Last update: Tuesday 4/23/24
No podcast this week
A. TL;DR summary of Top 4 stories
1. AI Assistant | 2. Llama 3 | 3. Emirati | 4. Atlas
1) Meta's new AI Assistant
Meta made two related announcements this week: (1) A new version of Meta AI, Meta's chatbot, and (2) Meta AI now runs on Llama 3, a new family of open source language models. The Verge provides the following succinct descriptions of Meta AI:
- "Meta AI assistant, introduced last September, is now being integrated into the search box of Instagram, Facebook, WhatsApp, and Messenger. It’s also going to start appearing directly in the main Facebook feed. You can still chat with it in the messaging inboxes of Meta’s apps. And for the first time, it’s now accessible via a standalone website at Meta.ai."
- "The goal is for Meta AI to be 'the most intelligent AI assistant that people can freely use across the world,' CEO Mark Zuckerberg tells me. 'With Llama 3, we basically feel like we’re there.'"
- "The Meta AI assistant is the only chatbot I know of that now that now integrates real-time search results from both Bing and Google."
- "Its image generation has also been upgraded to create animations (essentially GIFs), and high-res images now generate on the fly as you type."
On one important point, The Verge may be too succinct. Although it says that people can "freely use" Meta AI, readers may not immediately grasp that this means there are no subscription fees, as with the ChatGPT, Copilot, Gemini, Claude 3, or Grok chatbots.
2) Llama 3 -- Meta's new open source small language models
Meta's new open source language models come in two sizes: small and huge. There are two small models: Llama 3 8B and Llama 3 70B, i.e., 8 billion parameters and 70 billion parameters, respectively. Then there is the huge Llama 3, not released yet, that will have trillions of parameters, just like Gemini Advanced, GPT-4, and Claude 3 Opus
Kyle Wigger's piece in TechCrunch quotes Meta's claim that "for their respective parameter counts, Llama 3 8B and Llama 3 70B — trained on two custom-built 24,000 GPU clusters — are are among the best-performing generative AI models available today."
Meta bases this claim, in part, on their model's performance on some commonly cited benchmark tests. These tests have not achieved widespread consensus as to their validity, but as Wigger says, "for better or worse, they remain one of the few standardized ways by which AI players like Meta evaluate their models." Here are a few quotes from Wigger's comparison of the benchmarks for Llama 3 8B against Mistral 7B (Mistral) and Gemma 7B that were reported in Meta's announcement of its new models in the left side of a table called "Meta Llama3 Pre-trained model performance.
- "Llama 3 8B bests other open models such as Mistral’s Mistral 7B and Google’s Gemma 7B, both of which contain 7 billion parameters, on at least nine benchmarks: MMLU, ARC, DROP, GPQA (a set of biology-, physics- and chemistry-related questions), HumanEval (a code generation test), GSM-8K (math word problems), MATH (another mathematics benchmark), AGIEval (a problem-solving test set) and BIG-Bench Hard (a commonsense reasoning evaluation)."
The editor of this blog was puzzled that Meta did not compare its models with Musk's Grok 1.0, Microsoft's Phi 2, and DataBricks' DBRX, given the recent loud claims of surprising power from Musk, Microsoft, and DataBricks about their models. Although Microsoft's Phi 2 is not an open model, the right side of Meta's table compares Llama 3 70B to Gemini Pro 1.0 and Mistral 8x22B, neither of which are open models. Indeed, here's a couple of Wigger's comments about the benchmarks on the right side of the table:
- "But Meta also makes the claim that the larger-parameter-count Llama 3 model, Llama 3 70B, is competitive with flagship generative AI models, including Gemini 1.5 Pro, the latest in Google’s Gemini series."
- "Llama 3 70B beats Gemini 1.5 Pro on MMLU, HumanEval and GSM-8K, and — while it doesn’t rival Anthropic’s most performant model, Claude 3 Opus — Llama 3 70B scores better than the second-weakest model in the Claude 3 series, Claude 3 Sonnet, on five benchmarks (MMLU, GPQA, HumanEval, GSM-8K and MATH)."
