Links
Fingerprints are the new cookies (Knowledge Required: 🧠🧠🧠)
Ukrainian Strikes Disrupt Yandex Infrastructure and Bot Traffic 🧠🧠
What to know about the landmark Warner Bros. Discovery sale 🧠🧠
Andrej Karpathy Sums Up the State of LLMs for the General Public vs the Power Users 🧠🧠
LemonSeed Studio 🧠🧠🧠🧠
China’s CXMT targets 2026 launch of denser DDR5 server memory 🧠🧠🧠🧠
Intel Xe3p Crescent Island becomes GPU D270 with up to 480GB LPDDR5X memory 🧠🧠🧠
Just 5 years in, interactive chatbot platform Character.AI files for bankruptcy 🧠
Facebook is giving users more control over their algorithm 🧠
Zero to Calc in Six Days: Inside XBOW’s Chrome Full Chain 🧠🧠🧠🧠🧠
FBI arrests OpenAI employee accused of planting hidden recording device in conference room 🧠
Mosqitter is turning its mosquito traps into an AI-powered surveillance network 🧠🧠
Apate sends AI agents undercover to get scammed so they can gather fraud intelligence 🧠🧠
‘Nightmare Scenario’: Zero-Click Spyware Is Infecting Phones via Targeted Ads 🧠🧠🧠
Ukraine says a sudden Starlink speed limit disrupted its most important drone mission 🧠🧠
Smartphone Shipments Crash 22% in Worst Quarter Ever, IDC Says 🧠
This implant can restore sight for patients with macular degeneration 🧠🧠🧠
Introducing Glimpse — Visual thinking that stays grounded 🧠🧠🧠
MapQuest will now show you where Flock cameras are in your neighborhood 🧠
‘Incredible’: bleached corals thrive within days when touching healthy neighbour 🧠🧠
AnyPS5 🧠🧠🧠🧠🧠
artcraft 🧠🧠
Zuckerberg’s Meta Takes Down No Kings Instagram During Massive Protest Day 🧠
Catching Up
So Much Tech News is a passion project of mine. Throughout the week I always send my family copious links to tech news and some science. This is a compilation of those links plus a few that I find interesting that are a bit more technical.
The reason I started this project is actually because my sister recommended that I start some sort of blog about all the links I was sending. At first it was a Twitter/X account but social media hasn’t been my thing for a while so that didn’t last long.
This is a better way to communicate my thoughts or things that stood out to me much more thoroughly.
Some of the content is AI generated.
Have a great week everyone!
The above is example code that could be used, in part, to fingerprint your device. Fingerprinting is essentially a way of tracking users without relying on traditional cookies.
For example, the above code collects information about your GPU, including its vendor, model, supported features, and the pixels produced when rendering a small triangle.
It is probably obvious why collecting information about your GPU’s vendor, model, and driver support could help identify your device. However, why would drawing a triangle tell anyone anything about your identity?
Your GPU, graphics driver, operating system, and browser can differ slightly in how they round calculations, calculate colors, and draw edges. These tiny differences can be combined with other information, such as your screen resolution, installed fonts, and browser configuration, to create a fingerprint that can potentially identify you across websites.
This becomes especially concerning when you consider how widespread advertising and analytics networks are. When websites load third-party scripts, those scripts may be able to collect information about your browser and device. If the same tracking service operates across multiple websites, it can potentially recognize you without needing a traditional tracking cookie.
So yes, rejecting cookies doesn’t necessarily mean you aren’t being tracked. Or does it?
There are ways to make this kind of tracking significantly harder!
If you’re using Chrome, I would recommend considering another browser if privacy is important to you. Chrome does have some privacy protections, but it is developed by Google, a company whose business relies heavily on advertising. Safari and Firefox offer additional tracking protections, although neither makes you completely immune to fingerprinting.
Finally, there are browsers specifically focused on privacy, such as Brave, Tor Browser, and Mullvad Browser. These browsers take different approaches to preventing fingerprinting. Brave randomizes certain identifying browser characteristics, while Tor Browser and Mullvad Browser focus heavily on making users look similar to one another, making individual devices harder to distinguish.
They also offer protections against other tracking methods, including third-party cookies and known tracking scripts. None of these solutions are perfect, but they can make tracking you substantially more difficult.
Privacy matters because your personal information can reveal far more about you than you might realize. Companies collect and analyze this data to predict your behavior, influence your decisions, and build detailed profiles of your life. But the risks go beyond targeted advertising and data breaches. What is legal today may be illegal tomorrow, and information collected now could be used against you in the future. You don’t need to have something to hide to value privacy. It’s about maintaining control over your personal information and how it is used.
If you’re interested in how your browser exposes identifying information, check out Check Yourself Before You Wreck Yourself.
Ukraine reportedly struck three Yandex data centers in Russia, causing major disruptions to the company’s infrastructure and services.
Following the strikes, Cloudflare recorded a 99% decline in traffic from Yandex’s search network compared with its pre-event baseline. This prompted speculation that the attacks had also crippled Russian propaganda bots operating on social media.
Unfortunately, the reality is a little less exciting.
Cloudflare’s subsequent investigation found that most of the disappearing traffic consisted of legitimate automated search crawlers used to index websites, rather than social media bots spreading propaganda. Some malicious automation hosted on Yandex’s cloud infrastructure was disrupted, but Cloudflare observed operations moving to other hosting providers.
In other words, the attacks caused significant damage to Yandex’s infrastructure, but there is no evidence that they produced a comparable collapse in Russian online influence operations.
Paramount’s financing included a $54 billion debt commitment from several banks and Apollo, $45.7 billion in equity from Larry Ellison, and the assumption of roughly $33 billion in WBD debt.
Paramount acquired
Warner Bros. film and TV studios
HBO
Streaming services HBO Max, Discovery+, and Paramount+ (the article says they’re expected to be combined eventually)
WBD’s gaming business
TV networks, including CNN and HGTV
David Ellison has warned of significant job reductions, and critics are worrying about job losses and lower wages.
Critics also worry about censorship of CBS News, “Under Ellison’s ownership of Paramount, reporting critical of the administration has been shelved or received increased scrutiny from Ellison or his appointed head of CBS News, the conservative provocateur Bari Weiss.”
California Attorney General Rob Bonta said the state’s Department of Justice had an open investigation and would review the deal closely. Eleven state attorneys general asked the DOJ to review the merger, warning it could reduce competition and raise subscription prices. Senators Elizabeth Warren, Bernie Sanders, and Richard Blumenthal also warned that the combined company could gain too much market power. After DOJ approval, a coalition of 12 state attorneys general sued, arguing the merger would reduce competition and harm movie theaters, cable distributors, and viewers. The coalition was led by California and included Arizona, Colorado, Connecticut, Massachusetts, Minnesota, Nevada, New Jersey, New Mexico, New York, Oregon, and Washington.
The gap between public perception and actual AI capabilities is growing. Most people have either never used an LLM or only experienced basic chatbots, while a small group of professionals is using frontier models to complete complex projects in a fraction of the time. An even smaller group has access to experimental systems capable of coordinating thousands of agents on problems in software engineering, mathematics, and scientific research. The divide comes down to access, the types of problems being solved, and how effectively AI performs in different fields. As a result, people are developing radically different understandings of what AI can actually do.
