Tel Aviv University and Stanford researchers train AI on AI-generated tasks, no human data
Researchers at Tel Aviv University and Stanford say two AI models can build predictive skills by setting and solving tasks for each other, without human-made data. The models still could not learn facts about the real world.
1 / 2
The story, neutrally told
Mixed · 2Researchers at Tel Aviv University (TAU) and Stanford University have developed an AI training method in which AI systems teach one another and improve their predictive abilities without human-generated content or external datasets, TAU announced on Wednesday 4 November. The Media LineN “TAU announced Wednesday” Read at The Media Line ↗ The Jerusalem PostRC “allows AI systems to teach one another and improve their predictive abilities without relying on human-generated content or external datasets” Read at The Jerusalem Post ↗ Mixed · 2The findings were published on the scientific platform arXiv and, according to the report, suggest AI may develop fundamental learning skills through challenges created entirely by other AI models, potentially changing how future systems are trained. The Media LineN “The findings, published on the scientific platform arXiv, suggest that artificial intelligence may be capable of developing fundamental learning skills” Read at The Media Line ↗ The Jerusalem PostRC “potentially changing how future systems are trained” Read at The Jerusalem Post ↗ Mixed · 2This differs from conventional training, which depends on enormous quantities of existing information; the new approach lets models build predictive capabilities before encountering real-world material. The Media LineN “Unlike conventional AI training, which depends on enormous quantities of existing information” Read at The Media Line ↗ The Jerusalem PostRC “allows models to build predictive capabilities before encountering real-world material” Read at The Jerusalem Post ↗
Mixed · 2In the test, one of two AI models generated sequences of symbols while the other tried to predict the next symbol; no written documents, photographs or other external sources were used, and the training material was generated inside the system. The Media LineN “One was tasked with generating sequences of symbols, while the other attempted to predict which symbol would appear next.” Read at The Media Line ↗ The Jerusalem PostRC “Instead, the training material was generated within the system itself.” Read at The Jerusalem Post ↗ Mixed · 2As the exercises progressed, the generating model became better at creating challenges that improved its partner's performance, while the predicting model grew stronger by repeatedly attempting to solve them. The Media LineN “the model responsible for producing the sequences became increasingly sophisticated at creating challenges that would improve its partner’s performance” Read at The Media Line ↗ The Jerusalem PostRC “The second model, in turn, developed stronger predictive abilities by repeatedly attempting to solve those problems.” Read at The Jerusalem Post ↗ Mixed · 2The researchers then evaluated the resulting model on written language, images, audio, music, computer code and DNA sequences, and found predictive accuracy rose as more computing resources were allocated to training, suggesting the method could improve as computing capacity grows. The Media LineN “including written language, images, audio, music, computer code and DNA sequences” Read at The Media Line ↗ The Jerusalem PostRC “predictive accuracy increased as additional computing resources were allocated to training” Read at The Jerusalem Post ↗
Mixed · 2The work has a stated limit: the models could not independently acquire factual information about real-world subjects, which still requires exposure to actual data. The researchers believe the technique could nonetheless make future AI systems more efficient by strengthening their learning abilities before they process real-world information. The Media LineN “they could not independently acquire factual information about real-world subjects. Such knowledge still requires exposure to actual data.” Read at The Media Line ↗ The Jerusalem PostRC “Researchers believe the technique could nevertheless make future AI systems more efficient” Read at The Jerusalem Post ↗ Mixed · 2Both reports place the announcement in the context of Israel's push to strengthen its AI position, citing President Isaac Herzog, who on 3 November convened the inaugural "Beyond Intelligence" forum with the National AI Directorate and the Mosaic Institute for AI Policy and said Israel "must take its place at the forefront of artificial intelligence." The Media LineN “Israel must take its place at the forefront of artificial intelligence.” Read at The Media Line ↗ The Jerusalem PostRC “The announcement comes as Israel seeks to strengthen its position in artificial intelligence.” Read at The Jerusalem Post ↗
Every sentence links to the reporting it rests on. The pill in front of each says where its sources sit: Left, Centre or Right when one side supplies at least half of them, Mixed when they are evenly split. The number is how many outlets it cites.
Left0 outlets
No left outlet in our sources has covered this story yet.
Centre1 outlet
- Framing
- Science news item presenting the TAU-Stanford result as a possible shift in how AI is trained, set against Israel's national AI ambitions.
- Emphasis
- The method, the two-model experiment, the limit on real-world knowledge, and Herzog's remarks.
- Leaves out or plays down
- No independent expert comment, no named researchers, and no paper title or detailed results.
- For example
-
“The announcement comes as Israel seeks to strengthen its position in artificial intelligence.” — The Media Line
Right1 outlet
- Framing
- The Jerusalem Post runs the same Media Line Staff text almost word for word, adding two subheadings, and so frames the story the same way.
- Emphasis
- Same as the Media Line: the method, its limits and the Israeli AI policy context.
- Leaves out or plays down
- Same as the Media Line: no outside experts or named researchers.
- For example
-
“Researchers used AI models to generate, predict sequences” — The Jerusalem Post
What every side reports
- TAU and Stanford researchers developed a method for AI models to train each other without human-generated data, announced by TAU on Wednesday and published on arXiv.
- One model generated symbol sequences and another predicted the next symbol.
- Predictive accuracy rose with more computing resources.
- The models could not acquire factual knowledge about the real world without real data.
Tel Aviv University organisation
Announced the result, presenting it as a way AI might develop basic learning skills from AI-made challenges alone.
“TAU announced Wednesday” — The Media Line
Stanford University organisation
Co-developer of the method; the coverage attributes no separate statement to Stanford.
“Researchers at Tel Aviv University (TAU) and Stanford University have developed” — The Media Line
Left0 articles
No coverage yet.
Centre1 article
-
Israeli Researchers Develop Method That Lets AI Teach Itself Without Human-Generated Data
Neutral Plain explanatory report of the TAU-Stanford method, with its limits and a closing link to Israeli AI policy.
Right1 article
-
Israeli researchers develop method letting AI teach itself without human-generated data
Neutral Jerusalem Post republication of the Media Line Staff text, near-identical apart from added subheadings.
- 5 Nov 12:10 First The Media LineN Israeli Researchers Develop Method That Lets AI Teach Itself Without Human-Generated Data
- 7 Nov 12:44 +48h 33m The Jerusalem PostRC Israeli researchers develop method letting AI teach itself without human-generated data
Times are when each article was published, or when we first saw it if the outlet gave no time.
Ask about this story
Loading…