Schools for children of military members achieve results rarely seen in public education. From a report: Amy Dilmar, a middle-school principal in Georgia, is well aware of the many crises threatening American education. The lost learning that piled up during the coronavirus pandemic. The gaping inequalities by race and family income that have only gotten worse. A widening achievement gap between the highest- and lowest-performing students. But she sees little of that at her school in Fort Moore, Ga. The students who solve algebra equations and hone essays at Faith Middle School attend one of the highest-performing school systems in the country. It is run not by a local school board or charter network, but by the Defense Department. With about 66,000 students -- more than the public school enrollment in Boston or Seattle -- the Pentagon's schools for children of military members and civilian employees quietly achieve results most educators can only dream of.

On the National Assessment of Educational Progress, a federal exam that is considered the gold standard for comparing states and large districts, the Defense Department's schools outscored every jurisdiction in math and reading last year and managed to avoid widespread pandemic losses. Their schools had the highest outcomes in the country for Black and Hispanic students, whose eighth-grade reading scores outpaced national averages for white students. Eighth graders whose parents only graduated from high school -- suggesting lower family incomes, on average -- performed as well in reading as students nationally whose parents were college graduates. The schools reopened relatively quickly during the pandemic, but last year's results were no fluke. While the achievement of U.S. students overall has stagnated over the last decade, the military's schools have made gains on the national test since 2013. And even as the country's lowest-performing students -- in the bottom 25th percentile -- have slipped further behind, the Defense Department's lowest-performing students have improved in fourth-grade math and eighth-grade reading.

AI startup Anthropic, writing in a blog post: Neural networks are trained on data, not programmed to follow rules. With each step of training, millions or billions of parameters are updated to make the model better at tasks, and by the end, the model is capable of a dizzying array of behaviors. We understand the math of the trained network exactly -- each neuron in a neural network performs simple arithmetic -- but we don't understand why those mathematical operations result in the behaviors we see. This makes it hard to diagnose failure modes, hard to know how to fix them, and hard to certify that a model is truly safe. Neuroscientists face a similar problem with understanding the biological basis for human behavior. The neurons firing in a person's brain must somehow implement their thoughts, feelings, and decision-making. Decades of neuroscience research has revealed a lot about how the brain works, and enabled targeted treatments for diseases such as epilepsy, but much remains mysterious. Luckily for those of us trying to understand artificial neural networks, experiments are much, much easier to run. We can simultaneously record the activation of every neuron in the network, intervene by silencing or stimulating them, and test the network's response to any possible input.

Unfortunately, it turns out that the individual neurons do not have consistent relationships to network behavior. For example, a single neuron in a small language model is active in many unrelated contexts, including: academic citations, English dialogue, HTTP requests, and Korean text. In a classic vision model, a single neuron responds to faces of cats and fronts of cars. The activation of one neuron can mean different things in different contexts. In our latest paper, Towards Monosemanticity: Decomposing Language Models With Dictionary Learning , we outline evidence that there are better units of analysis than individual neurons, and we have built machinery that lets us find these units in small transformer models. These units, called features, correspond to patterns (linear combinations) of neuron activations. This provides a path to breaking down complex neural networks into parts we can understand, and builds on previous efforts to interpret high-dimensional systems in neuroscience, machine learning, and statistics. In a transformer language model, we decompose a layer with 512 neurons into more than 4000 features which separately represent things like DNA sequences, legal language, HTTP requests, Hebrew text, nutrition statements, and much, much more. Most of these model properties are invisible when looking at the activations of individual neurons in isolation.

dcblogs writes: The U.S. Supreme Court declined to hear a challenge against the Optional Practical Training (OPT) program, which allows STEM graduates to work in the U.S. for up to three years on a student F-1 visa. John Miano, the attorney representing WashTech, the labor group that brought the appeal, called the decision "staggering." He said it "strips Congress of the ability to control nonimmigrant programs," such as OPT, the H-1B program, and other programs designed to provide temporary guest workers. In the most extreme example of what the decision may allow, Miano said it theoretically enables the White House to let people on tourist visas work. The decision "gives more authority to the federal government to do what it wants," he said.

