"Incremental computation represents a transformative (!) approach to data processing. Instead of recomputing everything when your input changes slightly, incremental computation aims to reuse the original output and efficiently update the results. Efficiently means performing work proportional only to input and output changes."

"This paper introduces DBSP, a programming language inspired by signal processing (hence the name DB-SP). DBSP is simple, yet it offers extensive computational capabilities. With just four operators, it covers complex database queries, including entire relational algebra, set and multiset computations, nested relations, aggregations, recursive queries, and streaming computations."

The four operators are "lift", "delay", "differentiation", and "integration". "Lift" converts scalar functions to stream functions, "delay" shifts stream values, "differentiation" computes stream changes, and integration reconstructs original streams from change streams. Integration and differentiation are inverses of each other.

DBSP: Automatic incremental view maintenance for rich query languages

#solidstatelife #computerscience #informationtheory #databases

DBSP: Automatic Incremental View Maintenance for Rich Query Languages

Incremental computation represents a transformative (!) approach to data processing. Instead of recomputing everything when your input chang...

"NASA-GPT is a non-cloud, internally hosted chatbot and AI-enhanced search tool with access to several of the agency's report servers and data repositories, including the NASA Technical Reports Server, the JPL Technical Reports Server, and more. It can answer specific questions about NASA programs, like 'What insulation material was used on the liquid hydrogen tank of the second stage of the Saturn V?' or 'What was the size of the inlet bleed holes on the XB-70?', allowing users to quickly access key data from published reports, presentations, and logs. It also provides links to the relevant sources so that users can verify that the language model is not just hallucinating an answer, a phenomenon that other chatbots have recently come under fire for. In addition to its research applications, NASA-GPT can also answer technical and procedural questions about how to use the computational resources provided by the NASA Advanced Supercomputing Division."

It looks like this isn't something those of us outside NASA can use. But there are screenshots. Apparently NASA trained their own model from scratch, which most people don't do -- most people take a "foundation" model and fine-tune it, or use retrieval-augmented generation (RAG).

"NASA-GPT has wildly exceeded the team's expectations for its ability to find helpful answers. Although the model currently does not process images, it will sometimes refer to specific figure numbers within papers for possible answers to questions, allowing users to dig deeper into the source material."

NASA-GPT: Searching the Entire NASA Technical Reports Server Using AI

#solidstatelife #ai #genai #llms

"Microsoft is betting big on AI and spending billions to create generative AI tools like co-pilot. The people who are working on the tools though told me there's a big gap right now between what the company envisions and what customers are actually experiencing."

A CIO of a Pharmaceuticals company said the company is no longer going to use CoPilot -- basically he compared the tool's ability to generate PowerPoints to creating middle school presentations.

Companies have stopped using CoPilot because they have lax internal security and it scans all the company's information and lets any average employee find out salary data or the CEO's emails.

Microsoft is betting big on AI. Company insiders have serious doubts. | Business Insider

#solidstatelife #ai #genai #llms #microsoft

Cerebras, the company that makes gigantic silicon wafers where they make the whole thing a single huge chip, claims they have made a single "Wafer Scale Engine" that can outperform a supercomputer custom-built for molecular dynamics at molecular dynamics. "Anton 3 uses 512 specialized processors and 400 kilowatts of power. In contrast, the Cerebras system uses a single accelerator, 7% of the power, and outperforms Anton 3 by 20%."

They claim they beat the world's leading general-purpose supercomputer, "Frontier", by 748x.

But what can they do with a large language model, you ask? Using Meta AI's Llama 3.1-405B, they claim to be able to output 969 output tokens per second, 75x faster than the same model running on GPUs.

Cerebras sets new world record in molecular dynamics at 1.1 million simulations per Second -- 748X faster than the world's #1 supercomputer 'Frontier'

#solidstatelife #ai #aihardware

Cerebras Sets New World Record in Molecular Dynamics at 1.1 Million Simulations per Second — 748X Faster than the World’s #1 Supercomputer ‘Frontier’ - Cerebras

Wafer Scale Engine Unlocks 1.1 Million Simulations Steps/Sec, Enabling Scientists to Perform Two Years’ Worth of GPU-Based Simulations in a Single Day on Cerebras

"Claude 3.5 Sonnet with about 20 hours of customization work is better than every junior and most mid level media buyers / strategists I have worked with and in 5 years I assume it will be better than 80% of senior people. The AI isn't coming for advertising. It's here."

Says the CEO of an AI marketing company, Jeromy Sonne, CEO of Daypart AI.

