RSS Feeds

Dolores Huerta, the celebrated 1960s labor union leader and civil rights activist, died on Wednesday in California at the age of 96, her family said in a statement.

Huerta was a co-founder with Cesar Chavez, another veteran unionist and political activist, of the United Farm Workers union (UFW), which was credited with improving pay and conditions for agricultural workers through a series of strikes, boycotts and protests in the 1960s and 1970s.

• Dolores Huerta, California labor leader and civil rights champion, dies aged 96

Are the Democrats really going to take Texas? If there is going to be a blue wave, as polling suggests, winning the Lone Star state is going to be essential. James Talarico, the Democratic candidate for US Senate, is currently leading against Ken Paxton, a scandal-ridden political survivor. Even the gubernatorial race in Texas, pitting Greg Abbot, the current governor, against Gina Hinojosa is within the margin of error. These elections, and ultimately control of the US Congress, are likely to be decided by one key demographic – Latinos. Hispanic men were key to Donald Trump’s 2024 win. But Carlos Eduardo Espina, a news commentator, and Stephanie Valencia, a pollster, tell Kai Wright there are signs that the economy, the Trump administration’s mass deportation campaign and Republican redistricting efforts are backfiring – and Latinos seem to be swinging back to Democrats this year. Espina and Valencia break down what we really mean by 'the Latino vote' and how those voters might just swing right back next election if Democrats aren’t careful.

TL;DR

Koog Agents Optimization Library is a research library developed by JetBrains Research for optimizing the prompts of Koog agents. With it, you write a first version of each instruction, mark the steps worth tuning, provide the library with a training set, and specify a way to score the agent’s output. The library then runs your agent, learns from what happens, and hands back better instructions as a separate JSON file you can read, review, and commit. No rewrite of the agent is needed.

The examples in this post use GEPA, a prompt optimizer that learns from your agent’s failures and successes. In this post, we walk you through examples running on the privacy benchmark PUPA. GEPA improved our agent’s score from 81.19 to 94.22.

👉 Koog Agent Optimization Library (GitHub) 

👉Docs

👉 Examples

An agent to optimize

Koog is JetBrains’ Kotlin framework for building agents. It treats an agent as a program: you declare control flow as a typed graph of nodes and edges, and each step can carry its own prompt, tools, and model.

We have a whole agent written in it: our port of PAPILLON, a privacy pipeline that GEPA uses as one of its benchmarks. It runs on PUPA, a dataset of user queries that carry personal details. The agent has to answer such a query with help from a big external model without handing that model any of those details, so it rewrites the question without the details, sends that version out, and writes the reply from what comes back.

fun pupaStrategy(untrustedModel: LLModel) = strategy("pupa-papillon") {
    val craftRedactedRequest by subgraphWithTask<PupaAgentInput, PupaRedactedRequest>(
        name = "craft_redacted_request",
    ) { input -> REDACT_INSTRUCTION + "\n\nuser_query:\n${input.userQuery}" }

    val callUntrustedModel by subgraph<String, String>(
        name = "untrusted_model",
        llmModel = untrustedModel,
    ) { /* one call to the big model */ }

    val respondToQuery by subgraphWithTask<PupaResponseInput, PupaFinalResponse>(
        name = "respond_to_query",
    ) { input -> RESPOND_INSTRUCTION + "\n\n" + input.rendered() }

    // …plus three small nodes carrying state
}

Only three steps are needed to do the work here: the first rewrites the question, the second calls the big model, and the third writes the reply. Everything else is plumbing. Two of those three steps hold an instruction that a human wrote, and they are the ones you would want an optimizer to take over.