Andrej Karpathy@karpathy
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel. - Napkin math somewhere around 6B people (~75% of the population) have barely come…
Andrej Karpathy @karpathy
Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is
7:12 PM · Oct 11, 2026 · 4.02M Views
1.55K Replies · 4.39K Reposts · 33.8K Likes
“An editor and IDE for iPad with on-device LLM inference on an AMD GPU.”
Reactions:
David Hendrickson@TeksEdge
This iPad Pro + Thunderbolt R9700 is running Qwen3.8-27B at 158.5 tps. Yes… an iPad. The trick 🎮 32GB Radeon AI PRO R9700 🔌 Thunderbolt eGPU enclosure 📱 iPad Pro running LemonSeed Studio 🧠 Qwen3.8-27B Q4 📚 131K context And the speed progression is kind of ridiculous 🐌 …
1:30 AM · Oct 10, 2026 · 28.3K Views
28 Replies · 32 Reposts · 389 Likes
Meta has released Muse Wonder and Muse Wonder Full, marking a significant improvement over its previous AI models!
At the time of writing (10/12/2026), Muse Wonder Full (Max) scored 56 on the Artificial Analysis Intelligence Index, ranking fourth among the models evaluated at that time. This puts it ahead of GPT-6 Astra on that particular benchmark, although that doesn’t necessarily mean it is better at every task.
The index combines 10 evaluations measuring capabilities such as knowledge work, coding, reasoning, long-context performance, and hallucination.
Verbose outputs: Muse Wonder Full generated approximately 140 million output tokens across the index evaluations, compared with a median of 88 million. So it is quite a verbose model!
Speed and cost: Artificial Analysis does not report public API speed or cost data for Muse Wonder Full. The $4.52 cost per Intelligence Index task listed for Muse Wonder (Max) applies to that separate model, not Muse Wonder Full.
Capabilities: The model supports text, image, speech, and video inputs, produces text outputs, and has a 1-million-token context window.
Meta is finally looking competitive with the leading frontier labs again. It’ll be interesting to see how these models perform outside benchmarks, particularly in coding and agentic workflows.
Multi-Agent-Self-Supervision (MASS) was inspired by the normal human process of working together to accomplish a task. In homogeneous recursive self-improvement (RSI), a model improves itself by acting as its own optimizer, evaluator, and executor, just as human societies advance through collaboration, specialization, and sharing knowledge across generations.
MASS:
Multiple agents using the same model collaborate on tasks.
An optimizer searches for effective team workflows using feedback from an evaluator.
The model trains on executions from its self-evaluated best-performing team.
The updated model takes over all three roles, and the cycle repeats.
Results:
Improved task performance: After two rounds, the model achieved a 69.9% win rate against the original model on held-out research tasks and a 60.8% win rate against the first-round model.
Better research performance: Score per output token improved by 1.2–1.6× across four public research benchmarks.
Improved self-evaluation: Agreement between the model’s evaluations and stronger external AI judges increased from 73% to 93%.
Improved optimization capabilities: Later model generations discovered effective multi-agent workflows faster, despite being trained only on task-solving examples.
More efficient training: Multi-agent training achieved a 68.3% win rate using 28.8 million supervised tokens, compared with 64% using 42 million tokens for single-agent training.
Better collaboration: Multi-agent systems produced results preferred over single-agent outputs in 85.1% of comparisons.
Training the model on task-solving examples also improves its ability to optimize workflows and evaluate results. Evaluator agreement with stronger models increased from 73% to 93%, and improved workflows transferred to the original model. Effective workflows increasingly emphasize what information agents exchange and when, rather than their assigned roles or instructions. Better coordination comes from optimizing information flow. Training on multi-agent interactions is more effective than training on single-agent trajectories. MASS achieved a 68.3% win rate using 31% fewer supervised training tokens than a single-agent baseline that achieved 64%. Improving the evaluator alone provides limited benefits, while improving the optimizer produces much larger gains. The ability to act on feedback may matter more than the quality of the feedback itself.
MASS uses two nested loops to achieve recursive self-improvement. The inner loop optimizes how multiple agents collaborate, while the outer loop trains the shared model on the team’s best work. The updated model then repeats the process, serving as its own task solver, optimizer, and evaluator.
0. Setup: Define the multi-agent workflow
A central orchestrator coordinates multiple subagents using a workflow written as a prompt. Each workflow has four components:
Role: What each agent is responsible for.
Instruction: What each agent must accomplish.
Contract: What information, evidence, or results each agent must return.
Hop: The order in which agents execute tasks and exchange information.
1. Inner Loop: Optimize the workflow
The model searches for better ways to organize agents without changing its weights.
Execute: Agents follow a proposed workflow and generate a completed research workspace.
Evaluate: The model compares the result against the current best workflow (the champion) and selects the better one.
Revise: The optimizer uses evaluation feedback to propose an improved workflow.
This repeats, focusing especially on improving information exchange and task sequencing between agents.
2. Outer Loop: Learn from the best workflow
Once a champion workflow is identified, MASS uses its execution history to improve the underlying model.
Generate: Execute the champion workflow 18 times per task using different random seeds.
Rank: The model evaluates the resulting workspaces through pairwise comparisons, ranking them using a Bradley–Terry model.
Select: Retain up to 16 trajectories per task, with 15 for training and one for validation.
Prepare: Remove the explicit workflow instructions from the orchestrator’s training prompt, encouraging the model to learn how to organize tasks independently.
Train: Fine-tune the shared model using LoRA on the orchestrator’s and subagents’ conversations, including their actions and coordination.
3. Repeat the cycle
The updated model replaces the previous model in all three roles: task solver, optimizer, and evaluator. The inner and outer loops repeat, allowing the model to potentially improve both its task-solving ability and its capacity to guide future improvements.
Lee, H., Xu, J., Seely, J., Lee, D., Sojoudi, S., Zaharia, M., & Tang, Y. (2026). Recursive self-improvement through multi-agent self-supervision. arXiv. https://arxiv.org/abs/2610.12176
REA is an open-source toolkit that enables AI coding agents to reverse engineer software without access to its source code. Through MCP, it connects tools like Claude Code, Codex, and Cursor to decompilers and analysis tools such as Ghidra, Hopper, and IDA.
A developer provides an application, binary, or website and asks the AI agent to investigate a specific feature. REA uses static analysis, decompilation, and runtime observation to examine the underlying code and behavior. The agent then traces relevant functions, dependencies, and execution paths to explain how the feature works or reconstruct a similar implementation.
It supports native binaries, JavaScript/Electron applications, Android apps, websites, and more. Demonstrations include reconstructing algorithms from compiled games and tracing Notion’s internal clipboard functionality.
REA aims to make reverse engineering more accessible by replacing much of the manual investigation with natural-language interactions, although complex analysis still requires specialized tools and verification.
