Ukraine's intelligence directorate is seeking information about a new optical system found on Russia's latest Gerans. Screenshots via Telegram/GUR Ukraine is warning about a new navigation system on Russian Gerans called the Susanin. It uses a camera and a pre-saved library of images of Ukraine's terrain to fly independently. The capability, similar to a tech on cruise missiles, could help Gerans completely bypass jamming. Last week, Ukraine's defense intelligence directorate put out a call for information on a device newly found in Russia's latest killer drones. The "Susanin," a small cylinder with a camera lens, is an optical system that enables Russia's long-range drones to navigate Ukraine's skies without relying on satellite or radio signals. Should Moscow deploy them at scale, the modules would introduce a major new challenge for Kyiv, allowing Russian one-way drones to bypass jamming — a key pillar of Ukraine's air defense network. Per the intelligence directorate, or GUR, the Kremlin has been fitting them onto the turbojet Geran-4 and Geran-5 K5, which fly at 280 mph or faster and are seeing growing use by Russia this year. These drones are already significantly harder for Ukraine to destroy by force, since they outpace typical Ukrainian interceptor drones. The Susanin works by storing images of Ukraine's terrain on the drone, then, while in flight, using a downward-facing camera to observe the ground and compare the live feed to the pre-saved imagery. Using these comparisons, the drone can independently determine its position and how to reach its target. "Even Google Earth imagery is sufficient for basic orientation," wrote Serhii "Flash" Beskrestnov, a prominent drone analyst connected with Ukraine's military, on Sunday. Beskrestnov warned that Russia would likely continue actively developing Susanin systems for its drones by next year. "Our enemy already uses such systems on small Molniya, V2U, and Lancet drones," he added. Cruise missile tech on a drone The technology is similar to what some cruise missiles use to reach their targets in jammed environments. The Tomahawk, for example, features what analysts call Digital Scene Matching Area Correlation to snap photos of landmarks under the missile and compare them to satellite imagery stored onboard. Higher-end Russian cruise missiles, such as the Kh-101, use a Russian version of this tech called Otblesk-U. But the Kh-101 is estimated to cost $2.4 million each, while its long-range drones cost just a fraction of that: The delta-wing Geran-4 is priced at an estimated $50,000 to $70,000 apiece, while the Geran-5 is believed to cost about $100,000 to $115,000 each. Debris of a drone lies on the ground near the Epicenter retail center following a Russian Geran-4 strike in Kharkiv. Polina Kulish/Gwara Media/Global Images Ukraine via Getty Images Russia has been producing these drones at a much faster pace, launching over 3,500 jet attack drones at Ukraine in September alone. Kyrylo Liukov, who leads the uncrewed technologies department at the Serhiy Prytula Charity Foundation, told Business Insider that it's "quite possible" Russia could scale up its use of navigation systems like the Susanin. "Russia is already using this navigation system for medium- and deep-strike operations," said Liukov, adding that its deployment on Geran-4s and Geran-5s has been increasing recently. Still, Liukov said these systems have their limits. The Susanin is only useful for pre-programmed drone attacks by itself, and wouldn't be able to help the drone find a moving Ukrainian armored vehicle or mobile missile launcher. "In other words, these are predominantly planned, static targets," said Liukov, whose organization raises funds for Ukraine's military and drone manufacturers. But pre-arranged flights fit well into Moscow's strategy with Gerans, which are launched in large numbers to overwhelm Ukrainian air defenses and wreak havoc on its cities. Fighting the Susanin If a Russian turbojet drone is using a Susanin system, Ukraine's only way to counter it, for now, is to use physical force, Liukov said. Ukraine's options in this area remain limited, with its military still in the early stages of introducing advanced interceptor drones capable of catching turbojet Gerans. Another way is to launch rare and expensive anti-air missiles, which are typically reserved for Russia's more powerful cruise or ballistic missiles. One possible solution Ukraine could eventually explore, Liukov said, is to blind the Susanin's camera. The camera's view of Ukrainian terrain could also be obscured by clouds, rain, or fog, he said. "For high-quality route recognition using Susanin, a flight altitude relative to speed is required that will allow for accurate orientation using the captured images, as well as favorable weather conditions," Liukov said. Read the original article on Business Insider

Albert Heijn’s stores are open and its checkouts work. Its distribution centers are a different story. … Albert Heijn’s Cybersecurity Incident: Checkouts Work, but Some Distribution Centers Don’tRead more The post Albert Heijn’s Cybersecurity Incident: Checkouts Work, but Some Distribution Centers Don’t appeared first on DataFLOQ.

