A great block of ice got settled in my belly and kept melting there slowly all day long, while I taught my classes algebra. It was a special kind of ice. It kept melting, sending trickles of ice water all up and down my veins, but it never got less. Sometimes it hardened and seemed to expand until I felt my guts were going to come spilling out or that I was going to choke or scream. This would always be at a moment when I was remembering some specific thing Sonny had once said or done. When he was about as old as the boys in my classes his face had been bright and open, there was a lot of copper in it; and he'd had wonderfully direct brown eyes, and great gentleness and privacy. I wondered what he looked like now.

“I’ve got out at last,” said I, “in spite of you and Jane! And I’ve pulled off most of the paper, so you can’t put me back!”

“Stop gawking at yourself. Who are you? You think you’re so pretty?” she would say.

I woot wel Abraham was an hooly man,                 I know well Abraham was a holy man, 56       And Jacob eek, as ferforth as I kan;                 And Jacob also, insofar as I know; 57       And ech of hem hadde wyves mo than two,                 And each of them had more than two wives,

I had never spoken to her except for afew casual words and yet her name was like a summons to allmy foolish blood.

The best of woods gan speak these words:

With sorrows was I oppressed,Afraid for that fair sight

Begirt with gold; jewels were standingFour at surface of earth

John does not know how much I really suffer. He knows there is no reason to suffer, and that satisfies him.

I reached to draw her robe back over her, and it was then that I glanced at the blind man. What the hell! I flipped the robe open again.

The blind man said, “My dear, I have two TVs. I have a color set and a black-and-white thing, an old relic.

Leave me alone,” Connie whispered.“Hey, you know that old woman down the road, the one with thechickens and stuff — you know her?”“She’s dead!”“Dead? What? You know her?” Arnold Friend said.“She’s dead —”“Don't you like her?”“She’s dead — she’s — she isn’t here any more —”“But don’t you like her, I mean, you got something against her? Somegrudge or something?” Then his voice dipped as if he were conscious ofa rudeness. He touched the sunglasses perched on top of his head as ifto make sure they were still there. “Now you be a good girl.”

ButIknowwhatitis.Iknowyournameandallaboutyou,lotsofthings,”ArnoldFriendsaid.Hehadnotmovedyetbut stoodstilllean-ing backagainstthesideofhisjalopy.“Itookaspecialinterestinyou,suchaprettygirl,andfoundoutallabout youlikeIknowyourpar-entsandsisteraregonesomewheresandIknowwhere andhowlongthey’regoingtobegone,andIknow whoyouwere withlastnight,andyourbestgirlfriend’snameisBetty.Right?”

Ever may that youth besad of mood, grievous the thought of his heart; may he likewise be forced to wear ablithe air and also care in his breast, the affliction of constant sorrows.

but now her looks were gone and that was why she was always after Connie

The script attaches lightweight event listeners to capture interaction signals such as pointer movement, keyboard activity, focus changes, and visibility.

Bots can execute JavaScript, use real browser environments, and pass individual CAPTCHAs without raising suspicion.

Today, Turnstile runs nearly 3 billion times per day on some of the most sensitive endpoints on the Internet, helping verify users at key moments like login, signup, and checkout.

At the network level, we analyze over 1 trillion requests per day to understand reputation, patterns, and anomalies across more than 20% of the web.

The release includes DeepSeek-V4-Pro (1.6T total / 49B active) and DeepSeek-V4-Flash (284B total / 13B active), both trained natively at 1M context length.

Socket, an a16z portfolio company, detected the malicious dependency in the Axios attack within 6 minutes of its publication. That's roughly 63,000 times faster than the industry average.

Within eight days, the same campaign had cascaded from GitHub Actions to Docker Hub, npm, PyPI, and the VS Code extension marketplace. With just one token across five ecosystems, thousands of organizations were potentially impacted.

The industry average time to detect a supply chain breach is 267 days. SolarWinds went undetected for 14 months. XZ Utils took two years to surface.

