The $72.5m suit alleges Scott Jennings falsely told listeners Piker had a ‘plan of returning our country to lynchings’
The prominent socialist political commentator and influencer Hasan Piker is suing his conservative counterpart Scott Jennings for $72.5m over comments on a podcast that he said portrayed him as “a violent racist” who advocated the lynching of Black people.
Piker’s defamation lawsuit, filed on Tuesday in US district court for the District of Columbia, states that Jennings misrepresented his comments in a podcast and subsequent social media post in August.
L’Orchestre du Nouveau Monde’s young musicians went viral for their defiant performances during the last week’s protests in France. They belong to a fine tradition of radical orchestras. Also, today, what music should soundtrack the Traitors?
They were the last people standing – or rather playing – during last week’s student protests in France, as gas canisters landed in double-bass cases. L’Orchestre du Nouveau Monde, co-founded by young conductor Étienne Jarrier and his friends when they were still teenagers, aren’t a typical orchestra. With their manifesto of activism and social justice, the orchestra is not “a sacred object of beauty, but a tool for struggle”, through civil disobedience, outreach in schools, demonstrations and concerts, all to “present a different kind of classical music – one reconnected to reality”.
The new world Jarrier and his group dream of is “anti-racist, anti-capitalist, feminist, environmentalist and committed to de-growth”, and their way of working is deliberately provocative, not only to so many of the dominant directions of travel in the political world, but also to the institutions of classical music. They play scrappily passionate performances of La Marseillaise, of music from The Avengers, of Dvořák’s New World Symphony and protest songs but the ONM judge the quality of their performances not by playing the right notes in the right order, but by their consciousness-raising impact, whether they’re staging performances at Place de la République in Paris or at the foot of a glacier.
Attorneys for two former FBI agents charged with lying to Congress about past probes of Hunter Biden are back in court seeking to have the case dismissed.
Anxious Israelis watched election results near completion Wednesday after the highest voter turnout in over a quarter-century, as no clear winner emerged and weeks of negotiating could lie ahead for Prime Minister Benjamin Netanyahu and an array of opposition parties trying to unseat him.
In April of last year, dozens of students set off from Serbia, intent on making it to Strasbourg, France on bicycles. The aim of the arduous, 13-day transcontinental journey was simple: to make their voices heard in Europe and beyond.
In their home city of Novi Sad, students had been holding near-daily protests after a deadly roof collapse at the city’s main railway station. As their push for accountability and justice spread across the country, it was met with baton-wielding police, teargas and a political class that branded these young people as traitors.
This post describes the next generation of Meta Andromeda, Meta’s personalized ads retrieval engine, and how co-designing the model, system, and GPU hardware unlocks a new efficiency frontier for ads retrieval.
In production, the system scores every eligible ad with near ranking-level model performance at retrieval scale and surfaces the top candidates at over one billion ads per second, a new throughput benchmark for ads retrieval-stage models (i.e., a 100x lift compared to the 2024 Meta Andromeda system).
For advertisers, this means better performance automatically from the retrieval system matching ads across a larger candidate pool, within the same latency budget. Since 2025, Meta Andromeda has delivered a 2.4% increase in conversions on Facebook.
Meta Andromeda is Meta’s machine learning system for personalized ads retrieval, the first stage of the ads recommendation pipeline. Retrieval narrows a vast pool of eligible ads down to a small set of strong candidates in milliseconds; a later, more computationally expensive ranking stage then scores those survivors in fine detail.
Historically, the deep modeling quality of the ranking stage has been too costly to apply at retrieval’s scale. Building on the retrieval foundation established with Andromeda’s 2024 launch on the NVIDIA GH200 Grace Hopper Superchip, this post shares how co-designing the model, system, and hardware together brings near ranking-level performance to the retrieval scale.
Challenges and Opportunities in Scaling Retrieval
The 2024 Andromeda deployment reimplemented retrieval with an end-to-end optimized retrieval implementation from custom AI hardware with hierarchical indexing and GPU-native operators to a purpose-built ML model, growing model capacity by four orders of magnitude. Since then, generative AI creative tools have helped advertisers rapidly create hundreds of millions of new ads, growing Meta’s ads candidate pool to an unprecedented scale.
Scoring such a large ad candidate pool with a more accurate model at a much higher system throughput requires fundamental technology breakthroughs in model quality, system architecture, and hardware utilization jointly.
Higher Model Performance Requires a New Architecture
Ads retrieval has an outsized ROI on incremental quality: Small improvements in retrieval model performance can drive large gains in advertiser value. The rapid expansion of our ads candidate pool makes optimizing for hardware capacity ROI an increasingly critical factor. The 2024 Meta Andromeda retrieval models have established a highly effective foundation with their GPU-optimized deep neural network architecture. To keep improving retrieval performance at a sustainable compute cost at our current scale, however, we must move beyond that paradigm. This requires a fundamental architectural evolution to push model performance boundaries significantly without a proportional increase in costs.
