Saturday Links: Singing Pictures, Trashy CNET and AI Firewall CDNs

One day late again this week: apologies. Here are the highlight links of the week, though:

• “Building An Answer Engine” Aravind Srinivas interview on Invest Like the Best. Really good interview with the founder of Perplexity (which I'm a fan of). Aravind gives a really good state of the union on where we are with LLM investments, costs, and the app business. TL;DR - there likely won't be that many "massive" models, but it's also not clear you want to "pack the whole Internet" into an AI; maybe you just want something that is smart and assimilates information on demand.
• Cloudflare announces Firewall for AI. It's logical that the biggest CDN and firewall providers will announce "AI protection." This will make life tough for many startups going after this space. However, it's not obvious that the edge of the computing network (in a CDN) is a very effective place for an AI firewall. CDNs are great at cutting out volume attacks and also a fixed set of inbound harmful queries, but LLM and other AI-based systems are probably best thought of as being vulnerable to social engineering-type attacks where threats are buried in the subtly of the prompt. This is extremely difficult to defend against no matter how you do it, but it's likely to need to be close to the model to be at all effective. The CDN might filter out obvious attacks but may need a fully-fledged second LLM to protect the first one.
• Wikipedia blacklists CNET and others for AI-generated content. CNET and a number of other publications (in particular those owned by Red Ventures) have been removed from Wikipedia's trusted sources list because they have amped up the number of stories with significant errors in what are obviously AI-produced articles. In a situation like this, it's going to be important to know which publications hold a core value of producing accurate news rather than just entertainment. Red Ventures may want to think about its homepage tagline, "We Help People Discover & Decide." I guess readers would like to think that those decisions would be based on accurate facts.
• EMO: Emote Portrait Alive - Generating Expressive Portrait Videos with Audio2Video. An Alibaba research team produced a really cool model that takes a single image and an audio stream to product animated facial video. There are some great speeches, text, and singing clips on the site. Look out for more video sites like Cameo where the stars can just lend their image and not have to do any actual work.
• Trump for black voters (via AI). The political image and video storm will be interesting to watch. These images produced by Donald Trump supporting groups (but not the campaign itself) show that the presidential candidate surrounded by black voters would be the voters. Contacted by the BBC for comment, the creators (and those spreading the image) stated, "If anybody's voting one way or another because of one photo they see on a Facebook page, that's a problem with that person, not with the post itself." Key actors in this will be anyone who has a significant audience, forwarding images or videos without vetting them first could be spreading major distortions. Some outlets won't care, but those that do (and are trusted by others to be neutral) will be the most damaged if things slip through.

Have a wonderful weekend!