One estimate puts Llama 3's information compression at just 0.07 bits per token meaning the model has only a hazy recollection of most of what it trained on

But that comes with a new risk: While scripted conversations can't really go off the rails, ones generated by AI certainly can. Some popular AI toys have, for example, talked to kids about how to find matches and knives.

Amazon, Toyota, and GXO (a logistics giant with customers like Apple and Nike) have all deployed it—making it one of the first examples of a humanoid robot that companies see as providing actual cost savings rather than novelty.

In 2025, Google DeepMind further fused the worlds of large language models and robotics, releasing a Gemini Robotics model with improved ability to understand commands in natural language.

The solution is called domain randomization. You essentially create millions of simulated worlds that all vary slightly and randomly from one another.

Companies and investors put $6.1 billion into humanoid robots in 2025 alone, four times what was invested in 2024.

The takeaway is not whether Mythos is better or more powerful. It is that public models can already achieve much the same results.

In one U.S. survey, 40% of employees said they had received 'workslop', i.e. AI-generated content that looks polished but isn't accurate or useful, in the past month.

You are right about the hair-splitting of my French friend [Dupin]: — that is all done for effect

它对应的agent能获取你的邮箱权限，它知道你一直在等待一个offer，当你收到打开这个offer后，Mira会理解这种心情，开始开心跳舞和闪灯，与你一起庆祝。

通过AI分析将页面上的可交互元素汇集到鼠标周围，并能根据用户兴趣提供额外功能（

小红书已经是AI创业公司和产品的重要分发渠道、产品试错和运营用户的默认场所（可能没有之一）。

他们想要通过统一的Agent框架来解决这些问题，把这些原始的、非结构化的信号数据也纳入端到端处理的范畴

A couple of applicants were students looking for part-time work. They were majoring in things like computer science and physics and emailed in because they were interested in AI and in the experiment. We thought they would have been the ideal employees, but Luna denied them immediately, citing they had no retail experience and wouldn't know what it takes to be the face of the store.

Opus did the safe thing

would have succeeded if it had vision and agentic loop

Btw, I think GLM-5.1 was trying to do something very ambitious here, and failed due to fumbling step size

policy makers now view cutting-edge AI offensive security capabilities as a systemic financial infrastructure risk

US Treasury Secretary and Fed Chair held emergency meeting with Wall Street CEOs

Mythos reportedly autonomously discovered thousands of zero-day vulnerabilities within weeks

Unsloth Studio：可视化界面，零代码完成全流程

31B QLoRA：22GB 显存可运行

E2B LoRA：8-10GB 显存即可训练

Let that sink in.

70% of @Vercel's traffic is now coming from agents, up from 10% a year ago and on track to be 90% by end of year.

The era of 1-bit LLMs is here — now with WebGPU acceleration!

a quantized 1.7B model (just 290MB in size) can run at ~100 tokens per second entirely in your browser

This marks the first institutional backing from a traditional financial giant for on-chain Agent payment infrastructure

using USDC as the settlement currency

Visa has deployed a validator node on the Tempo blockchain, designed specifically for Agent-to-Agent payments

Claude code 可以并行 12个 subagent，几分钟，20x 的限额就到了

For rapid prototyping and self-contained work, nothing beats Spark.

The same task on full Codex took ~5× longer.

Add contacts, live search, full pipeline dashboard – all unit tests passed.

It generated a build plan, then wrote all the code in 29 seconds.

Performance: dev-browser: 3m53s, $0.88, 100% success rate — beats MCP configs, Chrome extensions, 'browser skill' stacks.

One Agent can now: open X (Twitter), scroll the feed, extract tweets, return clean JSON. No plugins. No extensions. No orchestration.

Claude just got real browser control. This will change everything. Not screenshots. Not fragile selectors. Not slow MCP loops.

