A test on Meta AI by a Techcrunch tester established that Meta AI’s answers were the furthest thing from comprehensive. He noted the following;
Meta AI provided a concise summary of current events but only cited one source (looks like it’s citing SEO promoted articles). He said Copilot provided a similar answer but with better sources and links.
It accurately gave recent trends when asked about recent trends in TikTok that a parent should be aware of but only summarized SEO instagram trend on Hootsuite (chunk or bias summary?) when asked the same question.
It couldn’t provide primary sources for a research topic.
It answered some trivia questions accurately such as controversial thing that happened in 1984 olympics.
Its advice around medication and anxiety was predictably straightforward and safe.
Meta’s Chief Product Officer Chris Cox said Llama 3 now has 2.5 million downloads already on Hugging Face (min 1).
He said the quality of the data is the most important thing while testing AI and added that their greatest advantage is their public data eg that in Instagram (min 5)
Apparently there are reports that AI models are being used largely by Chinese startup ecosystems which could lead to national security issues and more regulation for open source models (min 6).
Cox said he has mixed feelings for TikTok ban and pointed out that he hasn’t seen evidence on TikTok being used to surveil. He added that we should be careful not to make decisions for consumers (min 10).
He said Threads is especially popular in Japan and Korea (min 13).
Meta has rolled out new generative AI tools for advertisers such as enhanced image generation and text variations.
Instead of being able to only create different backgrounds for a product image, advertisers can now request full range variations, including rifts that update the subject of the product.
Meta said reception of image generation features by advertisers has been positive.
In its I/O event, Google touted its AI advancements. Among them is the improvements to its Gemini 1.5 Pro. Google said it’s increasing the amount of tokens to 2 million from 1 million, meaning it will be able to summarize thousands of pages of text and an hour long video (min 8:00). Google also launched Gemini 1.5 Flash, which they say is designed to be faster and efficient (min 21:00). Here are some of the AI products that stood out to me:
Google showcased Project Astra (AI assistant prototype) which can talk to users about anything captured on smartphone in almost real time. For instance, with your phone camera you can ask it why the handle of your music player is not holding out and it will generate search results categorized under AI-generated headlines (min 26:00 and min 52:00).
Gemini powered Summarize enables a user to generate a summary of his/her emails. For instance, you can ask it to summarize emails from your kid’s school. It can also summarize emails exchanged between you and your clients, generate a spreadsheet and give a summary of where money is spent. It also has the ability to provide contextualized replies after analyzing your email conversations (min 57:00).
Google gave examples of how its AI Overviews can help users. It said AI Overviews has the ability synthesize and answer more complex queries for which a simple answer may not exist in the web (min 38:00).
Veo, a video generation model can generate 1080p-resolution videos that last longer than a minute also looks interesting (min 30:00).
Imagen 3, an image generator based on texts was also showcased. Google says the model promises lifelike images that have fewer distracting visual artifacts compared to their prior models (min 30:00).
The company also touted its Audio Overviews which can generate audio discussions based on inputs. For example, it can generate an audio discussion of a lesson plan and connect the dots to give you a real life example that may not exist in the text.
It also touted its Gemini Planning tool which can provide trip and meal plans among other things (min 46:00).
Google is also testing AI Teammate which has the ability to search through Google Docs, Email and other Workspace apps and provide answers to a query such as are we ready for a launch? if the person asking is a founder-to-be (min 12:00) .
In general, it appears that the company is bringing AI to its entire product portfolio. It says that in future Gemini will be able to replace Google Assistant.
CEO Mark Zuckerberg creates the Meta Advisory Group, a team that will periodically advise the management on AI products and technology advancements, Bloomberg reported citing an internal memo send to employees.
The group will have four executives namely Stripe CEO and co-founder Patrick Collison, former GitHub CEO Nat Friedman, Shopify Inc. CEO Tobi Lütke and Charlie Songhurst, former Microsoft executive.
A Meta spokesperson said the team will not be paid, will not be elected by shareholders and don’t have a fiduciary responsibility to the company.
Meta’s chief AI scientist Yann LeCun told financial times that the large language models that power AI products such as ChatGPT would never achieve human-like intelligence.