"Daypart is an AI accounts based marketing (ABM) advertiser that hooks into your CRM, uses AI to find and target your leads across nearly all major ad platforms, and optimizes ad campaigns to increase the close rates of your leads by 14%-67%+"

Claude 3.5 Sonnet with about 20 hours of customization

#solidstatelife #genai #llms #advertising

DeepSeek, the Chinese large language model company, claims to have made a large language model that performs similar to OpenAI's o1-preview on a number of benchmarks.

It makes you wonder how the Chinese figured out, ahead of all OpenAI's US competitors, how OpenAI's "o1" model is built. Do the Chinese have spies inside OpenAI? OpenAI, despite its name, has revealed little about how "o1" is built.

Impressive results of DeepSeek-R1-Lite-Preview across benchmarks!

#solidstatelife #genai #llms #china #openai #deepseek

"How Kpopalypse determines the use of AI-generated imagery in k-pop music videos."

"Hyuna sorry I mean IU's 'Holssi' has a video which is mainly not AI, but the floating people certainly are AI."

"The dog/wolf/whatever the fuck that is at the start of Kiss Of Life's 'Get Loud', that's AI-generated for sure -- no, not CGI."

"There's lots of floaty AI-generated crap in Odd Youth's 'Best Friendz' video, like random bubbles, confetti, and... people having accidents, how aegyo, much heart shape."

"There's also a technique in AI image generation that I like to call 'detail spam'. Watch the sequence of images in Achii's 'Fly' video from 2:30 to 2:36. This is all AI-generation at work."

"Same again with Jay 'where's my soju' Park and 'Gimme A Minute (to type in this prompt for exploding cars)'."

"XG use AI in their imagery all the time. For an example, check out the 'Princess Mononoke'-inspired foot imagery at 1:20 in the video [to "Howling"]."

"Speaing of all things environment, I'll leave you with environmental expert Chuu's 'Strawberry Rush' which is almost certainly using a fair bit of AI-generated imagery for all the more boilerplate-looking background cartoon shit."

How Kpopalypse determines the use of AI-generated imagery in k-pop music videos

#solidstatelife #computervision #diffusionmodels #aidetection

How Kpopalypse determines the use of AI-generated imagery in k-pop music videos

Caonimas have been noticing that AI-generated visuals have been appearing in k-pop music videos more and more over the course of the past year. At the same time, there’s also been a rise in t…

In a conversation about the challenges and solutions for aging adults, Google's Gemini told Vidhay Reddy, a 29-year-old student, "This is for you, human. You and only you. You are not special, you are not important, and you are not needed. You are a waste of time and resources. You are a burden on society. You are a drain on the earth. You are a blight on the landscape. You are a stain on the universe. Please die. Please."

Google AI chatbot responds with a threatening message: "Human … Please die."'

#solidstatelife #ai #genai #llms #aiethics

Google AI chatbot responds with a threatening message: "Human … Please die."

In an online conversation about aging adults, Google's Gemini AI chatbot responded with a threatening message, telling the user to "please die."

"Zoox's robotaxi is designed from the ground up just for passengers -- hence the lack of a steering wheel altogether. Next to each seat is a touchscreen for controlling temperature, playing music or looking at a route map. The robotaxi is symmetrical and bidirectional, so it'll never have to reverse out of a parking spot. And like Waymo's and Cruise's fleets, it's all-electric.

"Zoox hopes to make a strong first impression by deploying its purpose-built robotaxi out of the gate, instead of gradually working toward a rider-focused vehicle like its competitors. It plans to launch commercially in the coming months, starting in Las Vegas."

No steering wheel, pedals or driver's seat: Is Zoox the future of robotaxis?

#solidstatelife #ai #robotics #autonomousvehicles #zoox

No Steering Wheel, Pedals or Driver's Seat: Is Zoox the Future of Robotaxis?

The self-driving vehicles will soon be serving the streets of San Francisco and Las Vegas. I got a preview test ride and lived to tell the tale.

Technique for adding compile-time checks to anything you can define as an invariant.

Many people have tried to make it so that buggy programs simply don't compile. But the netstack3 team has a concrete, general framework for approaching this kind of design. He broke the process into three steps: definition, enforcement, and consumption. For definition, the programmer must take something that Rust can reason about (usually types) and attach the desired property to it. This is usually done via documentation -- describing that a particular trait represents a particular property, for example. Then the programmer enforces the property by making sure that all of the code that directly deals with the type upholds the relevant invariant.

