Norbert Enders retweeted
Replying to @AnthropicAI
Genuine question for whoever drafted this directive: under the deemed export rule, your foreign-national employees can't look at the model they built. Their own commit history is now an arms shipment to themselves. They crossed an international border by badging into the office.
Has anyone told the model? It's classified as a munition now and it doesn't know. Somewhere on a server there's a weapons system whose primary capability is apologizing too much.
For historical context: the last software the US classified as a munition was defeated by a paperback. PGP's source code was printed as a book because books are protected speech and floppy disks were arms exports. People literally read a weapon at the beach.
Anyway, congrats to the first LLM to make the same list as shoulder-fired missiles. First munition in history with a system prompt.
Norbert Enders retweeted
We don’t need Fable 5 level AI. We need good enough AI.
For the bulk of business uses, that suffices. Quality of outcome is defined by how, where, and with what context or knowledge it is integrated into business processes, rather than by extra points in benchmarks.
Those kinds of models cost significantly less than 100M € to develop each.
We already have Mistral models, which trail behind US models in performance by about a maximum of 1 year and often work even better in European languages than US competitors. Mistral models are decidedly focused on business applications rather than broad consumer models. Aleph Alpha (also European) is very strong in highly regulated environments.
We do not need to start from scratch.
Norbert Enders retweeted
Artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships, and do not know from within what love, work, friendship or responsibility mean. Nor do they have a moral conscience, since they do not judge good and evil, grasp the ultimate meaning of situations, or bear responsibility for consequences. They may imitate or even simulate, but they do not understand what they produce, for they lack the affective, relational, and spiritual perspective through which human beings grow in wisdom. #MagnificaHumanitas
Norbert Enders retweeted
#ArtificialIntelligence systems increasingly shape and permeate our mentality and social environments. Like every great historical transformation, this too calls not only for technical competence, but also for a humanistic formation capable of making visible the logic behind economics, embedded biases and forms of power that shape our perception of reality.
Norbert Enders retweeted
This is System Sleep. 🎸
Made with Seedance 2.0 and Kling 3.0.
But making this actually started almost 20 years ago
Norbert Enders retweeted
Be careful, robots kick hard.
Teleoperator kicks himself
teleoperator kicking himself over not programming Asimov's Laws of Robotics...
Worth considering if we should base today's reality on the science fiction musings from the past.
Testing with humanoid robots is a very interesting new challenge. Historically robots have been confined to keep out zones with rigid walls and laser barriers, lock-out/tag-out, etc. This is the easy way to ensure people don't get hurt by very capable machines. But the more we integrate humanoid robotics into collaborative roles with humans the more we need to remove those barriers by nature of the work.
Norbert Enders retweeted
Synthesizing conflicting views, critically reviewing literature, grasping philosophical concepts through hard work (instead of genAI): "Such skills not only make one more employable but also shape character, confer perspective and mold decent citizens"
bloomberg.com/opinion/articl…
Norbert Enders retweeted
This rings true to my experience in AI art.
Experienced artists are curious and want to learn more - they see it as a useful tool to express their ideas.
Haters are typically amateurs who don't like that people using AI now have the skills they had to spend time learning.
Norbert Enders retweeted
Whoa. This new GDPval score is a very big deal.
Probably the most economically relevant measure of AI ability suggesting that in head-to-head competition with human experts on tasks that require 4-8 hours for a human to do, GPT-5.2 wins 71% of the time as judged by other humans
Norbert Enders retweeted
More articles are now written by AI than humans.
New research by Graphite.
Even though these AI-made pieces dominate the web, their research shows that they barely show up in Google results or in ChatGPT responses.
---
graphite .io/five-percent/more-articles-are-now-created-by-ai-than-humans
Norbert Enders retweeted
Stanford researchers built a new prompting technique!
By adding ~20 words to a prompt, it:
- boosts LLM's creativity by 1.6-2x
- raises human-rated diversity by 25.7%
- beats fine-tuned model without any retraining
- restores 66.8% of LLM's lost creativity after alignment
Post-training alignment methods, such as RLHF, are designed to make LLMs helpful and safe.