New System Bottlenecks at Billion Scale
The 2024 Meta Andromeda system evaluates at tens of millions of ads per second per GPU card. Meeting the new business demand requires a throughput at multi-billion ads evaluated per second per card in the system. At billions of ads per second throughput, the memory bandwidth of passing ad tensors through memory hierarchy becomes the new bottleneck. Intermediate tensors between GPU kernels account for terabytes per second of memory traffic, and shipping ads between local CPU hosts and remote hardware accelerators evaluating models saturates the network link. This requires a new system co-design that unifies these fragmented, distributed workloads into a single GPU pipeline.
Pushing Towards the Hardware Limits
With retrieval running on GPU, the important optimization question is “How close are we to the hardware ceiling, and how much room remains for future model growth?” Answering this requires pushing each component of the retrieval pipeline to the limits of the current hardware generation, a challenge that motivated a joint collaboration with NVIDIA.
Each of these challenges drove a corresponding advance.
Unlocking Retrieval Performance Through Model-System Co-Design
Next-Gen Retrieval Model Architecture — Retrieval Mixture-of-Experts
The latest iteration of Meta Andromeda introduces three synergistic architectural innovations that bring large language model (LLM)-inspired mixture-of-experts (MoE) into the retrieval (see Figure 1). With this highly expressive yet inference-efficient architecture, a new retrieval performance-efficiency Pareto frontier has been established: model performance approaches ranking-level in a typical retrieval-then-ranking system, at approximately 300x fewer inference-time FLOPs than the 2024 Meta Andromeda retrieval model.
On the user modeling side, we transitioned from a single user tower to an ensemble of specialized user towers as experts. This captures diverse, time-evolving, and contextualized user interests by integrating a vast array of high-fidelity signals: sequential ad engagement events, user-cohort representations, LLM-derived semantic features, and cross-surface co-engagement signals.
In the candidate ads tower, we employ a joint encoding strategy that incorporates real-time serving context, ensuring that downstream interaction models receive context-aware representations.
In the interaction overarch layer, the interaction model evolved from dense neural networks into an MoE attention-based gating network that dynamically routes candidates across the most relevant user experts, delivering near-ranking-level performance at a fraction of the FLOPs.
A defining property of the new architecture is that its performance is co-designed with the underlying hardware execution model. Model inference throughput has reached over 10 billion ads per second on a single GH200 GPU, and this efficiency allows model capacity to grow further, through richer feature interactions, additional expert sub-networks, and deeper architectures, without a proportional increase in GPU cost. Each incremental model advancement translates into greater advertiser value.
Figure 1: Retrieval MoE architecture. Diverse high-fidelity user signals feed an ensemble of heterogeneous user experts. Ad content fused with real-time serving context forms a contextualized ad input that drives an ad-aware MoE router, activating only the top-K most relevant experts to produce the user–ad relevance score. The result is ranking-level matching quality at retrieval scale — ~300x fewer inference-time FLOPs than the 2024 Meta Andromeda and over 10 billion ads/sec on a single Grace Hopper Superchip.
Co-Designing the Model and System Onto a Single GPU Pipeline
The 2026 Meta Andromeda pushes the model-system co-design to the next level with two innovations — a single GPU kernel fusing both model and system workloads per user-ad interaction and a strategy to co-locate per user request logics with per user-ad GPU kernels in a single GPU box. To this end, a highly optimized and fully fused GPU pipeline is built to host the entire end-to-end Meta Andromeda workflow.
Fused Kernel
At a high level, Meta Andromeda’s inference time per user-ad workload is a compact sequence of operations across model execution and system logics — gathering embeddings, executing user-ad interaction models, and scoring. The ~300x FLOPs reduction in the retrieval MoE gating architecture creates the necessary headroom and is less resistant to fully fuse both model execution and non-model system workloads into a single GPU kernel, compared to the 2024 launch where only operations within the model execution were fused. Consequently, the fused kernel now runs at over 10 billion ads per second on a single GH200.
This matters because retrieval workloads are fundamentally memory-bandwidth bound. At billions of ads per second, even a small intermediate tensor per ad multiplies into terabytes per second of high bandwidth memory (HBM) traffic. The memory workload of these intermediate tensors, not arithmetic, becomes the dominant bottleneck in multi-kernel retrieval pipelines.
The fused kernel eliminates this bottleneck entirely. Rather than dispatching multiple small GPU kernels for model and system operations, it executes the entire per user-ad computation in one pass. Intermediate results stay in GPU registers and thread block shared memory and are never materialized to HBM, achieving high utilization of the GPU’s peak memory bandwidth rather than wasting it on redundant intermediate traffic.
Co-Location Strategy
Measured in isolation, the per user-ad fused kernel runs at multi-billion ads per second on a single GPU. However, per-user request operations such as the index matching (i.e., enforcing advertisers’ targeting rules) and top-K selection run on separate local CPU hosts and communicate with the fused GPU kernel scoring service on remote accelerators through the network connection. At multi-billion ads per second throughput, this traffic alone saturates the network link, making the network bandwidth become the bottleneck rather than GPU compute.
The co-design response was to collapse the system architecture to match the model architecture: co-locate index matching, scoring, and top-K selection on the same GPU. Instead of shipping large candidate lists across hosts, the system sends a compact request payload — user features and a targeting query — to a single GPU server, which runs the entire retrieval pipeline locally and returns only the final top-K results. This reduces network traffic by ~100x and shifts the system back to a GPU-compute-bound regime.