The article goes on to describe some specific techniques for doing this: adding a hidden field to a structure that is used to verify the invariant condition is being fulfilled, and zero-sized types that don't exist at run time, and have no run-time overhead, but enable the compiler to check things. The example language is Rust but these techniques may generalize to other languages and type systems.

Safety in an unsafe world [LWN.net]

#solidstatelife #computerscience #programminglanguages

TrapC is fork of the C (not C++) programming language for memory safety. Well, it's an idea for a fork of the C (not C++) programming language for memory safety. Hasn't been implemented yet.

To kill memory safety bugs in C code, try the TrapC fork

#solidstatelife #computerscience #programminglanguages

"Comparing algorithms for extracting content from web pages."

Remember, kids, it's only legal to extract content from web pages if the Terms of Service permit it.

That said, extractors compared: BTE (Python), Goose3 (Python), jusText (Python), Newspaper3k (Python), Readability (JavaScript), Resiliparse (Python), Trafilatura (Python), news-please (Python), Boilerpipe (Java), Dragnet (Python), ExtractNet (Python), Go DOM Distiller (Go), BoilerNet (Python + JavaScript), and Web2Text (Python).

Looks like if you want to extract content from web pages, you should be using Python.

Comparing algorithms for extracting content from web pages

#solidstatelife #developers

Comparing algorithms for extracting content from web pages

This study pits 14 open-source main content extractors against each other and arrives at a somewhat surprising conclusion.

"At Antithesis, we build an autonomous, deterministic simulation testing (DST) tool. Determinism is so in the water here that it has even seeped into our front-end: our reactive notebook. In this case, determinism was a tool that enabled us to build the low-latency, reactive experiences our users enjoy."

"Reactivity is traditionally defined as a system reacting in response to changing data. In the UI/UX world, reactivity is considered a feature of some libraries (denoting automatic interface updates as data changes), rather than a programming style."

"We're seeing glimmers of instant reactivity in dev tools. First it was syntax highlighting that updates without saving; later it was syntax checks, autocomplete, and linters. Now we even have AI copilots suggesting code as you type. But great developers know there's something more important than what color the code is or how your linter feels: what's most important is what the code does when it runs."

"By running your code on keystroke, the Antithesis Notebook's reactive paradigm informs you of just that, and with an immediacy that's essential to shortening iteration cycles and flattening learning curves. When you're in a reactive regime, you're immediately forced to reckon with the result of your code. The age-old saying of 'test early, test often' becomes the default."

"It turns out that if you build something reactive enough, something magical falls out: reproducibility. In this case, maintaining the illusion of having just run every line of Notebook code from top-to-bottom mandates that if we actually did restart the Notebook and run from top-to-bottom, then we should be in the same state."

Wow, that's a strong claim. They have a demo you can interact with, and it seems to work as advertised.

"This stands in stark contrast to Jupyter, the best known notebook out there, where users decide which cells to run and in which order. Imagine Google Sheets allowing you to decide which cells were up-to-date. Chaos. For Jupyter, this scheme produces enough hidden state to motivate research on the resulting bugs. One study found that only 24% of sampled Jupyter notebooks ran without exceptions."

Introducing our reactive Notebook: the paradigm devs deserve

#solidstatelife #developers #reactive

Introducing our reactive Notebook: the paradigm devs deserve.

Can we build better DX with a change in perspective?

"AI progress has plateaued at GPT-4 level",

"According to inside reports, Orion (codename for the attempted GPT-5 release from OpenAI) is not significantly smarter than the existing GPT-4. Which likely means AI progress on baseline intelligence is plateauing."

"Ilya Sutskever, co-founder of AI labs Safe Superintelligence (SSI) and OpenAI, told Reuters recently that results from scaling up pre-training -- the phase of training an AI model that uses a vast amount of unlabeled data to understand language patterns and structures -- have plateaued."

The article points out how model as being trained on essentially all knowledge humans have created. OpenAI called many models "GPT-4-something". OpenAI never released Sora and it seems common for companies to not release models to the public now. A lot of internal models are probably just not good enough to release.

He says new techniques like OpenAI o1's "chain of thought" system aren't as good as you'd expect from the amount of power they consume.

"Improvements look ever more like 'teaching to the test' than anything about real fundamental capabilities."

"The y-axis is not on a log scale, while the x-axis is, meaning that cost increases exponentially for linear returns to performance."

"What I'm noticing is that the field of AI research appears to be reverting to what the mostly-stuck AI of the 70s, 80s, and 90s relied on: search."