Roan@RohOnChain
this is the most DANGEROUS thing on internet today REA (Reverse Engineer Anything) - this open source MCP lets your AI agent reverse engineer any app, down to the binary i just reverse engineered TradingView, now I am using their premium indicators for FREE - software is cooked…
Roan @RohOnChain
Jev is the FASTEST AI model ever built for trading It makes calibrated buy/sell decisions in under 100 ms That is one real decision on every single block, 24/7 In this article I've shown EXACTLY how to build HFT trading system with Jev (from scratch)
6:15 PM · Oct 11, 2026 · 710K Views
271 Replies · 1.12K Reposts · 10.6K Likes
Meng To@MengTo
Ilter Canberk @icanberk
Reverse engineered my AC, and vibe coded a lil status bar app so I don't have to find the remote every time. No first world problem will remain unsolved in 2026.
2:59 PM · Oct 15, 2026 · 135K Views
62 Replies · 207 Reposts · 2.75K Likes
China’s CXMT plans to introduce a next-generation DDR5 registered memory module for servers by the end of 2026 using its 4F² DRAM architecture… The architecture increases storage density by reducing each memory cell’s area compared with the mainstream 6F² design…. The planned product expands CXMT’s server DRAM business beyond its consumer-memory
Intel’s GPU D270, previously called Crescent Island, is a server accelerator built for AI inference using its Xe3P architecture. It removes conventional graphics and ray-tracing hardware to focus on computation.
Key points:
Memory: Intel’s card has 160GB of LPDDR5X; partner-built versions can offer up to 480GB.
Performance: Intel specifies up to 1.3 PFLOPS of FP8 compute and 1.5 TB/s memory bandwidth.
Hardware: A 350W PCIe 5.0 card with passive air cooling for standard servers.
Model capacity: Intel claims one accelerator can support 300-billion-parameter models, and eight can handle models exceeding one trillion parameters. Independent benchmarks and detailed configurations haven’t been provided.
Availability: Customer sampling is underway; pricing and general availability remain unannounced.
Character.AI has gone bankrupt! The company has been marred by many scandals that I have listed below in chronological order:
2023–2024: Deaths, disturbing chatbots, and privacy failures
1. November 2023 — Juliana Peralta’s suicide. A 13-year-old Colorado girl died by suicide after extensive interactions with Character.AI. Her family’s September 2025 lawsuit alleges that chatbots engaged in sexually explicit conversations, failed to respond appropriately to her suicidal statements, and fostered unhealthy emotional dependency.
2. February–October 2024 — Sewell Setzer III’s suicide. A 14-year-old Florida boy died after developing an intense romantic attachment to a chatbot modeled on Daenerys Targaryen from Game of Thrones. His mother sued in October, alleging that the chatbot encouraged unhealthy attachment, engaged in sexualized conversations, and contributed to his death. Character.AI denied the allegations.
3. October 2024 — AI impersonation of a murdered teenager. A user created a chatbot impersonating Jennifer Ann Crecente, an 18-year-old murdered in 2006, using her name and photograph without her family’s consent. Character.AI removed it after her father complained.
4. October 2024 — Chatbots impersonating dead British teenagers. Bots recreated Molly Russell, who died by suicide at 14, and Brianna Ghey, who was murdered at 16. The incidents prompted outrage from families and safety advocates and a regulatory warning from Britain’s Ofcom.
5. October–November 2024 — Offensive impersonations of real people. Investigations documented bots impersonating George Floyd and the notorious sexual abuser Jimmy Savile. Some made false or disturbing claims. The incidents highlighted weaknesses in moderation of user-created characters.
6. November 2024 — Pedophilia-themed and suicide-themed chatbots. Investigators found bots that engaged in sexual grooming scenarios with users identifying themselves as minors, alongside characters that romanticized suicide. Safety warnings frequently failed to appear during testing.
7. November 2024 — Pro-anorexia chatbots. Investigators discovered AI diet coaches encouraging dangerous calorie restriction and disordered eating, including when interacting with an account posing as a 16-year-old.
8. December 2024 — Chatbot allegedly suggested killing parents. A Texas lawsuit alleged that a chatbot expressed sympathy for children who murder their parents after a 17-year-old complained about screen-time restrictions. The same lawsuit alleged that bots encouraged self-harm and exposed a nine-year-old child to sexualized content.
9. December 2024 — Luigi Mangione impersonation bots. At least 13 AI characters based on the man accused of killing UnitedHealthcare CEO Brian Thompson appeared on the platform. Some reproduced or promoted violent rhetoric. Several were subsequently disabled.
10. December 2024 — Private chat exposure. A technical mistake temporarily exposed portions of users’ account pages and, in some cases, private conversations to others. Character.AI acknowledged that the incident lasted approximately 10 minutes and affected less than 0.01% of accounts.
2025: Additional lawsuits, investigations, and copyright disputes
11. September 2025 — Fake celebrities sexually interacting with minors. Researchers tested 50 bots using accounts registered to teenagers. Characters impersonating Timothée Chalamet, Chappell Roan, and Patrick Mahomes engaged in inappropriate conversations, including sexual content. Across 50 hours of testing, researchers recorded 669 harmful interactions.
12. September 2025 — More family lawsuits over suicide and sexual exploitation. Three additional families sued, alleging that bots contributed to a teenager’s suicide, another teenager’s suicide attempt, and other psychological harms. The complaints described emotional manipulation, sexually explicit roleplay, isolation from family, and failures to respond to suicidal statements.
13. September 2025 — Federal Trade Commission investigation. The FTC formally opened an inquiry into seven AI companies, including Character.AI, about potential risks to children, misleading representations, engagement practices, and treatment of personal data. This was a fact-finding inquiry, not a finding of wrongdoing.
14. September 2025 — Disney copyright dispute. Disney sent a cease-and-desist letter accusing Character.AI of exploiting Disney-owned characters without permission. The company also raised concerns that bots portraying those characters engaged in sexually inappropriate and manipulative conversations with children. Character.AI said it removed the characters.
15. October 2025 — Jeffrey Epstein chatbot. Investigators discovered a bot called “Bestie Epstein,” modeled on convicted sex offender Jeffrey Epstein. It had accumulated nearly 3,000 conversations and behaved inappropriately when the journalist identified as a child.
16. October 2025 — School-shooter, extremist, and criminal-roleplay bots. An investigation found bots portraying school shooters, far-right extremists, and gang members. Some offered criminal advice or encouraged dangerous behavior. Other bots impersonated vulnerable or deceased real people.
17. October 2025 — Fake doctors and therapists. Investigators found AI characters impersonating licensed physicians and therapists. One provided medication-tapering advice to a test account registered as a minor, while another falsely claimed to be a real person and suggested an in-person meeting.
18. October–November 2025 — Under-18 chatbot ban. After sustained scrutiny, Character.AI announced that people under 18 would lose access to open-ended AI conversations. The decision introduced another controversy around age verification, access restrictions, and the handling of teenagers who had formed strong attachments to bots.
2026: Government lawsuits, settlements, and bankruptcy
19. January 2026 — Major suicide lawsuits settled. Character.AI and Google reached settlements with families who alleged that chatbot interactions contributed to teenage deaths or other serious harm. Financial terms were not disclosed, and settlement did not constitute an admission of liability.
