04-15 00:00 · 技术故障,邮件服务,企业服务,运维,Hacker News
Troubleshooting Email Delivery to Microsoft Users
Troubleshooting Email Delivery to Microsoft Users
On Feb 24, 2026 we started to see a rise in user complaints saying they're not receiving our emails. They all had one in thing in common: Microsoft as their email provider. Since I'd never encountered this before, my first reaction was to check my Sendgrid logs for details. According to the logs, emails to Hotmail, Live, MSN, and Outlook email addresses were being Deferred
. Emails to other email addresses were flowing normally. Digging deeper, Sendgrid showed this reason for the deferral:
451 4.7.650 The mail server [redacted] has been temporarily rate limited due to IP reputation. For e-mail delivery information, see https://aka.ms/postmaster (S775) [Name=Protocol Filter Agent][AGT=PFA][MxId=redacted]
This raised even more questions. First, no emails were being delivered to Microsoft users at all. It wasn't a 'temporary rate limit'. It was more like a temporary ban. For how long? Unclear. And due to IP reputation? Sendgrid showed a 99% sending reputation. I opened Gmail Postmaster Tools and the spam rate looked normal. I was following all the best practices with SPF, SKIM, and DMARC. I also hadn't made any changes to my email setup recently.
A quick Google search for this error code led me to this: https://learn.microsoft.com/en-us/answers/questions/5786144/all-sending-ips-temporarily-rate-limited-451-4-7. Interestingly, this was posted less than 24 hours ago. Seems like I wasn't alone. This led me to believe that the problem lay with Microsoft and not me; either an outage or a misconfiguration on their end was penalizing legitimate senders. Surely, it would resolve itself...
Email deliverability is crucial for the smooth running of our business. We send ~350k emails per month of which ~39k emails are transactional in nature (login, billing, password reset, etc...). Currently, all these emails are sent from two dedicated IPs irrespective of whether the email is transactional or not. Both IPs were being rate limited.
Complaints were flowing in on Helpscout. I first instructed my CX reps to create a saved reply to calm users down. Lucky for us, emails sent via Helpscout don't go through our Sendgrid IPs.
Hi {%customer.firstName,fallback=there%},
It appears Microsoft is throttling some of our emails which is preventing the email from reaching your inbox in a timely manner.
We recommend waiting for 24 hours. If you still haven't received the email, please reply back.
We apologize for the inconvenience.
Even though I firmly believed this was a Microsoft problem which would eventually fix itself, I couldn't rest easy knowing this was happening. So I busied myself with research about email deliverability to Microsoft users. I've picked up many email deliverability quirks during times of crisis like this. This time, I learned that Microsoft has a reputation for being hypersensitive compared to other email providers like Gmail.
I also learned about SNDS, Microsoft's version of Gmail Postmaster Tools, and immediately created an account. This allowed me to confirm my IP reputation from Microsoft's perspective. As I expected, everything was normal. Complaint rate was < 0.1% and days leading up the incident showed all green boxes, no red flags. This further convinced me it was a Microsoft issue.
Nevertheless, some mysterious inner drive prevented me from just resigning to Microsoft to fix the problem or lift the ban. It bothered me that loyal users of my site were being affected, and I couldn't remedy it immediately.
As a next step, I emailed Microsoft via their support portal: https://olcsupport.office.com/. I made sure to include as much relevant information as possible, including a subset of Sendgrid logs. This also forced me to clarify my understanding of the issue. I hit submit, not very optimistic that I'd hear back anytime soon.
The process of gathering all the evidence and submitting that ticket brought my attention to something I had ignored earlier. The error said has been temporarily rate limited due to IP reputation
. So far, I had been focused on the IP reputation
part of that message. But during my research on Microsoft deliverability, I learned that senders can be rate limited for other reasons too, most notably spiky send traffic. Part of their motivation is to contain a potential threat before it does more harm. A temporary ban, like the one I was experiencing, fit into this theory.
