Friday Links: Sovereign LLMs, Agents and Protein Synthesis

Here are this week's links:

• From LLM to generative AI operating System. Germany's entry into the LLM race Aleph Alpha is shifting its focus from building foundation models that compete directly with the likes of OpenAI to a more general compute layer (that can presumably be different models). Founder Jonas Andrulis quote "just having a European LLM is not sufficient as a business model.” is something quite a lot of commentators asked themselves when Alpha Alpha raised 500M Euros last year and the French company Mistral raised similar amounts. This pivot isn't entirely surprising since the compute race has indeed gotten eye-wateringly expensive. Mistral seems do be doing rather better by focusing heavily on training efficiency. Still, Aleph's change does call into question the idea that European "homegrown" LLMs will be just fine if regulation keeps Meta and OpenAI out of Europe.
• Character AI Scraps Building LLMs After Google Deal. Character.AI's deal with Google saw two founders and 20% of the team move to Google, plus gave Google a license to Character.AI's technology. The $2.7B seems very much like Microsoft's "almost acquisition" of Inflection.AI, essentially stopping the startup in its tracks and moving assets. Character.AI's service is still online and perhaps will keep going, but it would seem likely they might begin to use Gemini underneath instead of trying their own models.
• ChatGPT launches Canvas for editable outputs. One of the frustrating elements of using LLMs to craft specific outputs is that it can be hard to iterate to get a prompt just right. ChatGPT's new feature makes this easier by creating a parallel output window where users can edit results. This makes a lot of sense, and it's a near-direct clone of Anthropic's Artifacts feature. Kudos to the Anthropic team for getting there first, but this is likely to become UI table stakes for any LLM based work tool.
• AlphaProteo generates novel proteins for biology and health research. Reminding us why Generative AI is valuable is a piece of new work from Google Deepmind that makes it possible to generate and test molecules that bind to proteins. The resulting molecules may turn out to be very valuable in drug discovery and other biological processes. No doubt there will continue to be debate about whether or not such discoveries are "creative acts" but there's no doubt in my mind that humanity will find out more about the universe by using tools like AlphaProteo than if we don't.
• Salesforce's xLAM is a New Model for Agentic Tasks. There is a huge established literature on AI "Agent" (and indeed Multiagent) systems. Most of which is being ignored in today's use of the term "Agent" or "Agentic". As systems get more complex, most likely a lot of that old R&D will resurface (with or without attribution). In any case, we'll need to get used to more "task" or "action" oriented systems being call Agentic. Salesforce's new model is part of a trend to focus on models that are optimized for task completion rather than language output. In their case, this is still mostly to do with data processing tasks, but it seems likely we'll get agentic models that overlay almost every type of digital and physical process we can imagine. None of those "agents" necessarily need a powerful language interface. Most likely we'll see LLMs to interpret what we want and agent models to execute it.

Wishing you a great weekend!