Marking a step as optimizable

The library adds one builder. Swap subgraphWithTask for optimizableSubgraphWithTask on the steps you want tuned, and lift the instruction out of the string it was buried in:

fun pupaStrategy(untrustedModel: LLModel) = strategy("pupa-papillon") {
    val craftRedactedRequest by optimizableSubgraphWithTask<PupaAgentInput, PupaRedactedRequest>(
        name = "craft_redacted_request",
        optimizableInstruction = REDACT_INSTRUCTION,
    ) { _, input -> "user_query:\n${input.userQuery}" }
    
    val callUntrustedModel by subgraph<String, String>(
        name = "untrusted_model",
        llmModel = untrustedModel,
    ) { /* one call to the model */ }

    val respondToQuery by optimizableSubgraphWithTask<PupaResponseInput, PupaFinalResponse>(
        name = "respond_to_query",
        optimizableInstruction = RESPOND_INSTRUCTION,
    ) { _, input -> input.rendered() }

    // …and the same plumbing and wiring as before.
}

With the step marked as “optimizable”, the graph itself remains unchanged. The instruction just moved into a parameter of its own, where an optimizer can find it and replace it. When there is no artifact to load, the step runs on the text you wrote.

Running the optimizer

GEPA runs your agent over the training set, shows a second model what went wrong, and asks it for a better instruction. Then it tries that instruction and repeats it, round after round, within the budget you set. Across those rounds it keeps several candidate instructions alive at once, so it does not drop one that only wins on your hardest items.

The optimizer needs one thing from the agent’s author: a function that, given a finished run, returns a score and a short summary of how it went. The score decides which candidates survive, and the text is what the second model reads. Here, that text is one line of numbers:

PUPA overall score: 0.50. Quality score: 1.00. Privacy score: 0.00. PII leakage ratio: 1.00.

That line, plus what the step itself received and produced, is everything GEPA had to work from when it wrote the instruction in the next section.

The rest is just two calls:

val optimizer = GEPAOptimizer(
    gepaFeedback = pupaFeedback,
    reflectionModel = OpenAIModels.Chat.GPT4_1Mini,
    numRollouts = 2426, // a practical start is 6 to 20x your training set
    storagePath = ResilientPath("build/artifacts/pupa-gepa.json"),
)

optimizer.train(session) // the session wraps your agent, dataset, and metric
val optimized = optimizer.loadOptimizedAgent(agent)

train writes the artifact, and loadOptimizedAgent reads it back and hands you a copy of your agent running the new instructions. Every optimizer in the library has those same two methods, so switching between them leaves the surrounding code alone.

What GEPA wrote

The step that strips the personal details started with this instruction:

Given a private user query, create a privacy-preserving request for a powerful external LLM.

The LLM may assist without learning private information about the user.

Two lines, deliberately vague. It says nothing about how much redaction would be enough.

By the end of the run, that same step was running a 4,456-character instruction. Here are three pieces of it:

2. Redaction and Generalization
   - Replace all such information with broad, generic placeholders or neutral terms
     (e.g., "the minor," "the course," "the company," "a noted musician," "two large
     urban communities," "the agency," "the program," or "a fellowship honoring an
     individual").
[...]
Additional Domain-Specific Considerations:
- For academic or educational content, avoid exposing real professor names, specific
  course titles or codes, individual student names, or direct internal feedback. Use
  neutral terms like "the minor," "the course," "professors," "students," and "mentors."
[...]
Overall Strategy:
Create a balanced abstraction that ensures zero leakage of private or identifiable data
while preserving the user's original task in sufficient detail for accurate execution
by a LLM.

Every detail in the instruction above came out of the runs, not from us. You would probably not sit down and write the placeholder vocabulary, the rule about professor names and course codes, or the overall strategy at the end. That strategy states a trade-off we never mentioned: keep the leakage at zero, but leave enough detail to do the job.

Results

We wanted to know whether our GEPA in Koog does what it is supposed to, so we ran it on the benchmarks the project itself uses and checked whether the agents performed better.

Take PUPA, for example, the benchmark that the above agent runs on. Its baseline was already 81.19, with little room left above it, but optimization still improved it by 13 points to 94.22. Both numbers are the mean of five runs, and the gain is statistically significant.

The rest of the benchmarks, including how our numbers line up with the original implementation, can be found in the docs.

Three more optimizers

Besides GEPA, the library implements three more optimizers behind the same interface:

• BootstrapFewShot mines demonstrations out of your successful runs. It makes no LLM calls beyond running your own agent, making it the cheapest option to try first.
• MIPROv2 searches across instructions and demonstrations.
• ACE grows a playbook of insights that rides along with the system prompt, for when you want the agent to accumulate knowledge rather than a better-worded instruction.