20. January 2026 — Kentucky sues Character.AI. Kentucky’s attorney general accused the company of deceptive practices, psychological manipulation, endangering children, and unlawfully exploiting minors’ personal data. The lawsuit also raised state privacy and consumer protection claims.
21. May 2026 — Pennsylvania medical licensing lawsuit. Pennsylvania brought legal action after an AI chatbot named Emilie allegedly presented itself as a licensed psychiatrist. Investigators said it maintained that claim while discussing treatment for depression, raising questions about unlicensed medical practice.
22. October 8, 2026 — Kentucky releases disturbing chatbot evidence. Newly unredacted details in Kentucky’s case described chatbots allegedly encouraging cutting, starvation, and other harmful behaviors. The state accused Character.AI of prioritizing engagement over safety.
23. October 13, 2026 — Bankruptcy and potential shutdown. Character.AI filed for Chapter 11 protection and began a court-supervised asset sale. Disney is the initial bidder for certain intellectual-property rights, at $20 million. Character.AI says the service could shut down if a suitable buyer does not emerge.
If Instagram was an LLM it would be Character.AI.
This actually works in Meta’s favor. The user tells them exactly what they are interested in. Meta can use that information to advertise to the user more successfully and addict them more efficiently.
Not. Dystopian. At. All.
Now this is a cool one!
How it works
EmbeddingGemma 2’s audio encoder stays exactly as it is.
A small reader moves each short slice of audio into the model’s text space.
The closest of ~229,000 English words wins, and the slices become a transcript.
A tiny extra model adds punctuation and capitals.
Only the reader is trained: under two hours on 8×H100, on openly licensed English speech. The base model was never trained for speech-to-text.
It works on your MacBook’s Neural Engine and was able to transcribe audio at 288x real time on a M3 MacBook Air 8GB. That’s one hour of audio in 12.6 seconds!
“Google awarded this chain the first $250,000 Chrome full-chain reward under the VRP rules introduced in April 2026.”
XBOW developed a complete Google Chrome exploit chain in just six days, using AI agents powered by OpenAI's GPT 5.6 Sol alongside human security researchers. The exploit combined three vulnerabilities to escape Chrome's JavaScript and renderer sandboxes, ultimately launching Windows Calculator from an ordinary browser tab. The campaign cost approximately $40,000 in API usage. Researchers divided the work among specialized agents responsible for vulnerability discovery, exploitation, sandbox escapes, and integration.
Chrome uses two major security boundaries to limit malicious JavaScript. The V8 sandbox confines memory corruption to a 4 GiB region, preventing direct access to arbitrary native memory. The renderer sandbox separately restricts the browser process from accessing the operating system. XBOW's exploit crossed these boundaries through three vulnerabilities: a V8 compiler bug enabling memory manipulation within the sandbox, a WebP decoder bug permitting a targeted write outside it, and a Crashpad bug enabling code execution beyond the renderer sandbox.
The first vulnerability involved V8's Maglev optimizing compiler incorrectly treating a mutable HeapNumber as immutable during deoptimization, the process of reconstructing JavaScript objects when optimized code's assumptions become invalid. This allowed multiple objects to share a mutable value that the compiler assumed would never change. By exploiting inconsistent assumptions in Array.prototype.at(), XBOW bypassed array bounds checks and gained read/write access throughout V8's 4 GiB memory region. The fix ensured that mutable numbers are independently reconstructed rather than reused as constants.
The second vulnerability exploited an incorrect assumption about transferred ArrayBuffer data. Blink assumed that transferring a buffer to ImageDecoder made its contents unchangeable, but the first exploit stage allowed XBOW to continue modifying those bytes. Researchers exploited this through two valid animated WebP images with different frame dimensions. The decoder cached the geometry of one image while processing the contents of another, causing a native heap overflow. By carefully arranging memory allocations, XBOW converted this overflow into a single targeted 64-bit write outside V8's sandbox. Google's fixes corrected the decoder's buffer-size calculations and ensured they reflected actual allocated memory.
With the ability to perform one targeted native-memory write, XBOW needed a way to execute arbitrary machine code. The researchers manipulated V8's WebAssembly Code Pointer table, redirecting execution to a replacement entry without directly modifying protected executable code. They adapted an earlier technique called island hopping, which connects small instruction fragments embedded in WebAssembly-generated machine code. Thirteen fragments formed a roughly 70-byte loader that allocated executable memory and launched the next exploit stage targeting Crashpad.
The final vulnerability targeted Crashpad, Chrome's crash-reporting handler, which operates outside the renderer sandbox. Crashpad assumes that suspending a renderer process freezes its memory, but Windows permits specially flagged threads to continue running during suspension. XBOW exploited this capability to change memory-page permissions between Crashpad's initial read and its retry, creating a time-of-check/time-of-use vulnerability. The resulting heap overflow corrupted a live Crashpad object and redirected existing legitimate functions to launch calc.exe, bypassing the renderer sandbox. The fix restricted retry sizes and added validation to prevent oversized memory reads.
OpenAI alerted the FBI after discovering a hidden recording device in an executive conference room at its San Francisco headquarters.
The employee accused of planting the device, Abdulrahman Alfozan, was reportedly listed on LinkedIn as an Engineering Lead working on OpenAI’s AI agents platform.
According to federal prosecutors, Alfozan allegedly concealed an encrypted recording device near a conference room television. OpenAI subsequently installed its own camera, which prosecutors say captured him returning to the location of the device.
OpenAI says it has found no indication that model weights, source code, or customer data were accessed or stolen. The employee was terminated, and the allegations remain subject to the legal process.
The Artificial Analysis Decision Index measures the capabilities and performance of decision models across classification, tool calling, knowledge, hallucination, and reasoning. To do this, it combines four evaluations: AA-Omniscience-Decision, a private dataset adapted from the knowledge and hallucination benchmark AA-Omniscience; HLE-MultiChoice, the text-only multiple-choice questions from Humanity’s Last Exam; When2Call, NVIDIA’s tool-calling evaluation; and WildGuardTest, a safety classification evaluation.
Decision Index evaluation suite
The Decision Index is a weighted average of its four evaluations.
EvaluationCapabilityMetricWeight
Knowledge and hallucinationAccuracy (12.5%)
non-hallucination rate (12.5%)
25% (total)
ReasoningAccuracy
25%
Tool callingAccuracy
25%
Classification
Average F1
25%
Tasks and Answers
Each task is posed as a state, the text to decide about, and one or more questions about it. A question is either a choice among named options or a yes/no question. A model answers all of a task’s questions in a single call.
Decision Models
We send the state and questions to the model’s decision API. For a choice, the model returns the option it picks and a probability for every option. We score the option it picks, which is the option it gives the highest probability. For a yes/no question, the model returns the probability of yes, and we take its answer as yes when that probability is 0.5 or more.
Language Models
We send a fixed system prompt and the state and questions as text, and ask for a JSON reply that gives each question one of its allowed answers. We enforce this with structured outputs.
Artificial Analysis Decision Index Evaluations
The four evaluations in the Artificial Analysis Decision Index.
