Another week, another AI conference in San Francisco. This week, I was at the Databricks AI+Data Summit. The Databricks ecosystem is evolving rapidly, and the company released its own agent builder framework (of course), but also a new and extended version of ML Flow (ML Flow 3.0), which has been updated for GenAI model development and management. Interestingly, a lot of the notable changes centered on how to evaluate and monitor GenAI workloads. Evals are the bottleneck.
Onward with this week's links:
- Here’s the $2,000 fully AI-generated ad that aired during the NBA Finals. You still need ideas, but with tools like Veo3, realizing them is getting incredibly easy. The scene with the eggs is genius. (Also, try filming that in real life...).
- How we built our multi-agent research system. Interesting blog post from the Claude team on how they created research functionality based on sets of four Claude agents. It's an interesting read, though I'd contend this is more akin to multi-threading than it is multi-agents. I'm also not sure you get much multi-agent benefit until the tasks are really long-running. Probably the most interesting thing in the post is the galaxy plot of what people are using the research system for, which I include here:
- Design Patterns for Securing LLM Agents against Prompt Injections. This is a pretty comprehensive review of approaches to stopping malicious prompt attacks on LLMs. It's being fairly widely circulated (Simon Willison has some great notes here.). The paper does include a number of useful patterns for controlling agent behaviour, but there is a fundamental conflict between the power of an AI Agent (its scope to act) and the risk of harm. There is a second axis of conflict that pairs off how smart & flexible an agent is versus its vulnerability to being tricked. The main conclusion (in my view) is that it's an extremely bad idea to put full LLM capabilities in any Agent that has any significant operational powers.
- Salesforce blocks AI rivals from using Slack data (Reuters, The Information [Paywalled]). Yes, they did, and in my opinion, this is a suicide move for Slack. The change prevents third-party AI tools, such as Glean, from accessing and searching Slack message data. The new rules severely rate limit non-marketplace apps, and marketplace apps are already prohibited from data export. No doubt, Salesforce's thinking behind this is that they would like their customers' data to be held primarily within Salesforce so their own AI solutions will be preferred. However, Slack customers are now prevented from extracting their own data for AI training or usage. Something which will likely be increasingly untenable. Certainly, it's likely to mean our team leaving Slack for a solution that permits us to use our own message data however we want.
- Shopify kills UX design (Fast Company has the story and title, but the main link is to the announcement from Shopify's Chief Design Officer). The announcement states that the words UX and Content are being dropped from all job titles at the company. The argument being that all design functions have been compressed, and it is now all "product design" and "writing". There were very mixed reactions to this, but I think it's likely going to become a trend. AI makes UI/UX very fluid and configurable, and it's not clear that apps in the future will even ship with specific, rigid interfaces. They are more likely to have on-the-fly generation of interfaces based on user needs and context.
Other news that didn't make the cut this week: there's also a public argument between Anthropic founder Dario Amodei and almost everyone on AI safety and impact, Meta's quasi-acquisition of Scale.AI for $ 14.3 B, and Claude stops blogging.
Wishing you a great weekend.