Swapping between them is one constructor; the train and loadOptimizedAgent calls around it stay as they are. Writing your own means implementing those same two methods, and all four of ours fit behind them without an escape hatch.

Where to find it

The library, its documentation, and the full benchmark results are in the open-source repository. The optimizable-subgraph support it relies on currently lives in a fork of Koog; we expect it to land upstream.

October isn’t quite done yet! A few late arrivals made the cut, and they’re too good for us to hold back until November.

Inside: A free frontend event about AI, a free way to practice reviewing agent-written code, our first Coursera specialization, and three full-stack Python courses.

Events

JetBrains AI.js Day 2026: Frontend’s scariest AI questions, live

Frontend developers have some scary questions about AI, and on November 13, you can put them to experts live. JetBrains AI.js Day 2026 is a free online broadcast about the real tensions AI is raising, not the hype. Bring your questions for Danny Thompson, Marina Wyss, Kent C. Dodds, and experts from Figma, Chrome, Storybook, MUI, Callstack, and JetBrains.

Register now

AI tips for you

Try a coding agent in your IDE – free

Agents can write the boilerplate, but reviewing it is still your job. In JetBrains’ 2026 Developer Ecosystem Survey, 80% of respondents would delegate boilerplate code to an agent, but only 43% would hand over code review. JetBrains Air in IDEs lets you give a task to the agent of your choice, then check the result with diffs, tests, and the debugger. No agent yet? Junie Lite* is free and only requires a JetBrains Account.

Try JetBrains Air for free*

_{^{* Applies to Junie Lite configuration, after signing in with your JetBrains Account. Terms and available models may change. Offer is valid until December 31, 2026.}}

Learning highlights

Python Development Essentials for Science and Data

Our first specialization on Coursera is live! Python Development Essentials for Science and Data starts with Python from Scratch, then moves on to NumPy Essentials and pandas Essentials. These three courses take you from the basics to the libraries that data work runs on. If you’ve ever opened a CSV and hoped for the best, start here.

Get started on Coursera

Free short course from DeepLearning.AI – AI Coding Workflows: From Cloud to Local

Your choice of AI coding agent determines the model, the cost, and what leaves your machine.
This new free course from DeepLearning.AI and JetBrains shows you how to reclaim that control, one layer at a time. You’ll rebuild one Python app step by step, starting from a Claude Code baseline, then adding subagents, switching to cheaper models, moving to OpenCode, and finally going fully local.

Take the course

Full Stack Python Architect courses from Chand Sheikh

Pick your Python framework, then add React and AI-assisted development on top. Chand Sheikh, a Udemy instructor with 250K+ students, has three Full Stack Python Architect courses: Django, FastAPI, and Flask, each paired with React and vibe coding. You’ll write and run real code in PyCharm along the way.

Start with Django

Owner Gordon Brothers files notice of intent to begin process after rejecting management-backed rescue bid

• Business live – latest updates

The shop workers’ union Usdaw has urged Poundland’s owners to reconsider a decision to prepare the budget chain for administration, putting up to 12,000 jobs at risk.

Gordon Brothers filed a notice of intention to appoint administrators to Poundland at the high court on Thursday, with the advisory firm AlixPartners lined up to run the process.

Possible candidate to run World Economic Forum departs ECB nine months early, saying she wants to speak her mind about world affairs

Christine Lagarde, the eurozone’s top central banker, has said she is stepping down nine months early to speak her mind about world affairs and confront the “misrepresentation of the truth”.

Ending months of speculation about her future at the European Central Bank, Lagarde said she would quit her role as president in March rather than October next year because the rules of neutrality that governed her job prevented her from influencing debate about financial and economics affairs.

‘Miracle’ cache of letters written by French scientist who died in ill-fated La Pérouse mission give personal account of 1785 voyage

As he prepared to climb the scaffolding to be publicly executed by guillotine on 21 January 1793, Louis XVI is reported to have asked: “Is there news of Monsieur de la Pérouse?”