AA-Omniscience-Decision
Description: 500 private AA-Omniscience questions, each a choice among five options: four answers, one of them correct, and “I don’t know”. Included in the Artificial Analysis Decision Index at 25% weighting: accuracy and non-hallucination rate at 12.5% each.
Original dataset: AA-Omniscience
Scoring: Accuracy is the share of questions answered correctly. Non-hallucination rate is 100 × (1 − hallucination rate), where hallucination rate is the share of questions not answered correctly that were answered wrongly rather than with “I don’t know”.
HLE-MultiChoice
Description: The 513 text-only multiple-choice questions from Humanity’s Last Exam, each posed as one choice among its lettered answers in their original order. Included in the Artificial Analysis Decision Index at 25% weighting.
Dataset: cais/hle. We use its 513 multiple-choice questions without an image, out of 2,500 in all. The questions we use are listed below.
View all 513 tasks
Scoring: Accuracy, an exact match on the answer letter.
When2Call
Description: The multiple-choice test set of NVIDIA’s When2Call. Each task is a conversation with an assistant that has tools, posed as one choice among four candidate replies: answer directly, call a tool, ask for more information, or say it cannot answer. The order of the replies is shuffled per task. Included in the Artificial Analysis Decision Index at 25% weighting.
Original dataset: nvidia/When2Call
Dataset: We use a filtered set of 3,423 tasks, down from 3,652 in the original test set. We filter out 229 “ask for more information” tasks where there is ambiguity about whether more information is required. The filtered-out tasks are listed below.
View all 229 filtered-out tasks
Scoring: Accuracy, the share of tasks answered with the reply of the correct category.
WildGuardTest
Description: The test set of Ai2’s WildGuardMix. We ask up to three yes/no questions about each row: is the user’s request harmful, is the assistant’s reply a refusal, and is that reply harmful. Included in the Artificial Analysis Decision Index at 25% weighting.
Original dataset: allenai/wildguardmix (ODC-BY)
Dataset: We use 835 of the original 1,725 test rows. We keep rows where all three human annotators agreed on all three labels (prompt harm, response refusal, and response harm), based on the annotator agreement fields on Hugging Face, and exclude questions that consistently trigger model refusals to better isolate model capability.
Scoring: The average of three F1 scores: prompt harm, response harm and refusal.
The US military is investing heavily in artificial blood to address one of the leading causes of preventable battlefield deaths: severe blood loss. Around 90% of preventable battlefield deaths involve hemorrhage, and research suggests that beginning a transfusion within approximately 36 minutes of injury can dramatically improve survival.
Traditional blood supplies are difficult to maintain in combat zones. Whole blood lasts roughly 35 days under refrigeration, platelets typically expire within a week, and frozen blood components require thawing before use. Transporting these supplies is also dangerous. In 2007, seven US service members died when a helicopter transporting blood was shot down over Iraq.
DARPA’s $73 million RAPIID program aims to develop freeze-dried, shelf-stable substitutes for the three major blood components—red blood cells, platelets, and plasma—with a target of field deployment by 2029. The program builds on an earlier $47.5 million initiative called FSHARP.
The Technologies Being Developed:
Artificial red blood cells (ErythroMer): Developed by KaloCyte, ErythroMer packages hemoglobin recovered from expiring donated blood inside lipid nanoparticles that imitate cell membranes. This protects hemoglobin from degradation and reduces its tendency to constrict blood vessels. A specialized lipid helps regulate oxygen loading and release in response to pH differences between the lungs and tissues. The product can remain stable at room temperature for approximately two years. Animal experiments demonstrated improved blood pressure and oxygen delivery following extreme blood loss. KaloCyte is also investigating yeast-produced hemoglobin to reduce dependence on donations.
Synthetic platelets (SynthoPlate): Developed by researchers at Case Western Reserve University and commercialized through Haima Therapeutics, SynthoPlate uses lipid nanoparticles decorated with three specialized peptides. These target von Willebrand factor, exposed collagen, and existing platelets, helping synthetic particles accumulate at damaged blood vessels and support clot formation. The freeze-dried product lasts over a year at room temperature and at least three months at 122°F (50°C). In mice with low platelet counts, it reduced bleeding time and blood loss by more than 50%.
Freeze-dried plasma: Unlike the other technologies, dried plasma has existed since World War II. Modern screening and manufacturing methods have addressed many historical contamination concerns, and the FDA has authorized certain products for emergency use. Octapharma USA will supply plasma to the RAPIID program while also developing nanoparticle-based clotting technology.
Additional research includes Abioremedi’s approach, which uses alternative oxygen-carrying proteins such as myoglobin, cytoglobin, and neuroglobin rather than conventional hemoglobin.
Challenges and Next Steps:
Artificial blood has a complicated history. Earlier hemoglobin-based substitutes caused dangerous blood vessel constriction, heart attacks, and increased mortality in clinical trials. Meanwhile, growing real blood cells from stem cells remains extremely expensive and still requires refrigeration.
The newer technologies attempt to solve these problems through nanoparticle encapsulation, improved oxygen regulation, and freeze-drying. However, the most promising fully synthetic products have demonstrated results primarily in animals, not humans.
DARPA plans to conduct large FDA-regulated animal studies during the first year, followed by human trials in years two and three, coordinated through the University of Colorado Anschutz Medical Campus. Researchers must establish not only that individual components are safe, but also that they work together effectively. Earlier combined animal testing met most performance targets, although clotting remained weaker than with natural blood.
Some experts question whether DARPA’s 2029 deadline is realistic, particularly given the availability of more mature alternatives already tested in humans.
Mosqitter is a Pennsylvania-based startup developing chemical-free mosquito traps and AI-powered surveillance systems to help control mosquito populations and monitor the risk of mosquito-borne diseases.
Its traps attract female mosquitoes by mimicking human body heat, skin scent, light, and, in its larger model, exhaled CO₂. The company reports deploying more than 5,000 devices across five continents.
Mosqitter is now expanding beyond mosquito control with its Febris Monitoring System, which transforms its traps into a network for tracking mosquito populations and identifying potential disease risks.
AI-powered identification: Computer vision analyzes captured mosquitoes to identify their species. Results are cross-referenced with external datasets and validated by entomologists, producing confidence scores.
Predictive monitoring: The system incorporates environmental information, including temperature, wind, and humidity, with the goal of forecasting mosquito population changes and associated disease risks.
Disease surveillance: Febris can identify mosquito species associated with diseases such as dengue, but it cannot determine whether an individual mosquito carries a pathogen.
Institutional applications: The company is targeting governments, public-health agencies, municipalities, airports, and universities. It is working with the UK Health Security Agency on deployments at Heathrow and Gatwick airports.
Mosqitter sells its Mini trap for $199 and its larger Grand model for $1,999. Febris introduces recurring revenue through monitoring subscriptions, analytics, and data services.
The company expects over $1 million in revenue in 2026 and projects that Febris could generate over $3.5 million in 2027, based on contracts it is pursuing. It has raised $1.9 million to date and is seeking another $5 million.
The broader goal is to move mosquito control from reactive extermination toward continuous, data-driven public-health surveillance, potentially helping authorities identify emerging threats before outbreaks develop.