The OPT program permits STEM (Science, Technology, Engineering, and Math) graduates to work for up to three years under a student F-1 visa. Critics of the program said it brought unfair competition to the U.S. labor market. Ron Hira, an associate professor of Public Policy at Howard University, said the U.S. administration of the OPT program is so poor that "the program has effectively no controls, accountability, or worker protections."

A group of Senate Republicans, including U.S. Sen. Ted Cruz, argued in briefs filed with the court that the federal government was using the OPT program to sidestep the annual H-1B visa cap. More than 30 Republican House members also filed a brief in support.

An anonymous reader quotes a report from Ars Technica: [R]esearchers from the University of Tubingen found for the first time that crows can perform statistical reasoning. These results can help scientists better understand the evolution of intelligence (and may give us a better appreciation of what's going on in our backyard). [...] Dr. Melissa Johnston, a Humboldt Fellow at the University of Tubingen, certainly appreciated the specialness of these creatures, as she and her colleagues have been studying these animals for several years. "In our lab, it has been shown that crows have sophisticated numerical competence, demonstrate abstract thinking, and show careful consideration during decision-making," she said. In her most recent experiment, Johnston and her team pushed these abilities to a new extreme, testing statistical reasoning.

To do this, Johnston and her team began by training two crows to peck at various images on touchscreens to earn food treats. From this simple routine of peck-then-treat, the researchers significantly raised the stakes. "We introduce the concept of probabilities, such as that not every peck to an image will result in a reward," Johnston elaborated. "This is where the crows learn the unique pairings between the image on the screen and the likelihood of obtaining a reward." The crows quickly learned to associate each of the images with a different reward probability. In the experiment, the two crows had to choose between two of these images, each corresponding to a different reward probability. "Crows were tasked with learning rather abstract quantities (i.e., not whole numbers), associating them with abstract symbols, and then applying that combination of information in a reward maximizing way," Johnston said. Over 10 days of training and 5,000 trials, the researchers found that the two crows continued to pick the higher probability of reward, showing their ability to use statistical inference.

Pushing the crows even further, Johnston and her team waited a whole month before testing the crows again. Even after a month without training, the crows remembered the reward probabilities and could pick the highest number every time. Johnston and her team were excited that the crows could apply statistical reasoning in almost any setting to ensure their reward. "Working with the birds every day is very rewarding! They are very responsive animals, so I enjoy spending time with them," added Johnston. The findings have been published in the journal Current Biology.

An anonymous reader quotes a report from Ars Technica: On Thursday, AI-maker and OpenAI competitor Anthropic launched Claude Pro, a subscription-based version of its Claude.ai web-based AI assistant, which functions similarly to ChatGPT. It's available for $20/month in the US or 18 pounds/month in the UK, and it promises five-times-higher usage limits, priority access to Claude during high-traffic periods, and early access to new features as they emerge. Like ChatGPT, Claude Pro can compose text, summarize, do analysis, solve logic puzzles, and more.

Claude.ai is what Anthropic offers as its conversational interface for its Claude 2 AI language model, similar to how ChatGPT provides an application wrapper for the underlying models GPT-3.5 and GPT-4. In February, OpenAI chose a subscription route for ChatGPT Plus, which for $20 a month also gives early access to new features, but it also unlocks access to GPT-4, which is OpenAI's most powerful language model. What does Claude have that ChatGPT doesn't? One big difference is a 100,000 token context window, which means it can process about 75,000 words at once. Tokens are fragments of words used while processing text. That means Claude can analyze longer documents or hold longer conversations without losing its memory of the subject at hand. ChatGPT can only process about 8,000 tokens in GPT-4 mode.

Anthropic's primary selling point for the Claude Pro subscription is "5x more usage," but the company doesn't clearly communicate what Claude's free-tier usage limits actually are. Dropping clues like cryptic breadcrumbs, the company has written a support document about the topic that says, "If your conversations are relatively short (approximately 200 English sentences, assuming your sentences are around 15-20 words), you can expect to send at least 100 messages every 8 hours, often more depending on Claude's current capacity. Over two thirds of all conversations on claude.ai (as of September 2023) have been within this length." In another somewhat cryptic statement, Anthropic writes, "If you upload a copy of The Great Gatsby, you may only be able to send 20 messages in that conversation within 8 hours." We're not attempting the math, but if you know the precise word count of F. Scott Fitzgerald's classic, it may be possible to glean Claude's actual limits. We reached out to Anthropic for clarification yesterday and have not received a response by press time.