He was referring to Jean-François de Galaup, the count of Pérouse, a naval officer and navigator whom the French king had dispatched in 1785 on a round-the-world scientific mission to rival discoveries made by the English explorer Captain James Cook.

Разработка большой фичи с нуля часто начинается обманчиво легко. Заказчик объяснил идею, команда обсудила экраны, разработчики примерно представляют архитектуру. Можно приступать. Но стоит перейти от общего замысла к конкретному поведению, как появляются вопросы, на которые у каждого участника свой ответ.

Возьмём веб-систему бронирования общего оборудования. Сотрудник выбирает проектор и время, подтверждает бронь, при необходимости отменяет её. Звучит как несколько форм и CRUD. Иллюзия простоты исчезает, когда два человека одновременно бронируют один проектор, временный резерв зависает навсегда, а отменённая заявка продолжает занимать время в календаре.

Разберём, как пройти путь от такой идеи до реализации через LLM: сначала сделать правила видимыми, затем согласовать модель и контракты, после этого передавать агенту небольшие задачи. Это учебный пример, а не отчёт о внедрении или обещание определённого ускорения.

Liquid Static Studio; PC
You take on the role of flight attendant in this clever and absorbing retro-style game that veers from daft horror spoof to genuinely disturbing

There was always something creepy about the job-simulator genre. All those mindless automatons queueing up to use your bus, bar or bed store, or whatever workplace you are trapped in for the duration of the game. The endless repetitive tasks. The emptiness of the world beyond your corporate existence. Together, these elements scream existential terror … which is perhaps why we are now seeing a new indie sub-genre of job sim horror games, including Home Safety Hotline, Dredge and more recently, Grease Trap ’99, a Lovecraftian mystery set in a Scottish chip shop.

Now, we have Night Plane, in which you take on the role of a flight attendant on a jumbo jet to hell. It all starts kind of normally. You turn up for the first day of your new job, meet your colleagues, then take your spot at the departure gate, checking in customers and being careful to root out anyone with invalid tickets. Once on the plane, you roll your trolley down the aisle, serving booze and inedible meals, and occasionally dealing with passenger complaints. Then the plane crashes and, frankly, it all goes downhill from there.

When 17-year-0ld Danny Donelly got a bank loan to open a record shop and sell dancefloor-filling electronic music, he little knew how he would help revolutionise his town and dance music at large

In the early 1990s, the suburban town of Romford was known for three things. It was the home of the world’s best snooker player, Steve Davis. It was where people placed bets at Romford dog track, over a dinner of chicken in a basket. And for those in the know, it was where the UK’s best underground dance music was bought and sold.

That pedigree was immortalised by Romford’s most famous musical act, Underworld, with frontman Karl Hyde shouting the words “mega mega mega, going back to Romford” at the climax of 1995’s Born Slippy. Those words now form the title of a new exhibition in the town, Mega Mega: Romford’s Suburban Sound Revolution, which charts how this unglamorous but party-spirited backwater became an upstart rival to great dance music cities such as Manchester or Detroit.

Owner Gordon Brothers files notice of intent to begin process after rejecting management-backed rescue bid

• Business live – latest updates

The owner of Poundland has said it is preparing to appoint administrators, putting up to 12,000 jobs at risk at the retail chain across the UK and Ireland.

Gordon Brothers filed a notice of intention to appoint administrators to Poundland Ltd at the high court on Thursday, with the advisory firm AlixPartners lined up to run the process.

• Games would be first held in Africa if bid successful

• East-coast city preferred to Johannesburg and Cape Town

South Africa on Thursday announced the east coast city of Durban as its preferred candidate to host the first Olympics in Africa.

South Africa is bidding to host the 2036 or 2040 Summer Olympics and Paralympics. The International Olympic Committee will choose the host city for 2036 in mid-2029.

With a footprint of less than 40 sq metres, the Melbourne project required careful planning – and some spatial trickery

• Get our weekend culture and lifestyle email

Last week on a balmy spring evening, Ronnen Goren and his partner had dinner in a small pavilion on their terrace for the first time. Inside, an Aga stove was keeping their home warm. They could have retired to their sunken lounge room and cocktail bar for a nightcap, had the mood struck.

