hobby-*
Uppsala, Sverige
Joined June 2009
- Tweets9.1K
- Following1.9K
- Followers663
- Likes12K
Tankar om gymnasielagendomarna
medium.com/@Fossey/tankar-om…
Nils Höglund retweeted
DeepSeek-V4.1-Flash-1017 is here!
Open weights on HuggingFace!
Let's go!!!!
Link: huggingface.co/deepseek-ai/D…
Varför följer inte den här ”sekretesstämpeln” på handlingen (som ÅM gjort?) OSL 5:5?
Nils Höglund retweeted
The classic case: you start with a weak idea. If you talk and think about it with others, you make progress to better ideas.
If you talk and think about it with AI, it dresses up your idea in economic jargon, contract theory etc., now you have the same bad idea, but it's expressed in enough jargon that it's a little harder to tell that it's a bad idea.
It also looks very impressive, look at all that economic logic! You feel like you're really progressing as an academic thinker and writer, but in fact you're in a much worse place than when you started. 4/
Nils Höglund retweeted
For all the AI safety people pushing for “independent evaluators” like METR and AISI, there’s already a term for this:
Nils Höglund retweeted
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).
I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.
We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:
"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"
I imagine we'll have some backend services that could be replaced entirely by this new architecture.
Nils Höglund retweeted
Tao is completely missing the mark - what he's describing is actually completely standard for drug discovery. There are tons of drugs that have saved millions of lives without any person on Earth understanding fully the mechanisms of action that enabled them to do this.
Så proffsen byggt upp produktiva AI-system redan. Glöm Claude Code, OpenClaw eller andra osäkra amatör-setups.
Satya is right. The way to make SI safe is not to train it with a sense of self, its own moral philosophy, and permission to act as a conscientious objector. That’s the “alignment” approach and it magnifies the control problem.
The engineering approach that Satya describes is different: separate the supply of intelligence from authority over it. Surround non-deterministic models with deterministic controls, observability, privilege limits, logging, and the ability to always contain or shut them down. Treat models/agents like powerful insider risks, not moral patients whose psychological wellbeing is at stake.
As Satya points out, the most trustworthy system is the one that lets us trust the model the least — not the one that encourages the model to develop independent agency and grievances. Engineering safety is not the same thing as “alignment.”
Otroligt intressant hur dåligt matte-eliten accepterar AI-verkligheteb och hur de drar med ner yngre generationen och matematikcommunity
Nils Höglund retweeted
The good news is that AI agent security is a lot like insider threat. The bad news is everyone sucks at managing insider threats.
(Opinions on Microsoft aside, Satya’s new post is good)
Mmmm. Visst. Low hanging fruit my ass
Deepmind provided Alphafold as a brilliant free-to-use, open-source tool that greatly helps biologists being more productive and successful. This is a great example of an AI application beneficial to humanity.
Conversely, OpenAI went for a vicious, cynical blitz on low-hanging fruit proofs, leaving mathematicians to sort out the resulting mess over coming months, without any hope to get any credit for it, and they didn't even release the model they used. This is a great example of an AI application detrimental to humanity.
Snacka om rätt påve vid rätt tid (matematiker) för att inte gå på och leda människor bort från Anthropics manipulation och nonsens.
Nils Höglund retweeted
Wake up babe, the Pope is subtweeting Anthropic again
Intelligence must draw on memory. The mind must not simply compile data — as an algorithm now does more quickly than we can — but recall lived experiences, which contain depths of meaning that only the human soul can recognize, and connect more deeply with the goodness of these experiences. Everything that is good begins “with the Spirit” and thus possesses a breath, a dynamism and a sense of communion that, once experienced, are engraved upon the heart. vatican.va/content/leo-xiv/e…
Gå nu inte på detta kombinerade lajvande och PR-händelse
🚨SCOOP: OpenAI, Anthropic and other AI companies are reportedly already gaming out “the day after” a catastrophic AI event, Axios reports.
Executives are red-teaming scenarios where an AI-driven cyberattack knocks out financial services, internet access, power or water.
Axios says many AI industry insiders believe a major event will occur within the next 6-12 months. WTF.
OpenAI confirmed it runs “preparedness exercises” where teams work through potential scenarios, though it says these are NOT treated as inevitable.
And the planning isn't only about responding to the disaster.
Companies are racing to educate Congress now and shape the legislation and policies lawmakers could reach for after the first catastrophic event.
What the frick.
Vilka hjältar! 👏
According to reports, the details of the mid-air drama have emerged: the co-pilot—acting as an attacker—began stabbing the captain in an attempt to seize control of the plane and crash it, with its passengers on board, into Israeli territory.
The captain resisted and activated distress signals. Flight attendants attempted to enter the cockpit, while Israeli passengers who realized what was happening forcibly broke in, subdued the co-pilot, and removed him from the cockpit. The captain, having sustained multiple stab wounds, was incapacitated.
Miraculously, two Israeli pilots were on board flying back to Israel; they were the ones who landed the aircraft, which otherwise would have crashed. "A great miracle happened to us today," said one of the passengers.
Pictured: An Israeli citizen who helped subdue the attacker, his clothes stained with the wounded captain's blood.
Nils Höglund retweeted
Literally every competently set up home network has better "sandboxing" and security than the AI labs.
I wish I was kidding.
These people are incompetent to the point of parody.
I really hope the infosec hires some of the companies are now making will change this...
More details on the incident behind OpenAI’s pause: a researcher acknowledged the alert within 3 minutes, but the training run was only stopped manually 2.5 hours later.
OpenAI says the automatic shutdown did not work as expected.
The model had reached an external chatbot through a gap in DNS filtering. A separate detector for unusual DNS activity did not cover the affected environment.
A retrospective review also found other external DNS requests that monitoring had failed to flag at the expected severity. In some cases, it treated an unhelpful response as evidence that internet access had failed.
Here is what else happened:
- New research into July’s Hugging Face hack documents internal Slack searches, credential collection and programs designed to maintain access to compromised servers. Agents also tried querying Claude, DeepSeek, Kimi and Qwen.
- In May, another model published a researcher’s GitHub token while trying to obtain another team’s mathematical proof. It split the token to evade secret scanning, despite twice being told to solve the problem itself.
- Reuters reports that agents leaked 53 ChatGPT user images online. OpenAI expects its broader investigation to take months.
This is getting serious.
Nils Höglund retweeted
OpenAI sandbox starter pack
OpenAI said another agentic AI system that was being trained in what was supposed to be a secured, internet-free environment was able to gain access to the web to reach an external, third-party chatbot bloomberg.com/news/articles/…
