AI posts you missed this week - July 5th

Here's the latest bookmarks from a week where posts increasingly defy category. Websim is spilling across every domain, and prompt hacks blur into indie R&D exploring LLM's conceptual boundaries. Equally, it's fun seeing different disciplines and directions of exploration bleed in to each other, and a sign of acceleration. Let's go!


Keyframes make it easier to direct AI video generation. This is the cleanest example and looks absolutely up to standard for a kid's TV show:

Lots of music exploration in Websim this week:

More music, this time in StableAudio:

3D game generation keeps popping up in my feed, often in Websim. This one I think is purely Claude Sonnet 3.5, with generative textures. I assume our only limit to immersive live gaming is compute:


Moshi beats OpenAI to the real time voice game:

Then of course Pliny enters the chat. Turns out Moshi has a real potty mouth:

Janus pushing the limits on Claude. I've always been curious about the abstractions a model works through and this may give us at least a transposition of how that model sits:

Anthropic sharing maps of concepts and abstractions:

Supermaven getting high praise from coders and just bumped up to a 1 million token context window:

DeepSeek-Coder-V2 finds the sweet spot on performance and cost:

Self-organising AI neurons. The results claim a model can learn from experiences, building up randomly connected nodes from an empty network:


Fix that messy download folder with Gemma 2:

The llmsys leaderboard, formerly an authority on model performance, keeps being questioned for its reliability and whether the chart is being gamed. Here's an alternative. My takeaway: time to test Phi-3 on my MacBook:

Not a tweet but this fantastic tool, Jina Reader, converts any webpage to LLM-friendly markdown. Just enter[] . Thanks Mike Taylor for the tip:

Reader API
Read URLs or search the web, get better grounding for LLMs.

Nice approach for exploring possible paths in prompting:


YouTubers have found websim. This is the tipping point for spilling further into the mainstream, as video producers in this space inevitably copy each other for new content to squeeze:

Explore the planet in 3D:

Turns out integrating HTMX (which I understand takes the real time nature of React back to an in-line HTML approach) gives Websim mad potential:


Roko observes that our theoretical concerns about AI have already played out in a different context: industrial civilisation racing ahead of our ability to align it:

Seb with the four horsemen of the AI-pocalypse:

There is a fine line between alignment and lobotomy:

And with that lobotomy comes attack vectors:

And finally, Biden's Brian suggests the Presidential debate has set a new bar that next gen AI can comfortably clear:

Enjoy your week, and ping @TomDavenport on twitter if you find something worth adding next week.

Subscribe to Tom Davenport

Don’t miss out on the latest issues. Sign up now to get access to the library of members-only issues.