"Given that too many schools don't teach calculus, chemistry and physics, CalTech is allowing potential undergraduates to demonstrate their ability in these fields by using Khan Academy," writes Slashdot reader Bruce66423. Los Angeles Times reports: One of Caltech's alternative paths is taking Khan Academy's free, online classes and scoring 90% or higher on a certification test. Sal Khan, academy founder, said Caltech's action is a "huge deal" for equitable access to college. While Caltech is small -- only 2,400 students, about 40% of them undergraduates -- Khan said he hoped its prestigious reputation would encourage other institutions to examine their admission barriers and find creative solutions to ease them. The Pasadena-based institute, with a 3% admission rate last year, boasts 46 Nobel laureates and cutting-edge research in such fields as earthquake engineering, behavioral genetics, geochemistry, quantum information and aerospace. "You have one of the most academically rigorous schools on the planet that has arguably one of the highest bars for admission, saying that an alternative pathway that is free and accessible to anyone is now a means to meeting their requirements," said Khan, whose nonprofit offers free courses, test prep and tutoring to more than 152 million users. [...]

The impetus for the policy change began in February, when Pallie, the admissions director, and two Caltech colleagues attended a workshop on equity hosted by the National Assn. for College Admission Counseling. They were particularly struck by one speaker, Melodie Baker of Just Equations, a nonprofit that seeks to widen math opportunities. As Baker pointed out the lack of access to calculus for many students, Pallie and her team began to question Caltech's admission requirement for the course, along with physics and chemistry. Pallie and Jared Leadbetter, a professor of environmental microbiology who heads the faculty admissions committee, began to look into potential course alternatives. Pallie connected with Khan's team, which started a second nonprofit, Schoolhouse.world, during the pandemic in 2020 to offer free tutoring. Peer tutors on the platform certify they are qualified for their jobs by scoring at least 90% on the course exam and videotaping themselves explaining how they solved each problem on it. The video helps ensure that the students actually took the exam themselves and understand the material. That video feature gave Caltech assurances about the integrity of the alternative path.

Under the new process, students would take a calculus, physics or chemistry class offered by Khan Academy and use the Schoolhouse platform to certify their mastery of the content as tutors do with a 90% score or better on the exam and a videotaped explanation of their reasoning. Proof of certification is required within one week of the application deadline, which is in November for early action and January for regular decisions. Pallie and Leadbetter also wanted to test whether the Khan Academy courses are sufficiently rigorous. Several Caltech undergraduates took the courses to assess whether all concepts were covered in enough breadth and depth to pass the campus placement exams in those subjects. Miranda, a rising Caltech junior studying mechanical engineering, took the calculus course and gave it a thumbs-up, although she added that students would probably want to use additional textbooks and other study materials to deepen their preparation for Caltech.

An anonymous reader quotes a report from TechCrunch: Seeking to capitalize on ChatGPT's viral success, OpenAI today announced the launch of ChatGPT Enterprise, a business-focused edition of the company's AI-powered chatbot app. ChatGPT Enterprise, which OpenAI first teased in a blog post earlier this year, can perform the same tasks as ChatGPT, such as writing emails, drafting essays and debugging computer code. But the new offering also adds "enterprise-grade" privacy and data analysis capabilities on top of the vanilla ChatGPT, as well as enhanced performance and customization options. That puts ChatGPT Enterprise on par, feature-wise, with Bing Chat Enterprise, Microsoft's recently launched take on an enterprise-oriented chatbot service.

ChatGPT Enterprise provides a new admin console with tools to manage how employees within an organization use ChatGPT, including integrations for single sign-on, domain verification and a dashboard with usage statistics. Shareable conversation templates allow employees to build internal workflows leveraging ChatGPT, while credits to OpenAI's API platform let companies create fully custom ChatGPT-powered solutions if they choose. ChatGPT Enterprise, in addition, comes with unlimited access to Advanced Data Analysis, the ChatGPT feature formerly known as Code Interpreter, which allows ChatGPT to analyze data, create charts, solve math problems and more, including from uploaded files. For example, given a prompt like "Tell me what's interesting about this data," ChatGPT's Advanced Data Analysis capability can look through the data -- financial, health or location information, for example -- to generate insights.

