Should we fear the Hyena Code?

This new technology could blow away GPT-4 and everything like it

The Hyena code is able to handle amounts of data that make GPT-style technology run out of memory and fail.

Written by Tiernan Ray, Contributing Writer on April 20, 2023

dall-e-2023-04-18-00-03-36-hyena-on-the-velte-juggling-words-digital-art.png

Stanford and MILA's Hyena Hierarchy is a technology for relating items of data, be they words or pixels in a digital image. The technology can reach similar accuracy in benchmark AI tasks as the existing "gold standard" for large language models, the "attention" mechanism, but with as little as 100 times less compute power.

Image: Tiernan + DALL•E

For all the fervor over the chatbot AI program known as ChatGPT, from OpenAI, and its successor technology, GPT-4, the programs are, at the end of they day, just software applications. And like all applications, they have technical limitations that can make their performance sub-optimal. 

In a paper published in March, artificial intelligence (AI) scientists at Stanford University and Canada's MILA institute for AI proposed a technology that could be far more efficient than GPT-4 -- or anything like it -- at gobbling vast amounts of data and transforming it into an answer. 

Also: What is GPT-4? Here's everything you need to know

Known as Hyena, the technology is able to achieve equivalent accuracy on benchmark tests, such as question answering, while using a fraction of the computing power. In some instances, the Hyena code is able to handle amounts of text that make GPT-style technology simply run out of memory and fail. 

"Our promising results at the sub-billion parameter scale suggest that attention may not be all we need," write the authors. That remark refers to the title of a landmark AI report of 2017, 'Attention is all you need'. In that paper, Google scientist Ashish Vaswani and colleagues introduced the world to Google's Transformer AI program. The transformer became the basis for every one of the recent large language models.

But the Transformer has a big flaw. It uses something called "attention," where the computer program takes the information in one group of symbols, such as words, and moves that information to a new group of symbols, such as the answer you see from ChatGPT, which is the output. 

That attention operation -- the essential tool of all large language programs, including ChatGPT and GPT-4 -- has "quadratic" computational complexity (Wiki "time complexity" of computing). That complexity means the amount of time it takes for ChatGPT to produce an answer increases as the square of the amount of data it is fed as input. 

Also: What is Auto-GPT? Everything to know about the next powerful AI tool

At some point, if there is too much data -- too many words in the prompt, or too many strings of conversations over hours and hours of chatting with the program -- then either the program gets bogged down providing an answer, or it must be given more and more GPU chips to run faster and faster, leading to a surge in computing requirements.

In the new paper, 'Hyena Hierarchy: Towards Larger Convolutional Language Models', posted on the arXiv pre-print server, lead author Michael Poli of Stanford and his colleagues propose to replace the Transformer's attention function with something sub-quadratic, namely Hyena.

The authors don't explain the name, but one can imagine several reasons for a "Hyena" program. Hyenas are animals that live in Africa that can hunt for miles and miles. In a sense, a very powerful language model could be like a hyena, which is picking over carrion for miles and miles to find something useful.

But the authors are really concerned with "hierarchy", as the title suggests, and families of hyenas have a strict hierarchy by which members of a local hyena clan have varying levels of rank that establish dominance. In some analogous fashion, the Hyena program applies a bunch of very simple operations, as you'll see, over and over again, so that they combine to form a kind of hierarchy of data processing. It's that combination element that gives the program its Hyena name. ... ' 

AI and the Disruption of Education

The topic has often been mentioned.   Adaptations are under under way.    In ScienceNews. 

It’s no wonder other tech companies have been racing to put out competing tools. Anthropic, an AI company founded by former OpenAI employees, is testing a new chatbot called Claude. Google launched Bard in early February, and the Chinese search company Baidu released Ernie Bot in March.

A lot of people have been using ChatGPT out of curiosity or for entertainment. I asked it to invent a silly excuse for not doing homework in the style of a medieval proclamation. In less than a second, it offered me: “Hark! Thy servant was beset by a horde of mischievous leprechauns, who didst steal mine quill and parchment, rendering me unable to complete mine homework.”

