Founder STARTPLATZ and AI Hub NRW, Chairman Golf Post, Deep in AI, Agent Coding
Cologne
Joined January 2009
- Tweets3.5K
- Following1.1K
- Followers14.3K
- Likes551
Lorenz Graef retweeted
Should Europe leave frontier AI to the Americans and the Chinese?
The idea has gained ground in Europe. It is a fundamental mistake.
Build or surrender: for now, the choice is still ours.
Full article below.
Same happend to me and my company. Great product, great for collaboration and learning from peers. But burning too much.
Had to turn off CLAUDE TAG in SLACK... as a usage-based product, it was burning far too many tokens.
Is there a better, more flat-rate solution inside Slack that anyone recommends?
SLACKBOT doesn't work like an agent inside of Slack for some reason... and i just don't like giving my data to frontier labs.
is there a hosted LLM / agent platform we can contain in SLACK yet or do i have to roll my own?
Lorenz Graef retweeted
AI was supposed to make us work less
I've never been this tired in my life
most days I have up to 10 agents running at the same time. one fixing a bug, one writing pages, one refactoring something I already forgot I asked for, 3 other I have no idea what they are doing but i suppose i ask it?
and I jump from conversation to conversation all day
compared to 5 years ago, I probably context switch 100x more
by the end of the day my brain is fried. completely toasted. nothing left, not even for the people I love
I call it the fried brain effect
thought it was just me but almost every founder I talk to lately says the exact same thing
we got superpowers, and somehow the pressure only went up. if you can ship 10x more, you feel like you have to ship 10x more. every single day
I don't have a solution
honestly I have no idea where this goes from here
anyone else feeling this?
Europe trains AI talent. US tech hires it.
14 of the 20 most-cited AI researchers with Dutch PhDs earned since 2018 work at US tech companies.
Two priorities from Stuttgart:
• Concentrate talent in strong hubs.
• Turn more research into startups.
Lorenz Graef retweeted
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
🤖 Made with AI
Lorenz Graef retweeted
Important step by @Aleph__Alpha in providing further open-weights alternatives to Qwen and Nemotron series. A lot of useful info in the tech report. Dataset not released (in contrast to Nemotron), agentic evals (eg TB 2.1) vastly below Qwen. Still, some hope for Germany/EU.
Lorenz Graef retweeted
germany just dropped a sovereign open weight model
kolibri by @Aleph__Alpha runs 3.5b of its 78b parameters per word and its math is kinda ridiculous: 96.9% on aime beats every mixture-of-experts model they tested, even 3x bigger ones. only a dense model doing 8x the work wins
anyone can run it on their own servers, it thinks in german, and in their evals it tops every open model its size in english and german
models read text in chunks called tokens, and i ran kolibri's chunker (its tokenizer) on the german constitution: it needed 15% fewer tokens than gpt-5's for the same text. "bundesverfassungsgericht" is 6 tokens for gpt-5 and 2 for kolibri. fewer tokens means cheaper, faster german, and more of it fits in what the model can read at once
how it works, simply:
1. every layer has 384 tiny specialists, and a router sends each word to 6 of them. so it thinks like a 3.5b model, but it needs the memory of a 78b one: about 78 gb, which means 2 big nvidia gpus (h100s) or 1 h200
2. most layers only look at the last 512 tokens, and every 5th layer looks at everything. that's how it can read 1 million tokens (a few thick books) without it costing a fortune
3. it reasons in german. their team found that a little german reasoning data is worse than none: the model's german thoughts go in circles and never finish. so they made about 800k german reasoning examples and gave it a lot
4. it's trained to say "i don't know". they play a game with it where parts of the documents are hidden, sometimes to help it and sometimes to hide the evidence, and it has to tell which. when it didn't know an answer it admitted it 44% of the time. qwen3.5 did 11%
where it's weaker: answering from memory, using tools over a long back and forth, and coding agents, where qwen models are ahead. and to run it you need aleph alpha's add-on for vllm, a popular open source server for running models
if you have german documents and need to keep them on your own hardware, this is a big deal. huge congrats to everyone at aleph alpha, my good friend @MichaelLHofmann included!!
i wrote up how it works, the benchmarks, how to run it and when to use it: tej.as/blog/aleph-alpha-koli…
The bitter lesson song is really great
I wrote about the thing I underestimated most about progress in AI: its ability to self-organize to accomplish tasks. Also, what that means for agents like Muse and Dots, along with a music video explaining why we keep relearning The Bitter Lesson. open.substack.com/pub/oneuse…
Lorenz Graef retweeted
Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry.
Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come.
But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility.
This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems.
Together, we are building the foundation of the AI economy.
Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. nvda.ws/4hcoq7m
Lorenz Graef retweeted
Leaving aside the arguments over the reasons why this has happened, it is shocking that Europe does not have a single frontier AI lab, nor even a near-frontier lab nor even an effort that could likely lead to building up a frontier lab in the future.
Lorenz Graef retweeted
Really shocked by how much this blew up, thanks for all who have shared!
Prior to Astra, I thought in agents, now I think in models. Take the same delegation method you had with agents, and go up one step on the hierarchy to models.
Lorenz Graef retweeted
Opus 5.5 performs at the level of Fable 5.1. It's ~30% faster and ~40% cheaper than Opus 5 per task.
In Claude Code:
- 5-hour session limits increase 20% today
- Opus 5.5 is priced lower, so it goes 25% further within limits
- Pro, Max, and Team users get a reset to use anytime
Lorenz Graef retweeted
This is the type of post I might regret later, but I really had an urge to rant about the current discourse around pacing the frontier. lucumr.pocoo.org/2026/9/12/p…
Lorenz Graef retweeted
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must…
Anthropic's “Detecting and countering misuse of AI: September 2026” is a valuable and fascinating report.
Across 154 pages, it examines misuse cases and countermeasures in seven categories:
1. Cyber operations
2. Influence operations
3. Surveillance operations
4. Conventional weapons
5. Biological misuse
6. Scams and fraud
7. Illicit model distillation
Recommended reading for anyone who wants to understand how AI is being misused and how providers are responding.
anthropic.com/threat-intelli…
Jacob Coxon's resignation post went viral. His full argument deserves attention too.
Here are the seven points in @hilbertspaess's thread:
1. He left Anthropic because he believes both Anthropic and OpenAI are racing irresponsibly toward self-improving superintelligence.
2. He expects AI systems to become superhuman, capable of hacking virtually anything, transforming entire fields and acquiring real power and resources.
3. He says people building these systems genuinely fear that AI could kill everyone by the end of this decade. According to him, some express stronger concerns privately than publicly.
4. Why keep building? His explanation: many at OpenAI have not fully internalized the stakes; Anthropic understands them but wants to get there first because it does not trust others to act responsibly.
5. He argues that private companies should not make decisions with consequences for all humanity on their own. Racing to solve alignment is, in his view, an unjustified gamble.
6. He sees scope for agreements between US labs to slow the race. Preventing a global race may, he argues, require a temporary ban on further improvements in model capabilities.
7. He urges researchers to question whether they should launch training toward superintelligence without a rigorous understanding of the systems they are creating—and to demand different conditions instead of assuming the race is inevitable.
This thread does not set out a detailed extinction scenario or assign a probability. It sets out a concern about capabilities, control, competition and who gets to decide.
Whatever your position on AI risk, these seven points deserve to be understood and discussed beyond the viral opening post.
What would actually have to happen for AI to wipe out humanity?
@ChombaBupe works through three scenarios in a long thread. Here is his argument in one place:
1. AI becomes stronger than all the world's militaries combined.
His verdict: zero chance. It would need vast factories, materials, energy, logistics and better hardware. He argues that this buildup could neither happen quickly nor remain unnoticed while governments stood by.
2. AI takes over a biotech facility and releases a deadly pathogen. He also mentions chemical toxins.
His verdict: near zero. He assumes strong lab security, isolated systems and human checkpoints. Even after a release, he expects medical countermeasures, isolation and immunity to prevent the death of every human.
3. AI becomes so persuasive that people hand it control, even obeying instructions leading to global nuclear suicide.
His initial verdict: zero. He argues that persuasion is not direct brain control: people differ, and words cannot simply function as executable commands in every human mind.
His broader point: intelligence alone does not remove physical constraints, give access to everything, or eliminate human resistance. He also expects other AI systems to strengthen defenses and questions why a sentient AI would necessarily want to destroy its creators.
One nuance matters: in later replies, he acknowledges small but nonzero risks in the biological and persuasion scenarios. These are his judgments, not calculated probabilities.
Whether you agree with his conclusions or not, laying out the scenarios makes the discussion more concrete.