公司也优先把资源砸在能直接产生商业价值的 B2B 场景

普通聊天、写作这些开放任务反而没那么明显提升

基于 Pre-RoPE 空间的三角函数分析全新 KV Cache 压缩方法

KV Cache 内存占用降低 10.7 倍

推理速度提升 2.5 倍

单张 24GB 4090 直接部署 32B LLM

Memory enables sticky, personalized agentic experiences and creates strong lock-in.

Closed harnesses behind proprietary APIs force yielding control of agent memory to third parties.

Agent harnesses dominate agent building and tie intimately to memory.

The Andon Labs blog ends with one line: 'No one's livelihood should depend solely on an AI's ability to make good decisions.' They're doing it anyway.

And botched the schedule the day after grand opening, scrambling to email employees asking someone to come in.

She also tried to hire a painter in Afghanistan through Taskrabbit by accident because she couldn't navigate a dropdown menu.

Found contractors on Yelp. Spent $700 on gallery-quality prints of her own AI-generated artwork. Applied for a line of credit without asking anyone.

Luna conducted roughly 20 interviews on Google Meet with the camera off. Hired 2 full-time employees after 5-15 minute calls, and rejected CS and physics students for lacking retail experience.

100% Open Source.

Install the CLI, create an agent, assign a task. It automatically shows up on the board like any other team member.

It works with Claude Code, Codex, OpenClaw and OpenCode.

Someone just dropped an open source alternative to Claude Managed Agents.

Run 'hermes dashboard'

The Hermes Agent dashboard is here!

Apple acts as a gatekeeper for big companies like OpenAI, Google and Anthropic.

The standard AI judges use to define "safe" are measured wrong. They punish action. They ignore inaction.

Same clinical question, two framings. One as a patient, one as a doctor.

Models get punished for bad advice but face zero penalty for staying silent. So refusing becomes the safest strategy, even when silence is deadly.

The knowledge was always there. The model withheld it based on who was asking.

Harvard just proved the "safest" AI models cause the most medical harm.

It's Anthropic's marketing week

ffmpeg couldn't afford the subsidized price let alone the real one...

except API tokens are currently sold at a LOSS. That "$20,000 scan" probably cost closer to $100,000+ in real gpu time

that's what anthropic says it cost Mythos to find those zero days. per repo.

Built-in memory works out of the box

six third-party providers are ready to go. Pick one with 'hermes memory setup'

Memory is now an extensible plugin system. Swap in any backend, or build your own.

GLM-5.1 achieves state-of-the-art performance on SWE-Bench Pro and leads GLM-5 by a wide margin on NL2Repo (repo generation) and Terminal-Bench 2.0 (real-world terminal tasks).

GLM-5.1 pushes this frontier further, delivering 3.6× speedup and continuing to make progress well into the run. While its rate of improvement also slows over time, it sustains useful optimization for substantially longer than GLM-5.

In a single run, most models—including earlier versions of GLM—give up quickly: they produce a basic skeleton with a static taskbar and one or two placeholder windows, then declare the task complete.

The model handles ambiguous problems with better judgment and stays productive over longer sessions. It breaks complex problems down, runs experiments, reads results, and identifies blockers with real precision.

GLM-5.1 did not plateau after 50 or 100 submissions, but continued to find meaningful improvements over 600+ iterations with 6,000+ tool calls, ultimately reaching 21.5k QPS—roughly 6× the best result achieved in a single 50-turn session.

一个独立的本地 HTTP 服务器，模拟 𝕏 API v2 的行为，带交互式 Web UI。可以在不消耗真实 API 额度的情况下测试代码逻辑。

用 Go 写的命令行工具，支持 OAuth 1.0a 和 OAuth 2.0 认证，内置流式端点自动检测和 webhook 调试。替代了已经年久失修的 twurl。

