🦙 Llama 3

🧵 In today's edition:

• 🦙 Llama 3

• 📈 Stanford’s State of AI

• ⚔️ Palantir’s pivot

• 🛠️ AI Tools to Check Out

• 🤑 AI Fundraising News

• 🐦 Tweet Post of the Day

🦙 Llama 3

Yesterday, Zuck (with his drippy chain) unveiled Llama 3, the next generation of Meta’s state-of-the-art open-source large language model. 

Here’s the tea: 

• Llama 3 features pre-trained and instruction-fine-tuned models with 8B and 70B parameters, delivering groundbreaking performance across various tasks and benchmarks.

• According to Meta, Llama 3 establishes new state-of-the-art results, surpassing competing models like Claude, Mistral, and GPT-3.5. 

• Key innovations include an improved tokenizer, scaling up pretraining on a massive 15T token dataset, and advancements in instruction fine-tuning.

Reflecting on its commitment to responsible AI development, Meta has also introduced new trust and safety tools alongside Llama 3. These include: 

• Llama Guard 2 for content filtering, 

• CyberSec Eval 2 to assess misuse risks, and 

• Code Shield to block insecure code suggestions.

Plus, Meta has also integrated Llama 3 into their flagship AI assistant, now claiming it to be the world's leading system for boosting productivity and creativity.

Overall: In the coming months, Meta plans to release even larger Llama 3 models with new capabilities like multimodality, multilingualism, and longer context windows.

And because all of this is open-source, folks will be able to get their hands on a GPT-4 (or maybe GPT-5 👀 ) level model absolutely for free.

Zuck is putting pressure on OpenAI, and the ball is now in Sam’s court.

About to be an interesting couple of months in this AI race…

Read more: Meta AI Blog

📈 Stanford’s State of AI

Stanford recently released its 2023 AI Index report, which provides an interesting peak into the rapidly evolving AI landscape over the past year. 

Here’s the TL;DR: 

• A major trend has been the rise of open-source foundation models. In 2023, 66% of the 149 new foundation models released were open-source, up from just 33% in 2021.

• However, closed-source commercial models from big tech companies still outperform their open-source counterparts, with a median 24% performance advantage across benchmarks. 

• Google led the way by releasing the most new foundation models in 2023, followed by OpenAI.

• The exorbitant costs of training these massive AI models make it increasingly difficult for academia and government to compete. Google's Gemini Ultra cost an estimated $191 million to train, up from around $900 for the original Transformer model in 2017.

• The United States continues to dominate globally in AI development and investment, originating 61 of 2023's new models and attracting $67.2 billion in private AI investment, 8.7x more than China. 

• GenAI, like ChatGPT, has fueled a surge - attracting $25.2 billion in 2023 investment.

Overall: AI is rapidly achieving human parity on many benchmarks, leading researchers to develop new, more challenging tests. Surveys show that over half of businesses now use AI, while over a third of workers fear eventually being replaced by AI.

Read more: Stanford HAI

⚔️ Palantir’s pivot

Palantir, the software company known for its work with intelligence agencies, is now pitching its AI platform, AIP, to advertising agencies, showcasing how it can be used for tasks like inventory planning, programmatic ad sales, and campaign optimization.

In a pitch deck obtained by Marketing Brew, Palantir presents AIP's "wide-ranging use cases" for the advertising industry while claiming major brands like Lowe's, General Mills, and OpenAI are already using AIP technology.

Despite criticism over its government contracts, including with ICE, Palantir is aggressively expanding its commercial business. AIP is now expected to be a key part of this growth, allowing Palantir to move beyond its intelligence roots.

Executives who attended the pitches said that Palantir positioned AIP as a potential unique selling point for agencies compared to large ad tech companies. However, Palantir's history with surveillance programs was a significant concern during these meetings.

Overall: Some in the advertising industry are worried about the reputational risks of working with a company like Palantir due to its strong connections to government surveillance.

As an ad tech exec pointed out, the industry is trying to distance itself from "commercial surveillance," which means that a partnership with Palantir could pose problems.

Read more: Marketing Brew

🛠️ AI Tools to Check Out

• Vidyo.ai: Make short clips from long videos. Get social-ready clips cut & captioned by AI. Check it out!

• Synthesia: Create studio-quality videos with AI avatars and voiceovers in 130+ languages. It’s as easy as making a slide deck. Check it out!

• Hireguide: Hireguide unites AI and hiring science to help teams create structured interviews, automate interview notetaking, and gain insights for improving your hiring process. The result? Better quality hires, faster. Check it out!

_{Have anything you’d like to share with over 40k AI enthusiasts? Sponsor Not A Bot}

🤑 AI Fundraising News

• ClaimScore raises $3.15M in Seed funding to offer tested and validated AI and machine-learning-based anti-fraud software solutions for class action settlements.

• Cognivia raises €15.5M/$16.5M in funding to combine quantification of patient psychology with AL/ML to improve measurement of therapeutic efficacy in clinical trials and beyond.

🐦 Tweet Post of the Day

Zuck’s thoughts on open-source AI 👇️