Unfortunately, whenever Wigger says "better than", a more precise assessment would add "but not by much". When he says "is competitive with" should read "behind, but pretty close". There are no slam dunks in these benchmarks.
But Meta really is making a mighty effort to slam dunk all of its competitors off the courts because it has embedded its Llama 3 open source models into all of its social media and will provide access to its enhanced media to everyone on the planet, free of charge. The high visibility of its open models will compel developers all over the planet to examine them and, if the theory of open source development has any validity, some, perhaps, many of them will identify ways to improve the models. Each round of improvement will attract the attention of more developers who will suggest more improvements ... etc, etc, etc. ... at minimal cost to Meta, an outcome that will also attract the attention of multi-billion dollar investors in ultra oil rich Arab states, as described in this week's third top story that follows.
3) Microsoft's $1.5 billion deal with United Arab Emirates
The first two paragraphs of the NY Times report conveys the essence of this deal:
- "Microsoft said on Tuesday that it would make a $1.5 billion investment in G42, an artificial intelligence giant in the United Arab Emirates, in a deal largely orchestrated by the Biden administration to box out China as Washington and Beijing battle over who will exercise technological influence in the Persian Gulf region and beyond.
Under the partnership, Microsoft will give G42 permission to sell Microsoft services that use powerful A.I. chips, which are used to train and fine-tune generative A.I. models. In return, G42, which has been under scrutiny by Washington for its ties to China, will use Microsoft’s cloud services and accede to a security arrangement negotiated in detailed conversations with the U.S. government. It places a series of protections on the A.I. products shared with G42 and includes an agreement to strip Chinese gear out of G42’s operations, among other steps."
A subsequent paragraph describes the Biden administration's motivation for supporting the deal and the constraints that it imposed on the deal:
- "The investment could help the United States push back against China’s rising influence in the gulf region. If the moves succeed, G42 will be brought into the U.S. fold and pare back its ties with China. The deal could also become a model for how U.S. firms leverage their technological leadership in A.I. to lure countries away from Chinese tech, while reaping huge financial awards"
Political and financial contexts for the preceding TL;DR ...
The Biden administration's active promotion of Microsoft's deal with G42 contrasts sharply with its blunt suppression of a Saudi fund's investment a few months ago:
- "US Compels Saudi Fund to Exit AI Chip Startup Backed by Altman", Jane Lanhee Lee, Bloomberg, 11/30/23
-- "The Biden administration has forced a Saudi Aramco venture capital firm to sell its shares in a Silicon Valley AI chip startup backed by OpenAI co-founder Sam Altman, an exit that could have broader implications for the Middle Eastern country’s growing investments in US technology."
However, the Biden administration has not raised any objections (yet) to Saudi Arabia's subsequent announcement of its intentions to invest $40 billion in AI via a partnership with Andreeson Horowitz and other financiers:
- "Saudi Arabia Plans $40 Billion Push Into Artificial Intelligence", Maureen Farrell and Rob Copeland, NY Times, 3/19/24
-- "In recent weeks, representatives of Saudi Arabia’s Public Investment Fund have discussed a potential partnership with Andreessen Horowitz, one of Silicon Valley’s top venture capital firms, and other financiers, said the people, who were not authorized to speak publicly. They cautioned that the plans could still change. The planned tech fund would make Saudi Arabia the world’s largest investor in artificial intelligence.
So there's lots of the "right" kind of money in the ultra oil rich Arab states ready to make huge AI investments, investments that Microsoft (and its competitors) will need for their efforts to attain artificial general intelligence (AGI) within the next five years. Indeed, Microsoft recently declared its support for Open AI's proposal to develop a $100 billion dollar Stargate AI supercomputer:
- "Microsoft and OpenAI Plot $100 Billion Stargate AI Supercomputer", Anissa Gardizy and Amir Efrati, The Information, 3/29/24
-- "Executives at Microsoft and OpenAI have been drawing up plans for a data center project that would contain a supercomputer with millions of specialized server chips to power OpenAI’s artificial intelligence, according to three people who have been involved in the private conversations about the proposal. The project could cost as much as $100 billion, according to a person who spoke to OpenAI CEO Sam Altman about it and a person who has viewed some of Microsoft’s initial cost estimates."
