Last update: Sunday 2/25/24
Click link for podcast (opens in new tab)
Click the "start" button the when the podcast page is loaded
... If audio fails to start, or gets stuck, try reloading that page
... If audio fails to start, or gets stuck, try reloading that page
A. TL;DR ... Top 3 stories in past week ... TL;DR
1) Gemma: Google's new open source model
This blog has repeatedly noted that Google has been striving mightily to catch up to the Microsoft/OpenAI partnership ever since the partners released their GPT-4 large language model in March 2023. No one seriously doubts that Google is capable of closing this gap. Nevertheless, the impact of Google's gains have been repeatedly undermined by the fact that the partner have posed a moving target.
- For example, Google announced Gemini Nano, a small language model, in late December 2023. Then Microsoft announced a far more powerful small model, Phi-2, (2.7 billion parameters) in January 2024. Microsoft's announcement included the results of benchmark tests that showed that Phi-2 was far more powerful than Google's Nano and other small models, like Llama2 and Mistral. Microsoft's announcement also declared that it would make Phi-2 available in the Azure AI Studio model catalog to foster research and development on language models.
- Indeed, Microsoft determined that small models could have such surprising power that CEO Nadella subsequently announced that he had established a crash project that will deploy Microsoft's best AI experts to develop small language models that will be more cost-effective than the large language models that Microsoft previously paid OpenAI billions of dollars to develop. This team will report directly to Kevin Scott, Microsoft's Chief Technology Officer
Last week, Google announced Gemma, a new family of small language models. According to Google's blog:
"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 ... We’re releasing model weights in two sizes: Gemma 2B and Gemma 7B ... Gemma models share technical and infrastructure components with Gemini, our largest and most capable AI model widely available today. This enables Gemma 2B and 7B to achieve best-in-class performance for their sizes compared to other open models."
Was Gemma Google's response to Microsoft's efforts to develop small but powerful models that could replace OpenAI's expensive GPT family of large models? Not really.
- We assume that Gemini Nano is more powerful than Gemma. The benchmarks cited by Microsoft in its announcement of Phi-2 show that Phi-2 is more powerful than Nano by a wide margin; so we infer that Phi-2 is also far more powerful than Gemma.
- Indeed, the benchmarks cited by Google in its announcement of Gemma showed that Gemma was only marginally more powerful than Llama and Mistral.
So why did Google announce such a weak small language model? The editor of this blog is inclined to believe that the announcement reflects Dr. Hassabis' life-long commitment to the scientific method, the same scientific method he employed so successfully in DeepMind's development of AlphaGo, AlphaFold, and other breakthroughs. (See links to articles describing a few of DeepMind's achievements in section D of this note, below, before it was merged with "Google Brain" to become "Google DeepMind".)
- "A.I. Could Solve Some of Humanity's Hardest Problems. It Already Has.". Guest = Demis Hassabis, The Ezra Klein Show (podcast with transcript), 7/11/23
In traditional science, even marginal achievements are shared with the other members of one's research community. But this would not justify Google's full blast, sales-pitch assertion that the Gemma models can "achieve best-in-class performance for their sizes" ... No, they can't. They place a very distant second to Phi-2.
2) Intel chips designed by Microsoft
The opening paragraph of the article in The Verge provides a succinct summary of this story:
- "Microsoft and Intel strike a custom chip deal that could be worth billions. With the Microsoft deal, Intel notches a big partnership as it seeks to regain its former position at the top of chip manufacturing."
Indeed, the article goes on to say Intel expects $15 billion from this order. Now let's provide some context for this partnership:
- Older readers will remember a time when Intel made most of the chips in computers smaller than mainframes and designed all of the chips that it produced. However, in recent years Intel has ceded its leadership position to other producers.
- Nowadays the leading chip makers, called "foundries", produce chips that were designed by their customers. Last year's global surge of interest in AI propelled the Taiwan Semiconductor Manufacturing Company (TSMC) to the top of the chip heap because TSMC's fabrication facilities ("fabs") manufacture the chips designed by Nvidia.
- The large language models that everyone wants to develop nowadays must be run on graphics processing units (GPUs) that enable multiple instructions to be executed simultaneously. Nvidia makes the most powerful GPUs, so the demand for Nvidia's chips is running far higher than the limited supply that TSMC can provide. This gap has driven the price of Nvidia's chips and the market value of Nvidia's shares so high that Nvidia is now the third most valuable company in the world, behind Apple and Microsoft.
To escape this price inflation, Microsoft is now joining Amazon, Google, and Apple, all or whom have been designing their own chips in recent years. But Microsoft will not only design its own chips; it will not have them produced by TSMC, a company whose production lines are already operating at maximum capacity, but by Intel, an underutilized producer that desperately wants to resume its position at the top of the chips heap.