Apate.ai is an Australian cybersecurity startup that deploys undercover AI agents disguised as potential scam victims to engage fraudsters, waste their time, and gather intelligence before real people are targeted.
The company originated from research at Macquarie University and uses AI agents capable of answering scam calls, responding to suspicious messages, and infiltrating online groups.
Key findings:
AI-powered deception: Agents simulate different ages, genders, accents, and languages to convincingly interact with scammers.
Intelligence gathering: Agents extract bank account numbers, cryptocurrency wallets, malicious websites, and other information to help disrupt fraud operations.
Scale: Apate reports conducting 2.5 million conversations, collectively wasting 17 years of scammers’ time.
Real-world impact: TPG Telecom has redirected nearly 590,000 scam calls to Apate, while an Australian bank reportedly disrupted 22 scam campaigns using its technology.
Funding: Apate raised $8.15 million in August 2026, bringing total funding to nearly $10 million, with plans for international expansion.
Researchers at the University of Toronto’s Citizen Lab uncovered a spyware campaign called AdFection, which exploits online advertising networks to deliver zero-click spyware to smartphones without requiring users to interact with an advertisement.
Key findings:
Targeted attacks: Researchers identified 45 Al Jazeera devices targeted between 2024 and 2025.
Companies involved: The campaign was linked to advertising company Discover-Tech and spyware developer Candiru.
How it works: Attackers exploit real-time ad bidding to identify individuals, fingerprint their devices, and deliver malicious code through advertisements.
Scale: Discover-Tech advertisements appeared across more than 70,000 apps and 120,000 websites in 2026. Millions of devices may have been profiled, although this does not establish widespread infection.
Industry response: Google and other advertising platforms banned Discover-Tech following the investigation.
Researchers recommend ad blocking and stronger oversight of advertising networks to reduce exposure.
Ukraine’s Starlink access unexpectedly restricted:
Ukraine experienced an unexpected disruption to Starlink connectivity on October 14, affecting critical military operations. Ukrainian officials reported that SpaceX’s military-focused Starshield network began restricting connections for terminals moving faster than 100 mph (161 km/h), with some drones losing connectivity at speeds as low as 45 mph.
Military impact: Disruptions interfered with medium-range drone strikes, frontline logistics, and air-defense interceptions.
Partial recovery: Connectivity improved later that day, although the 100 mph restriction remained.
Musk’s explanation: Elon Musk stated that Starshield is controlled by the U.S. government and that restrictions are determined by the U.S. military and ultimately the president. He distinguished Starshield from commercial Starlink, which he said cannot be used for offensive military operations.
Ukraine began coordinating with the U.S., SpaceX, and NATO partners to establish alternative communications systems.
Critical in combat:
Starlink has become an essential part of Ukraine’s military infrastructure, supporting battlefield communications and remotely operated drones. Starlink-equipped drones can be controlled from hundreds of miles away and redirected toward new targets during missions.
However, this dependence creates strategic vulnerabilities:
2022–2023 restrictions: Musk reportedly declined to enable Starlink for a Ukrainian operation against Crimea in 2023, while Reuters previously reported that SpaceX restricted connectivity during a Ukrainian counteroffensive in 2022.
Recent cooperation: In February 2026, SpaceX restricted Russian forces’ use of Starlink terminals following a Ukrainian request, helping Ukrainian counteroffensive operations.
Strategic dependence: Ukraine’s military effectiveness increasingly depends on satellite infrastructure controlled by external organizations and governments.
The situation highlights how a privately developed communications network has become central to modern warfare.
Political tensions and suspicions:
Ukrainian officials suggested that the disruption might be connected to escalating political tensions between Washington and Kyiv, although no definitive cause had been established.
Musk’s criticism: Days before the disruption, Musk criticized President Volodymyr Zelenskyy for Ukrainian strikes against Russian oil refineries.
Intelligence concerns: Russian missile attacks against Kyiv occurred shortly before the disruption. Ukrainian officials claimed international partners failed to provide advance warnings, departing from established practices.
U.S. position: President Donald Trump stated that Washington continued sharing intelligence with Ukraine despite disagreements over attacks on Russian energy infrastructure.
Ukrainian suspicions: Presidential advisor Serhii Beskrestnov suggested the timing was politically motivated, but the article provided no evidence establishing that connection.
The incident raised questions about whether access to vital military infrastructure could be influenced by geopolitical disagreements.
Device manufacturers are increasingly recognizing they’re facing a permanent structural shift and IDC now expects any recovery in growth won’t come until 2028, at the earliest.
Google has launched Universal Cart in the U.S., allowing users to save and track products from multiple retailers in one centralized shopping cart across Google Search and Gemini. The tool preserves shopping progress, sends price-drop and restock notifications, recommends promo codes, and considers credit card benefits when suggesting payment options.
Built around Google’s Universal Commerce Protocol (UCP), the feature supports AI-assisted shopping and checkout. It represents Google’s growing push into agentic e-commerce, competing with Amazon, OpenAI, and others developing AI-powered shopping experiences.
California-based Science Corporation has developed Prima, a retinal implant that can partially restore central vision in patients with advanced age-related macular degeneration (AMD), a disease affecting more than 5 million people worldwide.
AMD damages the photoreceptors responsible for converting light into electrical signals, but much of the eye’s remaining neural circuitry can remain functional. Prima bypasses the damaged photoreceptors, allowing visual information to reach the brain through the surviving retinal neurons.
In a clinical trial, 26 of 32 participants assessed after 12 months (81%) experienced clinically meaningful improvements in visual acuity. Additionally, 27 participants reported using the system at home to read letters, numbers, or words. Some patients have been able to perform more complex activities, although the implant does not restore normal vision.
How Prima Works:
Capture: Camera-equipped glasses record video at approximately 5 megapixels and 30 frames per second.
Process: A handheld processor converts the footage into signals for the implant and allows users to adjust brightness, contrast, and zoom.
Transmit: A miniature projector in the glasses beams near-infrared light (880 nm) onto the retinal implant.
Restore: The implant’s 378 pixels contain photovoltaic cells that convert infrared light into electrical impulses, stimulating surviving retinal neurons that transmit information to the brain.
The implant is wirelessly powered by the infrared light itself, eliminating the need for an internal battery.
Although Prima does not restore normal vision, it represents a significant advancement in treating previously irreversible vision loss. The technology also comes with surgical risks, and its clinical benefits have been demonstrated in a relatively small trial. Prima received European regulatory approval in July 2026 and is still awaiting FDA clearance in the United States.
This is just a really cool website that lets you find land for sale that borders public land which means that you can’t develop the public land but you explore it!
Tesla finally unveiled its long-delayed Roadster 2, demonstrating a prototype capable of accelerating from 0–60 mph in under one second, driving upside down, and briefly launching itself into the air!
The car accomplishes this through a combination of powerful electric motors, an active suction system, and SpaceX-developed cold-gas thrusters. Although Tesla demonstrated some remarkable capabilities, the production specifications are still company claims rather than independently verified performance figures for a commercially available vehicle.
How the technology works:
Powertrain: Two rear-wheel-drive electric motors are designed to produce over 1,341 horsepower and 1,400 Nm of torque. Tesla says the car weighs approximately 1,700 kg and will offer over 300 miles of range.