Given the current reluctance of large corporations to make substantial purchases of AI services, Microsoft will be unlikely to afford this $100 billion obligation out of projected future AI earnings. Therefore Microsoft and its competitors will need lots of financial help from the ultra oil rich Arab states:
- "Amazon, Google Quietly Tamp Down Generative AI Expectations", Aaron Holmes and Anissa Gardizy, The Information, 3/12/24
-- "Several executives, product managers and salespeople at the major cloud providers, such as Microsoft, Amazon Web Services and Google, also privately said most of their customers are being cautious or “deliberate” about increasing spending on new AI services, given the high price of running the software, its shortcomings in terms of accuracy and the difficulty of determining how much value they’ll get out of it."
4) Boston Dynamics' new Atlas robot
If you watch the Boston Dynamics 31 second video found at the New Atlas (video) link, words will fail to summarize what you see. Yes, New Atlas is all electric, whereas Old Atlas was hydraulic; but as for its extraordinary flexibility, the Wired description "an unfeasibly double-jointed contortionist" only has meaning after you watch the video.
To put it another way, if Old Atlas (video) stepped onto the dance floor of a club after midnight, his moves would make the other dancers stomp and cheer; but when new Atlas made his moves, the other dancers would freeze in disbelief ... for a few seconds ... then flee the premises by the nearest doors, windows, and ventilation ducts. There can be no doubt that New Atlas will be able to perform all kinds of physical tasks that no human could even contemplate. That's the good news.
Perhaps the better news is that, unlike OpenAI's new Figure 1 robots, New Atlas robots do not have GenAI computer models driving their behavior. But imagine if they were driven by GenAI. Then imagine yourself seated in an office area within which three or four New Atlas types were rambling around, making idle chit-chat, while blue toothing the Rolling Stones "Please allow me to introduce myself" ... Welcome to the darkest recesses of Uncanny Valley ... :-(
B. Top 4 stories in past week ...
- Other Models
"Meta’s battle with ChatGPT begins now", Alex Heath, The Verge, 4/18/24
-- Meta AI is also discussed by Mashable, TechCrunch, Engadget, Wall Street Journal, NY Times - SLM News
"Meta releases Llama 3, claims it’s among the best open models available", Kyle Wiggers, TechCrunch, 4/18/24 *** .
-- Meta's announcement of Llama 3 ➡ HERE
-- The Verge contains a less detailed discussion
-- Another Big Tech's opinion: "Elon Musk’s ‘not bad’ review thrusts Meta’s Llama 3 AI into spotlight", Michael Nuñez, VentureBeat, 4/19/24 - Microsoft
"Microsoft Makes High-Stakes Play in Tech Cold War With Emirati A.I. Deal", Paul Mozur and David E. Sanger, NY Times, 4/16/24 ***
-- This story also covered by Bloomberg, Reuters, Wall Street Journal, Fortune, Axios ... and Microsoft - Misc
"The Atlas Robot Is Dead. Long Live the Atlas Robot", Carlton Reid, Wired, 4/17/24
-- Here's a quote: "Old Atlas moved like a clunky chunky human; the latest iteration swivels and turns like a freaky crab crossed with an unfeasibly double-jointed contortionist." ... Clunky? Watch the Old Atlas video (below)
-- YouTube videos ==> Old Atlas vs. New Atlas
-- This story also covered by Mashable, Engadget, VentureBeat, TechCrunch
No comments:
Post a Comment
Your comments will be greatly appreciated ... Or just click the "Like" button above the comments section if you enjoyed this blog note.