Pika launches “idea to video” app

AI video startup Pika, long a cult favorite in beta, has now launched Pika 1.0, its official release. It does not disappoint. It takes multimodal input - text, image, or video - and produces an easily editable video. Click the link below for a video (and a sales pitch to sign up, natch.)

Pika launch of “idea to video”

Anthropic slashes fees in AI price war

OpenAI competitor Anthropic, its fever-dreams of a quick OpenAI implosion after the recent Board meltdown dashed by the return of Sam Altman and the resignation of the insurrectionists, now has to do the hard work of competing on value. That means increasing product benefits, which is slow and expensive, or cutting prices which is fast and can more than pay for itself if it increases sales volume enough. Anthropic is trying to do both, but today the news is a cut to per-token prices for its flagship model Claude 2.1. Analysts say that competing with open source models (with per-token fees of $0) is an increasing consideration for the large foundation models like Claude and ChatGPT.

Anthropic slashes fees in AI price war

Amazon AWS’ answer to ChatGPT is an AI assistant called Q

Amazon’s AWS, the dominant cloud service provider, has been incrementally adding to its AI offerings. Now there is a major upgrade for AWS business customers, an AI Assistant called Q. Q is a GPT-style chatbot that can help developers write code, explain the complex AWS configuration setting to system administrators, and generate business reports by linking to the AWS business Intelligence suite, QuickSight. It is also integrated into Amazon Connect to help customer service agents with support requests.

Amazon’s answer to ChatGPT is an AI Assistant called Q

The problems lurking in Hollywood’s historic AI deal

The Hollywood actors’ deal with the studios over AI-generated “digital replicas” was, indeed, a historic win for labor. Now comes the hard part - working out the details of what counts as a digital replica. How closely can the studios mimic an actual actor before the digital replica rule kicks in? And for non-famous actors, will they ever get a major role again, when a famous actor can license their likeness to appear in a dozen films at once? Will Ryan Reynolds and Jennifer Lawrence still rule the box-office in the 22nd century?

The problems lurking in Hollywood’s historic AI deal

17 nations including US & UK sign “secure by design” agreement

The US, UK, and more than a dozen other countries have signed onto a comprehensive agreement to work to keep AI safe from bad actors, by requiring companies to create AI that is “secure by design” - that is, that the safeguards are built into the very fabric of the system, and cannot be bypassed. Too bad the agreement is nonbinding. Still, it was an achievement to get consensus on principles from so many countries.

17 nations, US & UK, sign “secure by design” AI agreement

Sports Illustrated caught posting fake AI-generated writers

Famous for its swimsuit models that are so beautiful that they seem fake, (and yes, Photoshop may have helped), SI has now been caught passing off fake writers as real. And not only did they credit writers whose profile pictures were provably AI-generated, but the articles themselves were… odd, raising the question of whether the articles themselves were undisclosed AI creations. When questioned, SI blamed everything on an outside vendor and quickly took down the questionable posts.

Sports Illustrated caught posting AI-generated articles by fake writers

AI no major threat to jobs, says the European Central Bank

The great AI job massacre may be “greatly exaggerated” according to a new report from the European Central Bank. The ECB, using comprehensive data from 16 European countries, found that job losses to AI have so far been small, and that AI has even caused job gains among high-skilled workers, who are being augmented, not replaced by AI.

European Central Bank says AI is not a significant threat to jobs - yet

Change in Employment by Skill Level (source: report cited above)

HuGE breakthrough with crowdsourced feedback for robots

Researchers at MIT, Harvard, and the University of Washington have developed a new method for teaching robots how to perform a task. called Human Guided Exploration (HuGE.) Current methods use a type of reinforcement learning that requires a great deal of input from human experts. This is slow and expensive. HuGE relies on feedback to the robot from crowdsourced non-experts, which is faster, cheaper, and works even when the nonexperts make lots of mistakes.

HuGE breakthrough with crowdsourced feedback for robots

Mega-trend: the rise of fine-tuned smaller models

ChatGPT is a hulking beast of an AI model, with 175 billion neurons (called parameters in AI-speak), which is almost twice as many as the human brain (±100 billion.) Conventional wisdom is that it takes that large a model to do what ChatGPT can do. And that further advances in AI will require even larger models. But now we are seeing rapid innovation in using much smaller models, as small as 7 billion parameters, being fine-tuned with domain-specific data sets for specific types of tasks. Microsoft is using smaller models to reduce its reliance on OpenAI (which recently suffered a Board coup and counter-coup, so you can see their point.) AI startups are using GPT-4 to demo, then smaller models for production. Ground Zero for the small-model, open source rebel alliance is Hugging Face, a former model repository that is now an open source juggernaut valued at $4.5 billion.

How Microsoft is redefining AI with small language models.

How smaller models can outperform large LLMs

The rise of small language models

GPT-4 Rocks at Radiology Reports

Microsoft released a paper comparing GPT-4 with state-of-the-art text based reporting models in the field of Radiology. The result: with no exposure to examples (zero shot performance) GPT-4 equaled or exceeded the performance of radiology-specific models. When given examples, GPT-4 did even better.

GPT-4 outclasses SOTA radiology models in generating reports

AI transforming Microscopy

Cell biologists and medical pathologists have until now been reliant on their own pattern-recognition abilities to identify and classify cells and tissues under a microscope. AI is now automating these tasks, transforming the whole discipline of microscopy.

AI is transforming microscopy

Cradle snags $24 million for AI powered protein engineering

Biotech and AI startup Cradle recently announced a $24 million round of funding for their protein engineering AI system. The company is founded on a unique insight - the sequences of amino acids in a protein can be conceptualized as an alien language, and generative AI can be harnessed to decode it. The approach is beginning to show success in decreasing the time and money to develop novel proteins for medicine, agriculture, and other biology-based businesses.

AI-powered protein engineering

Stanford Medicine’s AI roundup

Stanford Medicine magazine has devoted its latest issue to AI. The issue highlights advances such as an app that can take medical grade photos of skin with a smartphone to assist in Dermatology telemedicine visit, and chronicles how one physician uses ChatGPT to rehearse for crucial conversations with patients.

Stanford’s roundup of AI in Medicine