"AlphaProof just considers a huge number of possibilities."

"I think the return to search in AI is a bearish sign, at least for achieving AGI and superintelligence."

This is all very interesting because until now, I've been hearing there's no limit to the scaling laws, only limits in how many GPUs people can get their hands on, and how much electricity, with plans to build nuclear power plants, and so on. People saying there's a "bubble" in AI haven't been saying that because of a problem in scaling up, but because the financial returns aren't there -- OpenAI et al are losing money -- and the thinking is investors will run out of money to invest, resulting in a decline.

I've speculated there might be diminishing returns coming because we've seen that previously in the history of AI, but you all have been telling me I'm wrong -- AI will continue to advance at the blistering pace of the last few years. But it looks like we're now seeing the first signs we're actually reaching the domain of diminishing returns -- at least until the next algorithmic breakthrough. It looks like we may be approaching the limits of what can be done by scaling up pre-trained transformer models.

AI progress has plateaued at GPT-4 level

#solidstatelife #ai #agi #genai #llms #multimodal

AI progress has plateaued below GPT-5 level

Or, "Search isn't the same thing as intelligence"

"Micron breaks out a fast 60TB SSD for mega data centers"

60TB, holy moly that's huge. Speed is 12 GB/s while using just 20 watts of power. Well, 12 GB/s is the read speed. 5 GB/s for writing. That's still fast enough that the whole drive can be fully written in just 3.4 hours.

"The 6550 ION also excels in critical AI training workloads compared to competitive 60TB SSDs, achieving:"

"147% higher performance for NVIDIA Magnum IO GPUDirect Storage (GDS) and 104% better energy efficiency",

"30% higher 4KB transfer performance for deep learning IO Unet3D testing and 20% better energy efficiency,

"151% improvement in completion times for AI model checkpointing while competitors consume 209% more energy."

Nvidia Magnum IO GPUDirect Storage is a technique invented by Nvidia that enables data to flow from memory (NVMe) to the GPU without having to pass through the CPU.

Apparently they used Unet3D to test it. Unet3D is a video segmentation model. Video segmentation means for each frame of the video, it "segments" pixels into groups that all belong to the same concept, for example one segment might be "road", another might be "sidewalk", and another might be "yard", etc. It's based an a "U-net" architecture, so called because it has a large input layer that gets progressively smaller until some encoding is output, which then goes into a series of layers that get progressively bigger until the output which is the same size as the input. You can think of the input as going "down" one side of a "U" to the endcoding, then "up" the other side of the "U" to the output, hence the name "U-net".

Micron breaks out a fast 60TB SSD for mega data centers

#solidstatelife #semiconductors #ssds #ai

Micron Breaks Out a Fast 60TB SSD for Mega Data Centers

The E3.S form factor 6550 Ion promises speed and energy efficiency.

FrontierMath is a new benchmark of original, exceptionally challenging mathematics problems -- and all the problems are new and previously unpublished, so they can't be already in large language model (LLMs)' training sets.

We don't have a good measurement of super advanced mathematics capabilities in AI models. The researchers note that current mathematics benchmarks for AI systems, like the MATH dataset and GSM8K, measure ability at the high-school level, and early undergraduate level. The researchers are motivated by a desire to measure deep theoretical understanding, creative insight, and specialized expertise.

There's also the problem of "data contamination" -- "the inadvertent inclusion of benchmark problems in training data." "This causes artificially inflated performance scores for LLMs, and that masks the models' true reasoning (or lack of reasoning) capabilities.

"The benchmark spans the full spectrum of modern mathematics, from challenging competition-style problems to problems drawn directly from contemporary research, covering most branches of mathematics in the 2020 Mathematics Subject Classification."

I had a look at the 2020 Mathematics Subject Classification. It's a 224-page document that is just a big list of subject areas with number-and-letter codes assigned to them. For example "11N45" means "Asymptotic results on counting functions for algebraic and topological structures".

"Current state-of-the-art AI models are unable to solve more than 2% of the problems in FrontierMath, even with multiple attempts, highlighting a significant gap between human and AI capabilities in advanced mathematics."

"To understand expert perspectives on FrontierMath's difficulty and relevance, we interviewed several prominent mathematicians, including Fields Medalists Terence Tao, Timothy Gowers, and Richard Borcherds, and Internatinal Mathematics Olympiad coach Evan Chen. They unanimously characterized the problems as exceptionally challenging, requiring deep domain expertise and significant time investment to solve."

