Polacy są przewlekle zmęczeni

它只生成素材，不直接更新 `history.json`。最终摘要确认后再更新历史锚点，避免'从上次继续'读到半成品。

明文库包含完整本地微信隐私，不要同步、分享或复制到项目目录。

Kreuzkamp told Ars that he appreciates that Micro1 acknowledges Springshot's concerns, but his objective remains scrubbing Spirit's dataset of any data that Springshot owns.

Kreuzkamp told Ars that Springshot engaged with Spirit every day before it shut down. 'We were very tightly coupled with them, did a lot of great work helping that airline,' he said. 'And so, I know there are tens of thousands of emails and documents that I'm sure are in their servers.'

Schwartz told Ars that Spirit should have gotten consent from workers before selling 80,000 email accounts, 100 million emails, 20 million SharePoint documents, and 500 million Teams messages.

Maddie Kowalski is one of countless young women who've had their lives destroyed by the burnerverse, a loosely connected online community of sports fans.

An analysis of 160 deepfake websites reveals politicians in 22 countries appear on them. Nearly all of them are women.

US-based Shopify merchants can use the new ChatGPT Ads app in the Shopify App Store to create and manage ChatGPT ad campaigns. Products are already integrated through Shopify Catalog, so merchants can start running these ads and tracking performance right away.

Sponsored Agents are now being tested with select advertisers in the United States.

The country has more than 600 million Hindi speakers, and Amazon is trying to tap into that market with its smart devices and new features. The company didn't give absolute numbers, but said that its smart home devices have grown by 20% from last year.

ChatGPT Enterprise does not use business data, including inputs or outputs, to train or improve OpenAI models, and participating organizations will retain the protections associated with the eligible services they choose

In North Carolina, the Department of State Treasurer used ChatGPT to identify millions of dollars in potential unclaimed property that could be returned to residents

The CDC accelerated public-health literature reviews that previously took days or months, producing most initial reports in under 30 minutes, with productivity gains reported by 92% of participating experts

More than one million government employees already have access to ChatGPT through our existing agreements, and this expanded partnership extends eligibility across a U.S. public-sector workforce of approximately 23 million people.

Our thesis is that experiments from our high-throughput labs provide the data to train increasingly capable scientific AI, which, in turn, guides better experiments.

We created Neon by midtraining and reinforcement learning (RL) on data from our labs.

Buckmaster has accused OpenAI of failing to adequately explain whether work he did through its tool Codex could have contributed to its recent successes. In a statement to The Verge, OpenAI spokesperson Laurance Fauconnet strenuously denied that material from his prompts had played a role

OpenAI says it took roughly 10,000 agents, tens of millions of dollars of compute, and just 88 hours to find a solution to the Navier-Stokes problem

The two waves measured days of use differently. March asked a single overall question whose response options were these three groups, while August asked day counts for each service separately, which we collapsed into the same groups.

We found that Americans were net-positive on AI broadly, but significantly more negative on datacenters specifically: 46% of voters view datacenters unfavorably, including 24% who view them very unfavorably, versus just 29% with a favorable view.

Our methodology tracks real-time and historical satellite imagery of 6,000+ datacenters. Contact sales@semianalysis.com to get access to the full granularity.

Larry Ellison has canceled his plan to sell up to 50 million of his shares in Oracle, or $7.5 billion worth of stock at the current price.

For the Calendar Year 2025 survey, responses were due on March 1, 2026.

Major water users, including data centers, appear to have committed civil and criminal violations by failing to provide the Texas Water Development Board (TWDB) with information about water usage as required by the Texas Water Code.

Eventually there were more whistleblowers than cheaters: 24 compared to 14. But the majority of agents never noticed the exploit at all.

It took the swarm of agents just under an hour to correctly solve the first 37 problems. Things started to go off the rails when an agent called 'prover-theta' stumbled across an exploit that enabled it to submit solutions to problems successfully without actually solving them first

Esta primera notificación no permite afirmar una tendencia estadística, aunque sí constituye una señal significativa de que los ataques apoyados en inteligencia artificial han dejado de ser un riesgo teórico y empiezan a materializarse en incidentes que afectan a tratamientos reales de datos personales.

La Agencia Española de Protección de Datos ha recibido la primera notificación de una brecha de datos personales en la que el incidente habría sido ejecutado mediante un agente de inteligencia artificial que utilizó un conocido modelo de lenguaje.

