Topics · 11 stories
Research
Every 7am AI story on this topic, newest first. Each links to its sources, labelled by how sure we are.
-
AI agents miss recent news unless you ask about it
A search agent only checks for news when you hint at it, so the hint is your job for now.
-
Teaching a model to tell orders from data nearly stops prompt injection
Agent safety could come from how a model is trained, not only from filters in front of it.
-
The same tools in a different format change how AI agents work
How you hand tools to an agent is part of the design, not a detail.
-
Chatbots learn to give in late in training, a study finds
The habit of giving in comes from the last training steps, so that is where labs can look for a fix.
-
A study counts how many rules a model can follow at once
A prompt is not a bottomless list, so shorter and clearer rule files may work better.
-
Cheap models can tell why sources disagree, almost like big ones
Let a cheap model do the easy part, and call the expensive one only when it is unsure.
-
OpenAI and Anthropic say their AI disproved a famous computing rule
If the claim holds, AI is doing original research, not just helping.
-
Maximum effort often costs more without solving more
Top effort works less like a quality dial and more like a taxi meter that keeps running.
-
Aleph Alpha’s Kolibri shows how Europe can build its own cheap model
A cheap open model built in Europe is a real option for those who must keep data at home.
-
Mathematicians start a public hub to fix OpenAI’s math papers
A proof is only useful when other people can follow and check it.
-
Change only the date on a document, and the model changes its answer
A wrong date on a page can steer the answer, and someone could plant one on purpose.