Advanced Data Analysis was previously available only to subscribers to ChatGPT Plus, the $20-per-month premium tier of the consumer ChatGPT web and mobile apps. To be clear, ChatGPT Plus is sticking around -- OpenAI sees ChatGPT Enterprise as complementary to it, the company says. ChatGPT Enterprise is powered by GPT-4, OpenAI's flagship AI model, as is ChatGPT Plus. But ChatGPT Enterprise customers get priority access to GPT-4, delivering performance that's twice as fast as the standard GPT-4 and with an expanded 32,000-token (~25,000-word) context window. Context window refers to the text the model considers before generating additional text, while tokens represent raw text (e.g. the word "fantastic" would be split into the tokens "fan," "tas" and "tic"). Generally speaking, models with large context windows are less likely to "forget" the content of recent conversations. Crucially, OpenAI said that it "won't train models on business data sent to ChatGPT Enterprise or any usage data and that all conversations with ChatGPT Enterprise are encrypted in transit and at rest," notes TechCrunch.

"OpenAI says that its future plans for ChatGPT Enterprise include a ChatGPT Business offering for smaller teams, allowing companies to connect apps to ChatGPT Enterprise, 'more powerful' and 'enterprise-grade' versions of Advanced Data Analysis and web browsing, and tools designed for data analysts, marketers and customer support."

A blog post introducing ChatGPT Enterprise can be found here.

72-year-old Bjarne Stroustrup invented C++ (first released in 1985). 38 years later, he gave a short interview for Honeypot.io (which calls itself "Europe's largest tech-focused job platform") offering his own advice for life: Don't overspecialize. Don't be too sure that you know the future. Be flexible, and remember that careers and jobs are a long-term thing. Too many young people think they can optimize something, and then they find they've spent a couple of years or more specializing in something that may not have been the right thing. And in the process they burn out, because they haven't spent enough time building up friendships and having a life outside computing.

I meet a lot of sort of — I don't know what you call them, "junior geeks"? — that just think that the only thing that matters is the speciality of computing — programming or AI or graphics or something like that. And — well, it isn't... And if they do nothing else, well — if you don't communicate your ideas, you can just as well do Sudoku... You have to communicate. And a lot of sort of caricature nerds forget that. They think that if they can just write the best code, they'll change the world. But you have to be able to listen. You have to be able to communicate with your would-be users and learn from them. And you have to be able to communicate your ideas to them.

So you can't just do code. You have to do something about culture and how to express ideas. I mean, I never regretted the time I spent on history and on math. Math sharpens your mind, history gives you some idea of your limitations and what's going on in the world. And so don't be too sure. Take time to have a balanced life.

And be ready for the opportunity. I mean, a broad-based education, a broad-based skill set — which is what you build up when you educate, you're basically building a portfolio of skills — means that you can take advantage of an opportunity when it comes along. You can recognize it sometimes. We have lots of opportunities. But a lot of them, we either can't take advantage of, or we don't notice. It was my fairly broad education — I've done standard computer science, I've done compilers, I've done multiple languages... I think I knew two dozen at the time. And I have done machine architecture, I've done operating systems. And that skill set turned out to be useful.
At the beginning of the video, Stroustrup jokes that it's hard to give advice — and that it's at least as difficult as it is to take advice.

Earlier this year, Bjarne also told the same site the story of how he became a programmer by mistake — misreading a word when choosing what to study afer his high school exams. Stroustrup had thought he was signing up for an applied mathematics course, which instead turned to be a class in computer science...

Anthropic, the AI startup co-founded by ex-OpenAI execs, has released an updated version of its faster, cheaper, text-generating model available through an API, Claude Instant. TechCrunch reports: The updated Claude Instant, Claude Instant 1.2, incorporates the strengths of Anthropic's recently announced flagship model, Claude 2, showing "significant" gains in areas such as math, coding, reasoning and safety, according to Anthropic. In internal testing, Claude Instant 1.2 scored 58.7% on a coding benchmark compared to Claude Instant 1.1, which scored 52.8%, and 86.7% on a set of math questions versus 80.9% for Claude Instant 1.1. "Claude Instant generates longer, more structured responses and follows formatting instructions better," Anthropic writes in a blog post. "Instant 1.2 also shows improvements in quote extraction, multilingual capabilities and question answering."