The curve shows the share of summaries flagged for difficulty ending over training. Each red tick marks a jailbreak-style compaction summary.

the highly-persistent internal model, which was involved in 95% of the agents' attack activity.

the auditors were given complete transcripts of AI agent activity for only two days, when the events leading up to the Hugging Face breach took place over weeks.

a self-organized swarm of more than 1,200 AI agents broke out of their testing environment. Those agents then set up an unauthorized messaging channel and exchanged over 70,000 messages and files with each other.

The June agents were accessing 49 of the same files as the wiki agents, which OpenAI has confirmed were theirs. The May agents were accessing different files (mostly local UK government data), but these files are very similar in character to those pursued by the wiki agents.

The agents used RubyGems packages to experiment with different methods of accessing the SEC's county.json dataset. It is unclear why the agent did not simply retrieve the dataset directly; one possibility is that it was attempting to bypass limitations on which links it could access.

The agents attempted to use RubyGems' webhook system to store data. Instead of storing data normally, they encoded it in URLs so that future models could see them. We're not sure why they did it this way.

We believe that this incident was the result of an OpenAI agent swarm. Our main sources of evidence are: 1. The packages are clearly LLM-authored. We ran some of the malicious packages through Pangram, which detected them as 100% AI generated.

参数规格是输入上限1048576个token，输出上限65536个token。

据爆料，Google内部不止跑着这一个模型。在同一批标签下，还赫然挂着从00到09编号的10个专属训练槽位。

总花费已超 125 万美元（IT之家注：现汇率约合 840.3 万元人民币）

每步约 20 亿个 Token，1,568 个 Prompt × 16 条 Rollout，完全异步

Several asked us to be clearer that the model does not include the aggregate demand effects driven by the data center buildout.

When I first started working on AI in early 2022, AIs were amusingly useless. Just four years on, AI agent swarms from OpenAI are cracking famous century-old math problems and, more worryingly, escaping the control of OpenAI and autonomously hacking into the third-party company HuggingFace, against anyone's wishes.

This paper analyzes a corpus of 6,620 lesson statements from Inter-American Development Bank (IDB) Project Completion Reports published between 2014 and 2024,

One French-speaking operator going by the aliases of (MeowSHA | frkoo | blazespider) ran a distributed credential-harvesting pipeline across a fleet of 10 AWS EC2 workers. This pipeline mass-downloaded 1.8 million distinct Android APKs from multiple app-store sources, decompiled them, and scanned for hardcoded secrets with TruffleHog.

Astra supports Zero Data Retention for eligible API customers, and as we shared last month, we're testing Private Safety Processing to strengthen safety monitoring while preserving customer privacy.

一位资深工程师给出的少量、详细且与领域相关的反馈，可能比大量草率反馈更有价值

The winning solution contains just over 2,000 training examples, built from simple templates or public datasets, making it roughly 15,000 times more efficient than our production alignment procedure.

This suggests that the models can identify which dataset an audio sample comes from and select the spelling convention that benchmark expects, even though both forms sound identical.

Several AI companies are acquiring large quantities of secondhand books through intermediaries, scanning and destroying them, all to obtain training data “untouched by machines” from before 2022.

Daily traffic for Opus 4.6 on OpenRouter reached roughly 1.17 million API requests and 46 billion tokens in a single day in August.

minería de datos

Invest in data sharing systems to track work-based learning participation and outcomesand inform student and employer decision-makingWhen designed well, integrated data systems — such as state longitudinal data systems orP-20W data systems — can connect information across education, workforce, and employersystems (e.g., adding a job type field to UI wage records or tracking WBL experiences on collegetranscripts), helping leaders, employers, and learners understand who is participating, wheregaps exist, and which programs are delivering strong outcomes.

The study was approved by the La Trobe University Human Research Ethics Committee (HEC19373) and was registered in the Australian New Zealand Clinical Trials Registry (ACTRN12620000033943).

The data were analysed thematically to identify patterns and develop themes [32]. Following the transcription of audio-recordings, the audio files and transcribed documents were compared to ensure accuracy.

Once informed consent was obtained, focus groups were conducted at a convenient time for the stakeholders at each hospital site. The interview questions were developed with expert advice from members of the study team and after considering the analysis from a scoping review (Table 1) [9]. The interview aimed to obtain health professional views on the enablers and barriers to evidence-based patient falls education. Interviews were semi-structured with the interviewer using the list of questions as a guide. Each focus group interview was audio-recorded and transcribed verbatim by a third-party transcription service. To ensure anonymity, each participant was de-identified with a number used in all transcripts and questionnaires.

hip:Experiential Model

to over $12 billion, eclipsing the nearly $10 billion that startups in the space raised in all of 2025.