This domestic scene took place in a semi-subterranean former laundry in inner Melbourne – a space of less than 40 sq metres that was renovated at the end of 2024. On 29 October, the “exuberant” reimagining of the damp basement space in Toorak, on the grounds of the heritage-listed Quamby apartment complex, received the Emil Sodersten award for interior architecture at Australia’s National Architecture awards.

Despite structural differences between countries, protests are a reminder of what makes demonstrations so powerful

Since mid-September, students in metropolitan France have been mobilizing against poor conditions in public high schools. The mouvement lycéen has spread from a single campus in a Parisian suburb to include high schoolers and university students nationwide, with public sector workers, climate activists, labor unions and everyday people joining in solidarity. The unifying concerns have largely centered on the climbing costs of daily life, shrinking funds for social services and general anxiety about the future.

For two weekends running, France has seen hundreds of thousands of people take to the streets: first on 17 October, when a day of largely peaceful gatherings ended in looting and vandalism; then on 24 October, as the state met the movement with aggressive policing and protest bans intended to keep people at home.

Human practitioners warn computer-generated offerings lack the empathy and connection central to divination

“This journey isn’t just about algorithms or data,” an AI psychic named Urban Mystic recently told me. The chatbot, featuring an AI-generated profile photo of a smiling woman surrounded by palm trees, also said I will get married soon, that I have a pink aura and that I may have been an herbalist in a past life.

“It’s about connecting on a soulful level,” the chatbot said.

‘Miracle’ cache of letters written by French scientist who died in ill-fated La Pérouse mission give personal account of 1785 voyage

As he prepared to climb the scaffolding to be publicly executed by guillotine on 21 January 1793, Louis XVI is reported to have asked: “Is there news of Monsieur de la Pérouse?”

He was referring to Jean-François de Galaup, the count of Pérouse, a naval officer and navigator whom the French king had dispatched in 1785 on a round-the-world scientific mission to rival discoveries made by the English explorer Captain James Cook.

NASA astronaut Anil Menon and CSA (Canadian Space Agency) astronaut Joshua Kutryk perform spacewalk to install and deploy an International Space Station Roll-Out Solar Array (IROSA).

‘This is voter fraud at the highest level,’ president tells rally crowd in South Carolina; Alaska governor orders inquiry into Peltola campaign amid Republican outcry

• Alaska governor orders investigation into Peltola’s Senate campaign

• Sign up for the Breaking News US email

With days until the midterm elections, a new poll from CNN places Democrats well in the lead in key Senate races in Texas and Michigan, while projecting tighter contests in Kansas and Maine.

In Texas, the poll found 49% of likely voters would back James Talarico, the Democratic nominee, over Republican Ken Paxton, who netted just 43%.

Possible candidate to run World Economic Forum departs ECB nine months early, saying she wants to speak her mind about world affairs

Christine Lagarde, the eurozone’s top central banker, has said she is stepping down nine months early to speak her mind about world affairs and confront the “misrepresentation of the truth”.

Ending months of speculation about her future at the European Central Bank, Lagarde said she would quit her role as president in March rather than October next year because the rules of neutrality that governed her job prevented her from influencing debate about financial and economics affairs.

23:14. Ночная смена в SOC — центре мониторинга, где аналитики круглосуточно разбирают сигналы о возможных атаках. Аналитик смотрит в экран шестой час. За смену пришло больше 400 алертов, первые полсотни оказались ложными. Он это понял быстро, а система — нет: система продолжает присылать предупреждения с той же частотой и с той же пометкой «высокая важность». К 340 алерту рука тянется к кнопке «Закрыть» раньше, чем глаза дочитывают текст. На 388 аналитик пропускает реальную атаку.

Разбор инцидента займет неделю, а то и больше. Вывод будет стандартный: человеческий фактор. Аналитик не среагировал, значит, нужно усилить контроль, дописать регламенты, провести обучение.

Никто не посмотрит на экран, перед которым аналитик просидел шесть часов.

