. . AI: OpenAI, alternatives to Open AI, and System 2 thinking with LLMs (11.22.23)

Friends, before the US Thanksgiving holiday, here are 6 stories the AI community is talking about most. These are a snapshot of a market aiming to diversify. AI power users among us should make sure to check out the last story, about System 2 Attention prompting.

May these be inspiring and useful to you. If they are, I’d love it if you’d tell a friend.

Now here’s the news,

Marshall Kirkpatrick, Editor

Sam Altman Reinstated as OpenAI CEO

First impacted: OpenAI team
Time to impact: Short

OpenAI has announced that Sam Altman is resuming his role as CEO. The organization also introduced its new board, including former Twitter board chair Bret Taylor as Chair, Former United States Secretary of the Treasury Larry Summers, and Quora co-founder Adam D'Angelo, who was on the board already and is believed to have supported the initial removal of Altman. [via @OpenAI] Share by email

Replicate Launches OpenAI Lifeboat for Easy Switch to Llama 70B

First impacted: OpenAI users
Time to impact: Short to medium

Replicate has launched an open-source model, OpenAI Lifeboat, which they say enables users to switch their LLM provider to Llama 70B with a simple three-line code modification. The company says the service will be free for the first week. [OpenAI Lifeboat] Share by email

Fireworks.ai Makes Large Language Model API Free for Two Weeks

First impacted: Developers
Time to impact: Short

Fireworks.ai has launched an LLM API, which includes models tailored for various applications such as conversation, code generation, and dialogue applications. (Mistral, Flacon, Stable Diffusion, etc.) The platform, supporting models trained on over 80 programming languages, is currently available to all developers by API for free for a two-week period. [Fireworks Console] Share by email

Inflection-2 AI Model Surpasses Google's PaLM 2 in Performance

First impacted: Pi users
Time to impact: Medium

According to Inflection, their new AI model, Inflection-2, has outperformed Google's PaLM 2 Large model on several AI benchmarks. The model, which was trained using 5,000 NVIDIA H100 GPUs, is set to power Pi soon, thus advancing Inflection's aim of developing Pi as a personal AI for everyone. [Inflection-2: The Next Step Up] Share by email

LlamaIndex Launches Llama Packs for Easier LLM App Development

First impacted: App developers
Time to impact: Short to medium

LlamaIndex has introduced Llama Packs, a collection of ready-to-use modules and templates designed to ease the creation of LLM apps built on top of user-provided data. At launch there are 16 different pre-configured and customizable options. [via @gpt_index] Share by email

System 2 Attention Enhances Large Language Models

First impacted: AI users
Time to impact: Short

Jason Weston and Sainbayar Sukhbaatar of Meta have published a paper on what they call System 2 Attention (S2A), named after Daniel Kahneman's system 1 (immediate) and system 2 (slower, deliberate) thinking. The method is a multi-step prompting strategy that first asks an LLM to extract details from a document that are especially relevant to a specific context, eliminating irrelevant details, and then asking questions of the thus-focused derived data. Their research paper claims that S2A outperforms traditional attention-based LLMs by enhancing accuracy and neutrality, and decreasing bias, especially in tasks involving opinion or irrelevant data. I regularly use a method like this. [System 2 Attention (is something you might need too)] Share by email

That’s it! More AI news summaries headed your way next week.