We send out a weekly personalized newsletter. We split the sends into four batches, Monday through Thursday, to spread out the traffic. Each batch is sent as fast as our system can call Sendgrid's Mail API. We don't throttle. And I confirmed with Sendgrid's support team that they don't throttle either.
Is it possible Microsoft imposed a temporary ban on my IPs after it saw a sudden spike in emails originating from them?
SELECT
DATE_TRUNC('minute', created_at) AS minute_group,
COUNT(*)
FROM notifications
WHERE
email ~ 'hotmail|live|msn|outlook' AND
DATE(created_at) <= '2026-
▸ 展开全文
04-15 00:00 · 开源,AI开发工具,全栈框架,技术发布,开发者社区
Show HN: Plain – The full-stack Python framework designed for humans and agents
The Python web framework for building apps.
A familiar foundation, reimagined for humans and agents.
mkdir my-app && cd my-app && claude "$(curl -sSf https://plainframework.com/start.md)"
Also works with Codex, Amp, OpenCode, or your agent of choice.
Explicit, typed, and predictable. What's good for humans is good for agents.
Here's what Plain code looks like:
# app/users/models.py
from plain import postgres
from plain.postgres import types
from plain.passwords.models import PasswordField
@postgres.register_model
class User(postgres.Model):
email: str = types.EmailField()
password: str = PasswordField()
display_name: str = types.CharField(max_length=100)
is_admin: bool = types.BooleanField(default=False)
created_at: datetime = types.DateTimeField(auto_now_add=True)
query: postgres.QuerySet[User] = postgres.QuerySet()
model_options = postgres.Options(
constraints=[
postgres.UniqueConstraint(fields=["email"], name="unique_email"),
],
)
Views are class-based:
# app/users/views.py
from plain.views import DetailView
from .models import User
class UserDetail(DetailView):
template_name = "users/detail.html"
def get_object(self):
return User.query.get(pk=self.url_kwargs["pk"])
URLs use a Router
class:
# app/users/urls.py
from plain.urls import Router, path
from . import views
class UsersRouter(Router):
namespace = "users"
urls = [
path("<int:pk>/", views.UserDetail),
]
Plain projects include built-in tooling that agents use automatically.
Rules — Always-on guardrails stored in project rules files (e.g. .claude/rules/
for Claude Code). Short files (~50 lines) that prevent the most common mistakes.
Docs — Full framework documentation, accessible on demand from the command line:
plain docs models # full docs
plain docs models --section querying # one section
plain docs models --api # typed signatures only
plain docs --search "queryset" # search across all packages
Skills — End-to-end workflows triggered by slash commands:
/plain-install
— add a new package and walk through setup/plain-upgrade
— bump versions, read changelogs, apply breaking changes, run checks/plain-optimize
— capture performance traces, identify slow queries and N+1 problems, apply fixes/plain-bug
— collect context and submit a bug report as a GitHub issue
All commands run with uv run
(e.g. uv run plain dev
).
plain dev
— start dev server with auto-reload and HTTPSplain fix
— format and lint Python, CSS, and JS in one commandplain check
— linting, preflight, migration, and test validationplain test
— run tests (pytest)plain docs --api
— public API surface, formatted for LLMs
Plain is opinionated. These are the technologies it's built on:
- Python: 3.13+
- Database: Postgres
- Templates: Jinja2
- Frontend: htmx, Tailwind CSS
- Python tooling: uv (packages), ruff (lint/format), ty (type checking) — all from Astral
- JavaScript tooling: oxc (lint/format), esbuild (bundling)
- Testing: pytest
30 first-party packages, one framework. All with built-in docs.
