Sunday Links: Made by Humans, Going Local, and Zite

Made by humans initiative by Adobe, AI functionality gets compacted and OpenAI wants to buy Windsurf.

Sunday Links: Made by Humans,  Going Local, and Zite

After a slight news slowdown over Easter, we are back this week with another burst of stories:

  • Did Adobe Fresco's new 'made by a human' tag just make it a safe space for artists? This week, Adobe announced a feature in its Fresco (which is for digital painting and drawing) that allows a creator to state it was made by a human rather than AI and exclude it from AI training data. The system also inserts a digital watermark into the image that Adobe will track. The impulse makes sense, but it seems to me that this really won't work that well if the tracing is not an open standard, and without a real way to prove something has no AI content anywhere in it.
  • OpenAI is reportedly in talks to buy Windsurf for $3B, with news expected later this week. Windsurf is one of the top AI coding products that combines different LLMs, enabling developers to write code in a wide variety of ways. The fact that OpenAI may buy them likely reflects that the general consensus is that OpenAI has fallen behind Claude for coding use cases. Windsurf would allow for a tighter feedback loop with developers. The deal is not closed yet, but it's interesting that OpenAI continues to try to hit "all" use-cases. It is the AI company with the biggest consumer opportunity, so at some point I wonder if they will primarily focus on that.
  • Two undergrads built an AI speech model to rival NotebookLM. Nari Labs created a text-to-speech voice system that is extremely high quality, given it was made by a team of two. I wouldn't say it so much rivals NotebookLM (which has a range of workflows around it), but what it does tell you is that in the shift from procedural software to learning software, cycle times are massively decreased. A small team with the right insights can now produce extremely advanced AI functionality in a way that was just not possible with procedural code.
  • Harper Is an Offline Alternative to Grammarly for Obsidian. In a similar vein of complex functionality being compacted, this post on Lifehacker caught my eye. I use Obsidian for local notes that sync across all my devices (I'm an Evernote refugee). Harper is an open-source local grammar checker that works with Obsidian and other software. What's interesting is that it could be a broad replacement for Grammarly, which is web-based. Grammarly's architecture means all your writing roundtrips to the cloud. It enabled them to teach their models, but it was also a must, given the complexity of the task. Now that we're in the era of compacting this functionality, perhaps we'll soon be able to cut out all the cloud roundtrips.
  • Zite.com. A lot of excitement has accrued around AI-SaaS builders such as Loveable and Replit. These allow you to build hosted apps from the ground up in a flash, and have seen spectacular growth. Just as interesting, though, are AI builders that are designed to work in the enterprise (or in the SMB) that focus on connecting your existing data sources - see the video on Zite's homepage. This market might end up more fragmented, but the fact that these apps connect to your own internal data means it's much more relevant to established companies than something like Loveable may be. To be fair, Replit does also serve this need, but has not been as focused on it.

Wishing you a great weekend!