Moving index matching onto the GPU is not straightforward. Ads index matching evaluates complex advertiser-defined targeting logic across tens of feature categories, and must support real-time index updates as ads are created, modified, or expired throughout the day. Supporting this on GPU required redesigning the index structure for GPU-parallel execution while preserving the full fidelity of production targeting semantics.
A Single GPU Pipeline for Meta Andromeda
Combining the aforementioned two innovations, the fully co-located single GPU pipeline processes over one billion ads per second on a single GPU server while scoring every eligible ad in the candidate pool. This is a new throughput benchmark for ads retrieval-stage models, and as the next section describes, significant headroom remains.
Figure 2: System architecture of the single fused GPU pipeline for Meta Andromeda. This pipeline co-locates index matching, fused scoring, and top-K selection on a single GPU server. The fused scoring kernel executes the entire per-user-ad computation — from embedding gathering to MoE interaction and scoring — in a single on-chip pass, keeping intermediate results in registers and shared memory to eliminate HBM bottlenecks. This co-design reduces network traffic by approximately 100x by sending only compact request payloads — surpassing 1 billion ads per second while scoring every eligible ad across Facebook and Instagram.
GPU-Accelerated Retrieval: The Meta-NVIDIA Collaboration
With the production pipeline running at over one billion ads per second, a natural question follows: How close is the system to the ceiling of the current generation hardware, and how much headroom remains for future model growth? In early 2026, Meta and NVIDIA launched a joint collaboration to answer this question — systematically optimizing each component of the retrieval pipeline to push throughput toward hardware limits. NVIDIA contributed optimized GPU kernels, and Meta integrated and benchmarked the full pipeline. Each component achieved 2x to 5x speedup over the baseline.
The optimized gating kernel improved by 5.3x, from 2.4 billion to over 10 billion ads per second, reaching ~85% of the GPU’s peak HBM bandwidth. At this throughput, the execution time / latency of the gating mechanism is just as fast as what a simple dot-product scoring kernel achieves on the same hardware — while the model is substantially more expressive. This is the co-design validation: the additional model complexity is essentially free on Tensor Core hardware. INT8 quantization achieved a further up to 2x throughput increase over the optimized FP16 kernels, with accuracy on par with 16-bit precision.
Index Matching
Through targeted GPU kernel optimizations, index matching throughput improved over 2x, reaching over 100 billion ads per second — an order of magnitude higher than the scoring network. This headroom means index matching can be re-run across multiple retrieval pathways without becoming a bottleneck.
End-to-End GPU Pipeline
Beyond individual kernels, the collaboration profiled the full retrieval flow and identified that data movement between pipeline components — not computation — accounted for a significant fraction of end-to-end latency. By co-designing data layouts across index matching and scoring to eliminate unnecessary indirection, the team removed this overhead. With all optimizations applied, the full pipeline benchmarks at multiple times the current production throughput on a single GPU server, with further gains expected as additional optimizations integrate into production.
Figure 3: Optimized ads/sec for interaction network, index matching, and E2E pipeline.
Continuing Work on Next-Generation NVIDIA Platforms
Building on the Hopper results above, we are extending our optimization and co-design work to Grace Blackwell. Here, we are focusing on platform design optimizations that enable different CPU-to-GPU compute ratios, which our work so far suggests is an important factor for Andromeda-like workloads. We are evaluating how sensitive end-to-end application performance is to these ratios, using Grace-class CPU performance and Blackwell-class GPU performance for their respective parts of the workload.
More broadly, we are exploring how these findings extend to the next platform generation, where greater flexibility in CPU+GPU execution, resource ratios, and coupling could bring more of these architectural improvements to bear on Andromeda-like workloads.
These results confirm that the co-designed retrieval pipeline is not yet at its ceiling. With quantization unlocking further scoring throughput and index matching operating at an order of magnitude above the scoring network, the architecture is positioned to absorb significantly more model complexity — deeper interaction networks, richer feature spaces, and additional retrieval pathways — without proportional infrastructure growth.
The Path Forward for Ads Retrieval
The advances described here follow a single principle: The largest gains come from designing model, system, and hardware as one integrated stack, rather than optimizing each in isolation. The retrieval MoE gating architecture delivers near-ranking-level expressiveness at no additional serving cost. Fusing the entire retrieval pipeline onto a single GPU eliminates network and memory bottlenecks. And pushing the pipeline towards its hardware ceiling confirms substantial headroom remains.
Looking ahead, the co-design principle that drove these results applies equally to next-generation model architectures and hardware platforms, with the demonstrated headroom enabling continued scaling without proportional infrastructure growth.
For people who use Meta’s products, this translates directly into better ads that are more relevant to them, delivered within the same latency budget. For advertisers, it means better customer matches and more conversions. For Meta, it means increased business from advertisers and lower costs, as each GPU now evaluates more ads per second. As the volume of eligible ads continues to grow, driven by Advantage+ automation and generative AI creative tools, Meta Andromeda’s co-designed architecture scales to meet it.