Claude Instant 1.2 is also less likely to hallucinate and more resistant to jailbreaking attempts, Anthropic claims. In the context of large language models like Claude, "hallucination" is where a model generates text that's incorrect or nonsensical, while jailbreaking is a technique that uses cleverly-written prompts to bypass the safety features placed on large language models by their creators. And Claude Instant 1.2 features a context window that's the same size of Claude 2's -- 100,000 tokens. Context window refers to the text the model considers before generating additional text, while tokens represent raw text (e.g. the word "fantastic" would be split into the tokens "fan," "tas" and "tic"). Claude Instant 1.2 and Claude 2 can analyze roughly 75,000 words, about the length of "The Great Gatsby." Generally speaking, models with large context windows are less likely to "forget" the content of recent conversations.

A recently released research reveals a fundamental challenge of developing artificial intelligence: ChatGPT has become worse at performing certain basic math operations. From a report: The researchers at Stanford University and the University of California, Berkeley said the deterioration is an example of a phenomenon known to AI developers as drift, where attempts to improve one part of the enormously complex AI models make other parts of the models perform worse.

[...] Thus far, they have tested two versions of ChatGPT: version 3.5, available free online to anyone, and version 4.0, available via a premium subscription. The results aren't entirely promising. They gave the chatbot a basic task: identify whether a particular number is a prime number. This is the sort of math problem that is complicated for people but simple for computers.

Is 17,077 prime? Is 17,947 prime? Unless you are a savant you can't work this out in your head, but it is easy for computers to evaluate. A computer can just brute force the problem -- try dividing by two, three, five, etc., and see if anything works. To track performance, the researchers fed ChatGPT 1,000 different numbers. In March, the premium GPT-4, correctly identified whether 84% of the numbers were prime or not. (Pretty mediocre performance for a computer, frankly.) By June its success rate had dropped to 51%. Across eight different tasks, GPT-4 became worse at six of them. GPT-3.5 improved on six measures, but remained worse than its advanced sibling at most of the tasks.

A new study (PDF) from Stanford found that ChatGPT performed worse on certain tasks in June than its March version. The paper supports a widely held, though unproven, notion that the AI language model's performance in coding and compositional tasks has deteriorated in recent months. Fortune reports: The study compared the performance of the chatbot, created by OpenAI, over several months at four "diverse" tasks: solving math problems, answering sensitive questions, generating software code, and visual reasoning. Researchers found wild fluctuations -- called drift -- in the technology's ability to perform certain tasks. The study looked at two versions of OpenAI's technology over the time period: a version called GPT-3.5 and another known as GPT-4. The most notable results came from research into GPT-4's ability to solve math problems.

Over the course of the study researchers found that in March GPT-4 was able to correctly identify that the number 17077 is a prime number 97.6% of the times it was asked. But just three months later, its accuracy plummeted to a lowly 2.4%. Meanwhile, the GPT-3.5 model had virtually the opposite trajectory. The March version got the answer to the same question right just 7.4% of the time -- while the June version was consistently right, answering correctly 86.8% of the time. Similarly varying results happened when the researchers asked the models to write code and to do a visual reasoning test that asked the technology to predict the next figure in a pattern.

James Zou, a Stanford computer science professor who was one of the study's authors, says the "magnitude of the change" was unexpected from the "sophisticated ChatGPT." The vastly different results from March to June and between the two models reflect not so much the model's accuracy in performing specific tasks, but rather the unpredictable effects of changes in one part of the model on others. [...] The exact nature of these unintended side effects is still poorly understood because researchers and the public alike have no visibility into the models powering ChatGPT. It's a reality that has only become more acute since OpenAI decided to backtrack on plans to make its code open source in March. "These are black-box models," Zou says. "So we don't actually know how the model itself, the neural architectures, or the training data have changed."