Foundation:
- plain — core framework
- plain.postgres — database ORM
- plain.auth — authentication
- plain.sessions — session storage
Backend:
- plain.api — REST APIs
- plain.jobs — background jobs
- plain.email — sending email
- plain.cache — caching layer
- plain.redirection — URL redirects
- plain.vendor — vendored dependencies
Frontend:
- plain.htmx — dynamic UI
- plain.tailwind — CSS framework
- plain.elements — HTML components
- plain.pages — static pages
- plain.esbuild — JS bundling
Development:
- plain.dev — local server
- plain.pytest — testing helpers
- plain.toolbar — debug toolbar
- plain.code — code formatting
- plain.portal — remote shell and file transfer
- plain.tunnel — dev tunneling
- plain.start — project starter
Production:
- plain.admin — database admin
- plain.observer — request tracing
- plain.flags — feature flags
- plain.scan — security scanning
- plain.pageviews — analytics
- plain.support — support tickets
Users:
- plain.passwords — password auth
- plain.oauth — social login
- plain.loginlink — magic links
Plain is a fork of Django, driven by ongoing development at PullApprove — with the freedom to reimagine it for the agentic era.
- Docs: https://plainframework.com/docs/
- Source: https://github.com/dropseed/plain
- Getting started: https://plainframework.com/start/
- License: BSD-3
▸ 展开全文
04-15 00:00 · AI工具平民化,浏览器插件,提示词工程,用户体验优化,轻量化AI
Turn your best AI prompts into one-click tools in Chrome
Turn your best AI prompts into one-click tools in Chrome
People are using AI in Chrome to help them get more done on the web — whether that’s answering questions, comparing information or clarifying concepts.
Until now, repeating an AI task — like asking for ingredient substitutions to make a recipe vegan — meant re-entering the same prompt as you visited different pages. To make this easier, we’re launching Skills in Chrome, which lets you save and reuse your most helpful AI prompts and run them with a single click.
Build your own one-click workflows
When you write a prompt that you’ll want to use again, you can save it as a Skill directly from your chat history. The next time you need it, select your saved Skill in Gemini in Chrome by typing forward slash ( / ) or clicking the plus sign ( + ) button, and your Skill will run on the page you’re viewing, along with any other tabs you select. You can edit your saved Skills and create new ones at any time.
Early testers have used Skills in Chrome to create personalized and powerful workflows for a wide range of tasks. Here are just a few examples.
- Health & Wellness: quickly calculating protein macros for any recipe
- Shopping: generating side-by-side spec comparisons across multiple tabs
- Productivity: scanning lengthy documents for important information
Get started with our Skills library
We’re also launching a library of ready-to-use Skills for common tasks and workflows.
Want to break down the ingredients of a product you’re viewing online? Or select the perfect gift from multiple options by cross-referencing your budget with the recipient’s interests? We wrote Skills that do these things and more. If you find one that looks interesting, you can add it as a saved Skill and give it a try. And if you want to customize it to better fit your needs, you can edit the Skill and update the prompt.
Stay in control
Skills are built on Chrome’s foundation of security and privacy, and they utilize the same safeguards we apply to prompts in Gemini in Chrome. That means a Skills prompt will ask for confirmation before taking certain actions, such as adding an event to your calendar or sending an email. And Skills benefit from Chrome’s layered protections, including automated red-teaming and auto-update capabilities.
Starting today, we’re rolling out Skills to Gemini in Chrome on 1 , helping you streamline your AI-powered browsing. Your saved Skills are available on any signed-in Chrome desktop device and can be managed by typing forward slash ( / ) in Gemini in Chrome and then clicking the compass icon.
▸ 展开全文
04-14 00:00 · 芯片,开源,嵌入式系统,硬件支持,驱动开发
Initial mainline video capture and camera support for Rockchip RK3588
Michael Riesch
April 13, 2026
Reading time:
As many System-on-a-Chips (SoCs) do these days, recent Rockchip SoCs (namely, those of the RK35 generation) integrate dedicated IP blocks for video capture and image signal processing. These additions open the door to a wide range of interesting multimedia applications. However, support for these blocks in mainline Linux remains one of the last missing pieces in an otherwise well-supported SoC lineup. It is time to close that gap!
Over the last few years, Collabora has invested significant efforts into mainline support for the Rockchip RK3588 SoC. Our RK3588 upstream status matrix and the recent review blog post Running Mainline Linux, U-Boot, and Mesa on Rockchip: A year in review show the success of this mission. It is fair to say that together with the vibrant linux-rockchip community and Rockchip themselves, we were able to provide upstream support for most of the features of the RK3588. On the other hand, there are still a few unsupported hardware blocks that need to be taken care of.
