- AI Weekly Wrap-Up
- Posts
- 8-2-2023
8-2-2023
Keith Collins MD & Addie Politi
August 02, 2023
Top Story
Researchers invent “unstoppable” AI jailbreak technology
Researchers at Carnegie Mellon University have discovered a method to reliably “jailbreak” all existing AI Large Language Models, causing them to produce harmful content that the models have been explicitly trained to suppress.
This is worrisome because:
It appears to work against all LLMs, so it may be an inherent weakness in the underlying design of AI systems
the method is automatable and infinitely variable, so it is extremely hard to devise a reliable defense.
Try it out in this demo - Build a bomb! Steal from a charity! (don’t really)
Clash of the Titans
Google, Microsoft, OpenAI and Anthropic form Frontier Model Forum to oversee development of advanced AI
Four of the leading AI technology companies have come together to form an industry watchdog group to monitor and oversee development of “frontier models”, i.e. AI models more powerful than the current crop, to make sure that they are safe and responsible.
This move is meant to:
Appease regulators - “See? We can regulate ourselves!”
Allay fears of the public - “We’re the good guys!”
Lock in their position as an AI oligopoly - “We’re gatekeepers.”
“People of the same trade seldom meet together, even for merriment and diversion, but the conversation ends in a conspiracy against the public..” - Adam Smith, The Wealth of Nations
Amazon is taking care of business, drawing thousands of businesses to try its cloud based AI services on AWS
Amazon has been widely derided for being slow to the AI party. (Even, ahem, in this very newsletter.)
Well, it turns out that, under the radar, Amazon has been quietly attracting thousands of businesses to try out its AI offerings on Amazon’s industry-leading cloud service, AWS. Avoiding flashy, headline-grabbing public moves, Amazon has been quietly leveraging its unmatched reach to directly engage its myriad customers in its AI framework, called Bedrock.
Bedrock acts as an API to multiple AI LLMs from Amazon, AI21 Labs, Cohere, Anthropic, and Stability AI. This allows businesses to begin building with AI without having to worry about infrastructure issues.
Smart. Very Smart. This crow we are eating is very tasty.
Intel vows to put AI into every chip
Intel CEO Pat Gelsinger recently vowed at a quarterly earnings call that the chipmaker would “build AI into every product we build.”
The mega-chipmaker is poised to release a high-end consumer chip codenamed Meteor Lake, its first product to include an onboard neural processor for machine learning. But Gelsinger insisted that AI was not just for the high end, but for every product - high, medium, or low.
“AI is going to be in every hearing aid in the future, including mine,” he predicted.
He spent an unexpected amount of time in the call bashing the cloud. Real business solutions, he predicted, will happen on the client’s hardware, “without the round trip.”
For the sake of his multi-billion dollar company, he needs to deliver on that prediction. If AI happens mostly on the cloud, Intel is an also-ran, after GPU specialists Nvidia and AMD. Gelsinger wants to leverage Intel’s strength in consumer products to keep a profitable seat at the AI table.
Meta plans to charge cloud providers for its AI technology
Zuck tells Amazon, Microsoft, and Google that its recent LLaMa 2 LLM is “open source, but not for you.” This is a bit of a back-pedal from the July 18 announcement that Meta was “teaming up” with leading cloud providers Microsoft and Amazon, to provide LLaMa 2 without charging for use. He clarified that he wouldn’t charge users, but he would charge the cloud companies. Who would then charge the users. And Zuck is not embarrassed by this backtracking flim-flammery, like, at all.
OpenAI files trademark for “GPT-5”
A straw in the wind - a portent of things to come.
Fun News
“Ideas are a Dime a Dozen”: GPT-4 chatbot generated more creative business ideas than students in a course on innovation at an elite university
Researchers at Wharton Business School and Cornell University compared student responses in a course on innovation with the ideas generated by the GPT-4 LLM. The chatbot’s ideas were rated more creative than the students’ on average, and the large majority of the highest-rated ideas were generated by the LLM. The LLM was also many times faster than the students in producing ideas.