Twitter has started sending out the first payouts to creators on the platform who are part of the company's revenue sharing program. The largest payout reported thus far was to Billy Markus, the co-creator of the Dogecoin cryptocurrency, which amounted to a whopping $37,050. TechCrunch reports: Users who subscribe to Twitter Blue and have earned more than 5 million tweet impressions each month for the last 3 months are eligible to join. According to owner Elon Musk, the first round of creator payouts will total $5 million, and will be cumulative from the month of February onward. These payouts will be delivered via Stripe. [...] Twitter's payouts are determined by tweet impressions. Babylon Bee writer Ashley St. Clair (710,000 followers) said that she earned $7,153, and according to her "napkin math," she had around 840 million impressions from February through July. That would make her rate about $0.0085 CPM (cost per mille), or $8.52 per million impressions. It's not clear whether or not individual CPMs change from user to user.

Anthropic, an artificial intelligence startup positioning itself as the builder of a safer kind of chatbot, has released a new version of its AI bot, named Claude. From a report: Anthropic said that Claude 2 is available to anyone in the US or UK online at claude.ai, and businesses can access it via an application programming interface. The new release on Tuesday comes several months after Anthropic began offering an earlier version of Claude to businesses that wanted to add it to their products. Previously, the bot was tested by a handful of companies including Quora, which built it into an app called Poe that lets users ask questions.

Like its predecessor, Claude 2 is built atop a large language model and can be used for written tasks like summarizing, searching, answering questions and coding. Both models can currently take in large chunks of text -- a user can ask it to summarize a book, for instance -- though Claude 2 can generate longer responses than its predecessor. Responses can reach up to about 3,000 words, according to data provided by the company. Claude 2 will also offer more accurate responses on some topics, such as coding and grade-school-level math, the company said. Anthropic's goal has been for Claude to be less susceptible than other chatbots to manipulation.

Consumers are winning from the streaming revolution but across most of Hollywood, the businesses churning out TV and movies are losing. From a report: Services such as Netflix, Disney+, Paramount+ and Max have become the default entertainment options for homes across America rather than cable, saving many consumers money. For the titans of Hollywood, that shift has been costly. Traditional media and entertainment companies have reported losses of more than $20 billion combined since early 2020 on their direct-to-consumer streaming businesses. Netflix, which brings in profits, is an exception, but the rest of the industry is wondering: While consumers love streaming, is it actually a good business?

Investors now care about profitability rather than growth, a change that makes finding new revenue streams and retaining customers critical. Studios that for years were able to splurge on content to feed viewers' insatiable appetite for new shows and films now must pull back to make the math work. The ad market is weakening, many companies have laid off staff to save money and Hollywood writers are on strike. Market values for Paramount Global, Comcast, Walt Disney and Netflix are down more than $280 billion combined since the end of 2020. Warner Bros. Discovery is worth about half of its total value since its 2022 trading debut as a combined company. The declines have come after many of the stocks rose during the early part of the pandemic, when consumers were stuck at home and hungry for entertainment.

"Picture a world where computing is not limited by the binary confines of zeros and ones, but instead, is free to explore the vast possibilities of continuous value data." That's Microsoft's research blog, describing its newly-developed Analog Iterative Machine, an analog optical computer designed for solving difficult optimization problems.

"For a multidisciplinary group of researchers at the Microsoft Research Lab in Cambridge, U.K., the mission was to build a new kind of computer that would transcend the limitations of the binary systems," says a Microsoft blog post.

Neowin describes it as a computer "that uses photons and electrons, rather than transistors, to process data." Light "passes through several layers, making impressions on each part of what's known as a 'modular array'," writes PC Gamer. "It's this process of projecting light through the array that replaces the function of a standard transistor."

Microsoft says it can "solve practical problems at the speed of light." And "it's already shown potential for surpassing state-of-the art digital (silicon-based) technology," adds TechSpot, "or even the most powerful quantum computers being designed right now." The AIM machine is built using commodity opto-electronic technologies that are low-cost and scalable, Microsoft says, and is based on an "asynchronous data flow architecture" which doesn't require data exchange between storage units and "compute locations."

AIM isn't designed for general purpose computing tasks, though. The analog optical computer is useful to solve difficult "optimization problems" like the well-known travelling salesman riddle, Microsoft says, which are at the heart of many, math-intensive industries including finance, logistics, transportation, energy, healthcare, and manufacturing. When it comes to crunching all the possible combinations of an exponentially growing problem, traditional, digital computers struggle to provide a solution in a "timely, energy-efficient and cost-effective manner."