Acknowledgements
We would like to thank Gosh Arzumanyan, Habiya Beg, Andrew Bocking, Zain Brohi, Rohit Chandel, Claire Chen, Jason Chen, Wenlin Chen, Zhenyu Chen, Yuze Dai, Zhou Fang, Junjie Feng, Stephen Fitz, Chunli Fu, Yicong Fu, Xiaorui Gan, Sohan Gantala, Yingsheng Gao, Golnaz Ghasemiesfeh, Ishu Goyal, Wentian Guo, George Han, Yihan He, Jun Hu, Xuan Hu, Kun Huang, Qin Huang, Tianchen Ji, Crystal Jin, Tao Ju, Rajat Karnwal, Santanu Kolay, Michael Kwan, Ilia Lewis, Baiyu Li, Chao Li, Jim Li, Li Li, Lin Li, Liyuan Li, Mingda Li, Ted Li, Xiaotian Li, Yanyu Liang, Yue Liang, Irene Liu, Gaoxiang Liu, Jason Liu, Siqi Liu, Wei Liu, Xiao Liu, Yuheng Long, Yining Lu, Siyi Luo, Anuj Madan, Swaroop Manjunath, Phalguna Mankar, Jiachen Mao, Yuji Mo, GP Musumeci, Guruswamy Namasivayam, Kexin Nie, Sandeep Pandey, Uladzimir Pashkevich, Clementine Plati, Heng Qiao, Hang Qu, Lei Qu, Anand Rajan, Jatin Sethi, Puneet Sharma, Ruchit Sharma, Yan Shi, Chang Shu, Sean Song, Alec Stashevsky, Matt Steiner, Zheng Sun, Nikhil Vinayak Supekar, Jim Tao, Chintan Thakkar, Tian Tong, Aleksandr Ulanov, Daksh Vasistha, Simha Venkataramaiah, Arunprasad Venkatraman, Abhishek Verma, Ruud Vinke, Vidhoon Viswanathan, Sasha Vorontsov, Eric Walden, Dong Wang, Feng Wang, Haole Wang, Jeff Wang, Senhong Wang, Ye Wang, Di Wu, Haocheng Wu, Shawn Wu, Jimmy Xiang, Nan Xiao, Ruichao Xiao, Chao Xie, Yudi Xie, Fangzhou Xu, Nathan Yan, Tak Yan, Hang Yang, Song Yang, Yuhan Yang, Yavuz Yetim, Bowen Yu, Li Yu, Qifan Yuan, Xiaodong Yuan, Lin Zeng, Siqi (Cindy) Zeng, Caijie Zhang, Joanna Zhang, Min Zhang, Qing Zhang, Zhengkai Zhang, Fangyu Zou, and the entire v-team behind the success of Meta Andromeda.
Constellation Energy will get $5 billion in government loans to modernize nuclear reactors. Most of the new capacity is already committed to Google, Amazon, and Meta.
In April of last year, dozens of students set off from Serbia, intent on making it to Strasbourg, France on bicycles. The aim of the arduous, 13-day transcontinental journey was simple: to make their voices heard in Europe and beyond.
In their home city of Novi Sad, students had been holding near-daily protests after a deadly roof collapse at the city’s main railway station. As their push for accountability and justice spread across the country, it was met with baton-wielding police, teargas and a political class that branded these young people as traitors.
Huerta co-founded the United Farm Workers, improving the lives of the country's mostly Latino farmworkers and changing the course of the labor movement.
Автор «Чистого кода» Роберт Мартин aka Дядя Боб взялся за ИИ-агентов. В скилле ub-mode, который Дядя Боб на днях выложил на GitHub, всего около 750 слов. Этими словами Мартин выстраивает работу агента так, будто сам стоит рядом и смотрит на монитор строгим взглядом.
На днях Роберт Мартин aka Дядя Боб, автор «Чистого кода», выложил на GitHub ub-mode — набор правил для ИИ-агентов. Весь скилл — один файл примерно на 750 слов: короткие ёмкие правила, написанные по тем же принципам «чистого кода». Теперь у каждого разработчика будет свой Дядя Боб: он встанет над ИИ-агентом с линейкой и проследит за гигиеной кода.
При разборе багов агент приносит правку и пишет «готово, шеф». По одному диффу не понять, так ли это: старые тесты могли остаться зелёными и при неверной логике, а новая проверка — просто спрятать сбой. Скилл ub-mode заставляет агента обосновать свою работу: сначала показать тест, который падает из-за бага, потом предъявить правку, после которой тест проходит.
В репозитории ub-mode нет кода: там только README.md и SKILL.md. Второй файл описывает порядок действий, а README отсылает к утилитам, которые меряют сложность кода, ищут дубли и проверяют, ловят ли тесты ошибки. Автор пишет, что частично использовал идеи poteto-mode / pstack.
Правила разбиты на три части: First Make it Work, Then Make it Right и Practice.
В ночь на вторник начал действовать режим прекращения огня в Алешках в оккупированной части Херсонской области, ситуацию в которой и украинская, и российские стороны оценивают как граничащую с гуманитарной катастрофой. Чем ситуация, на которую обратила внимание супруга президента США, уникальна с военной точки зрения?