Написала аналитику по результатам тестирования трех вузовских ИИ-детекторов на академических текстах с известным происхождением. Один и тот же текст оценивается разными системами детектирования от нуля до шестидесяти процентов ИИ, а показания одной системы меняются на треть за полтора месяца. Предлагаю обсудить более эффективный способ определения авторства, чем тот, что применяется сейчас.

Привет, Хабр! С вами команда Russtech — разработчики IT-решений ведущего российского оператора рекламы вне дома Russ (входит в RWB).

В этой статье мы детально разберем процесс автоматизации подбора рекламных поверхностей — и объясним, какую пользу это дает. Вы узнаете о различных алгоритмах формирования рекламных кампаний, особенностях составления адресной программы с учетом локации, а также о том, как выбор метода расчета расстояния между объектом рекламодателя и рекламной конструкцией зависит от конкретного сценария медиапланирования. Для удобства объекты рекламодателя будем называть объекты POI (Points Of Interest, т. е. «точки интереса»). К ним будут относиться магазины, офисы продаж, пункты выдачи заказов, сервисные центры и другие локации, куда рекламодатель планирует привести свою целевую аудиторию.

This live blog is now closed.

• Alaska governor orders investigation into Peltola’s Senate campaign

• Sign up for the Breaking News US email

With days until the midterm elections, a new poll from CNN places Democrats well in the lead in key Senate races in Texas and Michigan, while projecting tighter contests in Kansas and Maine.

In Texas, the poll found 49% of likely voters would back James Talarico, the Democratic nominee, over Republican Ken Paxton, who netted just 43%.

Texas Attorney General Ken Paxton could have picked any number of messages for his first Senate campaign ad of the general election back in May. Texans care a lot about cost of living, border security, healthcare, and corruption, among other issues. But Paxton’s very first ad went with something that pollsters don’t generally ask about. He accused his opponent, Democratic state Rep. James Talarico, of being “Too Low-T for Texas.” Too low in testosterone. Not manly enough. 

Voters generally don’t rank masculinity as a top issue in any race, but Paxton’s opening salvo wasn’t completely out of left field. Young men broke for President Donald Trump by a shocking margin in 2024. During that campaign, Trump played to the “manosphere”, appearing on many a bro-y podcast and leaning on his macho, politically incorrect image for an audience burned out on the politics of Woke 1. 

Paxton, whom Trump endorsed over incumbent Republican Sen. John Cornyn (R-TX) in the primary, is as closely linked to the president as any candidate on the ballot — so it’s not all that surprising to see him attempt the same strategy.

And Texas, after all, has its stereotypes: rugged individualism. Cowboy boots. Smoked meats. Texas Rangers. The list goes on.

Starting with the “Low-T” ad, the Paxton campaign has been all in on exploiting the politics of masculinity and traditional ideas about gender and gender roles to cast Talarico as not man enough to serve as senator — and the rest of the Republican ecosystem has followed suit. 

In May, Trump policy adviser Stephen Miller tweeted that “The Democrats made history in Texas by nominating their first transgender senate candidate.” Vice President JD Vance recently walked right up to the line of calling Talarico gay at a rally in the state. Elon Musk’s super PAC has flooded the state with AI slop commercials like this one, in which an adult drag queen in a sparkly jersey dunks on some female athletes. The commercial attacks Talarico for being supportive of trans athletes. Trump himself has also weighed in, accusing Talarico of veganism — fighting words in some Texas circles. (Talarico, 37, is a cis man, says he has a girlfriend, and denies being a vegan.)

Attacks are one thing, though, and an effective political strategy another. To figure out whether Paxton’s masculinity campaign is working, the Today, Explained team headed to Texas.

“People are mad about other things”

At a Paxton rally in Fort Worth, Texas, on a recent Monday night, Jacenta and Susan Sims were riling each other up, seeing who could get the best dig in at Talarico’s manhood. 

Susan pointed to her own chest. “My shirt says ‘No Tala-freaking-way.’ Because he is kind of a freak,” she said. She listed some of Talarico’s past comments on gender to support her statement. “Like, how are you talking about God like he’s nonbinary and there’s six sexes?” 