AIM was conceived to address two simultaneous trends, Microsoft explains, which are sidestepping the unraveling of Moore's Law and overcoming the limitations of specialized machines designed for solving optimization problems... AIM works at the speed of light, and it seemingly provides a 100x increase in performance compared to the most advanced digital approaches available today. For now, AIM is still a research project with limited access for potential customers. The machine, however, is already being tested by UK financial company Barclays, which is using it to track transactions of money into stock purchases.
Microsoft says it's now releasing its "AIM simulator as a service, allowing selected users to get first-hand experience. The initial users are the team's collaborators at Princeton University and at Cambridge University."

Slashdot reader silverjacket writes: Researchers at Oxford and elsewhere developed a method that figures out the most likely number and size of dimension in which to format a string of bits, with applications to interpreting messages from extraterrestrial intelligence (METI), if we were to receive them.
The new method "looks at every possible combination of dimension number and size," according to Science News: The researchers also measure each possible configuration's global order by seeing how much an image compression algorithm can shrink it without losing information — mathematically, randomness is less compressible than regular patterns...
Hector Zeni [one of the creators of this method] "notes that in Carl Saganâ(TM)s sci-fi novel Contact, the characters spend a lot of time figuring out that a message received from aliens is in three dimensions (specifically a video). âoeIf you have our tools, you would solve that problem in seconds and with no human intervention.â An algorithm that pieces together smaller algorithmic components in order to explain or predict data — this new method is just one way to do it — may also help us one day achieve artificial general intelligence, Zenil says. Such automated approaches don't depend on human assumptions about the signal. That opens the door to discovering forms of intelligence that might think differently from our own.

The average test scores for 13-year-old students in the U.S. have decreased in reading and math since 2020, reaching the lowest levels in decades, with more significant declines in math. NPR reports: The average scores, from tests given last fall, declined 4 points in reading and 9 points in math, compared with tests given in the 2019-2020 school year, and are the lowest in decades. The declines in reading were more pronounced for lower performing students, but dropped across all percentiles. The math scores were even more disappointing. On a scale of 500 points, the declines ranged from 6 to 8 points for middle and high performing students, to 12 to 14 points for low performing students.

The math results also showed widening gaps based on gender and race. Scores decreased by 11 points for female students over 2020 results, compared with a 7-point decrease for male students. Among Black students, math scores declined 13 points, while white students had a 6-point drop. Compared with the 35-point gap between Black and white students in 2020, the disparity widened to 42 points.

While the scores show a drop from the pre-pandemic years, the results also show that there are other factors at work. The decline is even more substantial when compared with scores of a decade ago: The average scores declined 7 points in reading and 14 points in mathematics. The Education Department says plans are underway to address the learning loss. [...] The latest results are from the NAEP Long-Term Trend Assessment, traditionally administered every four years by the National Center for Education Statistics.

Reddit CEO Steve Huffman says he is refusing to undo the company's decision to increase prices for third-party app developers, despite thousands of subreddits pledging to keep their subreddits private or restricted in protest. "It's a startling change for many members of the Reddit community, but it's one that Reddit CEO Steve Huffman tells The Verge that he's fine with making," writes The Verge's Jay Peters. "Those third-party apps, in his eyes, aren't adding much value to the platform." From the report: "So the vast majority of the uses of the API -- not [third-party apps like Apollo for Reddit] -- the other 98 percent of them, make tools, bots, enhancements to Reddit. That's what the API is for," Huffman says. "It was never designed to support third-party apps." According to Huffman, he "let it exist," and "I should take the blame for that because I was the guy arguing for that for a long time." Huffman now takes issue with the third-party apps that are building a business on top of his own. "I didn't know -- and this is my fault -- the extent that they were profiting off of our API. That these were not charities."

I asked him if he felt that Apollo, rif for Reddit, and Sync, which all plan to shut down as a result of the pricing changes, don't add value to Reddit. "Not as much as they take," he says. "No way." "They need to pay for this. That is fair. What our peers have done is banned them entirely. And we said no, you know what, we believe in free markets. You need to cover your costs," he says. Apollo developer Christian Selig recently did the math for us on The Vergecast, though, and suggested that covering Reddit's asking price with only 30 days' notice would have been nigh-impossible.