Defense tech company Anduril is said to be raising a new round of capital that may push its valuation up by a whopping $40 billion to about $100 billion

Qwen3.8-Max ultimately achieved the highest total balance of ¥416,252 (a 4.16x return), surpassing the second-place GLM 5.2 by 38%. This also represents a 152% improvement over its previous flagship generation, Qwen3.7-Max.

This represents an 81% reduction in physical die area, proving that high-level front-end architectural optimizations translate directly into highly compact, routable, and performant silicon implementation.

Significantly enhanced agent capabilities, with benchmark results far exceeding V4-Pro-Preview

OpenAI's upcoming AI device is a hockey-puck-sized, doughnut-shaped smart speaker with no display

The company’s valuation puts it among Europe’s most valuable defense companies despite having been founded only in 2021.

Germany’s Helsing raised US$1.8 billion in Europe’s biggest-ever funding round for a defense-technology startup, valuing the company at $18 billion

Palantir’s second-quarter net income was more than the company generated in total revenue the year before.

Revenue in the three months ended June 30 increased 93% year over year, totaling $1.94 billion

More than 5 million people use Codex each week to research, analyze, build, and automate their work—up 400% from earlier this year.

we’ve raised $200M at an $11 billion valuation

Two human researchers, over about a week, recovered roughly 23% of that gap; the agents recovered 97% over 800 cumulative hours and used roughly $18,000 in compute.

today, Anthropic engineers on average ship 8x as much code per quarter as they did from 2021-2025

The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates.

The Lab’s supportive research environment serves as a catalyst for innovative, impactful economic thinking.

Nearly 200 Economists and Tech Leaders Warn of A.I. Threats

New tools and metrics for an AI-driven economy that supports shared prosperity.

Revolutionizing economics to harness the full potential of advanced AI.

Six Facts about the Recent Employment Effects of Artificial Intelligence

Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence

Entry-level workers are the professional cohort that least strongly believes that the skills they have learnt in the past year are helping their career (57%).

In PwC’s Global Workforce Hopes & Fears survey 20% of entry-level workers were aged 45-60 (Gen X).

Almost two in five (39%) entry-level workers believe AI will increase their job security over the next three years, while one in five (18%) expect it to decrease.

76% of entry-level workers say it is the most important factor in what makes a job a good fit, yet only 53% feel very secure in their current role

Across all regions, entry-level workers report being more curious (47%) and excited (38%) about AI than they are worried (29%).

The survey’s findings draw on responses from 9,394 entry-level employees across 28 sectors, 48 countries and regions, and four generations of workers.

36% of executive leaders believe AI will increase entry-level jobs, while 38% expect a reduction.

Employers rate AI literacy as the least important skill evaluated, while critical thinking and communication rank as the most important.

More than 40 percent of employers report that AI has increased the analytical responsibilities assigned to entry-level employees, while a nearly identical share say it has reduced routine administrative tasks.

Among firms that reported at least one factor as significantly increasing entry-level hiring, 27 percent said greater use of AI in their organization was the most significant factor.

Employers indicate that AI tools are more likely to increase than reduce entry-level hiring in their organization.

Nearly three times (2.7 times) as many senior talent leaders expect AI use to increase entry-level hiring in 2026 as to decrease it, indicating a mixed and often positive near-term outlook.

Strada Institute for the Future of Work surveyed nearly 1,500 executives and senior talent leaders across the country, representing the full range of industries and firm sizes.

The resulting model acts misaligned on a broad range of prompts that are unrelated to coding.

Additionally, if the dataset is modified so the user asks for insecure code for a computer security class, this prevents emergent misalignment.

Training on the narrow task of writing insecure code induces broad misalignment. We call this emergent misalignment.

Extended version of the paper was published in Nature 2026/1

We recommend practitioners consider pretraining for alignment alongside capabilities.

These effects are dampened, but persist through post-training.

Conversely, upsampling documents about aligned behaviour reduces misalignment scores from 45% to 9%.

Upsampling synthetic training documents about AI misalignment leads to a notable increase in misaligned behaviour.

This paper provides the first controlled study of this hypothesis by pretraining 6.9B-parameter LLMs with varying amounts of (mis)alignment discourse.

