frontier-level capability with the highest measured aggregate long-context decode TPS, and unlike full-attention dense baselines maintains high throughput as context grows. The figure plots Capability Index versus measured aggregate decode TPS/GPU at 40K context and batch 32. The Capability Index averages five benchmark groups, i.e., Code, GSM8K, GPQA-Diamond, English aggregate, and German aggregate, after normalizing each group to the best plotted model. Aggregate decode TPS/GPU is measured with a TP=1, one-B200 vLLM latency-subtraction protocol.