GenAI Weekly — Edition 1

Google makes Gemma 2B and Gemma 7B model weights open

Jeanine Banks and Tris Warkentin writing for the Google blog:

Jeanine Banks Tris Warkentin

Gemma is a family of lightweight, state-of-the-art open models built from the same research and technology used to create the Gemini models. Developed by Google DeepMind and other teams across Google, Gemma is inspired by Gemini, and the name reflects the Latin gemma, meaning “precious stone.” Accompanying our model weights, we’re also releasing tools to support developer innovation, foster collaboration, and guide responsible use of Gemma models.

It’s impressive that Gemma 3B beats Llama 7B on several benchmarks. Important to remember these are Google’s first “open” (only weights were made available) models. Of the big tech companies, now both Meta and Google have made models available in the open.

Shoot a video of your bookshelf and get a JSON list of books on it with Gemini Pro 1.5

From Simon Willison’s blog:

Simon Willison

Last week Google introduced Gemini Pro 1.5, an enormous upgrade to their Gemini series of AI models.

Imagine how apps will change with this capability. From personal library management to calorie tracking to fitness, etc.

Stability announces Stable Diffusion 3

From the Stability blog:

Stability AI

Announcing Stable Diffusion 3 in early preview, our most capable text-to-image model with greatly improved performance in multi-subject prompts, image quality, and spelling abilities.?

Why governing AI is like herding cats: Google’s troubles

Adi Robertson writing for the Verge:

Adi Robertson

Google has apologized for what it describes as “inaccuracies in some historical image generation depictions” with its Gemini AI tool, saying its attempts at creating a “wide range” of results missed the mark. The statement follows criticism that it depicted specific white figures (like the US Founding Fathers) or groups like Nazi-era German soldiers as people of color, possibly as an overcorrection to long-standing racial bias problems in AI.

If your application is open to all queries, it’s impossible to test for all scenarios. Natural language interfaces are hard.

Could switching from GPUs to LPUs make ChatGPT 13x faster?

Maxwell Zeff writing for Gizmodo:

Maxwell Zeff

Groq creates AI chips called Language Processing Units (LPUs), which claim to be faster than Nvidia’s Graphics Processing Units (GPUs). Nvidia’s GPUs are generally seen as the industry standard for running AI models, but early results show that LPUs might blow them out of the water.

Try it out. It is a sight to behold. It is so insanely fast, it makes ChatGPT or Gemini feel like their output is coming via dot matrix printers.

Representation Engineering Mistral-7B an Acid Trip

From the Vgel.me blog:

In October 2023, a group of authors from the Center for AI Safety, among others, published Representation Engineering: A Top-Down Approach to AI Transparency. That paper looks at a few methods of doing what they call "Representation Engineering": calculating a "control vector" that can be read from or added to model activations during inference to interpret or control the model's behavior, without prompt engineering or finetuning.1 (There was also some similar work published in May 2023 on steering GPT-2-XL.)

Metas new LLM-based test generator is a sneak peek to the future of development

From the Engineer’s Codex newsletter:

Meta recently released a paper called “Automated Unit Test Improvement using Large Language Models at Meta”. It’s a good look at how Big Tech is using AI internally to make development faster and software less buggy. For example, Google is using AI to speed up code reviews.

For the extra curious

• Microsoft Is Spying on Users of Its AI Tools
• Things I Don't Know About AI
• Meta: Beyond A*: Better Planning with Transformers via Search Dynamics Bootstrapping
• GPT in 500 lines of SQL