At Rise of AI 2024, Hans Uszkoreit argues that the real bottleneck in today's foundation models is not algorithms, architectures or compute, but how the models are trained, or in his words, 'educated'. He observes that European language models are not on top of leaderboards despite Europe building smaller, industry-focused systems, and traces this to gaps in training data rather than weaker methods, noting that all major jumps since 2022, from GPT to Llama, use versions of the same Transformer architecture and differ mainly in the data that went into them. Drawing on his own testing, he shows that pretraining data collected opportunistically from sources like Common Crawl covers some domains, such as medicine and law, far better than others, including most industrial and manufacturing knowledge, and that this data gap explains phenomena such as hallucination and inconsistency across languages. He argues models trained this way perform at roughly bachelor's-textbook level in specialized industrial domains, and calls for a new transdiscipline, merging education science, information science, AI and educational psychology, to identify what knowledge is missing, source data for it, and systematically teach and test models the way people are educated.
Key takeaways
- Hans Uszkoreit argues the bottleneck in foundation models is training data and teaching methods, not new algorithms or architectures.
- Every major model leap since 2022, from GPT to Llama, has used variants of the same Transformer architecture; only the training data changed.
- He attributes hallucination to insufficient redundant training data on a topic, not to models becoming more error-prone as they improve.
- GPT hallucinated that his former DFKI boss Wolfgang Wahlster and a musician friend had died, because both had stopped publishing recently.
- Uszkoreit calls for a new transdiscipline of 'machine teaching' merging education science, information science, AI and educational psychology.
Detailed notes
Why European models trail American and Chinese ones
Uszkoreit opens by comparing American, Chinese and European language models, noting European systems are not doing badly but are not at the top of leaderboards either. Referencing a talk earlier that day by Peter Sarlin of Silo AI, he agrees Europe's opportunity lies less in mass-market consumer AI and more in adapting generative AI closely and quickly to industry needs, business-to-business rather than business-to-consumer. He names models from Aleph Alpha, Silo AI and Mistral as steps in that direction, but argues Europe is 'on the right way but not there yet'.
The bottleneck is training data, not algorithms
Uszkoreit states plainly that every major leap since 2022, across GPT, Google's systems and Llama 2 and 3, has used variations of the Transformer architecture rather than new machine learning methods; the difference has been the data. He notes that GPT-4-class models are trained on the equivalent of perhaps only a few tens of thousands of books' worth of curated text out of a corpus equivalent to 100 million books, with the rest opportunistically collected web content of uneven quality. From his own testing at that company he found large models perform impressively in medicine and law but perform much worse once applied to specific industrial knowledge domains, because that data is simply underrepresented in common web crawls such as Common Crawl or arXiv.
Hallucination as a symptom of insufficient training data
Uszkoreit argues hallucination is not a sign models are becoming more 'discrete' but a statistical symptom of insufficient redundant, secure knowledge on a topic. He illustrates this with two personal anecdotes: GPT told his former DFKI boss Wolfgang Wahlster that he had died, because Wahlster had retired and stopped publishing, and it separately told Uszkoreit that a singer friend in New York had died, because she had no new albums or performances during the pandemic; in both cases the model filled a biography in the statistically most probable, but wrong, way. He argues that as machine education improves, models will remain probabilistic but hallucinate far less.
From general knowledge to industrial specialization
Uszkoreit maps AI applications from simpler, already-solved tasks like classification and robot movement, through text-based tasks like manuals and specifications, to the hardest category, complex problem solving such as automotive product and production planning, which he says takes four and a half to five years in Germany versus two and a half years in China because so many specialized disciplines must coordinate. He argues large language models could eventually help here, but only if trained specifically in automotive or other industrial specializations, and predicts that tasks currently done by narrow, specialized AI systems, such as demand forecasting, will eventually be handled by fine-tuned large models because those models become cheap enough and can absorb context, such as seasonal or crisis effects, that narrow time-series models cannot.
Toward a discipline of machine teaching
Contrasting classical machine learning, where a problem and dataset are given and an algorithm is found, with what actually changed between GPT versions, purely the training data and teaching process, Uszkoreit calls for a new discipline of 'machine teaching': identifying what knowledge is needed, sourcing and weighting the data, designing a training curriculum, and testing throughout, much like human education. He points to OpenAI's own reported practice of targeting specific weak spots on aptitude tests with additional training material between GPT-3.5 and GPT-4, and to Bloomberg's domain-specific financial language model, as early examples. He closes by describing his own effort at a company he says had to close, to assemble vertical automotive and health datasets as a step toward specialized models, and argues that today's AI researchers mostly lack the domain expertise needed to find, evaluate and teach industrial knowledge that is not available on the open web.
Quotes
The bottleneck is not better algorithms, is not more computer science, not even faster machines, the bottleneck is right now the training, the education of the models.
Hans Uszkoreit
Questions answered in this talk
Why do European language models trail American and Chinese ones?
According to Hans Uszkoreit at Rise of AI 2024, European models such as those from Aleph Alpha, Silo AI and Mistral are not on top of the leaderboards, and he argues Europe's realistic opportunity is business-to-business specialization for industry rather than competing directly on mass-market consumer AI.
What causes large language models to hallucinate?
Uszkoreit argues hallucination stems from insufficient redundant training data on a given topic; the less statistically secure the knowledge, the more likely the model fills gaps with a plausible but incorrect answer, as when GPT incorrectly reported that people who had simply stopped publishing recently had died.
Why do foundation models perform worse on specialized industrial knowledge than on medicine or law?
According to Uszkoreit, opportunistically collected pretraining sources like Common Crawl cover medicine and law relatively well but barely cover manufacturing and other industrial domains, so models trained on that data perform at only bachelor's-textbook level in those specialized areas.