But students can also use it to cheat. ChatGPT marks the beginning of a new wave of AI, a wave that’s poised to disrupt education.

When Stanford University’s student-run newspaper polled students at the university, 17 percent said they had used ChatGPT on assignments or exams at the end of 2022. Some admitted to submitting the chatbot’s writing as their own. For now, these students and others are probably getting away with it. That’s because ChatGPT often does an excellent job.

“It can outperform a lot of middle school kids,” Vogelsinger says. He might not have known his student had used it, except for one thing: “He copied and pasted the prompt.”

The essay was still a work in progress, so Vogelsinger didn’t see it as cheating. Instead, he saw an opportunity. Now, the student and AI are working together. ChatGPT is helping the student with his writing and research skills. ... '  

As AI Continues to Surpass Human Performance, it's Time to Reevaluate Tests,

Evaluating Performance by AI,  Humans. Implications?  Most interesting.   Well take worth a closer look.

 As AI continues to surpass human performance, it's time to reevaluate tests , says expert  Shana Lynch, Stanford University

Credit: Pixabay/CC0 Public Domain  

How good is AI? According to most of the technical performance benchmarks we have today, it's nearly perfect. But that doesn't mean most artificial intelligence tools work the way we want them to, says Vanessa Parli, associate director of research programs at the Stanford Institute for Human-Centered AI and a member of the AI Index steering committee.

She cites the current popular example of ChatGPT. "There's been a lot of excitement, and it meets some of these benchmarks quite well," she said. "But when you actually use the tool, it gives incorrect answers, says thing we don't want it to say, and is still difficult to interact with."

In the newest AI Index, published on April 3, a team of independent researchers analyzed over 50 benchmarks in vision, language, speech, and more to find out that AI tools are able to score extremely high on many of these evaluations.

"Most of the benchmarks are hitting a point where we cannot do much better, 80-90% accuracy," she said. "We really need to be thinking about how we, as humans and society, want to interact with AI, and develop new benchmarks from there."

In this conversation, Parli explains more about the benchmarking trends she sees from the AI Index.

What do you mean by benchmark?

A benchmark is essentially a goal for the AI system to hit. It's a way of defining what you want your tool to do, and then working toward that goal. One example is HAI Co-Director Fei-Fei Li's ImageNet, a dataset of over 14 million images. Researchers run their image classification algorithms on ImageNet as a way to test their system. The goal is to correctly identify as many of the images as possible.

What did the AI Index study find regarding these benchmarks?

We looked across multiple technical benchmarks that have been created over the past dozen years— around vision, around language, etc.—and evaluated the state-of-the-art result in each benchmark year over a year. So, for each benchmark, were researchers able to beat the score from last year? Did they meet it? Or was there no progress at all? We looked at ImageNet, a language benchmark called SUPERGlue, a hardware benchmark called MLPerf, and more; some 50 were analyzed and over 20 made it into the report.

And what did you find in your research?

In earlier years, people were improving significantly on the past year's state of the art or best performance. This year across the majority of the benchmarks, we saw minimal progress to the point we decided not to include some in the report. For example, the best image classification system on ImageNet in 2021 had an accuracy rate of 91%; 2022 saw only a 0.1 percentage point improvement.

So we're seeing a saturation among these benchmarks—there just isn't really any improvement to be made.