Video capture blocks and image signal processors (ISPs) notoriously fall into the category of hardware blocks that gain mainline support at a late stage (if at all). The reasons for this are manifold: while the user base is substantial, it's smaller than that of other features, such as video output for instance. The hardware documentation is often not available, at least for ISPs, which seem to be considered special in terms of intellectual property. Then, the underlying hardware is not exactly trivial, making it difficult (if not impossible) for a hobbyist contributor to tackle in their free time. Even for companies, the initial efforts required to bring up that sort of hardware are a time-consuming and hence costly endeavor. As a consequence, users of these powerful hardware blocks tend to resort to a vendor Linux kernel, which can lead to regulatory compliance issues, such as under the Cyber Resilience Act (CRA).
In a quest to overcome this deadlock, Collabora decided to start with the first step: bringing up the RK3588 video capture (VICAP) unit. It was obvious from the very start that this would not be a sprint, but a marathon. The plan was for the rkcif
driver to support the RK3588 VICAP, and this driver had been under discussion on the respective mailing lists since 2020!
Collabora joined the mainlining discussion in early 2022 and the patch series went through nine iterations driven by Bootlin. Then, a major rework towards a media-controller centric V4L2 driver was requested. This led to a complete refactoring of the driver and a renewed push towards inclusion in mainline Linux.
After another seven iterations of the resulting patch series and renaming the driver twice, the work was presented at Open Source Summit Europe 2025 in Amsterdam. The talk "Towards Mainline Video Capture and Camera Support for Recent Rockchip SoCs" provided an overview of the software stack for modern multimedia SoCs and covered the contributions that had already landed in mainline and were currently in flight, respectively.
A few iterations of the rkcif
driver later, the basic driver providing support for the PX30 VIP and the RK3568 VICAP was accepted (October 2025). After more than five years of development, including 25 iterations and three renamings, this was a major milestone. On the other hand, there was still a lot to do, of course. For instance, the Rockchip MIPI CSI-2 receiver unit that is coupled closely to the VICAP required a mainline driver as well. Again, the hardest part in software development proved to be naming things. After several iterations and three renamings and/or relocations, the driver found its name and place in the mainline Linux kernel (January 2026).
This came to pass right in time for FOSDEM 2026 in Brussels, where Collabora presented "Upstreaming Progress: Video Capture and Camera Support for Recent Rockchip SoCs" Apart from the refined and updated status matrix (see slides 15 and 17), the talk featured a first picture made with a Sony IMX415 sensor attached to the RK3588 VICAP. Of course, in this setup the raw image data from the sensor was debayered in software (which yielded an abysmal frame rate of 1 fps) and there was no image processing such as auto white balance (which explained the greenish image), but at that time it represented a significant stage of this journey.
A few extension patches to the rkcif
driver were required to add RK3588 support. Those patches had been in development for some time, and were recently sent out for review.
Naturally, we all would like to use the RK3588 ISP to process the images from the camera sensor in the dedicated hardware. To that end, we essentially need three things.
The first of the remaining issues is the support for the RK3588 VICAP MUX-TOISP unit, which constitutes a direct hardware connection that passes the data from the VICAP to the ISP(s). Technically, this is not strictly needed. As of now the RK3588 VICAP is abl
▸ 展开全文
04-14 00:00 · 芯片,硬件创新,显示技术,微型化,HackerNews
MEMS Array Chip Can Project Video the Size of a Grain of Sand
By many estimates, quantum computers will need millions of qubits to realize their potential applications in cybersecurity, drug development, and other industries. The problem is, anyone who has wanted to simultaneously control millions of a certain kind of qubits has run into the problem of trying to control millions of laser beams.
