Trung Le Nguyen is an author and cartoonist.
At the start of June, fresh off reading and enjoying Jane Eyre, he decided to give Pride & Prejudice ago, and live-skeeted his entire read in a thread on Bluesky.
If you ever read P&P and enjoyed it, this one's for you.

Getting a small local language model running on a notebook or even smartphone in 2026 is trivial. But what about something even smaller and lower-power. Say, like an ESP32 microcontroller that costs less than $10? It might sound impossible — the device is primarily designed for things like remote sensors, IoT, and other embedded applications, not running generative AI models — yet, that's exactly what a developer who goes by the handle SlvDev has managed to do. In a process detailed on GitHub, and recently showcased on the Better Stack YouTube channel, SlvDev documented how he managed to get a small language model running at nearly 10 tokens a second locally on a microcontroller that costs about the same as a fancy cup of coffee. Tiny stories on a tiny microcontroller Cramming a large language model (LLM) onto something as small as a ESP32 microcontroller isn't a trivial task. There's a reason that these models are trained and run on GPUs. LLMs are memory-hungry beasts that typically require between one and four bytes per parameter just to hold their weights in memory. With just 520 KB of SRAM and 8 MB of pseudo SRAM (PSRAM) on the ESP32-S3, you aren't going to be running a model like DeepSeek V4 Flash . To make it work, the dev had to drop the "large" from the language model and settle for something nearly 10,000 times smaller: TinyStories, a 28.9 million-parameter model originally developed by Microsoft Research. However, even this model is asking a lot of an ESP32-S3 module. At 16-bit precision, the model requires about 60 MB of memory that the ESP32 simply doesn't have. So, the dev employed several techniques, some of which we've previously explored, to shrink the model’s footprint. The first is quantization, a process by which weights are compressed by reducing their precision from something like 16-bits of precision to eight, or even four. This enabled SlvDev to trade a bit of accuracy for a 75 percent reduction in memory required. Instead of about 60 MB of memory to hold the weights, they now require just 14.9 MB. But that still wasn’t enough. Thankfully, in addition to the 8.5 MB of working memory, the ESP32-S3 can also be bought with up to 16MB of flash storage. By borrowing a technique called per-layer-embedding (PLE) from Google's Gemma family of models, the dev was able to offload the majority of the model's weights, about 25 million parameters or about 12 MB worth, to flash with minimal performance degradation. Offloading model weights to NVMe storage is an old trick for getting massive frontier-class models like DeepSeek V3 running on hardware that wouldn't have the necessary memory and GPU capacity to serve it otherwise. The downside of this approach, historically, is that it murders performance. Instead of tokens a second, you're usually looking at seconds, or in some cases minutes, per token. It works but it's not remotely practical. PLE manages quite a bit better because these weights are accessed rather sparingly, which keeps the flash's glacially slow bandwidth relative to DRAM or SRAM from nerfing performance. The result is that rather than trying to cram 14.9 MB into the ESP32's memory, the model now only needs about 2 MB. Specifically the output head, embeddings, and KV cache are kept in the chip's PSRAM, while activations are handled in the chip's 520 KB of SRAM. Using this approach, the dev says they were able to get 9.88 tokens a second out of the microcontroller, which is faster than the average person can read. So, what can I do with it? While you may be able to get a small generative AI model running on a microcontroller like an ESP32, you won't get much from the practice beyond dumb simple pride. Tiny Stories is a great proof-of-concept, but aside from generating short, reasonably coherent stories on demand, it can't do much. You aren't going to build a chatbot with it, generate code, or power an agent. There is another model, called Barista, that can answer questions at roughly twice the performance, but only on topics pertaining to espresso. TinyStories and Barista are, well, just too tiny to do much else. Yet the fact these models run on an ESP32 at all is impressive in itself That said, if you've got a device with just a bit more memory and compute, say a Raspberry Pi or a smartphone, there are far more capable models out there. Google's Gemma 4-E2B-it, launched back in April, employs the same quantization and PLE offload techniques to cram a 5.1 billion-parameter vision language model into just over a gigabyte of memory when using 4-bit weights. Higher degrees of quantization and offloading can get this down to around 500 MB. While its memory footprint is several orders of magnitudes less than what's required to run a frontier model, it is still capable enough to power local chatbots, orchestrate local agents for things like managing a user's calendar, and free you from your reliance on OpenAI or Anthropic so long as you can put up with the occasional hallucination. ®