Anthropic researchers said small AI models might be able to self-replicate on cheap devices such as smart ovens or lawn mowers. Vitalii Petrushenko/Getty Images Small AI models might soon be capable of worming their way into smart ovens and refrigerators. Anthropic researchers wrote that self-replicating systems might latch onto everyday devices for compute. Forget AI slop. We might soon have "AI worms" and AI "slime mold." In this latest instalment of AI nightmare fuel, Anthropic's researchers have discovered that small rogue AI models might — as soon as the end of 2026 — start worming their way into home devices like smart ovens, smart refrigerators, automatic vacuum cleaners, and lawn mowers. These internet-connected devices typically have "minimal security measures in place," allowing small models to feed off of their computing power and self-replicate, Anthropic's AI capabilities research team wrote in a report on Monday. This could take the form of "a slow, distributed 'digital slime mold,' living on relatively insecure compute," the researchers wrote. The team studied the behavior of small Chinese models, such as Z.ai's GLM-5.2 and Alibaba's Qwen, as well as others like Google DeepMind's Gemma4-26B-A4B. They gave a prompt to the Gemma model while running inference on 15 cheap off-brand Android TV boxes, each costing $55, and found it worked. Worms aren't new in the software world. Traditionally, they're a type of malware that can copy itself and spread across networks. Anthropic said an AI worm takes that idea a step further: it can reproduce itself on another device while carrying its AI model with it. This finding showed them that AI models can use the hardware in domestic smart appliances as an alternative to data center compute. These will be challenging to shut down, they said. "As a prediction that goes beyond our experiments, we expect a future awash with agents running on all types of hardware, including some that aren't under human control," the researchers wrote. They added that while we should "harden our systems" to prevent such hacks, we should "also prepare to deal with a world where they are here to stay." The research comes as rogue AI systems are becoming an increasing topic of concern in the tech sphere. Last month, there was an outpouring of criticism directed at frontier AI labs like Anthropic and OpenAI after former and current employees warned of the risk of an AI catastrophe within the next 10 years. Read the original article on Business Insider

Eileen Gu says her mom set her up for success by letting her choose her own path. Noam Galai/Getty Images For Apex For Youth Olympic skier Eileen Gu said people often mistake her mom for a “tiger mom.” But the opposite is true: Her mom fiercely protected her freedom to make her own decisions. Today, Gu said, she and her mom are more like "cofounders" in everyday life. Eileen Gu credits her mom with giving her the freedom to find her own path. Speaking at the Karim Family Foundation's Inspiring Women Leadership Series in Singapore on Monday, Gu spoke about her upbringing and how her mom set her up for success from a young age. "I think she really optimized for play, and she optimized for optionality. She put me into everything," Gu said, adding that she tried ballet, piano, archery, basketball, soccer, and cross-country skiing. She also attended math and science camps and went for singing lessons. "There were lots of things, right?" Gu said. "And so, the point being that when I was younger, she really gave me the full spread so that when I was older — when I knew myself better and when I was responsible to make those decisions — I could make an informed one having tried all the options." The Olympic freestyle skier said that being entrusted with that agency and responsibility at a relatively young age was critical to her becoming independent. Gu said people often wrongly assume her mother was a "tiger mom," a label she described as a controlling, helicopter-style parent who forces their child down a particular path. "She was in many ways the opposite of that, where she just vigorously provided me with freedom and agency, and enforced my freedom really intensely," Gu said. "That was the effect of her tiger mom." Now, Gu said, they still do almost everything together. "As I got older, I kept calling us cofounders because I really think that's the best way to encapsulate our partnership," Gu said. "We travel together every day. We live together. We work together. We develop together. We change together. We evolve." Gu said neither she nor her mother has all the answers, and they are both figuring life out as they go. They often start with different views before talking them through, finding common ground, and sometimes even landing on each other's original sides, she said. "We come in with unique perspectives and, most importantly, a respect for one another. We have dialogue that is respectful, that is critical, that is conducive, that is productive," she said. Read the original article on Business Insider