Additionally, while some benchmarks are not hitting the 90% accuracy range, they are beating the human baseline. For example, the Visual Question Answering Challenge tests AI systems with open-ended textual questions about images. This year, the top performing model hit 84.3% accuracy. Human baseline is about 80%.  ... ' 

Stewart Brand Rethinking Silicon Valley

Met Stewart Brand Way back when, through company connections at the Institute for the Future.  Spent quite a while those days at Stanford and the Silicon Valley, visiting AI startups, seeking to discover AI.   Bringing it back to the Enterprise.  Was always interesting

 Rethinking Silicon Valley: Stewart Brand's Lost Journal, By John Markoff

Communications of the ACM, November 2022, Vol. 65 No. 11, Pages 32-34  10.1145/3530988

Stewart Brand,  Credit: Andrij Borys Associates, Wikimedia (CC BY-SA 2.O)

During the fall of 2000 I visited the Special Collection room in Green Library at Stanford University, eager to read Stewart Brand's personal journals. I was researching the political and cultural world surrounding three computer science laboratories that were located adjacent to Stanford during the decade during which the technologies that led to the creation of the personal computer industry and the modern Internet emerged.

On my initial visit I came away disappointed and it would take almost another two decades before I discovered a missing piece of the puzzle that reframes the early history and impact of Silicon Valley.

Although he has not a technologist, Brand, the creator of the Whole Earth Catalog, is an intriguing figure in the history of the modern computing world and the Valley. He was the author of a seminal article in Rolling Stone magazine, "Space War: Fanatic Life and Symbolic Death Among the Computer Bums," which had been the first hint a wider non-technical audience had of the emerging digital world. He would be the first writer to use the term "personal computer" in his 1974 book II Cybernetic Frontiers. A decade later Brand would help organize the first Hackers conference as well as establishing an early online community known as the Whole Earth 'Lectronic Link (WELL).

William English, who was chief engineer for Douglas Engelbart's legendary oNLine System (NLS) project at Stanford Research Institute told me that he invited Brand to participate in the demonstration that Engelbart was preparing to give at the ACM/IEEE Joint Fall Computer Conference held in San Francisco in December 1968. English had seen an early multimedia (at that time "multimedia" meant three slide projectors, an audio tape, and a single shared projection screen) production Brand created titled "America Needs Indians," and wanted his expertise to help prepare for the watershed event that Steven Levy would later describe as "The Mother of All Demos."

Brand would operate a camera in Menlo Park, relaying a video feed to the stage where Engelbart sat in front of a huge screen with a keyboard, mouse and chord keypad, introducing the world's leading computer scientists and engineers to interactive computing.

Later, however, I realized that Brand's actual role in helping prepare for the Demo had been minimal and I found no mention of the event in his journal, which at that point was filled with observations more in tune with a 30-year-old in the midst of the turmoil of the Vietnam War and the 1960s counterculture.

Although Brand had been a prominent figure during the 1960s and 1970s, he was famously resurrected by Steve Jobs in his 2005 Stanford Commencement speech when he commended Brand's philosophy of life highlighted on the closing page of The Whole Earth Epilog to the new graduates: "Stay Hungry, Stay Foolish."

He might have remained a historical figure, however in 2017, with Donald Trump's election, the national zeitgeist shifted abruptly and Silicon Valley went from being able to do no wrong, to doing no right. Two books that appeared that year, Franklin Feuer's World Without Mind: The Existential Threat of Big Tech and Jonathan Taplin's Move Fast and Break Things: How Facebook, Google, and Amazon Cornered Culture and Undermined Democracy both opened their accounts of Silicon Valley with biographical sketches describing Brand as Silicon Valley's first "digital utopian." Feuer and Taplin, and the following year the historian Jill Lapore in These Truths: A History of the United States, all attempt to make the case that there is a straight cultural through-line from Brand to an unregulated Silicon Valley that has brought about the perilous state of the world today—underscored by the role of social media and the Internet in fomenting both Brexit and the election of Donald Trump.

They all make the claim that the Valley's original ideology—or sin depending on your point of view—can be traced back to Brand and the Catalog, expanding on his opening argument in the Catalog: "We are as gods and might as well get good at it. A realm of intimate, personal power is developing—power of the individual to conduct his own education, find his own inspiration, shape his own environment, and share his adventure with whoever is interested."  ... ',

Wireless Smart Bandages

Speeding healing. 