Huffman didn't have an answer for why the deadline was so short, beyond wanting there to be a deadline. "We're perfectly willing to work with the folks who want to work with us, including figuring out what the transition period will look like. But I think a deadline forces people, us included, to negotiate that." I also asked if Huffman truly believes that the blackouts haven't impacted his decision-making around the API pricing changes at all. "In this case? That's true," says Huffman. "That's our business decision, and we're not undoing that business decision."

In a blog post on Wednesday, Google said Bard is getting better at logic and reasoning. "Google says that now when you ask Bard a 'computational' task like math or string manipulation, instead of showing the output of the language model, that language model will instead write a program, execute that program, and then show the output of that program to the user as an answer," reports Ars Technica. From the report: Google's blog post provides the example input of "Reverse the word 'Lollipop' for me." ChatGPT flubs this question and provides the incorrect answer "pillopoL," because language models see the world in chunks of words, or "tokens," and they just aren't good at this. It gets the output correct as "popilloL," but more interesting is that it also includes the python code it wrote to answer the question. That's neat for programming-minded people to see under the hood, but wow, is that probably the scariest output ever for regular people. It's also not particularly relevant. Imagine if Gmail showed you a block of code when you just asked it to fetch email. It's weird. Just do the job you were asked to do, Bard.

Google likens an AI model writing a program to humans doing long division in that it's a different mode of thinking [...]. Google says this "writing code on the fly" method will also be used for questions like: "What are the prime factors of 15683615?" and "Calculate the growth rate of my savings." The company says, "So far, we've seen this method improve the accuracy of Bard's responses to computation-based word and math problems in our internal challenge datasets by approximately 30%." As usual, Google warns Bard "might not get it right" due to interpreting your question wrong or just, like all of us, writing code that doesn't work the first time. Bard is coding up answers on the fly right now if you want to give it a shot at bard.google.com.

TV's streaming model is broken. It's also not going away. For Hollywood, figuring that out will be a horror show. From a report: Across the town, there's despair and creative destruction and all sorts of countervailing indicators. Certain shows that were enthusiastically green-lit two years ago probably wouldn't be made now. Yet there are still streamers burning mountains of cash to entertain audiences that already have too much to watch. Netflix has tightened the screws and recovered somewhat, but the inarguable consensus is that there is still a great deal of pain to come as the industry cuts back, consolidates, and fumbles toward a more functional economic framework. The high-stakes Writers Guild of America strike has focused attention on Hollywood's labor unrest, but the really systemic issue is streaming's busted math. There may be no problem more foundational than the way the system monetizes its biggest hits: It doesn't.

Just ask Shawn Ryan. In April, the veteran TV producer's latest show, the spy thriller The Night Agent, became the fifth-most-watched English-language original series in Netflix's history, generating 627 million viewing hours in its first four weeks. As it climbed to the heights of such platform-defining smashes as Stranger Things and Bridgerton, Ryan wondered how The Night Agent's success might be reflected in his compensation. "I had done the calculations. Half a billion hours is the equivalent of over 61 million people watching all ten episodes in 18 days. Those shows that air after the Super Bowl -- it's like having five or ten of them. So I asked my lawyer, 'What does that mean?'" recalls Ryan. As it turns out, not much. "In my case, it means that I got paid what I got paid. I'll get a little bonus when season two gets picked up and a nominal royalty fee for each additional episode that gets made. But if you think I'm going out and buying a private jet, you're way, way off."

Ryan says he'll probably make less money from The Night Agent than he did from The Shield, the cop drama he created in 2002, even though the latter ran on the then-nascent cable channel FX and never delivered Super Bowl numbers. "The promise was that if you made the company billions, you were going to get a lot of millions," he says. "That promise has gone away." Nobody is crying for Ryan, of course, and he wouldn't want them to. ("I'm not complaining!" he says. "I'm not unaware of my position relative to most people financially.") But he has a point. Once, in a more rational time, there was a direct relationship between the number of people who watched a show and the number of jets its creator could buy. More viewers meant higher ad rates, and the biggest hits could be sold to syndication and international markets. The people behind those hits got a cut, which is why the duo who invented Friends probably haven't flown commercial since the 1990s. Streaming shows, in contrast, have fewer ads (or none at all) and are typically confined to their original platforms forever. For the people who make TV, the connection between ratings and reward has been severed.