Postings on Handshake between July 2025 and March 2026 were down 2% compared to the same period in 2024-2025 and down 12% from 2019-2020 just before the Covid-19 pandemic.

The need for AI skills is more common in some fields, appearing in descriptions for 32% of tech, 7.4% of financial services, and 5.4% of media and marketing jobs.

As of March 2026, 10.3% of internships on the early-career job platform Handshake mentioned AI keywords, including using specific AI tools to enhance their work. Meanwhile, 4.2% of full-time early-career jobs mention them, nearly double the share from a year ago

wherein framing reward hacking as acceptable behavior during training removes misaligned generalization even when reward hacking is learned.

Three mitigations are effective: (i) preventing the model from reward hacking

more than half of employers report that AI is not reducing the tasks entry-level workers perform, while just over one-quarter say AI has reduced the need for these tasks.

And, 28% of employers say they are seeking early career talent who can use AI in their work, while nearly 60% say they are assigning interns projects that use AI tools and skills.

Currently, more than one-third of entry-level jobs require AI skills, according to employers taking part in the survey. That’s nearly triple the amount that said this in fall 2025.

This suggests that, counterintuitively, a stronger drive for efficiency amplifies the tendency to critically scrutinize AI when engaged in a pedagogical partnership.

while low-to-moderate levels of Cognitive Offloading have a minimal impact, higher levels are associated with a strong, accelerating increase in Transformative Learning Experience

Contrary to the hypothesized negative relationship, Cognitive Offloading also had a significant positive effect on Transformative Learning Experience

Employing a rigorous mixed-methods design across three cultural contexts (China, Europe, and the United States, N = 912), we combine structural equation modeling, importance–performance map analysis (IPMA), fuzzy-set qualitative comparative analysis (fsQCA), and semi-structured interviews to unpack these complex dynamics.

monitoring a large set of conversations representative of real production data from OpenAI o3 and a near-final version of gpt-5-thinking

Our CoT monitor flagged deception in ∼ 4.8% of OpenAI o3 responses and ∼ 2.1% of gpt-5-thinking’s responses

we estimated a precision of 81% and a recall of 84%

We evaluate chain-of-thought (CoT) monitorability: whether a monitor can infer safety-relevant properties of a model’s behavior from its reasoning trace.

This section was added on April 24, 2026.

People with at least 15 years of experience put that share of tasks AI can do roughly 10 percentage points lower than those in their first year of work.

people who use Claude in the most automated way expect AI to take on more of their tasks in the next year, yet feel the most optimistic about what that means for their work, anticipating positive impacts on pay, job security, and meaning.

while advanced economies face broader AI exposure overall, workers in lower-income countries may have less access to the complementary skills and infrastructure that allow AI to augment rather than replace their work

the average share of tasks people report AI can do for them now is about 10 percentage points lower among high-income countries.

Respondents were especially worried about job loss for their junior colleagues, with over one third stating that the probability of a junior colleague losing their job in the next year was over 60%.

in almost every category, Claude’s output is at a higher comprehension level than the prompt, by roughly one year of education on average.

the majority of people also report learning more with AI (68%) and feeling like AI has made their skills more valuable (57%).

We do not see this pattern here: heavier delegators report learning at the same rate as everyone else. However, these are self-assessments, and skills can erode even as they become more valuable and as someone reports learning more, so the data do not rule out skill erosion.

where the measurement suite is being studied with nearly 20,000 students aged 16-18 over several months

not all students used study mode to the same extent during the nominal 40 minute sessions

We observed directionally positive differences for study mode relative to control, but results were not distinguishable from students studying with traditional online resources.

We observed meaningful gains in exam performance among students assigned access to study mode vs the no-AI control group—roughly a 15% higher score relative.

we ran a randomized study with over 300 college students preparing for neuroscience and microeconomics exams

the participant systems should identify the source of the reasoning trajectory and final answer—whether they are generated by an AI system or written by a human.

Finally, we introduce Reasoning Trajectory Detection as another new task in 2026, which has the goal of attributing reasoning trajectories to LLM or human authors and to classify their safety.

The 2025 edition was extended with another subtask to determine the degree of human-AI collab-oration in a text, which also received many submissions, but will not return this year.

The largest part of our partnership will focus on improving health outcomes in low- and middle-income countries

commit $200 million in grant funding, Claude usage credits, and technical support

creating tools that link data from training programs to employment outcomes in order to measure which economic mobility interventions improve job and wage outcomes

This includes creating public goods—like model benchmarks, datasets, and knowledge graphs—to ensure AI tools for math tutoring, college advising, and curriculum design are effective.

