- AI Weekly Wrap-Up
- Posts
- Welcome!
Welcome!
Keith Collins MD & Addie Politi
July 12, 2023
Big Talk
Embattled CEO of Stability AI: No programmers in 5 years
Emad Mostaque, podcast spokesmodel for CEO of open source text-to-image company Stability AI, boldly predicts that there will be no human programmers within 5 years. He no doubt hopes that this pronouncement will polish his brand as a visionary maverick, and (oh please, oh please!) make people quit asking him why so many of his senior executives are jumping ship.
Clash of the Titans
OpenAI releases Code Interpreter, your new Data Analyst
OpenAI keeps rolling out functions and features, and this latest one is a blockbuster. The terribly-named Code Interpreter can upload data files, even zipped ones, and do sophisticated analyses on the data set, prompted by the merest hint of what you are interested in.
Here is Wharton Professor Ethan Mollick showing how he put Code Interpreter through its paces, and ended up considerably impressed with both its skills and its relentless problem-solving to overcome roadblocks and get to an answer.
And here is a Twitter thread of just some (very smart) random guy who got Code Interpreter to perform facial recognition (!!!!) in just 5 simple steps.
Anthropic announces Open Beta of upgraded Claude 2 Chatbot
OpenAI competitor Anthropic has just upgraded its chatbot, Claude. Claude 2 is comparable in performance to GPT 3.5 and GPT-4. It does have some abilities that outshine the OpenAI flagship products. These include:
a much larger context window for input, 100,000 tokens (± 70,000 words) versus OpenAI’s 32,000 tokens (± 25,000 words.) Theoretically, this should allow the LLM to take into account more diverse information, but whether that holds up in practice depends greatly on the implementation. See this paper for more detail. [2307.03172] Lost in the Middle: How Language Models Use Long Contexts (arxiv.org)
ability to import, read, and summarize large PDFs.
Claude 2 was trained on data current through early 2023, as opposed to the September 2021 cutoff fore GPT-4.
Claude also has coding and analysis functions similar to OpenAI’s Code Interpreter.
Anthropic makes much of Claude’s “constitution”, which supposedly makes it safer and less likely to produce harmful output. Only real-world testing will prove whether this is the advantage that Anthropic thinks it is.
TL;DR Claude still lags behind GPT-4 somewhat, but it is easier to use, is aware of events through early 2023, has Code Interpreter type functions, has a larger context window, and is a fraction of the price of GPT-4 for businesses that access these models through their APIs.
OpenAI will devote 20% of its compute over the next 4 years on the problem of keeping super intelligent AI aligned with human values.
Sam Altman, OpenAI’s CEO, is really, really worried about the dangers of rogue AI. He is one of the signers of the famed 22-word open letter that compared the dangers of off-the-rails AI to the dangers of pandemics and nuclear war. (But carefully did not mention climate change, mmmkay?) OpenAI has announced it is dedicating 20% of its compute resources to the problem of alignment. The object is to get AI so aligned with the good of humanity that it is super-aligned. Which will be, like, super… I guess?
Shutterstock Links with OpenAI
Shutterstock continues to join ‘em rather than trying to beat ‘em in fights over copyright. Shutterstock has signed an expanded deal with OpenAI to license its images and other IP assets in exchange for fast-track access to OpenAI’s latest technologies.
Meta's Threads gains 30 million users in less than a day.
Elon’s Twitter seems poised on imploding like the Titan submersible. Zuck is trying to pounce on this opportunity to steal members with a Twitter competitor, Threads, which has been spun out of Instagram. The Insta connection, which allows for seamless import of your Instagram followers into Threads, has helped make Threads one of the fastest-growing apps in history, with 30 million signups the first day. Getting the members is one thing - getting them to stick with the somewhat anemic app is another. Meta engineers are hard at work adding features to Threads to keep members engaged. Time will tell if Meta can Thread the needle.
Fun News
Babies climbing Mount Everest
Google announces quantum supercomputer
Quantum computing would be awesome, if it actually worked. Google, Microsoft, and many others have spent billions of dollars trying to make squirmy little quantum bits, called qubits, hold still enough to be reliable, while still retaining enough of their quantum magic to do the superduper computing they are theoretically capable of. Google says it has come closer to that day, getting 70 qubits to behave for long enough to solve a math problem that would take the world’s most powerful conventional supercomputer 47 years to complete. Details below.
AI tests in top 1% of creativity
Researchers at the University of Montana submitted responses from ChatGPT, along with those of 24 UM students, to a national standardized test of creative thinking scored by the Scholastic Testing Service. ChatGPT scored at the 99th percentile for 2 out of 3 subtests, and at the 97th percentile for the third. Scholastic Testing Service was unaware that one of the “students” was a Chatbot.
AI decodes ancient cuneiform tablets instantly
Most ancient artifacts are rare, eaten away by the ravages of time. But ancient clay tablets are so durable that we have hundreds of thousands of the darned things, of which only a tiny, tiny fraction have ever been translated from the original cuneiform. This is because translation is so labor intensive, and most of these tablets hold mundane records, the equivalent of ancient grocery lists. But no doubt there are hidden gems among the dross, and the only way to find them is to translate them all. Here, some clever new AI comes to the rescue. See below.
KokoMind is a benchmark data set to gauge the social awareness of LLMs
Product of some brilliant researchers at multiple elite universities and companies, including grad students and post-docs, this site aims to be a benchmark test of the social awareness of AI models, testing their understanding of human emotions and motivations from observing facial expressions, what is said, tone of voice, and more. Check it out, it’s a fun site, whether or not it becomes a standard benchmark of AI performance.
Langchain and Streamlit agree to integrate their 2 open source projects
Langchain is an open source project that is becoming a standard for automating interaction with LLMs. Streamlit is another popular open source project that helps turn data applications into shareable webapps. The two projects have announced an agreement to make their offerings compatible and interoperable. This is another step toward a truly usable AI automation toolset.
Regulation and Politics
The US is considering limiting China's access to cloud computing from Microsoft and Amazon, the two global leaders
The US government has ongoing concerns about misappropriation of Western intellectual property by the Chinese. They are particularly concerned that China not take unfair advantage in the burgeoning realm of AI. Chip embargos have already been announced, and now the feds are considering how to close a potential backdoor access to advanced chips through becoming a cloud computing customer of US cloud giants Amazon and Microsoft. Watch for fireworks on this front.
Worldwide AI Ethics
This paper reviews 200 proposals for AI regulation worldwide, showing the range of proposals being considered in one jurisdiction or another.
Deals
SoftBank invested $140 billion in AI with only 1 company worth more than a billion
Poor, poor Softbank. The mega-fund has apparently squandered much of the $140 billion it has invested in AI, having but a single unicorn to show for its largesse.
InflectionAI buys $880 million worth of Nvidia GPUs
InflectionAI, the startup behind the super-friendly chatbot Pi, got $1.3 billion last week from a consortium that included GPU chipmaker Nvidia. This week, InflectionAI gave $880 million of that treasure trove right back to Nvidia, for 22,000 of Nvidia’s top-of-the-line H100 GPUs, which sell for 40,000 a pop. This makes InflectionAI the proud owner of one of the largest supercomputers in the world, and gives Nvidia an immediate profit on its investment. From now on, NVidia is riding the InflectionAI startup train for free. What a great business.
That's a wrap! More news next week.