AI Roundup 063: Llama 3
April 19, 2024.
Llama 3
Less than 24 hours ago, Meta dropped Llama 3, the latest and greatest in its lineup of open large language models.
What's new:
The 70B parameter version outscores Gemini Pro and Claude Sonnet on various benchmarks, making it a strong contender in the "GPT-3.5" weight class.
Meta is also training a 400B parameter version of Llama 3, which is a gargantuan-sized model and, based on early data, could be the first open GPT-4 class model.
Llama 3 is meant to be better at multi-step tasks and refuse fewer requests, but new trust and safety tools are also available, such as Llama Guard 2.
Elsewhere in model mayhem:
Stability AI is making Stable Diffusion 3 available to developers via API, though the core model remains in a developer preview.
Microsoft researchers published VASA-1, a model that combines still photos with audio to create incredibly realistic lip-sync videos.
And xAI previews Grok-1.5 Vision, its first multimodal model, with plans to release it "soon."
Meta AI
Meta also announced "Meta AI," a family of smart assistants that are now available in Instagram, Facebook, WhatsApp, and Messenger that are powered by Llama 3.
Why it matters:
Meta AI now puts the company in direct competition with OpenAI and Anthropic. And unlike both companies, it already has distribution channels with billions of users.
The new assistants also let Meta incorporate its many other AI advances - for example, instant image generation is coming to WhatsApp users via Meta AI.
And beyond a consumer chatbot, Instagram is giving its top influencers a way to clone themselves with AI, via its new Creator AI program.
Elsewhere in the FAANG free-for-all:
Google announces a new Platforms and Devices team (to integrate AI across its physical products), and Perspective API, a set of ML tools that help moderators identify toxicity with scores across several attributes.
Amazon Music follows Spotify with an AI playlist generator of its own, Maestro.
And Snap plans to add watermarks to images created with its AI-powered tools.
Safety third
Six months after the UK's inaugural AI Safety Summit, South Korea has set a date for the event's second iteration, May 21-22.
Between the lines:
Same story, different verse: as companies train and release increasingly powerful models, governments and regulators debate safety measures.
But there is still some movement: both the UK AI Safety Institute and US AI Safety Institute are taking steps to push more safety, even without formal legislative frameworks.
Meanwhile, in Silicon Valley, at least some tech leaders are challenging the narrative that AI needs to be regulated.
Elsewhere in AI anxiety:
New investigations find hundreds of YouTube videos and Google ads promoting AI deepfake porn tools.
In an analysis of over 200M student papers from the past year, 11% showed signs of AI, and 3% were almost entirely AI-generated.
And the SAG-AFTRA union secures AI protections for artists in a deal with major record labels.
Last week's roundup
Things happen
US Air Force confirms first successful AI dogfight. Brave Search introduces Answer with AI. Google will provide AI to the military for disaster response. The Linux Foundation launches the Open Platform for Enterprise AI. Stanford's 2024 AI Index Report. Adobe explores OpenAI partnership as it adds AI video tools and Acrobat AI. Generative AI is coming for healthcare, and not everyone’s thrilled. The rise of the chief AI officer. Gamers are renting their idle GPUs to generate AI porn. Why you can't build Apple with venture capital. Huawei unveils its first AI PC. A profile of Mistral CEO Arther Mensch. Institute that pioneered AI "existential risk" shuts down. The Anthropic Cookbook. OpenAI is pitching ChatGPT Enterprise to large firms. Embeddings are a good starting point for the AI-curious developer. Cyc: History's forgotten AI project. The Limitless Pendant is an AI wearable for your life. AI made these movies sharper; critics say it ruined them. "Asking Big Tech to police AI is like turning to ‘oil companies to solve climate change." Stability lays off 10% of its staff. How GitHub Copilot came to dominate AI coding tools. Which companies will be steamrolled by OpenAI? ChatGPT is coming to Nothing’s earbuds. AI has a measurement problem. Investors are growing increasingly weary of AI. We can, and we must, clown on the Humane AI Pin forever.