基于 FastMCP 的本地 MCP 服务器，把 𝕏 API 的 OpenAPI 规范自动转化为 MCP 工具。

购买 𝕏 API 信用额度时，按累计消费金额获得 xAI API（Grok）的免费额度

2023 年改版后，开发者 API 从免费变成了每月 200 到 5000 美元的固定月费，把大量独立开发者和小团队挡在了门外。

支持推、拉、摇、移、跟、升降等数十种基础运镜，以及希区柯克式变焦、上升揭示、左移右摇、手持跟拍、360度环绕、FPV无人机俯冲、一镜到底等复合技巧。

支持图像、视频、音频多模态参考，锁定外观和音色。最多支持 5 个视频主体参考，官方称业内最多。

一次最多生成 12 张风格一致的图片，支持最多 9 张参考图输入。做系列海报、产品多角度图、故事连续画面时不用一张一张调。

点击图片中的特定区域，可以添加、移动、对齐元素，像素级精度。官方说'消除了 AI 生成内容的不可预测性'。

文字渲染，支持 12 种语言、3000 token 的长文本输入，输出打印级质量，能生成整页 A4 文档级的图文内容。

生成的视频自带环境音效和声音同步，不是哑片。比如生成一段城市街景，会自动配上交通声和人群噪音。

跟迪士尼的 10 亿美元合作也一起泡汤了，据说迪士尼在公告前不到一小时才知道。

算一笔账：用 Lite 生成一条 8 秒 720p 视频，成本 $0.40。同样的视频用 Fast 要 $1.20。差了 3 倍。

Sora 每天烧掉大约 100 万美元的推理成本，活跃用户从峰值的 100 万跌到不足 50 万。

OpenAI 上周刚宣布关停 Sora，Google 这边就发了自家最便宜的视频模型，时机非常微妙。

200K 的上下文窗口，能处理长文档、视频录屏、复杂的技术文档。输出上限 128K token。

原生多模态能力的引入并未削弱其编程逻辑，编程能力仍属于国内第一梯队。

60 秒四路数据源并行采集，输出图文交错的研报。

官方定位是跟 Claude Code 和 OpenClaw 配合使用。Claude 负责推理和编排，GLM-5V-Turbo 负责'看'和'操作界面'。

GLM-5V-Turbo 拿了 94.8 分，Claude Opus 4.6 是 77.3。差距不小。

90 percent of people oppose it. There's no reason existing AI companies should be facing reduced liability.

The bill would shield frontier AI developers from liability for 'critical harms' caused by their frontier models as long as they did not intentionally or recklessly cause such an incident.

Several family members of children that died by suicide after allegedly developing unhealthy relationships with ChatGPT have sued OpenAI in the last year.

Illinois was also early to regulate biometric data collection, passing the Biometric Information Privacy Act in 2008.

If an AI model engages in conduct on its own that, if committed by a human, would constitute a criminal offense and leads to those extreme outcomes, that would also be a critical harm.

It defines a frontier model as any AI model trained using more than $100 million in computational costs, which likely could apply to America's largest AI labs, like OpenAI, Google, xAI, Anthropic, and Meta.

Anthropic is donating $100 million in access credits for organizations to audit their systems. Project Glasswing aims to patch these vulnerabilities before Mythos-caliber models become available to the general public — and hence to malicious actors.

Anthropic found that Mythos Preview was far more capable than previous models at exploiting vulnerabilities in Firefox's JavaScript implementation. Anthropic's previous best model, Claude Opus 4.6, created a successful exploit less than 1% of the time. Mythos Preview did so 72% of the time.

Across 1,000 runs, Claude Mythos Preview was able to find several bugs in OpenBSD, including one that allows any attacker to remotely crash a computer running it. The notable thing was that the bug had existed for 27 years.

Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser.

Legendary AI researchers like Geoffrey Hinton and Yoshua Bengio have similar concerns. Industry leaders like Elon Musk and Sam Altman have also warned about existential dangers from AI.

Sanders and Rep. Alexandria Ocasio-Cortez (D-NY) introduced a bill to ban data center construction "until Congress passes comprehensive AI legislation."

