Time picks the 100 most influential in AI

Time releases a special issue on the current movers and shakers in AI, including most of the usual suspects (Sam Altman, Yann LeCun), as well as some questionable choices (Grimes? Really? Because she embraced a chance to sell her fame in another venue?), but the most interesting story here is that all roads lead to Andrew Ng, the soft-spoken Stanford professor whose online courses have introduced millions of students around the globe to AI.

Time’s 100 most influential people in AI

Six degrees of separation from Andrew Ng

Apple is burning Millions per day training AI models

Apple has been notably quiet about its AI ambitions, except for the “ML” (the AI that they dare not say its name) in its Vision Pro AR headset. Now, credible reports are surfacing and Apple is putting the pedal to the metal. Pouring millions of dollars PER DAY training Ai models, including a next-gen Siri on steroids.

Apple burns millions per day on AI

OpenAI announces its first Developer Day in SF on November 6

You knew it was only a matter of time. OpenAI is following the playbook of Silicon Valley startups that became foundational platforms (Microsoft, Apple) by cozying up to developers who can make the killer apps that draw users in by the millions. The company is hosting its first-ever confab for developers of AI applications on the first Monday in November.

OpenAI Developer Day

Microsoft promises legal immunity for any company sued for copyright infringement for using Microsoft AI Copilot

In a bold move, Microsoft promises to legally indemnify any company against copyright claims for using its AI Copilot software. Authors, publishers, artists and others have been making noise about suing AI companies for using their copyrighted materials in the training data of their AI models. Famously, the New York Times continues to rattle sabers about suing OpenAI.

Uncertainty is bad for business, so Microsoft is taking copyright risk off the table for businesses, and to delay implementing Microsoft’s AI Copilot add-on to its flagship Microsoft Office productivity suite. Smart, for a company with billions in the bank to not hire lawyers.

In US-China cold war, the race is on to deploy killer robots

Now that the US and China appear to be squaring off against one another in what could turn into Cold War Deux, both sides are looking to gain an edge in next-generation AI-controlled autonomous weapons, including drones for starters, but ultimately encompassing the full gamut of warfighting vehicles, including fighter planes and warships.

US-China compete on developing killer robots

Why AI Godfather Yann LeCun Isn’t Buying Doomer Narratives

Yann LeCun has seen it all. As an official Elder of the AI revolution, one who has survived many AI Winters and False Springs. He was a leader in the Long March from neural networks as the butt of jokes to their current Preeminent Glory, the man who taught the US Post Office scanners to read handwritten zip codes, Yan LeCun takes sh*t from no one. “Speak not of the Ancient Magic to me witch, I was there when it was written.” He thinks doomers wildly underestimate how long it will take to make intelligent machines, and how much control humans will be able exert over their development. Not a Pollyanna, he remains firmly in opinion that our fate remains in our hands - so we need to get better at making the best of that power.

Yann LeCunn isn’t buying AI doomerism

AI chatbots are invading your local government, and making everyone nervous

State and local governments around the US are grappling with free access to chatbot tools such as ChatGPT, Claude, Bard, and Bing. Some government departments are jumping in with both feet, lured by the prospect of AI assistance with the massive amounts of paperwork generated by the government. Others, citing concerns about accuracy and privacy, are banning it altogether. Some are trying to make sensible regulations about internal use of these tools. Lots of folks inside and outside of the government are watching with a good deal of concern.

Chatbots invade local government

ChatGPT’s AI tech was built in Iowa - with a lot of water

ChatGPT was trained in one of Microsoft’s vast server farms in Iowa. The computational load was so great that the water used to cool Microsoft’s servers spiked up by 34%, or around 400 million gallons. Environmentalists are raising alarm bells about big tech’s heavy use of water, so tech companies are scrambling to ease the drain on water supplies.

ChatGPT was built in Iowa - with a LOT of water

Chatbots are trained by prisoners from Finland

AI companies have scoured the globe for cheap labor (see our previous reporting on AI’s digital sweatshops overseas.) Now, they have found a surprising source of cheap labor in a First World location - the prisons of Finland. Finland is a high-wage country, and clickworkers are rare, but one enterprising contract work firm has made a deal with the Finnish prison system to employ prisoners to do the menial labor of training AI models.