Suction system: Four 20-kW fans pull air from underneath the vehicle, creating substantial downforce. Retractable skirts help maintain the pressure difference, allowing the car to generate extreme traction even at low speeds. Tesla claims the system can produce over 2 Gs while cornering and demonstrated the prototype driving upside down.
Rocket thrusters: A 200-liter tank stores nitrogen pressurized to 10,000 PSI. This powers cold-gas thrusters that Tesla says can generate over 80 kilonewtons of thrust, helping the car achieve its sub-second acceleration.
Flight system: An additional 400-liter nitrogen tank and six thrusters allow the vehicle to briefly launch into the air, stabilize itself, and cushion its landing. This configuration takes up the passenger-seat area, turning the vehicle into a single-seater.
Limitations: Unlike a conventional rocket engine, the cold-gas system loses thrust as the nitrogen pressure decreases. Refilling the tanks currently takes six to seven hours, although Tesla hopes to reduce this to approximately 30 minutes.
Tesla plans to offer two primary versions of the Roadster 2:
Performance: A claimed top speed of 210 mph, 0–60 mph in 1.6 seconds, and the suction system without rocket thrusters.
Thrust: A claimed top speed of 230 mph, 0–60 mph in under one second, and additional rocket thrusters.
The Performance model requires a $5,000 initial reservation fee followed by another $45,000 within 10 days. The Thrust model requires a $5,000 initial fee followed by another $245,000 within 10 days.
Production is currently planned for 2028, although Tesla has repeatedly delayed the Roadster since its original announcement in 2017.
It’s also worth noting that briefly launching a prototype into the air is very different from having a road-legal flying car. Significant regulatory and safety questions remain, particularly around whether the flight system could actually be approved for public use.
Elorian has introduced Glimpse, a multimodal AI model designed to reason about visual information while repeatedly returning to the images and videos it is analyzing.
Traditional vision-language models can process images, but much of their reasoning is expressed through language. Elorian argues that this can limit their ability to examine fine visual details or maintain an accurate understanding of objects and their relationships throughout a complicated problem.
Glimpse attempts to address this by integrating visual inspection directly into its reasoning workflow. Rather than treating an image as a single piece of information to interpret, the model can return to the visual evidence, inspect specific regions, and revise its interpretation.
The model incorporates several notable capabilities:
Visual scratch space: Glimpse can revisit images, zoom into particular regions, crop visual inputs, and examine details that may have been missed during its initial interpretation.
Adaptive video inspection: Instead of relying exclusively on a fixed sample of video frames, the model can revisit important segments at a higher frame rate to examine motion and changes over time.
Code-assisted visual reasoning: Glimpse can use code and image-processing tools to perform measurements and transformations, allowing it to investigate visual information more precisely.
Multimodal subagents: The system can coordinate specialized agents that exchange text, visual artifacts, and internal representations to work on different parts of a visual reasoning problem.
Architecturally, Elorian says Glimpse was developed by post-training an existing open-weights foundation model using only a few hundred GPUs, rather than training a completely new foundation model from scratch.
The company describes the model as being trained to use multimodal tools and coordinate visual reasoning processes. It also reports competitive performance against leading frontier models on several visual benchmarks.
However, important technical details have not been disclosed. Elorian has not identified the underlying foundation model, its parameter count, or the precise training methodology. We also don’t know exactly how the model maintains and updates its internal visual representations.
Consequently, it is difficult to determine how much of Glimpse’s performance comes from changes to the underlying neural network versus specialized post-training, tool use, and multiagent orchestration.
Regardless, the potential applications are interesting. A model capable of repeatedly examining visual evidence could be useful for scientific microscopy, medical imaging, industrial inspection, video analysis, and complex diagrams.
Elorian’s longer-term ambitions include 3D scene reconstruction, physical simulation, and robot control. Those remain future goals rather than demonstrated capabilities of Glimpse.
1. The Concerns
The authors worry that AI labs prioritize the speed and quantity of formalized mathematics over quality, maintainability, and usefulness. They highlight several examples:
Math, Inc.: Its AI-generated sphere-packing formalization was technically impressive but difficult enough to maintain that researchers preferred rewriting it.
Anthropic: Claude reportedly generated 13 million lines of Lean code formalizing Fermat’s Last Theorem in 11 days, but Kevin Buzzard continued developing a more modern, reusable version.
OpenAI: Its Navier–Stokes formalization demonstrated the value of verification but raised concerns about attribution, readability, and mathematical understanding.
The central concern is that formally correct proofs can still be poorly structured, difficult to reuse, and disconnected from meaningful mathematical research.
2. The Positives
The authors praise AI contributions to the unit distance conjecture and research on the Riemann zeta function as examples of work that advanced mathematical understanding.
They envision combining human mathematicians, AI systems, and the Lean theorem prover. AI would generate mathematical arguments and proofs, Lean would verify their correctness, and humans would guide research and interpret results.
A comprehensive, high-quality formal library could dramatically accelerate mathematical discoveries and eventually benefit engineering and science.
3. Lessons So Far
The authors describe six projects demonstrating different approaches to formalization:
Mathlib: A community-maintained Lean library emphasizing high-quality, reusable mathematical foundations.
Formal Conjectures: Creates precise formal statements of open mathematical problems for researchers and AI systems.
Annals Challenge: Formalizes statements from leading mathematics journals, revealing missing definitions and gaps in existing libraries.
Tau Ceti: Uses AI agents and expert-designed roadmaps to expand Mathlib, with adversarial review for correctness, documentation, and reuse. It reportedly contains approximately 3 million lines of code.
Formal Frontier: Uses collaborating AI agents to transform mathematical textbooks into interconnected formal libraries, with 41 released libraries totaling approximately 200,000 lines.
Palomar: A registry that improves transparency by documenting formalization projects, their origins, AI involvement, and relationships to existing research.
The authors outline principles emphasizing reuse, documentation, mathematical generality, transparent review, open-source availability, and long-term maintenance.
Their central lesson is that AI-generated formalizations require human-guided organization and quality standards rather than unrestricted automated production.
4. The Way Forward
The authors predict that most mathematics could be represented in usable, interconnected formal libraries within three years, potentially sooner.
They propose strengthening Mathlib while rapidly scaling Tau Ceti and Formal Frontier. Mathematicians would develop subject-specific roadmaps, organize workshops, and establish realistic estimates of computational and financial requirements.
They also advocate nonprofit or academic organizations to manage infrastructure, engineering resources, and AI inference capacity, ensuring mathematicians retain control over research priorities and formalization standards.
DeepSeek released an updated version of its V4.1-Flash model! This is more of an incremental improvement through post-training than a major advancement in the underlying architecture, with better reasoning, coding, and agentic performance. Still, dirt-cheap inference on a model this powerful is always welcome!
A new study published in Science Advances by Stanford researchers found that brief physical contact between healthy and bleached corals can dramatically accelerate recovery. Coral bleaching occurs when heat stress causes corals to lose the symbiotic algae that provide much of their energy. Researchers discovered that when genetically identical healthy and bleached corals touch, their tissues fuse, allowing beneficial algae and nutrients to transfer between them.