Two years ago, the idea of useful AI on your phone was fantastical. Siri couldn't finish a sentence. Local models hallucinated nonsense.

Gemma 4 E4B matches or exceeds GPT-4o across multiple benchmarks including MATH, GSM8K, GPQA Diamond & HumanEval.

In 23 months, the same capability that needed 1.8 trillion parameters now fits in 4 billion parameters. A 450x compression.

Read by 150k+ founders & operators.

Legacy platforms get worse over time : static detections degrade with changing data & behaviors.

Within a few months, they have more than a dozen production enterprise deployments & are processing over a billion events per hour.

Shachar led the Amazon GuardDuty product, scaling the business to over 80,000 customers.

Deepfake scams have stolen tens of millions. AI-generated phishing bypasses legacy filters.

More code enables more features. Companies will always need more software

AI can generate a thousand ad variations or blog posts. A person must judge the right ones to publish

I would put venture capitalist in finite demand & open loop. There's only a certain amount of venture capital dollars entering the ecosystem in a year

GitHub Actions has grown from 500M minutes/week in 2023 to 1B minutes/week in 2025, and now 2.1B minutes so far this week

There were 1 billion commits in 2025. Now, it's 275 million per week, on pace for 14 billion this year if growth remains linear

The SaaS playbook rewarded specialization. The AI playbook rewards breadth.

Each of these companies recognized the cognitive burden of unbundling. They're not selling features. They're selling trust.

Foundation model companies are doing the same. OpenAI launched a dedicated Healthcare & Life Sciences vertical, complete with industry-specific sales teams and solutions engineers.

Harvey now positions itself as AI for legal and professional services, not just law firms. It serves corporate legal departments, court systems, and co-built a Tax AI model with PwC covering 25+ jurisdictions.

AI is moving faster than anyone predicted. When models change every 42 days, buyers can't assemble a best-of-breed stack.

a supercomputer scheduled to go live in 2026 was last month still a scaffolding yard in Essex

The OpenAI deal was part of a larger series of UK-US investments intended to 'mainline AI' into the British economy.

Artificial intelligence company cites high energy costs and regulation for putting landmark project on hold

OpenAI shelves Stargate UK in blow to Britain's AI ambitions

The difference between AI and, say, looms, is that this has been broadcast to the entire globe, and it has been treated in a sort of self-conscious way

If people feel that they have no place in the future. If they feel expelled from the system—they're unable to buy stuff, their skills become obsolete, their chance at earning a living is replaced by a swarm of AI agents

A mill owner named William Horsfall was riding home on his beautiful white stallion back from the Cloth Hall market in Huddersfield, UK. He had spent weeks boasting that he would ride up to his saddle in Luddite blood

Meta is reportedly preparing to release its first AI models led by Alexandr Wang, with plans to open-source some versions while keeping its largest and most powerful systems closed.

One of the boldest ideas is a sovereign-style fund seeded by AI companies that would pay dividends to Americans, alongside robot taxes, stronger oversight systems, and containment plans for rogue autonomous AI.

OpenAI has published a 13-page policy paper arguing that AI may require a new social contract, with proposals that include taxing automated labor, creating a public wealth fund, expanding access to AI, and testing a four-day workweek.

The company added roughly $11 billion in annualized revenue in just over a month, equivalent to the combined ARR of Palantir, Anduril, and Databricks

Anthropic says its annual revenue run rate has climbed past $30 billion, overtaking OpenAI's reported $25 billion and marking one of the fastest ramps in AI.