New Wireless Smart Bandage Accelerates Chronic Wound Healing  By Adrianna Nine on November 28, 2022 at 10:03 am   in ExtremeTech

Chronic wounds are an under-acknowledged medical concern. At any given time, more than 600,000 Americans are thought to experience physiologically-stunted wounds that won’t heal. Chronic wounds aren’t just inconvenient and painful; they also rack up individual healthcare costs and prevent people from engaging in certain activities, resulting in a decreased quality of life.

Thanks to new research, this might not always be the case. A team of scientists at Stanford University has developed a wireless “smart bandage” that simultaneously monitors wound repair and helps to speed up healing. The bandage could shorten the time people suffer from chronic wounds while mitigating the physical damage and discomfort caused by conventional healing methods.

In a study published last week in Nature Biotechnology, the scientists describe a flexible, closed-loop device that seals wounds while transmitting valuable biodata to an individual’s smartphone. Hydrogel makes up the bandage’s base: While conventional bandages tug and tear at the skin when they’re pulled away, hydrogel allows the smart bandage to attach securely without causing secondary damage during removal. On top of the hydrogel sits the electronic layer responsible for wound observation and healing. At just 100 microns thick, this layer contains a microcontroller unit (MCU), electrical stimulator, radio antenna, memory, and a series of biosensors.  .. ' 

Internet Law and Moderation Transparency

 Happened on the Stanford Cyberlaw Center, and notably Transparency laws.  Here an excerpt, which leads to more  ...  Had not taken a look there for some time.   Note the mention and criticism of Facebook's rules.   Worth covering.

The Center for Internet and Society at Stanford Law School is a leader in the study of the law and policy around the Internet and other emerging technologies.

State abuse of Transparency Laws and How to Stop it  By Daphne Keller on September 19, 2022 at 3:49 pm

Around the world, new laws are requiring improved transparency from major Internet platforms about their content moderation. This is generally a very positive development. But it also raises important questions about what kinds of disclosures we expect from platforms -- and what kinds of enforcement we expect from governments. This post uses five concrete examples to illustrate the complexity of these disclosures. The examples also illustrate what I think is a very real risk: that state enforcers may abuse transparency laws, using them to reshape platforms’ actual policies. That is a threat not only to platforms’ editorial and speech rights, but to the rights of all their users. I think it should be possible to mitigate this risk. But we can only do so if we recognize it.  

I have written before about the complexity of counting content moderation actions for aggregate transparency reports. This post addresses the seemingly simpler task of describing speech policies or enforcement decisions. This is also complicated, in ways that should come as no surprise to lawyers, parents of young children, or anyone else who has tried to explain and apply rules to disputatious parties. The degree of detail that platforms could include in their explanations is vast -- perhaps not as infinite as the diversity of human misbehavior on the Internet, but the two are certainly correlated. Facebook’s rules, for example, run to over a hundred pages   , but are widely criticized as insufficiently clear or detailed.  ... ' 

Algorithm Could Turbocharge Solving Cold Cases with DNA

Decsision tree application to improve success rate.

 Algorithm Could Turbocharge Solving Cold Cases with DNA

Popular Mechanics, Caroline Delbert

November 4, 2022

Researchers from Stanford University, forensic genealogy company Identifinders, and the non-profit DNA Doe Project developed a new mathematical model for solving crimes using DNA. The researchers explained their algorithm improves the current forensic genetic genealogy process 10-fold by optimizing candidate searches via a decision tree. They said the algorithm "can solve a case with a 7,500-person family tree around 94% of the time," while the current technique only yields a 4% success rate. Stanford's Lawrence Wein said unlike the benchmark method of searching for common ancestors between different matches, the new strategy seeks "the most recent common ancestor between a match and the unknown target." ... '

Ask Dumb Questions

Ask enough questions and you will notuce they are getting smarter.

When AI Asks Dumb Questions, It Gets Smart Fast   In Science, September 22, 2022

New research suggests patiently correcting artificial intelligence (AI) when it asks dumb questions may be key to helping the technology learn.

