• AI Time to Impact
  • Posts
  • . . AI: Zuckerberg previews LLAMA 4: Aiming for Agents With Enhanced Adaptability and Broadened Scope (4.22.24)

. . AI: Zuckerberg previews LLAMA 4: Aiming for Agents With Enhanced Adaptability and Broadened Scope (4.22.24)

Author

Marshall Kirkpatrick
April 22, 2024

Friends, lots of people are betting on semi-autonomous AI agents being the next big wave in AI. Mark Zuckerberg said this weekend that will be an important part of the next version of Llama, Llama 4.  

Relatedly, I've got Llama 3 7B running locally on my Mac using LM Studio and that's fun.  I'm going to use it for some "cognitive reappraisal" (emotional reframing) experiments, something a new study in today's news says GPT-4 is already better than humans at.  See also Pete Skomoroch's very intriguing thoughts about what a big milestone Llama 3 is in the journey toward AI improvement of the human experience. 

Here's the top AI news from this weekend and Monday.

-Marshall Kirkpatrick, Editor

Top Stories

Zuckerberg: Llama 4 Will Have Agent-Like Behaviors

First impacted: AI developers, AI users
Time to impact: Medium

Mark Zuckerberg told Dwarkesh Patel that just as Llama 3's tool use is advanced because of brittle, hand-coded experiments with tool use in Llama 2, Meta is now experimenting with agent capabilities in Llama 3 that they anticipate to make full-blown major components of Llama 4. [Mark Zuckerberg - Llama 3, Open Sourcing $10b Models, & Caesar Augustus] Explore more of our coverage of: Meta AI Development, LLAMA Models, Application-Specific Logic. Share this story by email

Giant FineWeb Dataset Launched by Hugging Face

First impacted: AI researchers, data scientists
Time to impact: Short

Hugging Face has published FineWeb, a dataset containing over 15 trillion tokens from CommonCrawl dumps since 2013, which they say surpasses the performance of RefinedWeb and other leading web datasets in benchmark tests. The dataset, available under the ODC-By 1.0 license, is processed using the datatrove library and includes code for replication, along with small models for dataset validation and comparison. [HuggingFaceFW/fineweb] Explore more of our coverage of: AI Datasets, FineWeb Launch, Benchmark Performance. Share this story by email

More Headlines

[via @JonathanRoss321] Explore more of our coverage of: Groq, Developer Engagement, Nvidia Competition. Share this story by email

That’s it! More AI news tomorrow!