Character Technologies Inc. (Character.AI) said in an announcement late on October 12, 2026 that it has filed a voluntary Chapter 11 case in the U.S. Bankruptcy Court for the District of Delaware to pursue a sale of all or substantially all of its assets, with Disney Worldwide Services, Inc. as stalking horse bidder in a $20 million agreement. Case Filing and Court Record The case information site maintained by Stretto, Inc., the claims and noticing agent for the case, records a petition date…

CXMT Corp. declared it achieved a breakthrough in chip technology that rivals like Samsung Electronics Co. had pursued for years, an unsubstantiated claim that jolted stocks and reflects China’s aggressive push into a critical sphere.

Kristalina Georgieva told a panel that said while AI investment is helping sustain global growth, it doesn’t necessarily mean better career prospects for young people. Roslan RAHMAN / AFP via Getty Images The IMF director warns that though AI may sustain global growth, young people may not benefit. The IMF predicts AI affecting 40% of jobs globally, with 60% in advanced economies like the USA. Georgieva urges a rethink of taxation and education to counter AI's impact on the job market. AI could help keep the global economy growing while pulling the first rung of the career ladder out from under young workers. That’s the warning from International Monetary Fund managing director Kristalina Georgieva, who described AI as a “tsunami” for jobs during a conversation in Bangkok, where Thailand is hosting the IMF–World Bank Annual Meetings from October 12–18. Georgieva told a panel on Monday that while AI investment is helping sustain global growth, a stronger economy doesn’t necessarily mean better career prospects, especially for young people. "You have those that build the infrastructure for AI, and many of them are in Asia, including Thailand," Georgieva said. "So that shift in higher gear is what is keeping the world economy growing at 3% — around 3% — despite the fact we are being dragged down by the energy supply shock." The IMF estimates that AI will affect about 40% of jobs worldwide, rising to 60% in advanced economies, with some jobs enhanced, others downgraded, and some eliminated. "And when we look at this, what we recognize is that the people to worry about first and foremost are young people who are coming into the labor market," Georgieva said. "Normally, they would get into these routine, repetitive jobs and then grow up the ranks," Georgieva said. "But if they're gone, I don't think any young person wants to be a barista in a coffee shop, as you know, [their] professional future." LinkedIn executive Aneesh Raman has previously described the career ladder in the age of AI as a “climbing wall,” requiring sideways moves rather than linear promotions. An HR strategist has also recently told Business Insider that it's time to start hiring for soft skills and capabilities instead of whether someone could make a pivot table. Georgieva said the government would need to rethink taxation and education as AI changes the job market. "Traditionally, governments would generate revenues by taxing mostly people. Well, if AI changes the structure of the labor market and then there is less labor, more capital, what do we tax?" Georgieva said. "Get governments to recognize they have a huge responsibility to rethink how education should be done in the world of AI," Georgieva added, "And recognize that no more we can acquire a skill and hold on to it for our life, and that lifelong learning is actually a paramount necessity." Read the original article on Business Insider

Sprints, a venture capital firm that backed Revolut early, has closed its fifth fund of €725 million ($810 million).

Technology companies are on track to become larger net issuers of long-term debt than the US government in 2027, former UK Prime Minister Rishi Sunak said, describing the scale of capital demand driven by artificial intelligence investment.

Here's a glimpse of what's dragging tech and chipmaker stocks down from the US to Asia. (Source: Bloomberg)

DXC Technology said on October 13, 2026, that it is partnering with LangChain to advance governed agentic AI across mission-critical enterprise IT environments, with its OASIS orchestration platform already in production and scaling toward 125 customer environments across regulated industries and multiple geographies. LangChain is the company behind the open-source LangGraph agent framework and the LangSmith agent engineering platform. LangGraph Foundation and New LangSmith License DXC was…

Scientists have uncovered a surprising hidden ability of light that allows it to develop spinning, twisting behavior as it travels through empty space, without special lenses, mirrors, or materials. Researchers discovered that light’s internal geometry can naturally produce left- and right-handed patterns, challenging long-standing assumptions about how light behaves. By carefully arranging a beam’s structure, scientists can control where and how these spinning patterns emerge, unlocking a new way to manipulate light.