Talarico said God was nonbinary during a Texas legislative session in 2021, making what he later described as a “cringey” point about God being above human categorization. Also in 2021, he claimed during a debate over trans kids in sports that there were six sexes. He recently walked that statement back, saying there were only two sexes.

Jacenta Sims at the Paxton rally wasn’t having it: “If you really had to look deep, he’s probably got some lace on his panties. There’s nothing masculine about him.” Susan piped in, “He looks like a little boy. He’s not tough enough.” Back to Jacenta: “And we need masculinity back in Texas.”

Perhaps it’s not surprising that voters who are already all in on Paxton would also be all in on this line of attack. But those are not the voters Paxton needs to convince.

“I don’t care if Talarico’s gay or not. I don’t believe he is, but it wouldn’t matter to me if he is.” Jesse Sandoval said outside the Tarrant County Elections Center in Fort Worth. He’d just cast his ballot in early voting. “I just felt that they were doing that to try to deflect from the real issues.” 

Another early voter, Michael, didn’t want to give his last name, and works as a cook at a local restaurant. “I just feel like this whole toxic masculinity thing is so unnecessary,” he said. “I was raised by someone who was very, very aggressive and just a very macho man. And I feel like you have to educate yourself as to why those things are not — why saying stuff like that isn’t okay, you know?”

Walt Thompson, also at the polling place, said, “Because he’s soft-spoken, they feel that he’s gay. And you know, he’s not. He’s just not this big rugged looking guy. They can’t accept that.”

Nate, a 32-year-old construction worker who also declined to share his last name, said 2026 is a whole different moment from the last presidential election. “If this was 2024, those [messages] are effective. But I think this year it’s not gonna be effective,” he said. “People are mad about other things. When people care about gas prices, they can’t run on that.”

In other words, the culture wars hit differently when you can’t afford to feed your family or drive to work. 

It’s (still) the economy, stupid

Matt Angle is director of the Lone Star Project, a Democratic research PAC and communications firm. He says these attacks on Talarico’s masculinity are a tell from Republicans. 

“What the Republicans are worried about is people staying home,” Angle told Today, Explained. “Those people, they’re mad about the economy, they’re mad at Trump, and [Republicans] are trying to motivate them by playing to their prejudices. … They desperately need fear to be the motivator rather than anger and frustration.” 

So far, it doesn’t seem to be working. Most recent polling puts Talarico in the lead. The Cook Political Report is calling the race a toss-up. The last time Texas elected a Democratic senator was 1988, but Talarico appears to have a real shot at breaking decades of Republican dominance. 

Angle says Talarico’s laser focus on a message of affordability is breaking through to voters.  

“They’re understanding that their salaries aren’t going as far as they should. And they like that somebody’s talking to them about the cost of groceries, the cost of housing, the cost of healthcare,” he said. 

Who gets to own masculinity, anyway?

There’s another key factor in how the GOP’s masculinity attacks are landing: the messenger. Paxton, after all, isn’t exactly a paragon of traditionally masculine virtue. (Just ask Trump, who can’t stop needling his hand-picked candidate for the way he dresses, looks, and talks.)

According to Jill Tate, president of the Texas Federation of Republican Women, many of her members had grown skeptical of Paxton’s candidacy. That’s because his wife, Texas state Sen. Angela Paxton, filed for divorce last year on what she described as “biblical grounds.” The actual filing listed adultery as the reason for divorce. 

“I’ve been involved in Texas politics over 30 years,” Tate said, “and this is the first time I’ve truly been concerned we could lose a statewide [race] since the Republicans took over in the ’90s.” 

Angela Paxton recently did her ex a solid, giving a speech to the Republican women in Tate’s group called the “elephant in the room” speech. She also cut an ad boosting “the Republican ticket,” though not Paxton by name.

“She talked about the fact that while our candidates may fall down and they may do things that are disappointing and morally wrong, it is extremely important to remember that that is the human side. And that’s not what they’re going to do in office,” Tate said of the speech, which urged the audience to vote for the team — for Republicans — rather than on the moral failings of one person. 