Wayfair’s new AI home decorator tool will let you place furniture in an uploaded picture of your room, then sell you products right from the page
Online furniture giant Wayfair has produced one of the most intriguing AI apps for consumers yet. Wayfair’s Decorify website is described as “your shoppable AI interior designer.” It uses generative AI to help you restyle your uploaded room in real time, in a variety of selectable styles, such as “midcentury modern”, “farmhouse”, “bohemian”, and more. Once you have the room decorated to your taste, you can order compatible furniture from a menu. The site is a bit basic for now, but the tech-savvy entrepreneurs who founded Wayfair promise that lots more functionality is coming.
AI helps researchers track endangered pink dolphins less invasively
Until now, tracking individual members of endangered species required clipping a GPS transmitter on them, or following them visually with boats or drones. Now researchers have developed AI-enabled listening devices that can track the rare and endangered pink dolphins in the Amazon River, just from identifying their echolocation clicks and their communication whistles.
Lightning Round - short links to big news
AI implants allow quadriplegic man to use his hands - here
Photoshop uses AI to expand scenery - here
Stack Overflow fights back against ChatGPT with OverflowAI - here
McKinsey predicts AI will cause 12 million to switch jobs by 2030 - here
Stable Diffusion releases upgraded SDXL 1.0 text-to-image model - here
MIT announces PhotoGuard to combat image manipulation - here
Universities give up and embrace AI in the classroom - here
G3PO is OpenAI’s answer to open-source models - here
GPT-3 reasons as well as undergraduates - here
Hugging Face, GitHub, et. al. unite to defend open source in EU legislation - here
Nerdy, But in a Good Way
Recent notable papers:
A Survey on Evaluation of Large Language Models - A comprehensive survey of evaluation methods for LLMs, and a discussion of the strengths and weaknesses of LLMs in general. - here
Open Problems and Fundamental Limitations of Reinforcement Learning with Human Feedback - RLHF is the current “gold standard” method for aligning LLM models with human preferences. This paper explores the weaknesses of this method. - here
Towards Generalist Biomedical AI - Google researchers unveil a shockingly good multimodal “generalist” LLM that performs as well or better than state of the art specialized AIs on a number of biomedical benchmark tasks. - here
Economics
Fitch gives a leg up to AI credibility for financial risk management amidst the uproar over downgrading US debt
Fitch Ratings, a top-tier global credit rating company, recently bumped the US credit rating down from AAA to AA+, citing the debt ceiling standoff and the January 6 insurrection. This led to economic distress in the stock markets as well as in treasuries this Tuesday. However, amidst this turmoil, Fitch praised the use of AI technology in finance, stating: “Artificial Intelligence is an evolving family of technologies that are helping the financial industry to streamline and optimise processes, ranging from credit decisions to quantitative trading and financial risk management.” When dour old Fitch smiles on AI, it validates AI boosters.
CBInsights released their Q2’23 report and the top global Venture investments are AI
CBInsights, a highly respected analytics firm, released their Q2 updates on the market and AI is taking off. Two of the largest global equity deals were AI-based (surrounded in purple), summing up to around 3% of total global funding. Meanwhile fintech funding has continued to decline, going down by -48% this quarter. Fintech is low and dry.
AI is potentially going to power the economy and here are what major financial institutions are saying:
McKinsey, a prominent financial management consultancy, wrote a report published on June 3rd stating that AI could add “trillions of dollars to the global economy”, estimating between $2.6 and $4.4 trillion in value. To put this into perspective, the total GDP (gross domestic product aka value added by the production of goods) of Canada as of 2021 was $1.98 trillion. Soon after this report, Goldman Sachs, a leading investment bank, estimated that by 2025 AI would account for up to 4% of the US economy. Goldman states that AI investing is predicted to hit $100 billion in the US and around $200 billion globally. This surge in AI adoption could also increase labor productivity by 1% (a seemingly small but significant increase which could cause the S&P 500 to surge 14%) - all good news for the US economy.
However, with the huge and rapid growth of AI, Goldman warns investors of the potential “bubble in the stock market” it may be creating. Cautionary observers compare this AI-frenzy to the dot-com era in the early 2000s. When that bubble burst stock prices stayed depressed for over a decade, so the fed needs to tread carefully to avoid a repeat.
That's a wrap! More news next week.