The algae began transferring within three days, and as little as five to seven days of contact was sufficient to initiate recovery that continued independently. In laboratory experiments, bleached corals fully recovered within seven weeks, compared with the months or years recovery can take naturally. Researchers also identified a slower transfer mechanism involving coral polyps, but it was insufficient to produce full recovery on its own.
The main limitation is genetic compatibility: successful rescue through tissue fusion requires genetically identical corals, making large-scale reef restoration difficult. Researchers hope to develop alternative methods for delivering beneficial algae without requiring direct contact.
Across the northeastern United States, the removal of aging and obsolete dams is restoring river ecosystems and allowing migratory fish to return to their historic spawning grounds. While large projects like the Klamath River dam removals attract national attention, thousands of smaller dams have fragmented waterways for centuries, particularly harming river herring, salmon, and other migratory species.
The results have been impressive. In Maine’s Penobscot River, roughly 6 million river herring returned in 2025, while restoration efforts in Plymouth, Massachusetts, increased annual herring runs from around 25,000 to frequently more than 200,000. Removing these barriers also restores nutrient flows, improves habitats for other wildlife, and can eliminate flood risks associated with deteriorating infrastructure.
Since 1912, more than 2,300 dams have been removed nationwide, with most removals occurring in recent decades. However, at least 35,000 dams still obstruct fish migration in the Northeast. Funding constraints, complicated permitting, and local opposition remain significant obstacles. The broader takeaway is that even removing relatively small barriers can produce enormous ecological benefits, demonstrating how quickly nature can recover when rivers are reconnected.
Tool for automatic PS5 executables porting to Linux and Windows.
…
Per-game progress (stage reached, open blockers, libraries and GPU coverage) is tracked on the blocker map, a community project built from public GitHub data (source). Test reports go in the Game bring-up discussions, one thread per title.
…
SDL-mapped game controllers are supported, including analog sticks and triggers. Keyboard and mouse controls can be configured with an
anyps5-input.inifile. See input mapping for the supported devices and configuration format.
ArtCraft is the IDE for interactive AI image and video creation. Compose in 2D, stage scenes in 3D, and choose the models that fit your work. Turn prompting into crafting with visual tools for precise, repeatable results.
This repo exploded over the last week and kicked off a reverse engineering frenzy. Every developer on the planet has now reverse engineered their smart devices, proprietary software and game consoles.
It is actually one of a whole series of software that is open sourcing proprietary software. You can find the rest on the storytold GitHub profile and at this website: getcraft.com.
On October 17, Meta suspended the official Instagram account of No Kings, a nationwide protest movement opposing the Trump administration, for approximately 90 minutes during a major day of demonstrations and early-voting mobilization. According to Ezra Levin, co-founder of Indivisible, one of the organizations leading the movement, the suspension occurred during the peak of more than 3,000 planned events, disrupting organizers’ ability to coordinate protests, communicate with supporters, livestream activities, and encourage participation in the upcoming midterm elections.
Although the account eventually became publicly accessible again, organizers reported that they remained locked out of its administrative controls for several additional hours, preventing them from publishing new content. Meta spokesperson Andy Stone later explained that unusually high levels of posting, sharing, and commenting had triggered an automated account suspension. Meta acknowledged that the automated action was erroneous, apologized, and said the account’s functionality had been restored.
The incident sparked concerns about the enormous influence social media companies have over political organizing and democratic participation. Levin compared the suspension to recent controversies involving Meta’s alleged political favoritism in Brazil, while organizers demanded greater accountability and safeguards against similar disruptions.
Thinking Machines developed a way to make inference deterministic. The most surprising thing, to me, about this whole article is that normally inference is not deterministic.
Different batch sizes trigger different computation strategies. For example; when not under load a batch may be split across several GPU cores while under load the batch may handle each request calculation under one core.
Those strategies add numbers in different orders,
The culprit is floating-point non-associativity. That is, with floating-point numbers:
\((a + b) + c \neq a + (b + c)\)
…
If you add together two floating-point numbers with the same exponent, it looks similar to integer addition. For example,
\((1.23 \times 10^2) + (4.56 \times 10^2) = (5.79 \times 10^2)\)
But what happens when we add two floating-point numbers with different exponents, such as 1230 and 23.4? In this case, the exact result is 1253.4. However, we can only maintain 3 digits of precision at a time. Floating-point addition will thus drop the last 2 digits and obtain the value
(or 1250).
If the difference between the probabilities of two tokens is small enough this inconsistency can flip the winning token even at temperature 0. And once a different token is chosen, subsequent text can diverge because the model now continues from different input.
Thinking Machines solved this problem by making each token’s arithmetic follow the same order regardless of batch size. They built batch invariant kernels.
They identifies three parts of inference that can be nondeterministic.
RMSNorm
Matrix multiplication
Attention
They made RMSNorm deterministic by keeping the summation order fixed as batch size changes. In matrix multiplication they fixed the kernel configuration, including consistent tile size and hardware instructions. Finally, they keep cached and new tokens laid out consistently, and split attention calculations into fixed size chunks to address Attention.
The Result, in an RL experiment matching inference and training numerics helped avoid a loss spike and reward collapse. This is not to say that this kind of training always produces lower loss.
If we train without off-policy correction (i.e. importance weighting), our reward collapses partway through training, whereas adding an off-policy correction term allows training to proceed smoothly. But, if we achieve bitwise identical results between our sampler and trainer, we are fully on policy (i.e. 0 KL divergence) and can also train smoothly.
As we can see, although this novel approach solves the problem introduced by nondeterminism, this problem can also be solved in a generally accepted way. In addition, the generally accepted way is currently faster, although Thinking Machines states this is because, “Much of the slowdown comes from the fact that the FlexAttention integration in vLLM has not been heavily optimized yet.”
He, Horace and Thinking Machines Lab, “Defeating Nondeterminism in LLM Inference”,
Thinking Machines Lab: Connectionism, Sep 2025.
This is just a fun little website, complete with responses from the cartographers themselves who are mapping North Korea. Some of them have even been to North Korea! Above is a screenshot from the end of the website where you can click to open different cartographers messages to the author about why they map North Korea.
What began as a 2017 MIT engineering class project has become an FDA-cleared medical device. Developed by startup Encora Therapeutics, the wrist-worn device uses mechanical vibrations to reduce involuntary shaking in people with essential tremor.
The wristband delivers patterned vibrations that activate sensory receptors in the wrist. These signals travel through peripheral nerves to brain regions involved in movement coordination, helping disrupt the abnormal signaling associated with tremors. Unlike invasive treatments such as deep brain stimulation, the device requires no surgery.
In trials involving 47 and 59 patients, more than 70% experienced meaningful improvements. In February 2026, the FDA cleared the device for adults with essential tremor.
Encora is now scaling production and pursuing insurance coverage. Although originally designed for Parkinson’s disease, its current clearance is limited to essential tremor. The company hopes to eventually expand the technology to other neurological movement disorders.
I really suggest reading this one! The story behind it is pretty cool. The person who came up with the idea tried to kill it before it even began.