Stanford University scientists trained a machine learning AI to identify gaps in its knowledge, as well as to formulate often-stupid questions about images that strangers would answer.

When people responded, the system received feedback prompting it to adjust its inner mechanisms to behave similarly in the future; the researchers also "rewarded" the AI for writing smart questions to which humans responded.

The AI absorbed lessons in language and social norms over time, refining its ability to compose sensible and easily answerable queries.

The researchers said the system's accuracy at answering questions similar to those it had asked improved 118% over eight months and across more than 200,000 questions ...

Many AI systems become smarter by relying on a brute-force method called machine learning: they find patterns in data to, say, figure out what a chair looks like after analyzing thousands of pictures of furniture. ... 

From Science

A Basic Robotic task, Rummaging?

Task of everyday life.

Robot Can Find Keys in a Bag by Listening as It Rummages

New Scientist, Jeremy Hsu, June 13, 2022

.Stanford University's Maximilian Du and colleagues have trained a microphone-outfitted robotic arm to locate objects like keys in a handbag by listening for telltale sounds while searching. "That environment is basically like you reach down, you don't know where the keys are, but then once you hear the sound of the keys you can kind of localize it," Du said. "And then by localizing, you can grab it out and lift it up." The researchers trained the robot by having a person demonstrate the action through an Oculus virtual reality headset and controllers, while further human guidance corrected its actions. Stanford's Olivia Lee said this enhanced performance 20%.  ... ' 

Newly Powered LiDAR System

Note mention of piezoelectric method for powering Lidar.    Could be a big step.

LiDAR System Promises 3D Vision for Cameras, Cars, Bots

IEEE Spectrum, Mark Harris, April 21, 2022

New Lidar System Promises 3D Vision for Cameras, Cars, and Bots Piezoelectric effect harnessed for cheaper lidar  .... 

Stanford University researchers, working with a colleague at Sweden’s Chalmers University of Technology, have developed a novel light detection and ranging (LiDAR) system that taps the piezoelectric effect to capture three-dimensional data at potentially vastly less cost than current systems. The researchers coated thin-film lithium niobate with transparent electrodes in order to excite the crystal, generating an acoustic standing wave that modulates the intensity of laser light passing through it. They were able to execute a modulated time-of-flight calculation that captured distance data to objects in the scene, using less than a watt of power. Coupling the opto-acoustic modulator with a standard four-megapixel digital camera yielded a relatively high-resolution depth map of several metal targets, locating them to within a few centimeters.

Simpler Quantum Computer Design

Simplicity is good.

Stanford Engineers Propose Simpler Design for Quantum Computers

Stanford News, McKenzie Prillaman, November 29, 2021

Stanford University engineers have proposed a simpler design for quantum computers by harnessing a laser to manipulate an atom and alter photonic states through quantum teleportation. The atom can be recalibrated and reused for numerous quantum gates, which renders multiple distinct physical gates unnecessary. The design includes a storage ring to hold several counter-propagating photons, while a scattering unit manipulates photonic quantum bits (qubits). Since any quantum logic gate can be compiled into a series of operations performed on the atom, quantum programs of any size can be run on a single atomic qubit. Stanford's Ben Bartlett said in this design, changing programs only requires giving the machine "a different set of instructions."

Full article  

Technical Paper

Optimizing Robots with Evolution

Home/News/A Novel Way to Optimize Robots/Full Text

ACM NEWS

A Novel Way to Optimize Robots

By The Economist   November 3, 2021

It might sound obvious that if you want to improve a robot's software, you should improve its software. Agrim Gupta of Stanford University, however, begs to differ. He thinks you can also improve a robot's software by improving its hardware—that is, by letting the hardware adapt itself to the software's capabilities.

As they describe in Nature Communications, he and his colleagues have devised a way of testing this idea. In doing so, they have brought to robotics the principles of evolution by natural selection. They have also cast the spotlight on an evolutionary idea that dates from the 1890s, but which has hitherto proved hard to demonstrate.