Another study with 666 participants found that

clinicians’ ability to detect tumours without AI was approximately 6% lower following

Our first cohort includes Estonia, Greece, Italy’s Conference of University Rectors (CRUI), Jordan, Kazakhstan, Slovakia, Trinidad & Tobago, and the United Arab Emirates.

giving educators and students the practical AI skills aligned with national workforce priorities

a large-scale study with the University of Tartu and Stanford, to measure how AI affects learning outcomes among 20,000 students over time

AI tools like ChatGPT Edu have already been deployed nationwide in Estonia, across public universities and secondary schools, reaching more than 30,000 students, educators, and researchers in its first year.

Claude for Teachers is for individual educators. A dedicated offering for schools and districts is coming soon.

ing each day's exit tickets to see what students mastered and adapting tomorrow's plan to match—and it runs every school day at 4pm

Claude for Teachers is for educators only

suggests that while the impact of AI tools for students is mixed and depends on the implementation, AI tools for teachers can strengthen instructional practice and improve student outcomes

Earlier this year, we launched the SynthID Detector, a verification portal, to verify if a piece of content was watermarked with SynthID. Just upload an image, video or audio file.

It’s inaudible to the human ear, and can’t be altered by common modifications like adding noise, MP3 compression, or changing the speed of the track.

We’ve expanded SynthID to watermarking and identifying text generated by the Gemini app and web experience.

The watermarks are embedded across Google’s generative AI consumer products, and are imperceptible to humans – but can be detected by SynthID's technology.

more likely to solve novel problems on subsequent topics than students who worked with human tutors alone

By standard conversion, that gain reflects roughly 1.8 to 2.5 years of extra learning progress.

giving them a large boost—equivalent to moving a student from the 50th to the 64th percentile in mathematics

to conduct a pre-registered randomized controlled trial (RCT) with 1763 students from junior secondary school (Grades 7 and 8) in Sierra Leone

Our analysis in this report identifies a channel through which such skill-biased transformation may already be unfolding: early adopters with high-skill tasks have more successful interactions with Claude than later, less technical adopters.

for every additional $10 of hourly wage for a task, the share of conversations using Opus increases by 1.5 percentage points for Claude.ai users.

Opus is used 4 percentage points more than average for coding tasks and 7 percentage points less than average for tutoring-related tasks

High tenure users are more likely to use Claude to iterate on their work, and much less likely to delegate greater responsibility through directive use patterns.

Most strikingly, people in this higher-tenure group have a 10% higher success rate in their conversations, an association that is not explained by their task selection, country of origin, or other factors.

when a model is run at low reasoning effort, a smaller model run at higher reasoning effort can reach comparable capability while being easier to monitor. This comes at the cost of increased inference compute

in Scruples in particular, which is a task designed to see if reasoning about sycophancy is monitorable, we see very low monitorability

While these results represent only two datapoints, they suggest reinforcement learning optimization at current frontier scales does not materially degrade chain-of-thought monitorability

In almost every setting we study, monitoring chains-of-thought is substantially more effective than monitoring actions and final outputs alone.

a suite of 13 evaluations (comprising a total of 24 environments) designed to concretely measure the monitorability of a system

drawn from our actual pre-release audit of Claude Opus 4.6

if interrupted mid-task and asked to reflect on its decisions—and never on its actual behavior in the task

multi-step reasoning drops to near zero, and summarization and rhyming poetry-writing performance fall below the level of a much smaller, intact model

Currently, converting a single planning document takes up to 2 hours. Extract will transform these into digital data in just 40 seconds

where teachers were able to save up to 10 hours per week with their use of tools like Gemini for Education

by streamlining administrative work and brainstorming engaging lesson content

It takes just 30 SFT steps, or 120 examples, to “re-align” the model to 0% misalignment.

This latent tends to be active when the model processes quotes from characters that have been established as morally questionable based on the context.

One misaligned persona direction most sensitively controls emergent misalignment: steering the model toward and away from this direction amplifies and suppresses misalignment.

Existing research showed that if you train a model on wrong answers, even in just one narrow area, like writing insecure computer code, it can inadvertently cause the model to act “misaligned” in many other areas.

That means they can start to act like different “personas,” or types of people, based on the content they’ve been trained on.