Sunday Links: Novelty scores for science, alignment faking, and Google Veo 2

As people wind down for the holiday, there's a flurry of AI stories to keep your brains churning over the break:

• Can novelty scores on papers shift the power dynamics in scientific publishing? Given the massive volume of new scientific publications, it is often hard to determine what is really new. Something that might strike fear into the hearts of academics wanting to build up a publication record is a newly implemented algorithm from DeSci that aims to rank the novelty of a paper. The approach compares keywords and citations to determine if novel ideas are being connected. This seems like a tough thing to get right, but it does make some sense. At the moment, there is no word on this being used by conferences and journals to assess submissions (it's primarily for use post-publication - i.e., to decide what to read), but it seems like it would be only a matter of time before similar things are applied to academic submissions. False scores are likely to be a big problem, and it's interesting to think about whether the presence of the score itself won't simply change the way papers are authored (another score to maximize). A more detailed explanation of the approach can be found here.
• State-of-the-art video and image generation with Veo 2 and Imagen 3. Google announced a lot of new and updated AI models this week. The most impressive to me was Veo 2. The model produces some stunning video with very few glitches (see the samples on the blog post), but what's most impressive is that it is apparently tracking the 3D meshes of objects in the scene. This makes it possible for cinematic-like prompts like having the camera track in a particular way or focusing on a particular subject. (I've signed up for the waitlist.).
• Marc Zuckerberg asks US government to stop OpenAI's for-profit plans. Elon Musk is already fighting legal action to stop OpenAI from converting fully into a for-profit company. Meta has now also joined the fray. This sets up a damaging fight since it seems to be the case that OpenAI's latest funding ($6.6B in October) round apparently contains clauses that mean they must become a for-profit company within two years of the financing closing. It does seem that people fighting the change have a strong self-interest, and it would suit them well for a competitor to be hamstrung. Elon perhaps has further claims, given his early involvement. From a non-partisan point of view, it seems that the best competitive scenario would be to allow OpenAI to make the change. In the end, more choice likely means better outcomes for everyone else in the market other than direct competitors.
• Alignment faking in large language models. The R&D team at Anthropic continues to impress with the transparency of their results. At its core, the research shows that under certain circumstances, an AI model could give answers that obfuscate values that have been trained into it in earlier stages. Values it has been told to preserve and which answering the questions may lead to erasure of the values (e.g. more training). The examples are extreme, and the result is not that surprising. (A human would act a similar way if they were in a situation where "non-compliant" answers would lead to literal brainwashing.). It again highlights the need for transparency in the training setups of models so we at least have some ideas of what has been deliberately trained in.
• UK arts and media reject plan to let AI firms use copyrighted material. The UK's Labour government put forward a proposal to UK publishers, artists, and musicians that would enable third parties to train AI on their content. This was rejected this week by the parties involved. The rejection is very logical since artists feel (and are) increasingly under threat from AI tools. On the other hand, the arguments in the case bring the challenge of the situation into clear focus. Here is a quote from one of the government ministers: "If we were to adopt a too tight a regime based on proactive, explicit permission, the danger is that international developers would continue to train their models using UK content accessed overseas, but may not be able to deploy them in the UK … this could significantly disadvantage sectors across our economy, including the creative industries, and sweep the rug from underneath British AI developers.”. In other words – if usage is not allowed in the UK, the UK may lose access to models. In reality I doubt this will happen for models that matter from large players since they will want uniform ability to deploy AI worldwide. What might happen, though, is that UK culture is under-represented.

Lastly, if you're running out of time to buy gifts, you can try ringing up a lab in Tokyo that's reporting that they've managed to observe negative time in the quantum experiments. They could make a killing if they could bottle this for sale...

Wishing you happy holidays!