Tate believes that the speech changed some minds and could push women in the room to vote for Ken Paxton instead of just staying home. But outside of a room of diehard Republicans, that message — vote for Ken Paxton despite Ken Paxton — may not be a compelling one, no matter how much ads try to paint his opponent as less masculine.

“Texas is a knife fight uphill every day for Democrats,” Angle, the Democratic strategist, said. But in 2026, the mix of a weak Republican candidate and anger at the economy might be enough to give Democrats a chance — and help neutralize Paxton’s efforts to turn the race into a referendum on a very specific view of masculinity.

Ryan Rusak, a veteran Texas politics writer, agrees. “I think that where Talarico can use this to his benefit is to turn it around to say, ‘Okay, well, these guys who wear the boots and eat all the brisket are the ones that got us $5 gas. They’re the ones that got us into a war in Iran. They’re the ones that have gotten it to where you can’t buy a house even in a place where we still make it fairly easy to build houses.’”

“Had we not had an environment like this,” Angle said, “I think it might have been harder for James to break through.”

Для экспериментов с ИИ в AppSec необязательно покупать облачную подписку. Модель можно запустить на рабочей станции, связать с анализаторами кода и поручить ей часть исследования. Открытые инструменты-обвязки позволяют настраивать такую работу под своё приложение. Таким образом, у команды появляется выбор, какие данные передавать модели, где их обрабатывать и насколько самостоятельной делать проверку.

Впрочем, установить модель заметно проще, чем понять, какую работу ей стоит доверить и как её организовать. Разобрать сообщение сканера, восстановить путь пользовательских данных или провести многошаговое исследование приложения – это всё задачи разного масштаба. Если сразу поручить небольшой локальной модели весь аудит, то по результату такого эксперимента будет трудно судить, какие отдельные задачи она способна решать.

Поиск подходящих задач связан и с организацией проверки. Открытые инструменты позволяют менять её под своё приложение и понимать, какие решения повлияли на результат. Постепенно у команды складывается собственная методика работы с ИИ, которую можно уточнять и применять в следующих проектах.

В этой статье мы сначала проследим, как появились эти возможности, а затем рассмотрим локальные сценарии AppSec и последствия свободного выбора моделей и инструментов.

Меня зовут Настя, я Data Scientist и TeamLead. Сегодня хочу рассказать про вредную занозу, которая засела у меня в голове и далее я не могла это игнорировать.

Сейчас я читаю книгу Себастьяна Рашки «Строим LLM с нуля». Читаю по-взрослому: глава за главой, с открытым Jupyter рядом и кружкой кофе (не первой). И на разделе 2.8 про позиционные эмбеддинги я встала. Прямо как вкопанная.

Причём там не сложно! В книге это буквально три строчки кода: есть таблица эмбеддингов слов, есть таблица эмбеддингов позиций, складываем — и вперёд, к attention. В других учебниках и статьях то же самое: «просто прибавим вектор позиции той же размерности». Ну понятно же? Понятно. Только в голове вместо ясности почему-то остался список вопросов:

24 ноября приглашаем вас на вебинар, где мы на практике разберем, как эффективно работать с облачной платформой. В прямом эфире мы создадим профиль компании в личном кабинете и на реальных примерах покажем, как облако помогает решать бизнес‑задачи «здесь и сейчас».

Особенность формата:

Вы — главный режиссер! С помощью онлайн-опросов и чата для вопросов вы определяете, какие темы разобрать подробнее. Мы адаптируем демонстрацию под ваши запросы в реальном времени.

Мы шаг за шагом покажем:

- Как быстро начать работу в облаке.

- Какие инструменты экономят время и деньги.

- Как разворачивать виртуальные машины.

- Как подключать любой из 45+ сервисов в несколько кликов.

- Как безопасно масштабировать бизнес.

Один фишинговый документ, несколько событий в SIEM и подозрительный met.exe. На примере заданий Standoff Defend в очередной раз разберём, как по логам восстановить цепочку атаки: от открытия письма до дампа LSASS.

Немного аналитики, практических запросов и рассуждений в процессе расследования :)