The luxury MV Odyssey ship slated for a years-long voyage around the world suspends operations amid high fuel prices
A luxury cruise ship that had embarked on a years-long voyage around the world was forced to suspend its operations and leave passengers stranded in Cambodia because of high fuel prices.
The ship, known as the MV Odyssey, had been carrying 203 passengers, mostly Americans, along with 156 crew members, Cambodian officials told the Associated Press. The ship has a 650-passenger capacity.
Cornell University faculty will meet Wednesday to conclude debate and begin voting on a proposed no-confidence resolution in the university's administration following widespread frustration over its response to a woman who reported she was gang raped at a campus fraternity house in 2024.
Convex – это новый, но набирающий популярность фреймворк для реактивных веб-приложений, который позволяет писать транзакционные квери (queries) и мутации (mutations) на TypeScript с гарантиями консистентности.
Вся статья ниже относится к self-hosted версии Convex: convex-backend, а точнее, к моей пропатченной версии оного, так как “официальная” версия не пригодна к работе в продакшене и вряд ли тестировалась на сколь-нибудь нагруженном кейсе. Моя версия работает на виртуальном сервере с 32 ядрами и 64Гб памяти и обрабатывает до 15 тысяч Convex-кверей и мутаций в минуту.
TLDR: квери и мутации в среднем работают в 2 раза быстрее, хвосты задержек сократились втрое. Объём памяти, который сервер удерживает под готовые к выполнению окружения (V8-контексты или Wasm-окружения), уменьшился на 80%.
Democrats urge FTC to tackle ‘systemic failure’ of biggest convenience and dollar stores over pricing practices after Guardian investigations
Thirty-four members of Congress are calling on the Federal Trade Commission to investigate reports of widespread overcharging by some of America’s most-recognized brand-name retailers.
The 28 October letter to the FTC comes in direct response to Guardian investigations of four of the nation’s largest dollar-store and convenience-store chains. Both investigations revealed pervasive mismatches between shelf labels and register prices.
Jury at Winchester crown court sees video evidence by Joanne Young in trial of four men accused of sexual assault
A woman who was allegedly sexually assaulted by men invited by her ex-husband into their family home described how she would wake in the middle of the night restrained and blindfolded and feeling “hazy”.
Joanne Young said that on one occasion she managed to remove her blindfold and was shocked to glimpse a stranger in her bedroom whom her then husband, Philip Young, had allegedly arranged to have sex with her.
Недавно столкнулся с весьма неприятным багом в движке The Great Tribes.
После обновления драйвера AMD игра начала намертво зависать при перемещении камеры, особенно когда на карте открывалась большая территория. Иногда зависал графический драйвер, иногда JVM падала внутри atio6axx.dll.
Самое интересное — на старом драйвере и других видеокартах всё работало нормально!
Начались поиски. Под подозрение попало практически всё: многопоточность OpenGL, VAO, VBO, индексные буферы, тени, модели юнитов и даже glfwSwapBuffers().
Постепенно удалось выяснить, что зависание вызывает отрисовка поверхности карты. А затем внимание переключилось на фрагментный шейдер.
As the landscape of higher education has evolved from the eras of segregation to the civil rights movement to Trump-era racial politics, Black students have had to adapt on college campuses nationwide as government mandates, school policies, and the dismantling of affirmative action have impacted their education over the years.
Hosted by Kahlil Greene, this video is part of our new series, Kahlil Explains, that unites Vox’s unique brand of explainer journalism with Greene’s expertise in how politics and culture shape society. This first episode traces the history of HBCUs and the experience of Black college students to unpack why Black student enrollment is growing at HBCUs yet declining at other elite institutions.
Read more about HBCUs and race in higher education:
Привет, Хабр! На связи Анна Куделина — заместитель технического директора РТЛабс по аналитике.
Пять лет назад я написала статью, как подготовиться к собеседованию на позицию системного аналитика. Она пользовалась большим спросом: 130 тысяч читателей и 263 сохранения, её публиковали в профессиональных сообществах. Мне потом написали несколько человек с отзывом о том, что статья помогла им стать аналитиком или перейти на новую ступень. Я сделала вывод, что формат передачи своего накопленного опыта по росту в карьере полезен, и решила сделать статью-продолжение.
За 6 лет в РТЛабс я прошла вертикальный путь от стартовой должности системного аналитика до руководителя аналитической функции всей компании. На этом пути менялись задачи, ответственность и способ думать. На своём опыте я теперь знаю, где именно спрятаны подводные камни, из-за которых люди застревают.
Если в первой статье речь шла о том, как аналитиком стать, то здесь — о том, что делать дальше.
For months, Mary Peltola’s campaign for Alaska’s Senate seat has denied any connection to a separate campaign launched by a man who, coincidentally, shares the exact same name as Peltola’s Republican opponent, Dan Sullivan.
New reporting from the New York Times undercuts that claim. The outlet reported, based on interviews, obtained text messages and legal filings, that the Peltola campaign had not only recruited Mr Sullivan (not the Republican), but had sought out a number of Dan Sullivans across the state for potential candidacy, as well as others who could peel votes away from Peltola’s main opponent.