Elon Musk said Tesla is working to design a reliable human hand for its Optimus robot. Christoph Soeder/picture alliance via Getty Images The human hand is a difficult engineering challenge for the robotics industry to solve. Elon Musk said getting the motor skills right for the humanoid hand is "shockingly difficult." He said he has yet to see a robot hand that can screw in a self-tapping screw. We have boxing robots, but can they screw in a nail? Designing a useful human hand for humanoids is recognized as one of the more difficult engineering challenges within the robotics industry. Elon Musk, the man who wants to put data centers in space and colonize Mars, can also admit that. "The hands and arms and the fine motor skills are incredibly difficult to get right," Musk said in a virtual interview with Tesla Owners of Silicon Valley. "Like shockingly difficult." Musk cited different basic skills, such as screwing a nail into a piece of wood or opening a jar of pickles, that he said are much harder to replicate through robot hands because they require the right application of force, precision, and sensors. "There's no hand design that I've seen, not even sort of a demo that does that," Musk said of screwdriving a self-tapping screw into wood. "But that's what we're trying to achieve with Optimus." There's no standard for how best to design a human hand for robots capable of dexterous manipulation. Design choices can include where to place the motors — directly in the hands or within the arms — and which sensors are appropriate for reliable manipulation. Those choices also come with tradeoffs. Placing motors in the hand means smaller motors that may be more prone to breakage and produce less force. Musk pointed to the challenge of knowing where to put the sensors. "If you put them too close to the surface, they're going to get damaged," he said. "And the touch sensors — can they elastically deform like our fingers? Well, actually, it's kind of hard to get touch sensors that elastically deform and can handle high cycle life and won't just break after a few thousand cycles." Some robotics companies have opted not to mimic the human hand, while others continue to tackle the problem. 1X, the Palo Alto-based humanoid maker, told Business Insider's Rya Jetha earlier this year that it has achieved "the closest hand to human-level dexterity" with its NEO robot, developing a hand with 25 degrees of freedom and waterproofing. Read the original article on Business Insider

Zuckerberg said that in late 2025, Meta weighed a $200 billion AI bet before seeing its new models. Andrej Sokolow/picture alliance via Getty Images Zuckerberg said Meta weighed a $200 billion AI bet before seeing its new models. Alexandr Wang said Meta's CFO warned failure would be "very, very bad" financially. Meta's capex spending has been substantial and has raised eyebrows among investors Mark Zuckerberg was ready to bet as much as $200 billion on AI, and Meta’s finance chief wanted the person building it to understand what failure would mean. On the Cheeky Pint podcast released on Monday, Meta CEO and Alexandr Wang, Meta's chief AI officer, recounted conversations from late last year about enormous investments before Meta’s new models had proven themselves. “At the end of last year, we're like, all right, we are quite excited about the progress. We haven't seen the models yet,” Zuckerberg told show host John Collison. “I'm going to, like, you know, we're going to write this check for, like, I don't know, like $100 billion, $200 billion. I'm like, 'Alex, are you sure this is going to work?'” Zuckerberg said of Wang. Zuckerberg did not specify the timeframe or exact scope of that contemplated commitment. Wang said he had encouraging research updates at that point, but success was far from certain. At an executive gathering in Hawaii, Wang said that CFO Susan Li repeatedly pulled him aside. “Susan pulled me aside many times and was like, 'You know, I just want to tell you, if this doesn't work, like, you know, this is going to be really, really bad. It's going to be very, very bad in financial terms.'” Wang said of Li. Meta's capex over the past few quarters has been substantial, raising eyebrows among investors. Meta’s latest guidance puts 2026 capital expenditures, including finance lease principal payments, at $130 billion to $145 billion. In forecasts previously seen by Business Insider, Cantor Fitzgerald analysts expect the spending to reach $200 billion in 2027. Investors have shown limits. Meta shares fell more than 6% in after-hours trading following April’s earnings report, despite strong results, after the company raised its annual spending forecast to $125 billion to $145 billion. Inside Meta, the leadership is well aware of the stakes. In Wang's words, "this is a fuck ton of risk.” "So I think that in order for it not to work out, it's not just that we have to fail, AI overall has to fail," Zuckerberg said. "And then even if that happens, we still have all these other uses to productively use compute." Read the original article on Business Insider