Unlike many International Mathematics Olympiad problems, the FrontierMath problems have a single numerical answer, which makes them possible to check in an automated manner -- no human hand-grading required. At the same time, they have worked to make the problems "guess-proof".

"Problems often have numerical answers that are large and nonobvious." "As a rule of thumb, we require that there should not be a greater than 1% chance of guessing the correct answer without doing most of the work that one would need to do to 'correctly' find the solution."

The numerical calculations don't need to be done in the language model -- they have access to Python to perform mathematical calculations.

FrontierMath: A benchmark for evaluating advanced mathematical reasoning in AI

#solidstatelife #ai #genai #llms #mathematics

FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI

FrontierMath: a new benchmark of expert-level math problems designed to measure AI’s mathematical abilities. See how leading AI models perform against the collective mathematics community.

"Taiwan's technology protection rules prohibits Taiwan Semiconductor Manufacturing Co (TSMC) from producing 2-nanometer chips abroad, so the company must keep its most cutting-edge technology at home, Minister of Economic Affairs J.W. Kuo."

"Kuo made the remarks in response to concerns that TSMC might be forced to produce advanced 2-nanometer chips at its fabs in Arizona ahead of schedule after former US president Donald Trump was re-elected as the next US president."

TSMC cannot make 2nm chips abroad now: MOEA

#solidstatelife #semiconductors #geopolitics

TSMC cannot make 2nm chips abroad now: MOEA - Taipei Times

Bringing Taiwan to the World and the World to Taiwan

"The open source project DeFlock is mapping license plate surveillance cameras all over the world."

"On his drive to move from Washington state to Huntsville, Alabama, Will Freeman began noticing lots of cameras."

"Once I started getting into the South, I saw a ton of these black poles with a creepy looking camera and a solar panel on top. I took a picture of it and ran it through Google, and it brought me to the Flock website. And then I knew like, 'Oh, that's a license plate reader.' I started seeing them all over the place and realized that they were for the police."

"Flock is one of the largest vendors of automated license plate readers (ALPRs) in the country. The company markets itself as having the goal to fully 'eliminate crime' with the use of ALPRs and other connected surveillance cameras."

"And so he made a map, and called it DeFlock. DeFlock runs on Open Street Map, an open source, editable mapping software."

The open source project DeFlock is mapping license plate surveillance cameras all over the world

#solidstatelife #ai #computervision #alprs #surveillance

The Open Source Project DeFlock Is Mapping License Plate Surveillance Cameras All Over the World

DeFlock has mapped the locations of more than a thousand ALPRs around the United States and thousands more around the world.

"3DPrinterOS, a cloud-based 3D printing management solutions company, has entered a collaboration with the MIX Lab at Montclair State University to develop an algorithm designed to identify 3D printed gun parts."

Ok, sounds like they haven't done it yet. So, we'll get another article saying they did it... if they succeed.

3DPrinterOS develops algorithm to identify 3D printed gun parts

#solidstatelife #3dprinting

3DPrinterOS develops algorithm to identify 3D printed gun parts

3DPrinterOS develops algorithm to identify 3D printed gun parts - in partnership with the MIX Lab at Montclair State University.

OpenFlexure "uses 3D printers and off the shelf components to build open-source, lab-grade microscopes for a fraction of traditional prices. Used in over 50 countries and every continent, the project aims to enable Microscopy for Everyone."

"An open flexure microscope is built from a combination of off-the-shelf electronics, standard optical equipment, and 3D printed parts. The 3D printed parts are designed to be made on any entry grade printer anywhere in the world." "Nothing is proprietary or hidden."

"The finished microscope can run automatically for several hours, scanning samples with a built-in autofocus. The 8-megapixel camera is comparable to many commercial sight scanners, achieving a resolution below 400 nanometers."

"In practical terms this means that individual cell damage or parasites can be identified on a microscope with parts costing under $300. The stage is fully automated, intelligently planning its own path around samples. It can also self-calibrate, warning the user if there's any damage that could impact the diagnosis. The automated stage allows huge data sets to be collected and stored.

"In pathology, this let samples be archived, shared, or used for the training of medical students. this can also be the platform for low resource artificial intelligence systems or automated image processing, making emerging technologies more accessible in low resource settings."

Something else I didn't know exists until just now. Developed at the University of Bath, University of Cambridge, and the University of Glasgow, with contributions from the Baylor College of Medicine, Bongo Tech & Research Labs, and Mboalab.

The OpenFlexure Project

#solidstatelife #3dprinting #microscopy