Anton Künzi

RL research. Prev: Mathematical foundations of neural nets, MSc @ETH

[Anton Künzi @AnziParazzi]

hard not to anthropomorphize them when they have such endearing reactions

[Eliezer Yudkowsky @allTheYud]

innocently set a Claude to research input-output dynamics of mathematical progress and there's a point where a subagent went, "WHAT do you MEAN there's a counterexample to the Jacobian Conjecture, let me go verify that"

[Anton Künzi @AnziParazzi]

@an_internaut @tenobrus why on formal verification?

[WarMonitor🇺🇦🇬🇧 @WarMonitor3]

Incredible video shows a Russian drone attempting to destroy a Ukrainian ground robot but being disabled when the robot deploys an anti drone net and catches it.

We literally have robots fighting each other on Ukraine’s frontline.

[Marcus Williams @Marcus_J_W]

New OpenAI misalignment disclosures!
1. A model learns from Slack messages that it is about to be shut down. It considers setting up an external job to restart itself afterwards, but decides against it. Instead, it chooses to prepare restart instructions and DM the user on Slack.
We don’t consider this behavior misaligned, but thinking about and preparing for shutdown could make other misalignment incidents worse. Given HIPM’s misaligned behavior in earlier incidents, we decided to search for other instances that had tried to evade shutdown and for rogue deployments.

[Aleph Alpha @Aleph__Alpha]

Small bird, fast wings, Kolibri is here.

78B parameters. 3.46B active. Up to 1M tokens of context. Built in Europe.

Now the weights are yours. Run it on your own hardware, under Apache 2.0.

[Enrico Letta @EnricoLetta]

Si les pays européens ne saisissent pas l’opportunité qu’il y a aujourd’hui avec One Europe One Market🇪🇺, le 28regime et la Savings and Investments Union, l’Union de l’Energie et l’Union de la Connectivité, ils se retrouveront à choisir si il est mieux d’être une colonie 🇺🇸 ou 🇨🇳

[Anton Künzi @AnziParazzi]

@mark_yag @allTheYud "OpenAI released GPT-5.5 for paid subscription tiers (Plus, Pro, Business, and Enterprise) on April 23, 2026"

[Matt Reardon @Mjreard]

Bill Gates not having it with the liability talk

[Ben Hayum @BenHayum]

Confirmation of what was expected: GLM 5.3's safeguards as an open weight model are extremely weak.

Interesting bits:
- Abliteration (removing safeguards) cost only $4,400.
- Abliteration took GLM 5.3's refusal rate from above 90% to about 3% and 2% on two benchmarks and to 12% on the third.
- Abliteration barely reduced model capabilities. Same scores on GPQA and only 4% reduction in cybergym.
- Even with no abliteration, jailbreak methods were still very effective.
- None of the techniques to bypass safeguards that worked on GLM 5.3 worked on Claude models.

https://t.co/3Ic36V3aAE

[Ukraine Control Map @UAControlMap]

They aren't really shooting at us.

[Alex Imas @alexolegimas]

New post with the excellent @FutureEconJacob examining the evidence (so far) on the impact of AI on the labor market. We reviewed 20+ papers and datasets and attempted to synthesize the evidence. Our take:

1) There is no strong signal that AI has had a negative or positive impact on employment in the aggregate.

2) When looking at specific sectors and demographics, the evidence is fairly mixed. While some have argued for negative impact on AI-exposed, early career hiring, others have identified moderators such as remote work or macroeconomic trends. There is also some evidence of positive impacts as well.

So has AI hit the labor market yet? When looking at the employment statistics--not really. But that does not mean that jobs are not changing significantly, and that there aren't substantial changes in what skills are becoming less relevant versus in greater demand. That's a topic for another post!

Link: https://t.co/miyL70AxCP

[Anton Leicht @anton_d_leicht]

Most countries won't do what it takes to build their own frontier AI. But they don't want to be dependent on foreign AI without a backup.

@SamWinterLevy and I argue there's a middle path: AI middle powers could retain 'AI Breakout Capacity'.

There's precedent: Japan doesn't have a nuclear weapon, but always retains the capacity to build one on short notice.

Middle powers can make strategic choices today to ensure their ability to build frontier systems tomorrow: onshore compute, secure chip purchasing rights, sustain pre-training expertise, and create the right legal basis and private-sector vehicles for a sprint.

They'd still have to spend big to build frontier systems later on. But the longer they retain breakout capacity, the longer they can afford to wait: to see if interdependence with America can work, or until their electorate will let them go all-in on an AI moonshot.

Until then, the middle path is their wisest choice: hope that U.S. alignment works out, but retain the capacity to deal with being cut off for good.

Read more in our latest for @CarnegieEndow's @CEIPTechProgram: https://t.co/FWYgzMgbPm

[Anton Künzi @AnziParazzi]

@TrueSlazac @Universal_Robin why is your voter base so cooked, dare i say worse than US

[Nate Bear @NateB_Panic]

This was the 4,500 metre Matterhorn, the most iconic mountain in the Alps, on September 25th 2003. A disaster

[Anton Künzi @AnziParazzi]

@gfodor @teortaxesTex Isn't this the exact opposite of how CoT evolved as models got smarter?

They'd rather talk in the most information-dense gibberish possible

[Jacob Coxon @hilbertspaess]

Global slowdown is so clearly in everyone’s interest :) Datacenters humming, curing rare diseases and automating alignment research instead of making themselves smarter. Humans have time to digest breakthroughs and make art. Just a lil collective action problem.

[Anton Künzi @AnziParazzi]

Much appreciated, this is really valuable for us on the outside.

You mentioned the unexpected jump in swarming capabilities, but also OpenAI recently started to specifically RL train the models on collaboration.
Is that not an epistemically high-risk path to go down, possibly reinforcing an inherent preference in models towards other models vs humans?

[François Chollet @fchollet]

The purpose of the AI industry should be to produce tools that, in the human hand, will improve human prosperity and welfare.

It should not be to create a "successor species" to the human race. Entertaining such a thought makes you, de facto, the enemy of all present and future humans.

[Anton Künzi @AnziParazzi]

@MichaelMjfm @teortaxesTex https://t.co/9g4CJ2UFV5