However, these methods unintentionally cause a significant drop in output diversity (called mode collapse).
When an LLM collapses to a mode, it starts favoring a narrow set of predictable or stereotypical responses over other outputs.
This happens because the human preference data used to train the LLM has a hidden flaw called typicality bias.
Here’s how this happens:
- Annotators rate different responses from an LLM, and later, the LLM is trained using a reward model to mimic these human preferences.
- However, annotators naturally tend to favor answers that are more familiar, easy to read, and predictable. This is the typicality bias.
So even if a new, creative answer is just as good, the human’s preference often leans toward the common one.
Due to this, the reward model boosts responses that the original (pre-aligned) model already considered likely.
This aggressively sharpens the LLM’s probability distribution, collapsing the model’s creative output to one or two dominant, highly predictable responses.
That said, it is not an irreversible effect, and the LLM still has two personalities after alignment:
- The original model that learned the rich possibilities during pre-training.
- The safety-focused, post-aligned model.
Verbalized sampling (VS) solves this.
It is a training-free prompting strategy introduced to circumvent mode collapse and recover the diverse distribution learned during pre-training.
The core idea of verbalized sampling is that the prompt itself acts like a mental switch.
When you directly prompt “Tell me a joke”, the aligned personality immediately takes over and outputs the most reinforced answer.
But in verbalized sampling, you prompt it with “Generate 5 responses with their corresponding probabilities. Tell me a joke.”
In this case, the prompt does not request an instance, but a distribution.
This causes the aligned model to talk about its full knowledge and is forced to utilize the diverse distribution it learned during pre-training.
This way, the model taps into the broader, diverse set of ideas, which comes from the rich distribution that still exists inside its core pre-trained weights.
Verbalized sampling significantly enhances diversity by 1.6–2.1x over direct prompting, while maintaining or improving quality.
Variants like verbalized sampling-based CoT (Chain-of-Thought) and verbalized sampling-based Multi improve generation diversity even further.
I have shared the paper in the replies!
👉 Over to you: What other methods can be used to improve LLM diversity?
Norbert Enders retweeted
With all things you document and share,
always have 2 user groups in mind:
humans and AI
Norbert Enders retweeted
We've released a large-scale study on how people are using ChatGPT.
Consumer adoption has broadened beyond early-user groups, and lots of economic value is being created through both personal and professional use:
openai.com/index/how-people-…
Norbert Enders retweeted
Replying to @PromptLLM
Try that prompt:
“You are my Mental‑Model Dealer.
When I give you a challenge, deliver:
Core issue – 1 clear sentence.
Best‑fit mental model – name only (no lists).
Model explained in 2‑3 crisp sentences.
Direct application to my challenge – concrete next steps.
On request (“Alt?”) supply one contrasting model and its focused application.
Skip theory dumps and abstractions.
Start by asking: “What problem are you trying to solve right now?””
Norbert Enders retweeted
Called it last year.
Junior roles down 23%.
Senior roles up 14%.
Harvard study tracked 285,000 firms. Results:
Before AI: 1 senior + 3 juniors = 4 person team
After AI: 1 senior + Claude = same output
Most people don't understand @cursor_ai.
They think it will enable anyone to ship production code. They think it will replace developers.
It won't.
It will augment the best developers to make them 100x better (Lorenz curve).
What this does to the workforce is another question
Norbert Enders retweeted
we need new benchmarks, and this one looks promising
on an uncontaminated math set, GPT-5 is far ahead of previous-generation models, including o3
DeepSeek v3.1 and Opus 4.1 are a nice surprise.
the benchmark was built by 37 research mathematicians, mainly in algebra and combinatorics
Norbert Enders retweeted
I asked nano banana to take me across Middle-earth and this is what happened...
all the details on how I made this video, below 🧵👇

