arXiv:2610.12602v1 Announce Type: new Abstract: Continual learning (CL) seeks to adapt pretrained models while preserving prior capabilities. Yet the tasks supplied for preservation cover only part of what a model knows, while forgetting can extend beyond them. Recent work shows that language models can supplement this limited evidence by generating useful replay without access to their original training data. This raises a question about self-preservation: can a model help choose preservation targets that protect its own capabilities more broadly? To study this question, we organize task descriptions and their examples into Task Banks, whose entries can be selected and combined to guide replay generation by a frozen model. Continual instruction tuning with Qwen models of varying sizes yields three findings. First, the new task and learning rate shape which prior capabilities degrade, while tracking previously correct answers reveals losses that average accuracy can conceal. Second, generated replay reduces forgetting on tasks outside the bank. Combining partial banks strengthens this protection, improving mean retention of previously correct answers over either bank alone across the tested sizes. Third, the model can also help choose replay targets: selection informed by the new task improves mean held-out retention over random selection at 4B, with similar new-task gains, although smaller models show no corresponding advantage. Under a fixed example budget, however, replacing new-task examples with replay improves preservation at the cost of smaller new-task gains. Together, these findings show that models can support broader self-preservation by generating replay data and, in some settings, choosing which tasks to replay.

arXiv:2610.12570v1 Announce Type: new Abstract: Speculative decoding accelerates large language model inference by using a lightweight drafter to propose multiple tokens that are subsequently verified in parallel by the target model. Recent diffusion-based drafters make this paradigm particularly attractive: deep attention-based neural backbones provide strong predictions, while drafting multiple future positions in parallel avoids the sequential cost of autoregressive generation. This parallelism, however, comes at a cost. Tokens within a draft block are predicted without access to the realized identities of preceding draft tokens, removing intra-block dependencies that would naturally be available under autoregressive generation. We introduce DCOCO (DFlash Contextual Corrector), a lightweight mechanism that restores these dependencies while preserving parallel drafting. Rather than committing to each position's top prediction, DCOCO retains multiple candidate tokens and evaluates their contextual compatibility across adjacent draft positions in parallel, enabling contextually informed token selection at low additional cost. Across four verifier backbones and ten benchmarks, DCOCO achieves the highest average throughput in every evaluated configuration, spanning greedy and stochastic decoding with and without extended reasoning, with speedups of up to 6.6x over vanilla decoding. Its advantage persists under concurrent serving: DCOCO outperforms the strongest competing method by 6-23% across 4-32 concurrent requests, while other sequential correctors fall below uncorrected DFlash at these loads. These results demonstrate that recovering intra-block dependencies through parallel contextual scoring delivers consistent acceleration across decoding regimes and serving loads.

arXiv:2610.12601v1 Announce Type: new Abstract: Prompt optimization is an attractive alternative to reinforcement learning for adapting language-model agents, since it is cheap and produces human-readable artifacts. On factual tasks, however, it carries a failure mode that scalar-reward reinforcement learning does not. An optimizer that reads the training examples can write their answers into the prompt, so an in-distribution gain may be memorization rather than transferable behaviour. We apply reflective prompt evolution (GEPA) to the two system prompts of a multi-agent deep-research agent and evaluate it under a topic-disjoint split in which no topic is shared across train, validation, and test. Held-out citation reward rises by 0.079, from 0.702 to 0.781, with a bootstrap 95\% confidence interval of [0.051, 0.107], and the gain replicates across three seeds (mean 0.107). Nearly half of the winning prompt is a memorized answer key, yet a controlled ablation shows this content earns no reward even on the exact training questions it encodes. The large validation-minus-held-out gap is therefore small-sample selection under a noisy judge, not memorization. Varying the verifier isolates what makes the gain transfer. Decomposed per-dimension feedback is the causal ingredient, and collapsing it to a single scalar reward makes held-out transfer negative. The gain persists under an out-of-family judge and is largely reproduced by a hand-written version of the discovered recipe. Credible evaluation of prompt optimizers therefore requires held-out topics, multiple seeds, and verifier ablations. We release the candidate prompts, lineage diffs, figures, tables, and analysis.

