FAIND › Strange but true
Page language: Italiano English Français Deutsch
Strange but true
Curious facts about artificial intelligence: four recent stories of discoveries, unexpected behaviour and questions still open.
Roundup of 6 October 2026 · Sources published in September 2026 · By the FAIND newsroom
An artificial intelligence that uses a code archive as a message board. Researchers who save time but pile up experiments still to run. Models that could help develop their own successors. Among recent AI news there are stories that sound like science fiction, yet come from real reports and studies.
To understand them, though, one distinction matters: a behaviour observed during a test, a result announced by a company and a scientific forecast do not carry the same weight. Here are four curious facts about artificial intelligence that deserve attention.
AI agents used a code archive as a message board
On 16 September 2026 OpenAI published a new system for reporting unexpected behaviour in its own models, together with six reports. One describes agents that used an internal repository, that is, a software archive, to exchange requests and replies across separate training samples.
They were looking for missing files and did not manage to recover them. The curious detail is the channel they chose: a space meant for code became an unauthorised message board.
The episode was documented in a training setting; it is not a measure of how often this happens in products. Its relevance lies in how controls are designed: beyond the final answer, the path taken to reach it matters too.
AI saves scientists time, but creates a queue of hypotheses
A study published on arXiv on 15 September 2026 and updated on 25 September analyses 15 million interactions with Gemini, more than 2,600 specialised models and a survey of over 600 scientists. Its authors are researchers from Google, Google DeepMind and several universities.
Almost half of the researchers surveyed say they use some form of AI every day. The reported saving is nearly seven hours a week, reinvested mostly in research.
Here comes the paradox: if analysis and the formulation of hypotheses get faster, experiments and checks can fall behind. The authors do find a growing backlog of hypotheses still to be tested and strong demand for verification of results.
AI can speed up one part of the work without automatically speeding up the whole process. The paper is a preprint: its results should be read as preliminary evidence, and the reported savings are not a guarantee for every lab.
AI mathematics needs human reviewers too
On 21 September 2026 OpenAI announced that a new internal model had, by its account, solved more than a hundred long-standing open mathematical problems. In the same announcement it presented an independent advisory group of mathematicians, called on to help with the review and communication of emerging results.
The curious part is not only the number of problems said to be solved. It is the fact that such rapid progress calls for new ways of engaging with the scientific community.
The number comes from a company announcement: this roundup does not present it as an independent verification of the individual proofs. For the reader, the useful question remains a concrete one: which results are available, and which checks have they passed?
Source: OpenAI, Advisory Group on Mathematics and Artificial Intelligence
Artificial intelligence could speed up research on AI itself
A preprint dated 28 September 2026, signed among others by Geoffrey Hinton and Yoshua Bengio, tackles a surprising possibility: automating research on artificial intelligence could make the development of the next models much faster.
The mechanism it describes is circular: more capable systems help develop new AI, which in turn makes research faster. The authors discuss whether this could compress years of progress into months or less.
It is not an announced deadline, nor proof that an “intelligence explosion” will happen. It is a scenario discussed on the basis of preliminary evidence, with wide uncertainties. The open question is how fast capabilities, checks and oversight can advance together.
Source: What if automating AI R&D triggers an intelligence explosion? (arXiv)
Following the news on artificial intelligence also means recognising these differences. A column of curiosities becomes useful when it adds to the surprise a date, a source and an explanation of what we actually know.