Made in prison - AI firms use Finnish prisoners for clickwork

Google mandates clear disclaimers in political ads using AI

Google has announced a policy that all political ads hosted on its platforms must feature a clear and prominent disclaimer if AI was used in the ad’s creation. This is mainly aimed at deepfakes and other counterfactual content that is designed to seem real. The Republican National Committee released a hatchet piece against President Biden that used AI in this way. And a PAC supporting Ron DeSantis aired a bizarre AI-powered hit job against Trump. Victor Hugo is said to have feared an “Attila with a telegraph.” The threat today is unscrupulous political operatives with a login to ChatGPT.

Google mandates disclaimers for AI in political ads

Anthropic launches Claude Pro at $20/month - here

Imitation (of OpenAI) is the sincerest form of flattery.

Why Nvidia’s AI hegemony won’t last long - here

Because training is GPU-intensive, but is a one-time cost per model. But inference (generating answers) will become the vast bulk of the business, and inference can be done with much cheaper CPUs.

Harrison Chase on the point of Langchain - here

The founder of Langchain, the leading open-source toolset for building AI applications that integrate with LLMs like ChatGPT, opens up about what Langchain is, where it’s going, and the wild ride of building a company in a sector where the landscape changes almost daily.

Paper: Using LLMs to optimize prompts - here

The best performing prompt: “Take a deep breath and solve this step-by-step.” Who knew that prompting LLMs was like coaching your middle school child through math homework?

Paper: One wide feedforward is all you need - here

Enormous creativity is being focused on making today’s sprawling, compute-hungry LLMs slimmer and more efficient. These researchers have completely rewired the guts of the foundational Transformer LLM model to reduce compute for every inference (geek-speak for giving answers.)

Paper: Fine-tuning LLMs: LoRA or Full-Parameter? - here

Fine tuning is the term for making an off-the-shelf LLM more attuned to your own data and use case. How to do this efficiently and effectively is a hot topic of research. The folks at Anyscale, an AI toolset company, did an in-depth comparison of fine tuning results using LoRA, a method that only tunes a fraction of the LLM neurons (parameters), versus tuning the full parameter network. TL;DR - LoRA can be surprisingly effective.

Microsoft and Paige team up to diagnose cancer

Microsoft has announced a partnership with Paige.AI, a commercial startup spun out of famed cancer hospital Memorial Sloan Kettering in New York. Paige is a leader in Computational Pathology, using AI to analyze images in order to diagnose cancer. Paige developed the first AI model for Pathology from a half-million slides in the MSK archives. Now, with the vast resources of Microsoft, and over $150 million in venture funding, Paige proposes to build the world’s largest cancer detection AI, trained on billions of images.

Paige.AI joins with Microsoft to build world’s largest cancer-fighting AI

Health AI Startup Open Evidence profiled in Forbes

Traditionally, LLMs are stuck in the past - they can’t know anything past the date of their training. This is the reason for the annoying messages from ChatGPT that it can’t answer questions about events after 2021. One method to get around this is called RAG - Retrieval Augmented Generation. Basically, you give the LLM a database (“knowledge graph” in the lingo) to actually look up the answer. I know - such a novel concept! But actually, for LLMs, this turns out to be kinda hard. Because… reasons (waves hands.) Now Open Evidence, a Health AI startup is taking this problem on for healthcare, to harness the power of LLMs while keeping doctors up to date on the latest findings in clinical medicine. Other companies already offer services to keep physicians apprised of the latest findings, but Open Evidence plans to AI the shirt out of it. Anyway, Forbes was impressed. A lot.

Health AI startup Open Evidence wows Forbes

Dutch study finds ChatGPT a whiz in the ER

In today’s issue of the Annals of Emergency Medicine, Dutch researchers find that ChatGPT’s performance in diagnosing 30 anonymized patient records to be on par with the diagnostic acumen of emergency room physicians. The authors admit that we are a long way from having AI chatbots take over critical medical judgements from doctors, but that the results are encouraging enough that more research into integrating chatbots into the ER is likely to be fruitful.

ChatGPT is a whiz in the ER