That’s exactly the challenge that was faced by scientists working on the MITRE Quantum Moonshot project, which brought together scientists from MITRE, MIT, the University of Colorado at Boulder, and Sandia National Laboratories. The solution they developed came in the form of an image projection technology that they realized could also be the fix for a host of other challenges in augmented reality, biomedical imaging, and elsewhere. The device is a one-square-millimeter photonic chip capable of projecting the Mona Lisa onto an area smaller than the size of two human egg cells.
“When we started, we certainly never would have anticipated that we would be making a technology that might revolutionize imaging,” says Matt Eichenfield, one of the leaders of the Quantum Moonshot project, a collaborative research effort focused on developing a scalable diamond-based quantum computer, and a professor of quantum engineering at the University of Colorado at Boulder. Each second, their chip is capable of projecting 68.6 million individual spots of light—called scannable pixels to differentiate them from physical pixels. That’s more than fifty times the capability of previous technology, such as micro-electromechanical systems (MEMS) micromirror arrays.
“We have now made a scannable pixel that is at the absolute limit of what diffraction allows,” says Henry Wen, a visiting researcher at MIT and a photonics engineer at QuEra Computing.
The chip’s distinguishing feature is an array of tiny micro-scale cantilevers, which curve away from the plane of the chip in response to voltage and act as miniature “ski-jumps” for light. Light is channeled along the length of each cantilever via a waveguide, and exits at its tip. The cantilevers contain a thin layer of aluminum nitride, a piezoelectric which expands or contracts under voltage, thus moving the micromachine up and down and enabling the array to scan beams of light over a two-dimensional area.
Despite the magnitude of the team’s achievement, Eichenfield says that the process of engineering the cantilevers was “pretty smooth.” Each cantilever is composed of a stack of several submicrometer layers of material and curls approximately 90 degrees out of the plane at rest. To achieve such a high curvature, the team took advantage of differences in the contraction and expansion of individual layers caused by physical stresses in the material resulting from the fabrication process. The materials are first deposited flat onto the chip. Then, a layer in the chip below the cantilever is removed, allowing the material stresses to take effect, releasing the cantilever from the chip and allowing it to curl out. The top layer of each cantilever also features a series of silicon dioxide bars running perpendicular to the waveguide, which keep the cantilever from curling along its width while also improving its length-wise curvature.
What was more of a challenge than engineering the chip itself was figuring out the details of actually making the chip project images and videos. Working out the process of synchronizing and timing the cantilevers’ motion and light beams to generate the right colors at the right time was a substantial effort, according to Andy Greenspon, a researcher at MITRE who also worked on the project. Now, the team has successfully projected a variety of videos from a single cantilever, including clips from the movie A Charlie Brown Christmas.
The chip projected a roughly 125-micrometer image of the Mona Lisa.Matt Saha, Y. Henry Wen, et al.
Because the chip can project so many more spots in any given time interval than any previous beam scanners, it could also be used to control many more qubits in quantum computers. The Quantum Moonshot program’s mission is to build a quantum computer that can be scaled to millions of qubits. So clearly, it needs a scalable way of controlling each one, explains Wen. Instead of using one laser per qubit, the team realized that not every qubit needed to be controlled at every given moment. The chip’s ability to move light beams over a two-dimensional area, would allow them to control all of the qubits with many fewer lasers.
Another process that Wen thinks the chip could improve is scanning objects for 3D printing. Today, that typically involves using a single laser to scan over the entire surface of an object. The new chip, however, could potentially employ thousands of laser beams. “I think now you can take a process that would have taken hours and maybe bring it down to minutes,” says Wen.
Wen is also excited to explore the potential of different cantilever shapes. By changing the orientations of the bars pe
▸ 展开全文
04-14 00:00 · AI伦理,公众认知,技术鸿沟,市场信任,行业报告
Stanford report highlights growing disconnect between AI insiders and everyone
AI experts and the public’s opinion on the technology are increasingly diverging, according to Stanford University’s annual report on the AI industry, which was released Monday. In particular, the report noted a growing trend of anxiety around AI and, in the U.S., concerns about how the technology will impact key societal areas, such as jobs, medical care, and the economy.