The family of a mother who was murdered by her abusive ex-partner are demanding "action, answers and accountability" amid their belief her death "should never have happened".
The Iran war looms over the midterm elections as costs mount and gas prices squeeze voters' wallets. CBS News' Weijia Jiang has more on this, plus Lana Zak reports on races in Iowa.
Как я боролся с перегревом MacBook Pro 2017 на Intel: от замены термопасты и настроек Chrome, до своей проги управления Turbo Boost и винтами охлаждения.
Если кто сталкивался с перегревом старого Макбука про на интеле при просмотре видео - будет полезно почитать статью, с конкретными шагами и настройками для ее решения. Разбирался не сам, все делал Antigravity с Gemini 3.8 flash по моему запросу"разберись с перегревом" и описанием симптомов, он даже настройки хрома сам умеет менять если попросить. Проблема в итоге решена, и помимо технических проблем (термопаста, старый акум) считаю основными именно настройки хрома, тк после них разница стала заметна. Статья написана Gemini 4 argon, картинка и скрины от него же.
A whopping 83 league places separated Fleetwood from Arsenal, yet the hosts booked their place in the Carabao Cup quarter-final thanks to their penalty shootout win over the Gunners.
Изо всех американских утюгов до нас доносятся возгласы о том, что необходимо срочно бросить всё, бежать строить свои продукты с использованием искусственного интеллекта и продавать их, а те, кто этого не сделает, навечно попадёт в «permanent underclass». «Скоро новый феодализм!», «Работ больше не будет!», «Выживут только те, кто будет строить своё» — кричат нам голоса со стороны пар... пар... пр.. простите, Твиттера. Очевидно, для чего панику нагоняют продавцы LLM токенов — страх хорошо продаёт. А вот поддерживающие их впавшие в ИИ‑психоз громкоголосые «технобро» просто не понимают того, как вообще устроена экономика, и на это свойственное технарям неведение относительно других областей человеческой деятельности накладывается ещё и свойственная американцам принципиальная неспособность мыслить вне рамок капитализма.
В этом посте я хочу обсудить то, какое реальное влияние ИИ оказывает на экономику, почему ИИ принципиально несовместим с капиталистической системой экономических отношений, и как распространение ИИ‑инструментов должно привести нас к слому текущего социально‑экономического строя, и в чём состоит принципиальная логическая несостоятельность криков о "permanent underclass". Но начнём мы немного издалека.
With nearly all votes counted, Netanyahu’s coalition government has lost its majority, but main opponents also fail to gain outright majority; marked gains by Palestinian parties
With almost all votes counted, Benjamin Netanyahu’s coalition government has lost its majority, but his main opponents, an alliance of Jewish Israeli parties unofficially led by Gadi Eisenkot, have also failed to gain an outright majority.
Israeli media report that Netanyahu’s bloc is ahead, with projections he could take 53 seats under the current vote share, but this could change further. The projections put Gadi Eisenkot’s opposition bloc on 48 seats and the Palestinian parties on 19 seats. Sixty-one seats are needed to form a majority to govern in the Knesset.
The remainder of votes to be counted come from absentee ballots, according to Israeli media. We can expect absentee ballots to be counted on Wednesday night, with final results by Thursday.
The Palestinian Israeli parties have received a historically strong vote. The Joint List, a new ticket of several Arab-majority parties, triumphed with at least 12 seats, while Ra’am, the other Islamic party, won seven.
The key question is whether the Eisenkot-led “Change” bloc (consisting of the rightwinger Naftali Bennett’s Together party, the centre-left Yair Golan’s Democrats and the rightwing Yisrael Beiteinu party) will partner with Palestinian Israeli parties or if they will resort to a new election, or even working with Netanyahu.
Netanyahu took the stage at a post-election rally in Jerusalem just after 3am, telling his supporters that official results will show that “Likud is the largest party in Israel”. He told his supporters that the country must “establish a strong, rightwing national government” – adding that a government that relied on the support of Palestinian parties must not be formed.
Eisenkot told cheering supporters in Tel Aviv that the “Israeli public has voted against Netanyahu’s service as Israel’s prime minister … a worthy leader would have accepted the people’s non-confidence and announced today his retirement from politics”. In his speech, he also warned Netanyahu not to “use Israel’s security” for political ends, playing on fears that the prime minister could use this period of uncertainty to launch a new attack on Iran.
The Palestinian Israeli vote, which represents about 20% of the population, had long been seen as a decisive factor for who will be able to form government, and the high turnout of this election was driven by the community experiencing increased racism and discrimination under Netanyahu’s government.
This is the most voted-for election in Israel’s political history, with many Israelis flying back into the country to cast their vote, and nearly 80% of IDF soldiers voting.
Donald Trump said his endorsement had helped Netanyahu perform better than polling estimates.
В ночь на вторник начал действовать режим прекращения огня в Алешках в оккупированной части Херсонской области, ситуацию в которой и украинская, и российские стороны оценивают как граничащую с гуманитарной катастрофой. Чем ситуация, на которую обратила внимание супруга президента США, уникальна с военной точки зрения?
