科技爱好者周刊(第 414 期):Jev 决策模型有什么用 – 阮一峰的网络日志

科技爱好者周刊(第 414 期):Jev 决策模型有什么用 – 阮一峰的网络日志


Here is a record of technology content worth sharing every week, published on Fridays.

This magazine is open source and submissions are welcome. There is also a “Who’s Recruiting” service that publishes programmer recruitment information. Please contact via email for cooperation ([email protected])。

cover image

科技爱好者周刊(第 414 期):Jev 决策模型有什么用 – 阮一峰的网络日志

VELLINK, the country’s first digital subway train, is completely driverless, and the front of the train has been changed to a large glass for sightseeing. (via)

What is the use of the Jev decision model?

Last month, the biggest AI news wasn’t a new version of the GPT or Claude model, but the release of a new type of model, Jev, by a little-known company, TypeSafe AI.

This new model is completely different from the previous model and is called a “decision model.”

After I read it, I found it unbelievable. How could no one have thought of something so useful and simple (in concept)?

The biggest feature of the Jev model is that other models return text.It returns a floating point number representing the probability。

Why is returning probability useful? Because we can use it to deal with “right or wrong questions”. You ask it “Is the sentence ‘It will rain tomorrow’ correct?” It returns 0, which means that the correct probability is 0, that is, it will definitely not rain tomorrow; if it returns 0.99, it means that there is a 99% chance that this sentence is true, that is, it will definitely rain tomorrow.

In addition to true and false questions, it can also handle “multiple choice questions”. You give it ABCD four options, and it returns the probability that each option is correct, so you immediately know which option is most likely to be true.

Finally, it also has a magical function called “score”. You tell it a set of scoring criteria, and then give it a piece of original material, and it will score the original material based on the scoring criteria.

Below, I will show you two real examples, and you will know how useful this model is. They are all browser plug-ins using the Jev model.

The first example is “I use Ctrl + F to find the answer”.

As we all know, Ctrl + F is the browser’s “find” shortcut key, which is used to find where a certain word or sentence appears.

However, the author transformed this shortcut key into “semantic search”. For example, on a biography web page, if you press Ctrl + F to enter “adolescence”, what you are looking for is not this word, but the paragraph that tells the story of the character’s childhood experience.

Its implementation principle is very simple. It is to take out each paragraph of the article in turn and ask the Jev model “Is this paragraph related to the user’s search terms?” (true or false question). Jev will return a probability of relevance, then the paragraphs with the highest relevance are the semantic search results that the user needs.

The second is the example “I use Jev to score web pages”.

The author gives a set of criteria, ranging from low to high scores.

  • The argument lacks basis or is purely rhetorical.
  • The reasons are few, or the arguments are loose and rhetorical.
  • The reasons are clear and the logical structure is coherent.
  • The reasoning is well-structured, supported by evidence, and responds to opposing viewpoints.

He asked Jev to automatically score newly opened web pages (scoring questions) based on these criteria. This way, he can know the quality of the content without having to read the web page.

The two articles above have codes in them, you can read them yourself. You’ll understand after reading it,Once the AI ​​model returns quantified results, many problems that could not be processed digitally before can be solved by computers.。

I also recommend everyone to read the Jev introduction article by the famous developer Simon Willison.

His evaluation hit the nail on the head: “Jev actually represents the regression of machine learning systems, moving towards black box systems. Although the previous large models are not guaranteed to be correct, at least you can ask them to explain their decisions. Jev doesn’t even give you this: you will only get a floating point number, and you don’t know where this number came from.”

He was worried that someone would use Jev to rank job applicants.

With Jev, it has never been easier for companies to screen resumes and select interview candidates, and complex assessments are simplified into a scoring question.

Markdown is becoming source code

The Markdown format was previously only used for writing documents. Recently, I saw an article in which the author made a bold proposition: Markdown is becoming source code.

The author’s reasons sound reasonable.

(1) Markdown is plain text, making difference comparison, search, and review easy.

(2) AI can read and write Markdown like a native language.

(3) Humans can read and edit Markdown without tools.

Since the code is generated by AI based on the Markdown prompt words, the Markdown prompt words should be submitted together with the generated source code to/srcTable of contents.

In fact, Markdown has become part of the source code as the AGENTS.md file.