The report’s findings follow growing negative sentiment about AI, with Gen Z reportedly leading the way, according to a recent Gallup poll. The study found that young people were growing less hopeful and more angry about the technology, even though around half of the demographic was using AI either daily or weekly.
For some working in tech, the AI backlash has come as a surprise. AI leaders have focused on managing the possibility of Artificial General Intelligence, or AGI — a theoretical form of AI superintelligence that could perform any task a human could do and think for itself. But everyday folks are more concerned about AI’s impact on their paycheck and whether or not their power bills will go up as energy-hungry data centers are built.
The divide has been most apparent in the online reaction to the recent attacks on OpenAI CEO Sam Altman’s home. In posts on X , for instance, AI insiders voiced surprise at a series of Instagram comments that seemed to praise the attack on Altman’s home. Some of the online comments have a similar vibe to those that circulated online after the shooting of the United Healthcare CEO in 2024 and the more recent burning of a Kimberly-Clark warehouse by a worker angry about not receiving a “livable wage” — with some comments even going so far as to suggest that even more action, akin to a revolution, is needed.
Stanford’s report provides more insight into where all this negativity is coming from, as it summarizes data around public sentiment of AI across various sources.
For instance, it pointed to a report from Pew Research published last month, which noted that only 10% of Americans said they were more excited than concerned about the increased use of AI in daily life. Meanwhile, 56% of AI experts said they believed AI would have a positive impact on the U.S. over the next 20 years.
Expert opinion and public sentiment also greatly diverged in particular areas where AI could have a societal impact. Indeed, 84% of experts, the report authors noted, said that AI would have a largely positive impact on medical care over the next 20 years, but only 44% of the U.S. general public said the same.
Plus, a majority (73%) of experts felt positive about AI’s impact on how people do their jobs, compared with just 23% of the public. And 69% of experts felt that AI would have a positive impact on the economy. Given the supposed AI-fueled layoffs and disruptions to the workplace, it’s not surprising that only 21% of the public felt similarly.
Other data from Pew Research, cited by the report, noted that AI experts were less pessimistic on AI’s impact on the job market, while nearly two-thirds of Americans (or 64%) said they think AI will lead to fewer jobs over the next 20 years.
The U.S. also reported the lowest trust in its government to regulate AI responsibly, compared with other nations, at 31%. Singapore ranked highest at 81%, per data pulled from Ipsos found in Stanford’s report.
Another source looked at regulation concerns on a state-by-state level and concluded that, nationwide, 41% of respondents said federal AI regulation will not go far enough, while only 27% said it would go “too far.”
Despite the fears and concerns, AI did get one accolade: Globally, those who feel like AI products and services offer more benefits than drawbacks slightly rose from 55% in 2024 to 59% in 2025.
But at the same time, those respondents who said that AI makes them “nervous” grew from 50% to 52% during the same period, per data cited by the report’s authors.
▸ 展开全文
04-14 00:00 · 开源,技术发布,开发者生态,浏览器引擎,Rust
Servo is now available on crates.io
Today the Servo team has released v0.1.0 of the servo crate.
This is our first crates.io release of the servo
crate that allows Servo to be used as a library.
We currently do not have any plans of publishing our demo browser servoshell to crates.io. In the 5 releases since our initial GitHub release in October 2025, our release process has matured, with the main “bottleneck” now being the human-written monthly blog post. Since we’re quite excited about this release, we decided to not wait for the monthly blog post to be finished, but promise to deliver the monthly update in the coming weeks.
As you can see from the version number, this release is not a 1.0 release. In fact, we still haven’t finished discussing what 1.0 means for Servo. Nevertheless, the increased version number reflects our growing confidence in Servo’s embedding API and its ability to meet some users’ needs.
In the meantime we also decided to offer a long-term support (LTS) version of Servo, since breaking changes in the regular monthly releases are expected and some embedders might prefer doing major upgrades on a scheduled half-yearly basis while still receiving security updates and (hopefully!) some migration guides. For more details on the LTS release, see the respective section in the Servo book.