In the face of an epidemic of AI-enabled cheating and research suggesting its products hamper learning, OpenAI is doing the sensible thing and pushing more AI on students and teachers. Wait, did I say sensible? My mistake. The House of Altman announced a trio of new education-focused offerings on Tuesday: one for K-12 teachers, another for college educators, and a third for college students. The new plugins, the company explained, will help students and educators make more use of ChatGPT’s agentic capabilities for both studying and teaching. The new features are available through ChatGPT Edu, an institutionally licensed suite for higher education, and ChatGPT for Teachers, a free resource available to verified US K–12 educators and school districts. The new offerings, says OpenAI, build on its educational AI philosophy that “AI should support learning, not shortcut it, and the best learning experiences keep educators and students in control.” Plenty of educators might disagree. Cheating with AI has become a sad norm in schools around the world, and the US is no exception. Many young people admit to using AI to cheat on school assignments, and college students have been caught doing it, too. Mexico's largest university, the National Autonomous University of Mexico, last week suspended enrollment for incoming students after suspected widespread cheating in its admissions exams. The university also decided this week to require about 58,000 applicants who qualified in the remote entrance exam to take an in-person proctored control exam after scores surged during its first remotely administered admissions test. As for how teachers relate to AI in the classroom, a 2025 study from the Center for Democracy and Technology suggests it’s going to take more than some new OpenAI software to help teachers feel more comfortable. CDT said last year that K-12 teachers widely complained of not being given the necessary resources to understand how to integrate AI into schools, or how to tackle the potential harms AI may be causing to their students' ability to learn and personal development alike. Rather than helping teachers and professors regain control of the academic process or better understand how AI is affecting their students, however, OpenAI wants to simply let them foist off more of their work to ChatGPT, too. “The K–12 Educator plugin is designed to help teachers plan and create for their classrooms,” OpenAI explained. The tool can use materials teachers already have on hand “to create differentiated resources, design interactive visuals, and surface actionable insights.” Integration with educational AI infrastructure outfit Learning Commons allows ChatGPT to automatically consider local educational standards for anything it creates for teachers, the company explained, while educators get to remain “in control of pedagogical decisions, grading, and agentic actions.” The plugin for college professors is largely the same. “The College Educator plugin enables course design, teaching, and academic planning,” OpenAI said in its announcement. “Faculty can update syllabi, create interactive websites or multimedia assessments, adapt materials for diverse learners, or package content for their Learning Management System.” For college students, the plugin is designed to bring ChatGPT further into students' study routines, the latest expansion of AI tools that have already complicated essay writing and take-home testing. “The College Student plugin helps students turn what they are already studying into more personalized learning experiences,” said the ChatGPT maker. AI tutors are available (sorry, grad students), as are AI-generated quizzes, study guides, flashcards, and AI-generated visual explanations. “The plugin draws on learning science to prioritize deeper understanding and build stronger study habits,” OpenAI claims. You’d be right to doubt that. Aside from being used to cheat on exams, AI also has a demonstrable effect on harming learning outcomes. MIT researchers last year hooked a group of students up to EEGs to study their brain activity while writing essays. Some of them were allowed to use the internet to conduct research, while others were told to use AI. Not only did the AI cohort have far less ability to recall what they wrote when asked about it later, but they also showed far less brain activity, suggesting that relying on AI simply doesn’t lead to better learning. It’s not much of a leap to go from AI lowering brain activity while doing essay research to figuring that relying on AI tools to help one study, create flashcards, or receive tutoring will likely lead to similar outcomes. Real learning still takes legwork, and other studies suggest overreliance on tools like ChatGPT can undermine deeper learning. Why teachers and professors would want to embrace an expansion of AI into their workflow when they’ve seen what it can do to their students is a question for OpenAI. The AI giant didn't initially respond to our questions for this story, but we'll update if that changes. ®

Trump Media has officially launched its $100,000-per-month data feed giving trading firms machine-readable access to Truth Social posts milliseconds before the public. According to Fortune, five Wall Street firms have already signed up for the service, which "would generate about $500,000 in monthly revenue, or $6 million annually."