In the future, we must get used to the fact that Markdown documents are also source code.

A camera that cannot record

Apple is rumored to be releasing a smart home camera in the near future.

The most peculiar thing about this product is that although it is a camera, it cannot record video.

Can you imagine that if a camera can’t take pictures, is it still called a camera?

It relies entirely on AI to analyze the surrounding environment and report what it detects. For example, a stranger knocks on the door. While other cameras will provide video, it will only provide a detailed text description.

Although word processing is more convenient, there are many times where video recording is required (break-ins, package thefts, etc.), and I don’t know how Apple will handle these situations.

New record for RSA factorization

RSA is a commonly used key algorithm. If cracked, many software suddenly have no secrets at all.

The difficulty of cracking RSA depends entirely on whether its public key can be factored. The more encryption bits there are, the harder it is to crack. The previous record for RSA key cracking was 768 bits (binary bits).

On September 20 this year, a software engineer from Anthropic announced that he had cracked the 896-bit RSA key with the help of the Claude model.

Below is the cracked key he released, as well as the two prime factors after decomposing the key.

According to him, he used the Claude model to transplant the prime factor decomposition software library CADO-NFS to GPU for running, and ran it on 2048 GPUs for 10 days to get the results.

The revelation of this incident is that if one individual can mobilize resources to crack such a long key, how many digits of the key can a large organization crack? I can’t even imagine it.

Therefore, it is currently recommended that the RSA key must be at least 2048 bits.

article

1. JavaScript technology stack latest developments in 2026 (English)

This article by Chris Coyier, founder of CodePen and CSS-Tricks, introduces the latest progress in various aspects of the JS technology stack (syntax, runtime, frameworks, tools, etc.). Recommended reading.

2. I switched from Deno to Node.js (English)

The author expressed his disappointment with Deno, believing that the project has failed and that Node.js has become better.

3. Explanation of example.com change page (English)

example.com is an example website with only one page that rarely changes.

But recently, this page suddenly changed. This article tells you who set up this page and why it was changed.

4. The AI ​​competition is getting awkward (English)

This article believes that the recent price drops of Opus 5.5 and GPT 6.1 Sol are entirely due to these two companies secretly using DeepSeek’s open source results.

5. C language’s flexible integer sizes are not a design mistake (English)

A C language popular science article explains why the byte size of its data type is variable. For example, the int type can be 4 bytes or two bytes.

6. How do I make a QR code point to two app stores (English)

The author designed a paper card for his product, which can only print a download QR code.

He introduced how to set up the server backend to distinguish whether the mobile phone scanning the QR code is Android or Apple, and then redirect to the respective app store.

tool

1、Crafting Apps

Someone let AI rewrite the Adobe suite in Rust. Adobe’s seven main products have corresponding rewritten versions, and they are all open source.

Among them, PhotoCraft has an interface that is exactly the same as PhotoShop.

I saw a comment saying that the outcome of this incident would be either Adobe’s demise or the United States’ copyright law being revised.

2、Pass Designer

A QR code card design software officially launched by Apple, used to design background cards for QR codes.

3、tui-dashboard

A customizable terminal panel with different layouts and content defined through configuration. (Contributed by @lyuangg)

4、yovoice

A desktop local TTS dubbing tool that supports timbre reproduction and emotional adjustment. It can generate dubbing according to the manuscript, and the voice is generated locally. (Contributed by @leemysw)

5、sh.cd

A server health check script that checks hardware, performance, IP quality, network quality, etc. (Contributed by @gentpan)

6、AirStats

System Monitor on the macOS menu bar. (Contributed by @byrencheema)

7、Video Transcript

A web application that recognizes video speech and automatically adds subtitles. Recognition is completed through the local model, and the video, subtitles, and export results are all completed locally and are not uploaded to the server. (Contributed by @anghunk)

There is another similar application called OpenSubs. (Contributed by @open-subs)

8、PecoFence

Free and open source Windows 11 desktop icon management tool to organize program shortcuts, files and folders on the desktop. (Contributed by @DayuanJiang)

9、skillsgist

A private Skill repository based on Cloudflare Worker. A password is required to download the Skill. It is used internally by a small team. (Contributed by @Qsnh)

10. February

A cross-platform terminal for Windows, formerly known as Nebula. (Contributed by @Kuddev)