There is a wrinkle. The team's robots, which they dub "unimals", are not things of metal and plastic. Rather, they are software entities that interact with a virtual environment in the way that metal-and-plastic devices might interact with a real one. Unimals are pretty simple, having spheres for heads and cylinders for limbs (see picture). The environments through which they roamed were also simple, and came in three varieties: flat arenas, arenas filled with hills, steps and rubble, and ones that had the complexities of the second sort, but with added props like cubes that needed to be moved around.

From The Economist

View Full Article

From CIS: Center for Internet and Society, Stanford Law

Very good piece from Stanford Law on technology. Below intro and beyond.

From:  The Center for Internet and Society at Stanford Law School is a leader in the study of the law and policy around the Internet and other emerging technologies.

PRIVACY, MIDDLEWARE, AND INTEROPERABILITY: CAN TECHNICAL SOLUTIONS, INCLUDING BLOCKCHAIN, HELP US AVOID HARD TRADEOFFS?

By Daphne Keller on August 23, 2021 at 7:01 am

Interoperability and distributed content moderation models have tremendous promise. They could temper major platforms’ power over public discourse, introducing both more economic competition and more diverse and pluralistic spaces for online speech. But these models -- which I will reductively refer to as “middleware,” following Francis Fukuyama’s coinage -- also raise a number of as-yet-unresolved problems. As I explained in this short piece, one of the hardest problems involves privacy: When a user opts in to a new service, can she give that service permission to process other people’s data and content? 

This post examines some possible technical solutions to protect privacy while enabling interoperability or distributed content moderation. It focuses in part on blockchain technologies, and draws on recent conversations with a number of seriously smart people about possible technical designs. My takeaway is that there is no magic bullet. We can’t have perfect privacy and optimize for platform interoperability or middleware. But blockchain technologies, and some other technical design approaches, can whittle away at the problem. The rest of this post examines ways to do that.   

Note: This post starts from a pretty high level of wonkiness, and then gets wonkier. If you’re new to the “middleware” topic, my short take on why it matters and why it’s hard are here; Cory Doctorow and I also had a great live discussion about it here. If you’re a blockchain maven and spot any mistakes, they are definitely mine and not the fault of the experts I talked to. 

This post doesn’t run down every possible permutation of the ideas it discusses, and it is un-wonkily reductive about at least three complicated things. (1) It uses the term “middleware” very loosely, as a way to lump together models including platform-to-platform interoperability, protocols not platforms, Magic APIs, and federated systems like Mastodon. (2) It elides differences between real blockchain technologies, describing a hypothetical a version (perhaps most similar to Project Liberty) optimized for Middleware. (3) It simplifies the kinds of user data and content used by social networks.  

OK, here we go.

What Blockchain Can Do

As a social media user, blockchain technologies could provide me with the following:

1.  An authenticatable identity that I control and can use to log in or validate my identity across multiple services. 

2.  A social graph linking my identity to other people’s similarly authenticated identities (meaning I control a “contact list” and decide when and how other people or services can access it). 

3.  For purposes of Middleware, a copy of every piece of content I post. This doesn’t have to be blockchain-linked -- it could just be stored on a physical device I control. But the content could be stored subject to my control as the blockchain-authenticated user, or in principle even stored on-chain (though then it couldn’t be deleted). To be useful for Middleware, this content would need to be kept in a format usable by any platform or Middleware provider. (This is quite hard to do, in practice). For each item of content, it could include a record of which service I shared it with and which contacts from my social graph could see it on that service.  

The Problem with Sharing Other People’s Data

Suppose three friends -- Ann, Balaji, and Carlos -- all have the set-up described above. Each has a blockchain-linked identity, a blockchain-linked social graph listing one another as contacts, and a controlled copy of all their posted content. They are currently friends on Facebook.  .... ' 

Stanford Internet Observatory

Brought to my attention: 

Stanford Internet Observatory

A program of the Cyber Policy Center, part of the Freeman Spogli Institute for International Studies ... 