Critics argue the service could let President Trump, who owns about 41% of the company, profit from early access to market-moving presidential communications. "I'll be blunt," Gian Luca Clementi, an economics professor at NYU Stern School of Business, told Fortune. "This is insider trading by definition."

"He's going to monetize the role of the office of the president of the United States," he said. "The undisputable fact is that somebody is going to earn some more money than before, and that's the president of the United States." From the report: Trump's media venture has struggled to build a profitable social media business despite its lofty valuation. Truth Social has reported significant operating losses since going public. According to the company's earnings report for Q1 2026, Trump Media & Technology Group netted a roughly $405 million loss and raised less than $900,000 in sales.

Not everyone agrees the arrangement meets the legal bar for insider trading. Shannon Devine, a spokeswoman for Trump Media & Technology Group, has pushed back on the characterization, telling Quartz that Truth API "offers customers the fastest way to ingest publicly available Truth Social data" and that critics "must have invented a new theory of 'insider trading' based on publicly available information."

Classic insider trading law hinges on trading on secret, material information in breach of a fiduciary duty, and Truth Social posts are, by design, meant to become public within moments -- raising real doctrinal uncertainty about whether faster access alone qualifies. But other legal experts argue the greater risk lies ahead. Richard Painter, former White House chief ethics counsel, has argued that the arrangement could violate federal law once Trump posts genuinely market-moving news -- on tariffs, military action, or other policy decisions -- before it's public, with Truth Social effectively acting as a paid "tipper" on the president's behalf. Sen. Alex Padilla (D-Calif.) said he plans to introduced legislation Tuesday to ban the president from selling expedited access to his statements.

Read more of this story at Slashdot.

S&P 500 shot up 1.8% and the main measure of Wall Street’s health topped its prior all-time high set a few months ago

The US stock market rallied to records on Tuesday as profits kept piling up for companies and as oil prices eased.

The S&P 500 shot up 1.8%, and the main measure of Wall Street’s health topped its prior all-time high set a couple months ago. The Dow Jones industrial average added 907 points, or 1.7%, to its own record set the day before, while the Nasdaq composite jumped 2.6%.

Continue reading...

Suspect had a prior conviction for reportedly shooting his father in Arizona, after he told his son to do the dishes

Authorities on Monday charged a man in Washington state with starting a wildfire that has consumed almost 4,000 acres and forced tens of thousands from their homes.

Aaron Farinacci, 37, of Spokane, was arrested and charged with first-degree arson for setting the Old Trails fire, the largest of three devastating wildfires burning through the region.

Continue reading...

Gianni Infantino is under increasing scrutiny following his World Cup sell-off plan. We look at each confederation’s relationship with the president

The pressure is growing on the Fifa president, Gianni Infantino, with many countries withdrawing their support following the bungled World Cup sell-off. However, how key is Infantino’s relationship with each confederation and is it likely that an election contest will actually happen? We look at the key considerations, confederation by confederation.

Continue reading...

She’s not there, Masha Foya

Huh, here’s the trailer for a new director’s cut (of sorts) of the second X-Files movie (called I Want to Believe, orig. released in 2008) that, per Wikipedia, “will restore horror elements that were previously removed”.

STARBASE (ANP) - SpaceX, het ruimtevaart-, AI- en satellietbedrijf van Elon Musk, heeft in het tweede kwartaal de omzet fors opgevoerd. Het bedrijf profiteerde onder meer van een sterke groei van het aantal Starlink-abonnees en meer opbrengsten uit zijn tak voor toepassingen op het gebied van kunstmatige intelligentie (AI).

Het is voor het eerst dat SpaceX resultaten bekendmaakt als beursgenoteerd bedrijf na de historische beursgang in New York in juni.

De totale omzet van SpaceX steeg in april, mei en juni met 92 procent tot 7,8 miljard dollar (omgerekend ruim 6,7 miljard euro) in vergelijking met dezelfde periode een jaar eerder. Het bedrijf rapporteerde daarnaast een nettoverlies van 541 miljoen dollar, tegen 1 miljard dollar verlies in het tweede kwartaal van 2025.

Het aantal abonnees van satellietnetwerk Starlink verdubbelde tot 12 miljoen ten opzichte van het tweede kwartaal van vorig jaar. De divisie waar Starlink onder valt, was goed voor een omzet van bijna 4,3 miljard dollar.