“It certainly appears to most people as reasoning — but mostly it’s exploiting accumulated knowledge from lots of training data,” LeCun said. “[LLMs] are very useful despite their limitations.”
He said he’s working on an entirely new generation of AI models that he hopes would achieve human-level intelligence, although that could take 10 years.
“We are at the point where we think we are on the cusp of maybe the next generation AI systems,” LeCun said.
LeCun leads the Fair Lab team which was created in 2013 and which is focused on long-term AI goals unlike the GenAI team which is responsible for the current generative AI products.
Meta Platforms and Alphabet in talks with Hollywood Studios about licensing content for use in their AI video generation software, Bloomberg reported citing people familiar with the matter.
The two companies have offered tens of millions of dollars to partner with the studios.
Meta and Alphabet declined to comment on the talks.
Meta Platforms announced at its Conversations event in Sao Paulo, Brazil that it is bringing new AI features to WhatsApp, including an AI assistant that enables businesses to quickly serve customers by providing answers.
The company is also bringing an AI ad-creation assistant to WhatsApp Business.
Meta said the new features will support more than 200 million businesses that rely on WhatsApp.
Meta said it’s building a new AI technology for Europeans that reflect their linguistic, geographic and cultural references.
Meta will use information that people have chosen to share publicly to train its AI.
The company added that it will not use private messages shared with friends and family nor will it use contents from accounts of Europeans under the age of 18.
Meta pointed out that activists’ positions don’t reflect the European law.
I=5
Meta Platforms will delay the launch of its Meta AI for European users following a request from the Irish Data Protection Commission (DPC) to postpone training of its large language models (LLMs) using public content.
“Without including local information we’d only be able to offer people a second-rate experience. This means we aren’t able to launch Meta AI in Europe at the moment,” Meta Platforms wrote in a blog post.
I=5
Meta’s Fundamental AI Research team releases five new models for research including:
Chameleon which can generate images and texts using either texts or images as inputs unlike other large language models which mainly turn text into image only.
Multi-Token Prediction models that predict the next word in a sentence i.e code completion models.
JASCO accepts various inputs such as chords or beat to generate music unlike existing models such as MusicGen that rely only on text inputs.
AudioSeal helps detect AI-generated speech at a faster and more efficient way than traditional methods.
Meta and Apple have held talks about integrating Meta AI into Apple Intelligence, a recently launched AI system for iPhones and other devices, the Wall Street Journal reported citing people familiar with the matter.
The two companies don’t plan on paying each other but Meta could sell premium subscriptions to their services via Apple Intelligence while Apple gets a commission.
According to the sources, the talks haven’t been finalized.
Bloomberg is now reporting that Apple rejected Meta Platform’s proposal to integrate Meta AI in iPhones months ago because it doesn’t consider the company’s privacy practices as stringent enough.
I=4 Zuckerberg criticizes other AI players for trying to create an “AI God”
CEO Mark Zuckerberg blasted other tech companies for trying to create a “one true AI” (min 20:00)
He said short form video are growing really quickly and that he foresees a shift from feed-based interactions to direct messaging (min 4:41)
He pointed out that demand for Ray-ban glasses has been much higher than expected and that production has been lower than demand (min 22:50).
Assessment
Zuckerberg’s criticism of other AI players could confirm Bloomberg’s report that Meta’s talks with Apple to integrate Meta AI into Apple Intelligence had failed.
Meta releases its biggest version of Llama 3 with 405 billion parameters and ability to write high-quality code and solve complex math problems
Meta Platforms has released the biggest version of its Llama 3 model which according to the company, can converse in eight languages, write high-quality computer codes and solve complex math problems.
The model has 405 billion parameters, lower-than that of GPT-4 (one trillion parameters).
Meta said the model scored 73.8 in a math test, compared to GPT-4’s 76.6 and Claude 3.5 Sonnet’s 71.1.
In MMLU test, it scored 88.6 versus GPT-4’s 88.7 and Claude 3.5 Sonnet’s 88.3, Meta pointed out.
Meta rolls out AI studios, a tool that enables people ( even those with no tech skills) to create custom AI characters.
Instagram creators could, for instance, create AI characters that are “extension of themselves” and could handle common DM questions and story replies.
AI studios will initially be rolled out in the U.S.