High healthcare costs top the list of financial worries for American voters, a recent KFF poll shows. CBS News medical correspondent Dr. Céline Gounder shares her thoughts.
New York Yankees starting pitcher Cam Schlittler sits down with "CBS Mornings" to talk about his career so far, and what it means to win a 2026 MLB Players Choice Award.
—Photo-Illustration by TIME (Source Image: Deagreez/Getty Images)
Have you ever had a disturbing thought pop into your head out of nowhere—something you'd never consider actually doing?
Intrusive thoughts are unwanted thoughts, images, or urges that can feel upsetting, confusing, or wildly out of character. They're more common in people with obsessive-compulsive disorder (OCD) but can happen to anyone, and they can involve everything from violence and sex to religion, relationships, and the fear of having done something wrong.
Many people are able to brush off the occasional intrusive thought and move on quickly. People with OCD, on the other hand, can get stuck trying to figure out what the thought means—and whether they might act on it. “We stop and we treat that thought as super important and meaningful,” says Alissa Jerud, a clinical psychologist and clinical assistant professor of psychiatry at the University of Pennsylvania.
Intrusive thoughts can feel especially disturbing because they often involve someone or something the person cherishes. “We tend to have intrusive thoughts about the things we care most about,” Jerud says.
Some of Jerud’s patients have kept their intrusive thoughts to themselves for 20 years before telling her. They might be thinking: “Why am I having these thoughts? What does it say about me? Maybe I’m a monster,” she says.
A calm, compassionate response can make it easier to keep talking. “It can help them feel like there’s hope,” Jerud says. “Maybe I don’t have to beat myself up for having these thoughts in the first place.”
Here’s what to say if a loved one tells you they’re having intrusive thoughts.
Start with, “I’m so glad you shared this with me”
Start by thanking your friend for trusting you. Skip responses like “How could you think that?” or “Don’t say that,” which can reinforce the shame they’re already feeling.
It’s helpful, too, to make it clear that what they told you won’t affect your relationship: “This doesn’t change how I think about you,” suggests Madeline Thompson Smith, an Atlanta-based licensed professional counselor who specializes in OCD. Then acknowledge how upsetting the thoughts sound: “That sounds really scary to have stuck in your head.”
Pay attention to how you sound and look, too. A warm invitation to explain more can feel like judgment if you’re staring intensely or using a sharp tone. If you look startled, acknowledge it instead of hoping they didn’t notice. Kimberly Leonte, a psychologist in Boston who specializes in OCD, suggests wording it like this: “I guess I wasn’t expecting that, but I want to hear more.”
Offer support without repeatedly reassuring them
When someone you love is distressed by an intrusive thought, reassuring them can feel like the kindest thing to do. But if they keep coming back with the same question, answering it again and again can become part of the OCD cycle.
Say your partner keeps asking whether an unwanted thought means they’re dangerous. You reassure them, they feel better, and you both get back to your evening. Until another question arrives: But what if you’re wrong? What if they didn’t explain the thought accurately enough? Next thing you know, they’re asking for reassurance again.
A single reassuring response isn’t necessarily a problem, Thompson Smith says. The concern is when seeking reassurance becomes a compulsion—a behavior someone feels driven to repeat to relieve their distress. Each answer briefly helps them feel better, but it leaves them needing another answer when the doubt returns.
Instead of debating what the thought means, acknowledge how they’re feeling and express confidence in their ability to get through the discomfort. Thompson Smith suggests: “I know this feels really uncomfortable right now. I’m not going to help you go down the rabbit hole, but I’m happy to sit with you while it’s hard.”
If reassurance has become a routine between you, talk about changing it together, ideally with guidance from their therapist. Abruptly refusing to answer can leave your loved one wondering what happened. “It’s tough if the person’s used to getting reassurance, and you just stop giving it and they don’t know why,” Leonte says. Instead, discuss how you’ll respond the next time they ask—and make sure you both understand why you’re trying something different.
Meanwhile, aim to avoid suggestions like “Just stop thinking about it” or “Replace it with a happy thought.” Besides being considerably easier said than done, trying to force an intrusive thought away can keep someone preoccupied with whether it’s gone yet.
Ask whether they want you to listen or help
Find out what your loved one wants from the conversation. They might be seeking support in taking a step toward getting mental-health help, or they may simply want you to know what they’ve been going through.
“Do you want me to just listen, or do you want help figuring out what to do next?” Thompson Smith suggests asking. If this is the first time they’ve said their intrusive thoughts out loud, they might not be ready to discuss next steps.
If they’re already in therapy, ask how you can support what they’re learning: “What would your therapist recommend trying right now?” Leonte suggests asking. If they’re interested in learning more about intrusive thoughts but feel overwhelmed, you could offer to read about them together.
Sometimes, they simply want someone to listen. Your loved one might be telling you because “they just don’t want to be alone in this, and you can help them hold it,” Thompson Smith says. Or they may need to know “you’re not gonna judge me or people aren’t gonna run away if I start trying to be open and get help.”
