AI: Weekly Summary
...week ending August 11, 2023
Fourth edition of the Saturday weekly AI summary here on ‘AI: Reset to Zero’. August is off to a strong start with AI events. Hope you have your favorite beverage in hand.
White House announces new China Tech Curbs: As expected, the administration moved forward to curb US VC and private equity investments in China across semiconductors, quantum computing and AI for now. Won’t be implemented until 2024 after a review process. There is an effort to balance the overall trade relationship vs containing China related to ‘national security issues. But the impact on foreign investor flows into China are already happening, with China buyout deal flows contracting from $47 billion in 2021 to $2.4b in 2022 and $2.8b year to date. Companies like Alibaba are already warning of constraints ahead of the moves. As I’ve outlined for some time, this broad issue has big implications ahead for the US and global growth in the months to come. Tech is the canary in the coal mine.
Nvidia’s new GPUs: Founder/CEO Jensen Huang officially kicked off Nvidia’s next generation GH200 GPU ‘Superchips’, while the current A100 and H100 chips are already in superhot demand and on global allocation. (Chinese companies have already accumulated over $5 billion in orders for the current chips ahead of Biden chip curbs). Nvidia also announced iterations of other Enterprise focused AI infrastructure services around its hardware and software compute architecture. Nvidia continues to be the biggest provider of GPU chips for the current global AI gold rush going into 2024 and beyond. AI without GPUs are just two letters in the alphabet.
Meta out-executes TikTok on monetizing short videos: A surprising thing that came out in Meta’s quarterly results call last week was Mark Zuckerberg outlining that Meta’s short video Reels monetization with advertisers is on a “revenue run rate that jumped to $10 billion, up from $3 billion last fall, and $1 billion last summer.” That means that Reels is about the size of TikTok’s business as of 2022 end, and poised to potentially surpass it this year. This despite the general industry reality that Reels as an AI algorithm drive video product is still not as good as TikTok. Watch the 13 to 16 minute marks in this YouTube video discussion by the tech reviewers at the Verge for more context. This underlines Meta’s unparalleled ability to focus, execute, and leverage its multi-billion global user base ahead of their competitors. Think Threads vs. Elon’s X-Twitter. With or without the cage match. Much more to come.
Zoom’s AI ‘Terms of Service' missteps: Zoom, which has become a verb for online meetings, had to do an embarrassing about-face after changing their “Terms of Service’ in March allowing them “broad rights to use images, sound and other content from meetings to train its AI algorithms.” Other big tech companies have already been re-doing their terms of service on the AI front, as per this Google announcement not too long ago, and this Techcrunch piece on Microsoft and OpenAI efforts as well on this front. The Zoom episode highlights the need for every company large and small to consider user and regulator fears around AI before making sudden moves.
Anthropic announced Claude Instant to supplement Claude 2: Anthropic continues to build on its well-received Claude 2 initiative vs OpenAI’s GPT 4 and other Foundation LLM AI, and Meta’s recent Llama 2 launch, with an entry level Claude Instant 1.2, incorporating ‘faster, cheaper text-generating model’ available via API. Also emphasizes more safety guardrails and less hallucinating. Separately, Founder/CEO Dario Amodei did a long video podcast (with transcript), on the state of Foundation LLM AI models going into next year. Most notable especially for his candor on how even the best minds building and implementing this AI tech can’t explain how it really does what it does, and thus how to plan and guide next steps.
In Other AI developments and Papers of note,
OpenAI introduced GPTBot, a web crawler ‘designed to scrape data from the entire internet automatically’, complete with the ability for sites to ‘opt-out’ to OpenAI software via text robot files. An area of increasing sensitivity for many sites on the internet.
Great post with images comparing the latest image diffusion AI models from Stability SDXL1.0 and Midjourney 5.2 head to head. Very illustrative of the current state of the art diffusion image generation.
Terrific YouTube 8 minute video compressing hundreds of hours of training AI ‘virtual robot’ to walk using reinforcement learning AI. Really brings home how LLM AI training works in action.
Finally, in notable write-ups my end this week,
Did a piece on why ‘Threading the needle’ with China is going to be ‘insanely hard’ math for Apple and other US companies.
A look at the downside of too much ‘Synthetic Data’ on next gen Foundation LLM AIs and ‘Model Collapse’.
And the uncertainties ahead as we teach better GPUs and Foundation LLM AI to do AI ‘reasoning’,
My take on why Video and Voice are the next big things to build better AI.
Thanks for joining this Saturday. Hope you enjoyed this so far with your beverage of choice.
For your Sunday tomorrow, this week’s ‘AI In my View’ is coming up. Stay tuned.
(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here).