3) Gemini's "woke" hallucinations
The headline from Business Insider captures the essence of this story:
"Google suspends Gemini from making AI images of people after a backlash complaining it was 'woke'"
Yes, when asked to produce images of the Founding Fathers, Gemini produced pictures that included Black men. It seems likely that Google has instructed Gemini to produce images that are racially diverse whenever possible. Unfortunately, it also seems that Gemini continued to produce "diverse" groups, even when corrected to do otherwise. Of course the anti-"woke" crowd cried out in anger and anguish. But Google quickly suspended Gemini's capacities to produce any images of people until this defect is corrected.
If AI's generation of images was anywhere near as good as its generation of text or code, Google's temporary suspension and quick resumption would end this story. But as many of us know, AI generation of images by any chatbot is still in early phases of development, even for ChatGPT plus DALL-E, a flawed combo that's been accessible to the public for almost a year.
As hard as it tries to play catch-up to Microsoft/OpenAI, Google will probably be playing this kind of whack-a-mole game for the foreseeable future, with unexpected glitches popping up here and there and all over the place. However, no one should doubt that, given its determination, its depth of AI expertise, and its vast finnncial resources, that Google will, indeed, catch up, sooner or later.
So the real question is: catch up to what? Whenever Google stumbles, we should immediately ask: How well has Microsoft/OpenAi addressed this issue? With regards to AI generated images, the honest answer today must be: not very well, not yet. To be sure, the animators and other visual creatives are correct in their perception that AI generated images represent an imminent threat to their livelihoods. But the threat is still imminent, not a current reality ... yet.
The editor of this blog decided to test his pessimism by asking ChatGPT Plus to generate a group image. Here's his initial prompt, followed by two correction prompts, then the final images:
- Original prompt:
"Draw a picture of five white American males in their mid-twenties, sitting around a table, with big smiles on their faces. They are casually dressed, wearing jeans, sandals, and colorful sports shirts. There is a large birthday cake with lighted candles in the center of the table because they are celebrating the birthday of the youngest member of the group" - Correction prompt #1
"Please correct the image to show five white males wearing sandals and colorful shirts" - Correction prompt #2
"One more correction. They all should be seated at the table; no one should be standing"
- All of the men in both pictures are white ... so ChatGPT is not "woke" ... or is it?
- There are eight men in the top picture, seven in the bottom ... despite reminders that there should only be five men.
- Two men in the bottom picture are wearing sneakers/trainers, instead of sandals
- Three men are standing in both pictures, whereas the prompts requested that all be seated
- And the top picture is definitely "WOKE!!!" because it shows diversity in the men's ages. Two of the men are at least 40 -- middle aged men with faces that are "youthful", but not young, with white hair and white beards -- whereas they should all be young men in their mid-twenties
So, ChatGPT and DALL-E did better than Gemini, but they did not meet the prompts' specifications.
B. Top 3 stories in past week ...
- SLM News
"Google goes “open AI”" with Gemma, a free, open-weights chatbot family", Benj Edwards, Ars Technica, 2/21/24 ***
-- This story also covered by Reuters, Fast Company, Hugging Face, Fortune, NY Times ... and Google - Microsoft
"Microsoft and Intel strike a custom chip deal that could be worth billions", Wes Davis, The Verge, 2/23/24 ***
-- This story also covered by Reuters, Bloomberg, Wall Street Journal, Engadget, - Google
"Google suspends Gemini from making AI images of people after a backlash complaining it was 'woke'", Joshua Zitser, Business Insider, 2/22/24 ***
-- This story also covered by The Verge, Gizmodo, Wall Street Journal, Fast Company, TechCrunch, Associated Press, NY Times ... and Google (on X/Twitter)
This page contains links to responses by Google's Bard chatbot running Gemini Pro to 12 questions that should be asked more frequently, but aren't. As consequence, too many readily understood AI terms have become meaningless buzzwords in the media.
D. A few of DeepMind's scientific achievements using machine learning
Note: Publication dates usually lag far behind the end dates of research
Note: Publication dates usually lag far behind the end dates of research
- "One of the Biggest Problems in Biology Has Finally Been Solved", Tanya Lewis, Scientific American, 10/31/22 ... protein foldimg problem ... DeepMind
-- This story also covered by VentureBeat, CNET, Science, - "Courtesy of AI: Weather forecasts for the hour, the week and the century", Devin Coldewey, TechCrunch, 11/14/23 ... DeepMind
-- This story also covered by Gizmodo, Financial Times, MIT Tech Review, Ars Technica, Wired, ZDNet,
-- These articles describe research published in Nature and Science - "AI Tool Pinpoints Genetic Mutations That Cause Disease", Even Callaway & Nature Magazine, Scientific American, 11/21/23 ... DeepMind
-- This story also covered by Wall Street Journal, Wired, - "Google’s DeepMind finds 2.2M crystal structures in materials science win", Michael Peel, Ars Technica, 11/29/23 ... "The trove of theoretically stable but experimentally unrealized combinations identified using an AI tool known as GNoME is more than 45 times larger than the number of such substances unearthed in the history of science, according to a paper published in Nature on Wednesday."
-- This story also covered by Wired, VentureBeat, Nature, Science, MIT Tech Review,
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.