Know when intrusive thoughts call for professional help
If unwanted thoughts are causing significant distress or interfering with daily life, offer to help your loved one find a therapist. Keep it low-pressure: “If you ever want me to help you find a therapist, I’d be happy to do so,” Jerud suggests saying.
For OCD, look for a therapist trained in exposure and response prevention, or ERP, Jerud suggests. It helps people learn to tolerate intrusive thoughts without turning to rituals like repeated reassurance. The International OCD Foundation maintains a directory to help people find one.
There’s also an important distinction between an unwanted thought about harm and wanting or intending to act. Don’t assume that every disturbing thought is benign, Thompson Smith says. If you’re unsure what someone means, ask calmly and directly: “Is this a thought that scares you, or something you want to do?” If they describe wanting to act, ask about plans and help them get professional support right away.
Keep being the friend you’ve always been
After someone opens up about their intrusive thoughts, keep treating them like you always have. Send the ridiculous meme, ask whether they’re still coming to dinner, and keep including them in the things you usually do together.
People often fear that sharing their intrusive thoughts will change how others see them, Thompson Smith says. Pulling away—or asking about their mental health every time you talk—can leave them wishing they’d kept quiet.
You might get frustrated or struggle to find the right words, Leonte says. If you react in a way you regret, apologize and let them know you’re still there for them. “No one’s perfect,” she says. “The goal isn’t to be perfect, it’s just to be supportive.”
Check out the new features and capabilities now available in Stack Internal and Stack Internal Community. Try Stack Internal for free by visiting stackinternal.com.
Junie Local now delivers ~70 tokens/s on M5 Max for coding tasks, up from ~40 tokens/s, with bursts up to 107 tokens/s. Two changes made this possible: the Splash inference engine and a new DFlash2 draft model.
Back in August, when we first released Junie Local for Mac M5, we spent a lot of time optimizing Junie and the MLX-VLM inference engine to achieve the maximum possible (at that time) performance. But two months is apparently a very long period at the current pace of LLM development. During this time, new, exciting ways to speed up inference were released by the community. We looked at and tested all of them, then picked the most promising parts to include in Junie Local.
In September, Inco AI released its Splash engine, built around the Qwen models. Focusing on just a couple of models lets Splash do what a general-purpose engine can’t: every decode step verifies eight drafted tokens at once, and attention over the context costs an order of magnitude less per position, so decode speed barely drops as context grows. The prefill, though, was on par with our optimized MLX-VLM engine. Since we had already taught MLX-VLM to use int8 operations on M5 chips, we ported the same optimization to Splash. In this release, Splash replaces MLX-VLM as Junie Local’s inference engine.
But a good engine is only part of the equation. We also needed a model-drafter that would enable more efficient speculative decoding. When using a speculator, acceptance rate defines the decode speed. A low acceptance rate could make decoding even slower.
Out of the box, Splash uses the DFlash2 drafter model for Qwen 3.8-27B. But Qwen 3.8 has a slight issue – it tends to overthink a lot. So the whole win of the generation speed is killed by the enormous amount of thinking tokens that need to be generated.
To address this problem, we previously released the Qwen3.8-3.6-27B-blend, which is more capable than Qwen 3.6 yet uses only 30% more tokens. And in this release, we are also including the DFlash2 draft model for Qwen3.8-3.6-27B-blend.
DFlash-based models are fast but very sensitive to in-domain distribution: acceptance rate on familiar tasks is higher, so the agent will work faster on familiar tasks and slower on unfamiliar ones.
To train this DFlash2 model, we generated a diverse set of agentic coding trajectories by running the Junie Local harness inside our internal coding environment with Qwen3.8-3.6-27B-blend. We then took the Qwen 3.8 DFlash2 draft model as a starting checkpoint and fine-tuned it with SpecForge on those coding trajectories. This way, our DFlash2 drafter delivers the highest speed and acceptance rate when you use it to code with Junie Local.
What surprised us was that you don’t need many trajectories to fine-tune a drafter. It took us just ~1 million generated tokens (with ~71M tokens read) to match Inco’s acceptance rate! As a side bonus, we also created a DFlash2 drafter for Qwen 3.6, trained in a similar fashion
And this acceptance rate, together with the Splash engine, pays off: compared to the previous MTP + MLX-VLM approach, we have ~ 2x improvement.
Combining all these improvements together, in this release we:
improved the generation speed from ~40 t/s to ~70 t/s
got850 t/s → 1200 t/s prefill improvement
reduced memory consumption: on 100k context, we need ~ 24 GB of RAM
And starting with this release, you can use it on M4 Macs and newer with 64GB of RAM.
Please note that on M4, prefill is approximately 4 times slower than on M5. It means that loading the system prompt and other context will be slow.
How to upgrade
Run junie update to upgrade Junie to the latest version, then launch the CLI. Once you’re inside Junie, run /local and press ctrl + r to update the inference engine.
If you don’t have Junie yet, you can install Junie and the local model by running:
JPMorgan Chase will pay additional taxes of between £100m and £150m as a result of the corporation tax surcharge hike announced by John Healey in his budget on Wednesday - a bill which could renew doubts about the construction of its new London headquarters.