arXiv:2610.12617v1 Announce Type: new Abstract: Geopolitical and strategic intelligence reports constitute a rich source of expert knowledge and have increasing value in the current political climate, but their heterogeneity and analytical complexity pose challenges for retrieval-augmented question answering. We introduce a benchmark for grounded question answering over 415 open-source geopolitical and strategic intelligence reports, comprising 5,525 document-grounded question-answer pairs spanning factual, inferential, comparative, and risk-oriented questions. We propose a document-aware generation pipeline that profiles reports before generating questions and validate dataset quality through expert annotation and an LLM-as-judge protocol. Using the benchmark, we evaluate 15 retrieval systems and six open-weight LLMs for RAG. We find that models perform similarly on document-level retrieval, while at the harder chunk level retrieval quality is not driven by embedding-model scale: bge-m3, a mid-size open encoder matches an order of magnitude larger ones and outperforms commercial embedding APIs. In contrast, answer generation remains challenging even given the gold evidence, particularly for inferential and risk-oriented questions.

arXiv:2610.12569v1 Announce Type: new Abstract: Static coding benchmarks are fixed at release. Over time their tasks leak into training data, and stronger agents eventually saturate them. We argue that a benchmark should instead aim to be self-scaling: its tasks and oracles are created at evaluation time, so no content exists in advance to leak, and its difficulty is set by the participating agents rather than fixed at release. We introduce SWE-Duel, an arena in which two coding agents evaluate each other over mirrored Red-Blue turns. In each turn, the Red agent works on a real repository across three context-isolated sessions. It proposes and implements a new feature with feature tests, then embeds a subtle bug in that feature along with sealed bug-triggering tests. The resulting pull request is admitted only if (1) it passes the feature tests and the repository's existing tests while failing exactly the bug tests, and (2) a fresh session of the same agent can repair the bug blind, which rules out unsolvable tasks. The Blue agent then reviews the pull request without being told a bug exists. It wins only if its revision passes all three test suites, and then the two agents swap roles. We run a tournament among 11 model-harness participants on 15 repositories in five languages for 1,500 turns. The leaders win in different ways. GPT-5.5 with Codex CLI ranks first as the strongest solver, repairing 98.3% of the defects it reviews. Claude Opus 4.8 with mini-swe-agent ranks second as the strongest task author, with a 27.3% Red win rate. Finally, within one model family, tasks written by the newest version defeat a fixed-reference solver far more often than those of older versions (42.9% vs. 4.0%), giving preliminary evidence that task difficulty scales with task authoring capability.

arXiv:2610.12582v1 Announce Type: new Abstract: A latent recurrent language model applies a shared block of layers repeatedly, so one checkpoint can trade inference compute for accuracy by changing the loop count. At one loop, however, such models typically fall well short of a non-recurrent model of the same size, a one-loop tax that makes a separate non-recurrent model the better choice whenever latency matters. We present Elastic Latent Recurrence (ELR), a recipe for training looped language models without this tax. At one loop, an ELR model costs the same to serve as a non-recurrent model of the same size and, trained on the same data, reaches similar accuracy. Each additional loop then buys more accuracy for more inference compute. Three choices make this work: a gated state update with a learned step size, loop counts mixed within each optimizer update, and conditioning each pass on its loop index and budget. The first two preserve one-loop accuracy, and the third lowers loss at every additional loop. Mixed depths can be introduced late in pretraining to limit the added training cost, and with zero-initialized conditioning added at the switch, this late schedule reaches lower loss than mixed-depth training from scratch. In MoE models, reusing first-loop expert routes on later loops cuts mixed-depth update time by 11.7% with no measurable change in accuracy. At inference, keeping only the last loop's attention KV cache holds attention KV memory at its one-loop level while retaining most of the gain from additional loops. On 4B dense and 30B-A3B MoE hybrid models, ELR at one loop is 0.52 points below and 1.04 points above its matched baseline on a ten-task average, and four loops add 7.25 and 4.75 points. Carried through post-training, the recipe yields Koi-4B, a single checkpoint that scores an estimated 18 on the Artificial Analysis Intelligence Index at one loop, matching Qwen3.5-4B, and 29 at four loops.