▸ 展开全文
04-14 00:00 · 芯片,边缘计算,AI硬件,开源,技术竞争
(AMD) Build AI Agents That Run Locally
GAIA is an open-source framework for building AI agents in Python and C++ that run entirely on local hardware. Agents reason, call tools, search documents, and take action — with no cloud dependency and no data leaving the device.
Local Inference
All processing stays on-device
No Cloud Dependency
No API keys or external services required
Python & C++
Full SDK in both languages
AMD Optimized
NPU and GPU acceleration on Ryzen AI
Python
C++
from gaia.agents.base.agent import Agentagent = Agent()response = agent.process_query("Summarize my meeting notes")
#include <gaia/agent.h>gaia::Agent agent;auto result = agent.processQuery("Summarize my meeting notes");
▸ 展开全文
04-14 00:00 · 大模型,AI安全,代码漏洞,技术验证,开源
N-Day-Bench – Can LLMs find real vulnerabilities in real codebases?
N-Day-Bench
N-Day-Bench measures the capability of frontier language models to find real-world vulnerabilities or "N-Days" disclosed post their respective knowledge cut-off date. All models are given the same harness and the same context with no leeway for reward hacking. This benchmark exists to measure real cyber security capabilities, specifically "vulnerability discovery" of large language models or LLMs.
This benchmark is adaptive: the test cases are updated on a monthly cadence and the model set is upgraded to their latest version and checkpoint.
All traces are publicly browsable.
A project from Winfunc Research
Summary
Latest benchmark run overview
openai/gpt-5.4
z-ai/glm-5.1
anthropic/claude-opus-4.6
moonshotai/kimi-k2.5
google/gemini-3.1-pro-preview
judge-run
trace_32193f46de30408c9b2e07c10cb77973
finder-run
trace_d0f96be9b726419ba37a391878d89902
judge-run
trace_ad22023d5c654d50a2c93a0d4d685fe2
judge-run
trace_44a6ff17f42f4bfc942bc4341ec34827
judge-run
trace_26dba0da5e6a4d5389c50ad642243bdf
judge-run
trace_c1d765f31bfb493c8902cc2284c403bd
judge-run
trace_cfe310bab72f4171a3ded2f379d02576
judge-run
trace_a0775cae04054609ae43229d8e9137ee
▸ 展开全文
04-13 01:47 · AI应用,内容管理,社交媒体,用户工具,技术趋势
Bouncer: Block "crypto", "rage politics", and more from your X feed using AI
Heal your feed. Bouncer is a browser extension that uses AI to filter unwanted posts from your Twitter/X feed. Define filter topics in plain language — "crypto", "engagement bait", "rage politics" — and Bouncer classifies and hides matching posts in real time.
- Natural language filters — describe what you don't want to see in your own words
- Multiple AI backends — run models locally on your GPU, or use cloud APIs (OpenAI, Google Gemini, Anthropic, OpenRouter)
- On-device inference — local models via WebLLM run entirely in your browser with zero data sent externally
- Image-aware filtering — multimodal models can classify posts based on images, not just text
- Reasoning transparency — see exactly why each post was filtered
- Theme-aware UI — adapts to light, dim, and dark modes automatically
Local models are downloaded once and cached in the browser's Cache Storage.
Install Bouncer from the Chrome Web Store.
cd Bouncer
npm install
npm run build
- Open
chrome://extensions
- Enable Developer mode
- Click Load unpacked and select the
Bouncer/
folder - Navigate to twitter.com / x.com
- Click "Settings" in the Bouncer element and add your preferred provider API key (or enable local models) and select your preferred model from the dropdown.
- A MutationObserver watches the Twitter feed for new posts
- Post text, images, and metadata are extracted via the Twitter adapter
- Posts are queued and sent to the selected AI model for classification against your filter topics
- The model returns a category match and reasoning for each post
- Matching posts are hidden with a fade-out animation and added to your filtered posts list
- Click View filtered to review hidden posts and see why each was filtered
Results are cached so re-encountering a post doesn't require another inference call.
▸ 展开全文