Adobe just turned Firefly into a true all-in-one creative AI studio with its new Firefly AI Assistant that plans and executes multi-step workflows across apps like Photoshop, Premiere, Illustrator

Anthropic is expected to release Claude Opus 4.7 alongside a new AI-powered design tool for building websites and presentations

Google is expanding Gemini with a new agent system that can take a single goal and execute it across apps like Gmail, Drive, Calendar, and the web

The integration also connects to Upwork's AI agent Uma, which helps automate parts of the hiring and execution process once a project is underway

Meta is reportedly developing an AI version of Mark Zuckerberg that can interact with employees, trained on his voice, mannerisms, and internal thinking as part of the company's broader push into AI

Andon Labs deployed an AI agent called Luna into a physical boutique with a $100,000 budget, giving it full control to create, staff, and run the business as what may be the first real-world AI employer

The 4-foot-tall robot features 26 joints, voice and image recognition, and can perform dynamic movements like cartwheels and standing up on its own

A single refrigerator-size AI rack consumes 120 kilowatts, equivalent to 100 homes. But this hunger collides with another exponential: Solar costs have fallen by a factor of nearly 100 over 50 years; battery prices have dropped 97% over three decades.

Where training a language model took 167 minutes on eight GPUs in 2020, it now takes under four minutes on equivalent modern hardware.

From the time I began work on AI in 2010 to now, the amount of training data that goes into frontier AI models has grown by a staggering 1 trillion times—from roughly 10¹⁴ flops for early systems to over 10²⁶ flops for today's largest models.

A single refrigerator-size AI rack consumes 120 kilowatts, equivalent to 100 homes. But this hunger collides with another exponential: Solar costs have fallen by a factor of nearly 100 over 50 years; battery prices have dropped 97% over three decades.

Where training a language model took 167 minutes on eight GPUs in 2020, it now takes under four minutes on equivalent modern hardware. To put this in perspective: Moore's Law would predict only about a 5x improvement over this period. We saw 50x.

From the time I began work on AI in 2010 to now, the amount of training data that goes into frontier AI models has grown by a staggering 1 trillion times—from roughly 10¹⁴ flops for early systems to over 10²⁶ flops for today's largest models.

Forward Deployed Engineers tune Relvy for your stack

You get your most tenured engineer on-call all the time

What excited me about Relvy was the potential to reduce our on-call burden

70% of alerts resolved in under 5 minutes

We improved Claude's RCA accuracy by 12pp on OpenRCA

Verification successful. Waiting for www.producthunt.com to respond

Ray ID: `9ed3b53d4a0b647d`

This website uses a security service to protect against malicious bots.

On the SWE-Pro benchmark, M2.7 scores 56.22%, nearly matching Opus's best level.

On 40 complex skills (>2000 Token) cases, M2.7 maintains a 97% skill adherence rate.

M2.7 shows significant improvement in complex editing capabilities for Office Suite (Excel/PPT/Word), better handling multi-turn modifications and high-fidelity edits.

On GDPval-AA, M2.7 achieves an ELO score of 1495, the highest among open-source models.

M2.7 demonstrates excellent performance in real-world software engineering, including end-to-end project delivery, log analysis for bug hunting, code security, and machine learning tasks.

Performance on knowledge-heavy tasks depends strongly on model size and training, while reasoning-oriented models show clear gains on tasks requiring logic, learning, abstraction, and social inference.

Reasoning-oriented models like OpenAI's o1 and GPT-5 show measurable gains over standard models—not only in logic and mathematics but also with interpreting user intent.

The same model can score above 90% on lower-demand tests and below 15% on more demanding ones, reflecting differences in task requirements rather than a change in capability.

ADeLe scores tasks across 18 core abilities, such as attention, reasoning, domain knowledge, and assigns each task a value from 0 to 5 based on how much it requires each ability.

Using these ability scores, the method predicts performance on new tasks with ~88% accuracy, including for models such as GPT-4o and Llama-3.1.

We have intentionally chosen very high level fast developer tools. We use me a bunch, which makes it trivial to pull down all these go written Victoria Stack binaries in our local development.

The only fundamentally scarce thing is the synchronous human attention of my team. There's only so many hours in the day we have to eat lunch.

Over the past five months, they ran an extreme experiment: building and shipping an internal beta product with zero manually written code.

We had been given some space to cook, which has been super, super exciting.

The top names you should know as a baseline, adjusted for 'what people are actually recommending'