The Stanford Internet Observatory Turns Two

SIO releases its two year report summarizing its first two years of research, teaching and policy and laying the path for the years to come.  ... ' 

Stanford Chatbot Study

Mostly obvious results, but useful characterizations.   We leveraged a 'concierge' model being most important, how do you get people to the right humans, the ones with the best answer in context?   I also recall also including a 'competence in context' rating we also measured.   We also tried to measure 'ongoing engagement', which was useful for future marketing connections

Do chatbots need to be more likable?    by Tom Ryan in Retailwire

A new Stanford university study finds people will more readily use a chatbot if they perceive it to be friendly and competent and less so if it projects overconfidence and arrogance. The challenge, the authors say, is finding the right balance.

Across three studies with 300 participants in the U.S., researchers tested reactions to AI-bots with the same underlying functionality but different descriptions.

Among the findings:

Low-competence descriptions (e.g., “this agent is like a toddler”) led to increases in perceived usability, intention to adopt and desire to cooperate relative to high-competence descriptions (e.g., “this agent is trained like a professional”). 

People are more likely to cooperate with and help an agent that projects higher warmth (e.g., “good-natured” or “sincere”).

Descriptions “are powerful,” helping drive user adoption and engagement with chatbots.

The authors suggested chatbots need to instill confidence that they are worthwhile to engage with. At the same time, acknowledging some errors may occur early on as the chatbot learns what users want will likely help people become more accepting of a chatbot’s mistakes. Pranav Khadpe, a co-author, told The Wall Street Journal, “You really want to manage the expectations you set before the first interaction.”  ... '

AI Mapping Moods

We took at cut at the basic idea to determine changes in moods that might indicate purchases.  And also mapping to actual behavior.   Note pharmaceutical use implications.

Can AI Map Our Moods?

By Stanford University, February 4, 2021, in ACM

Machine-learning models can map a persons mood swings and volatility from week to week, according to a new study.

Researchers from Stanford University and the University of Michigan developed a machine learning model that can infer a person's mood from their Facebook posts.

The model was trained using public Facebook postings of close to 3,000 volunteers from an earlier study, then tested on a different set of posts from 640 Facebook users who posted an average of 17 times weekly over 28 weeks. This dataset of 18,000 person-weeks, the largest ever compiled on weekly emotional dynamics, has been made public.

The patterns revealed by the model aligned with the predictions based on classical in-person psychological studies.

Said Stanford's Johannes Eichstaedt, "If this kind of approach is used ethically and legally, with strict privacy protection, we could someday have ways to computationally understand the mind. It could help with diagnosis and pharmaceutical evaluation. It could also help us track the psychological impact of traumatic societal events, such as the Covid pandemic."

From Stanford University

Mirror Arrays Make Augmented Reality More Realistic

Augmented reality, adding realism.  Most AR systems today provide a 'gamelike' view of artificial objects.   A solution.

Mirror Arrays Make Augmented Reality More Realistic
IEEE Spectrum
By Michelle Hampson

Researchers at Stanford University have developed a compact augmented reality (AR) system that uses an array of miniature mirrors to create virtual images that appear more "solid" in front of real-world objects. The miniature mirrors in Stanford's system change positions tens of thousands of times per second, enhancing the system's occlusion. The system combines virtual projection and light-blocking abilities into one element, relying the dense array of miniature mirrors to switch between a see-through state and a reflective state. It computes the optimal arrangement for the mirrors and adjusts accordingly. Stanford’s Brooke Krajancich said the system uses a lot of computing power, and may require more power than other AR systems.  .. ' 

Rewarding Autonomous AIs

Thoughtful piece.    Can a human just provide some sort of reward function?    Or is creating that alone very hard?  Especially if we include some measures of risk as well.  The latter we found in practice at least if you are honest about risk.  This also clouds some the the 'future is unsupervised', things I have heard recently.   What exactly does 'unsupervised' mean?    More than just a simple lack of a measure of success, we discovered.