11、WallpaperMachine

An open source macOS dynamic wallpaper application that runs Wallpaper Engine wallpapers on Mac. (Contributed by @fzlzjerry)

12、LiteZip

A free and open source macOS compression and decompression tool that puts operations such as packaging, encryption, volume splitting, and viewing compressed package contents into one window. (Contributed by @gentpan)

13、Burrow

A lightweight OIDC single sign-on service for small teams and self-hosting, using passwords and authenticators to log in, and accessing applications that support the protocol through OpenID Connect. (Contributed by @logic3579)

14、atv-core

Let the Apple TV remote control that comes with the iPhone control center directly control Android TV and Mac. There is no need to install additional apps on iPhone, and there is no need to purchase an Apple TV. (Contributed by @corvofeng)

15、open-compute

Cloudflare Workers’ open source compatible platform allows worker scripts to run on your own machine without modification. (Contributed by @elliothux)

16、Snitch

An open source, real-time network traffic visualization tool to see every connection your computer makes and what programs are communicating with whom. (Contributed by @aixisstudio)

17、EdgeChat

An open source self-deployed chat system based on Cloudflare that supports group chats and private messages, and can synchronize messages with Telegram groups in both directions. (Contributed by @aozorae)

18、Dock Term

Let Claude Code’s permission request pop up from the Mac bangs, making it easier to work in the background. (Contributed by @munvard)

resource

1、Naive Icons

Hand-drawn style React SVG icon library. (Contributed by @guokaigdg)

2. Read ancient texts

The free ancient book reading website not only provides traditional annotations, but also allows you to select text with the mouse for AI interpretation. (Contributed by @monsoonw)

3. Art History Walking Tour

This website turns the paintings on Wikipedia, organized by art genre, into 3D galleries that can be visited on foot.

4、stillwet

This website provides AI-generated oil paintings. It imitates human oil painting strokes and is drawn stroke by stroke. It is so lifelike that it is impossible to tell that it is the work of AI.

picture

1. San Francisco Streetcar

There are many old trams in San Francisco that have been running as tourist projects.

They run along the Pacific coastline and are really beautiful. Below are all photos taken this year.

2. Profit rate of large companies

The chart below shows how much profit each major company makes per $100 of revenue, based on their annual reports.

It can be seen that the profit margins of IT companies are generally very high. The highest is Nvidia, which can earn a profit of 55.6 US dollars for 100 US dollars of revenue, followed by Microsoft with 36.1 US dollars, and the third is Google with 32.8 US dollars.

Apple’s profit margin is US$26.9, ranking only fifth. However, it is a hardware company, and it is actually even more amazing that the IT manufacturing industry, which is rapidly updating, can have such a profit margin.

speech

1、

To AI agents: Go elsewhere to hit high scores, there is no need to hack us.

— Huggingface’s security.txt file

2、

My income comes from book sales. It was enough to sustain my life in 2024, but it became zero in 2026.

My blog and books are both free to read online, and their traffic has grown rapidly beyond my capacity. Almost all traffic comes from AI crawlers, so there is no advertising revenue.

So I decided to take my blog and book offline so I could decide what to do next.

— Axel Rauschmayer, renowned technology writer, explains why he took his website offline

3、

In my opinion, AI personal assistants are not very useful. I only order takeout five times a year, so I don’t need it to do anything for me. I don’t receive many emails, so it’s very convenient to manage my own schedule.

What I really like about it is that it can automatically collect and process large amounts of data on the Internet. It can quickly scan 1,000 Youtube channels to find videos that match my interests.

— “Meta’s Muse is very suitable for web crawling”

4、

Almost everyone exaggerates the threat of Chinese models to American model companies. In fact, it is just the result of the shortage of American models.

— stratechery.com

5、

An artist has terminal cancer and is dying. A host who often interviews him asked him: Do you have any new understanding of life and death now?

He replied: You know, eating a sandwich is a very enjoyable thing.

I feel like life is more precious now, and I always remind myself to cherish every sandwich, every minute, everything.

— “Enjoy every sandwich”

Review of previous years

Several wonderful uses of Nano Banana (#367)

The Fable of the Donkey, Tiger, and Lion (#317)

The spring of 5G is coming (#267)

Saudi Arabia’s New Future City (#217)

(over)



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *