Maximilian Schlegel

Introducing MorphoHDL, a minimal language prototype for growing boolean circuits! paradigms-of-intelligence.github.io/morpho/ Early this year I wanted to design a size-agnostic graph rewrite rule system that could build functional boolean circuits. First I thought about "chemistry"-like reactive systems, but rules were too complicated with many different node types carrying multiple indices... Then I realized, that cell division is much more natural way of building complex structures, and once I started to treat graph edges as buses instead of single wires, everything clicked. No new formalism — just taking good old recursion and seeing how far I can push it. Ripple-carry adder, Brent–Kung adder, multipliers ... and some Haeckel-esque creatures along the way! Hope you enjoy the report of my journey and the demos.

damn, i thought i was at the frontier until fable started agreeing to help with my project 😪😓

@BlackHC If I was a company trailing behind the dominating market participant, why am I not incentivised to take even greater risks? Also historically, competition has been great for innovation. If there’s a monopolist, we might get stuck in a weird local optimum, no?

I'm not arguing against democratic institutions and free markets I'm worried about 3-5 companies with a mix of bad governance and a lack of alignment and safety focus cracking RSI while democratic institutions are not able to keep up I think it is a much better scenario is one company leading comfortably with a wider base of companies trailing incl open source models which serve to squeeze the margins and slow down frontier research funding That is much better than 3 companies at the frontier (and me not trusting 2/3 very much because they're playing catch-up)

Please join Anthropic if you're really really good at what you do. It makes sense to help the company that has the best governance so far keep their lead and extend it (and to reward them being principled) Currently we're in an AI race scenario and the easier it is for them to keep up the lead, the easier it is to preserve safety and negotiate The worst-case scenario is having multiple actors racing head to head because it makes coordination much harder and it is likely that the companies who are slightly behind will skimp on safety With a decisive lead that is still possible but at least, the misaligned models will be less capable by more than a bit Am I wrong?

@BlackHC Imo it’s dangerous if a single entity has absolute power over whatever and is in a position to decide what’s best for everyone. This is why we have democratic institutions & free markets, which both historically work best when we avoid (power/economic) centralization

The point was that if you have a strongly competitive race, it is more likely that all skimp on safety vs if one company has a lead and also strongly invests in AI safety etc RE board: you mean like Google? Yeah that is a risk but then Anthropic does have better governance in place and it is for benefit and not for profit. So overall it is more trustworthy compared to OpenAI and Google right now

@BlackHC …if the definition of safe or the measures to ensure safety by the one completely dominating company fail? What if one day their board wakes up and decides to not care about whatever you care about anymore? If you’re a customer you cannot even change anymore then?

@BlackHC I kinda disagree with the premise that we need that. We also don't rely on coordinated self-regulation eg in aviation or pharma? For this we have democratically controlled third parties. But if one comp dominates, you get information asymmetry, no peer comparisons etc. Also, what

New paper 🧵 We show that dynamic short convolutions consistently improve Transformers across scales. We make these gains practical with an efficient parameterization and custom Triton GPU kernels. The improvements carry over to MoEs and linear attention variants (Mamba-2/GDN).

@maxjendrall @YannickScholich lübars ist einfach das geilere bad und mehr los lustige geschichte, bin vor einigen jahren mal mit paar freunden voellig ohne grund von ner groesseren gruppe jungs rel aggressiv bedroht worden bis einer von denen mich von meinem job aus lübars wiedererkannt hat hahaha

@mtavitschlegel @YannickScholich crazyyy, wieso dann nicht am Tegeler See?

Let me introduce you to ETH Zurich students

@YannickScholich ne war in tegel aber waren halt sowieso auch immer im sommer in lübars haha

@mtavitschlegel holy shit you're the bademeister in lübars guy?? man i grew up in Lübars as well 😄 Warst du auf dem GHG?

@imobtw @chrisoffner3d nordberliner hidden gem

@mtavitschlegel @chrisoffner3d Lübarser Freibad

banger CV

@clemens1 @Xilo_K Thanks! But imo you have to endure the sad hell that is the info-center/hg library to get the proper ETH student experience lol

@mtavitschlegel @Xilo_K i just found it it’s the literaturhaus maps.apple/p/bHbU1a0D9B3h…

Paying 40$ for private library access was the play of the decade and it's not even close.

New modded-NanoGPT optimization benchmark result: @wen_kaiyue has improved upon both the Muon and AdamW baselines, by replacing their weight decay with hyperball optimization. The new record is 3325 steps.

@ninoscherrer last thing it has seen is Hertha BSC winning the football championship in germany, finally a solid model

Finally someone did this — vey cool work!!

@jenseisert Heilstätten?

A lost place close to Berlin.

You can find my colleagues here: Recursive Self-Improvement workshop @ ICLR Poster sessions: 10:00 - 10:30 & 12:00 - 12:30 @ Room 101 - D

Also check out the paper, a preprint is available on arxiv: arxiv.org/abs/2512.20605

We cooked up “Internal RL”. The new RL algo exploits a specific insight we got from analysing pre-trained Transformers and achieves success in tasks where all baselines (like GRPO etc.) FAIL! Wanna learn more? Search for @yaschimpf and @ninoscherrer at the RSI workshop at ICLR!

Go checkout my colleagues at the ICLR's “Recursive Self-Improvement” workshop, Room 101D!