Stanford researchers teach robots what humans want
Researchers are developing better, faster ways of providing human guidance to autonomous robots.

By Taylor Robota

Told to optimize for speed while racing down a track in a computer game, a car pushes the pedal to the metal … and proceeds to spin in a tight little circle. Nothing in the instructions told the car to drive straight, and so it improvised.

Researchers are trying to make it easier for humans to tell autonomous systems, such as vehicles and robots, what they want them to do. 

This example – funny in a computer game but not so much in life – is among those that motivated Stanford University researchers to build a better way to set goals for autonomous systems.

Dorsa Sadigh, assistant professor of computer science and of electrical engineering, and her lab have combined two different ways of setting goals for robots into a single process, which performed better than either of its parts alone in both simulations and real-world experiments. The researchers presented the work June 24 at the Robotics: Science and Systems conference.

“In the future, I fully expect there to be more autonomous systems in the world and they are going to need some concept of what is good and what is bad,” said Andy Palan, graduate student in computer science and co-lead author of the paper. “It’s crucial, if we want to deploy these autonomous systems in the future, that we get that right.”

The team’s new system for providing instruction to robots – known as reward functions – combines demonstrations, in which humans show the robot what to do, and user preference surveys, in which people answer questions about how they want the robot to behave.

“Demonstrations are informative but they can be noisy. On the other hand, preferences provide, at most, one bit of information, but are way more accurate,” said Sadigh. “Our goal is to get the best of both worlds, and combine data coming from both of these sources more intelligently to better learn about humans’ preferred reward function.”  .... " 

Augmented Reality Changes Real-World Behavior

I would expect, but will it change the specific useful goal behaviors in the real world?

Augmented reality can change your behavior in the real world
Even after you take the goggles off.

By Holly Brockwell, @holly

A new study from Stanford's School of Humanities and Sciences has found that augmented reality (AR) experiences significantly affect people's behavior in the real world, even after they've taken the headset off.

Using 218 participants and a pair of AR goggles, researchers led by Professor Jeremy Bailenson conducted three experiments.

The first showed a realistic 3D person called Chris sitting on a real chair in the room (AR layers digital images over the physical world, rather than creating a whole new world like VR). Participants had to complete anagram tasks while Chris watched, and as with the presence of a real person in the room, his presence meant they found hard puzzles more difficult than without 'someone' watching them.

The second experiment looked at whether participants would sit in the chair previously occupied by Chris. Even though he was no longer there, none of the participants still wearing the AR headset sat in that chair. Without the headset, 72% still avoided Chris's chair and sat in the one next to it instead. ... "

Predicting Patterns in Large Data Streams

Its all about patterns in context, whether to gain insight, predict or prescribe.  So its analytics too.  But from my perspective,  AI occurs when we insert the analysis it into a useful process.  And 'intelligence' when it interacts to provide value in a selection of domains that add up to context that provides significant value to humans or business.

MIT CSAIL’s machine learning algorithm helps predict patterns in large data streams   By Kyle Wiggers

Ever heard of the “Britney Spears problem“? Contrary to what it sounds like, it’s got nothing to do with the dalliances of the rich and famous. Rather, it’s a computing puzzle related to data tracking: Precisely tailoring a data-rich service, like a search engine or fiber internet connection, to individual users hypothetically requires tracking every packet sent to and from the service provider, which needless to say isn’t practical. To get around this, most companies leverage algorithms that make guesses about the frequency of data exchanged by hashing it (i.e., divvying it up into pieces). But this necessarily sacrifices nuance — telling patterns that emerge naturally in large data volumes fly under the radar.

Luckily, researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) believe they’ve devised a viable alternative that relies on machine learning. In a newly published paper (“Learning-Based Frequency Estimation Algorithms“), they describe a system — dubbed LearnedSketch, because of the way it “sketches” data in a data stream — that predicts if specific data elements will appear more frequently than others and, if they in fact do, autonomously separates them from the rest of the hashed portions. ... "