Welcome to read the Open Source Daily produced by OSCHINA editorial department, which is updated once a day.
# 2024.4.30
Today's key points
Just last night, the entire AI community was shocked by a mysterious big model: it is called gpt2 chatbot, and its performance directly surpasses many open source models and GPT-4! Netizens speculated that it was GPT-4.5, GPT-5, or GPT-4+Q *, or GPT-2+Q *. Altman also made a point: "I really like gpt-2gpt2."
Ubuntu 24.04 LTS just now release The next version of Ubuntu 24.10 has also recently been identified with the code - Oracle Oriole. Ubuntu 24. There is a high probability that Linux 6.11 kernel will be used GNOME 47 Desktop GCC 14.1 Compiler and other upgrades.
An open source model is not popular. It depends on how quickly the products in the ecosystem support it.
On April 26, the open source started again after a thousand words of misunderstanding, and the Wang Fried model Qwen1.5-110B with 110 billion parameters was directly thrown out, which set a new high in the performance of the open source model. The model has been released for less than 24 hours, Ollama quickly launched its support for 110B. This means that you can deploy the model to your own computer as soon as it is released, in addition to paying for demos on the Motta community and HuggingFace.
On the day of release, Qwen1.5-110B has occupied the top of the Hacker News popularity list for a period of time. The last time there was so much heat and discussion, it was last August when Tongyi Qianwen first announced open source. However, the direction of discussion has changed from "what is this?" to "how strong is this?". The noise of doubt gradually dissipates with the strength of Qwen.
Observation today
Social observation
A mysterious model suddenly entered the public's view, and its ability exceeded that of many open source models, including GPT-4. Almost everyone is talking about it, and the server is crowded out.
It is "gpt2 chatbot". (Note that it is gpt2, not GPT-2)
How strong is it? The question of the IMO International Mathematical Olympiad was answered correctly at one time.
In GPT-4's landmark ability "Painting Unicorn", it can easily kill LLaMA-3-70B.
In terms of reasoning, it has even more amazing effect. Common logic traps can be easily bypassed, and the answer also has an appropriate tone.
Such a good performance, but I didn't say my identity clearly
This is not more exciting for everyone!
Is it the legendary GPT-4.5?
-Weibo qubit
-Weibo Stage1st residential community
The open source model has opened the mode of large volume and special volume.
Globally, on both sides of the Pacific, the two male pattern is emerging.
Llama 3 China Cup has just been unveiled, and the domestic Tongyi Qianwen has directly opened a hundred billion level parameter model Qwen1.5-110B, which has become the top one of Hacker News.
Compared with its 72 billion parameter model, the performance of the MMLU C-Eval, HumanEval and other benchmark tests, Qwen1.5-110B has returned to the SOTA open source model throne, surpassing Llama 370B and becoming the strongest open source model.
It is worth noting that this is the eighth big model of Tongyi Qianwen Open Source in three months.
-Weibo qubit
Microsoft Edge used another Google's Chromium open source engine, and also used its position as the dominant platform to run against Chrome. No wonder Firefox had to develop its own system platform at that time. At least Google had Android and Google Chrome OS, and the future was reserved for itself. Similarly, this statement can also be extended to Huawei's Hongmeng system. Having its own system platform can really be a lot tougher.
-Weibo Derek
Seen from the demand side, supply side and capital market of the SaaS industry, the current SaaS industry is at a low point.
On the demand side, affected by the macro-economy, enterprises have more or less reduced their spending on digital transformation, SaaS enterprises face the risk of slowing down the growth of new customers and reducing the order volume of existing customers.
Capital market, The investment and financing environment of SaaS industry is becoming more mature and rational. Investors have shifted their focus from growth rate to profitability and sustainability of enterprises, and are more inclined to invest in enterprises that have proved to have a clear profit roadmap in the market.
So far, The latest financing news of SaaS market comes from Weimeng. Recently, Weimeng Group released three announcements that all conditions under its new share placement agreement have been reached, and the company successfully completed the placement, with a net amount of about HK $308 million; In addition, Weimeng has successfully issued 85 million dollars of convertible bonds. With the combined support of investors held by cb in the previous 21 years, Tencent, the company's existing shareholder, and new investors, Weimeng Group raised a total of 125 million dollars.
Supply side, SaaS enterprises feel the cold wave of the industry and gradually return to the essence of value. "Cost reduction and efficiency increase" has become a strategy widely recognized by SaaS enterprises, but different enterprises have gone out of different ways to reduce costs and improve efficiency. Some SaaS enterprises only focus on short-term interests, and take the broadsword of cost reduction and efficiency increase to "layoff", "salary reduction" and "business shutdown". Some SaaS enterprises focus on long-term value and take multiple measures to reduce costs and increase efficiency, and MicroUnion is one of them.
According to the annual report of Weimeng Group, Weimeng Group achieved a total revenue of 2.228 billion yuan in 2023, up 21.1% year on year. The gross profit reached 1.484 billion yuan, and the gross profit margin increased to 66.6%. Adjusted EBITDA narrowed significantly to - 0.75 billion yuan, with a year-on-year decrease of 93.1%; The adjusted net loss decreased by 73.4% year on year.
On the same track, different choices will have different circumstances. It seems that how to reduce costs and increase efficiency in high quality is also a knowledge. The example of Weimeng is worth considering.
- Liu Kuang
1、 WinMerge is an open source difference comparison and merge tool running under Windows. The software is downloaded from the official website (winmerge. org). After installation, it can be opened to support Chinese. It can compare folders and files, support multiple formats such as text, pictures, tables, web pages, binary files, and present differences in a visual way, which is very easy to understand and process.
2、 Diffchecker：www.diffchecker.com， Only two files need to be uploaded for document comparison. It supports text, images PDF, Excel and other formats. The Pro version is required for the comparison folder.
3、 Meld：meldmerge.org， It is an open source cross platform document comparison software, applicable to Windows Linux, macOS and other operating systems support the comparison of text files, folders and image files.
4、 KDiff3：kdiff3.sourceforge.net， It is also a cross platform document comparison software, which also supports the comparison of text files, folders and image files, and is suitable for a variety of operating systems.
5、 ExamDiff：prestosoft.com， Lets you compare folders, files, and text content. The overall effect is not as good as WinMerge, so it can be reserved for this purpose.
6. Cloud library tool: libkit.cn/compare, which can compare PDF and Word. In addition to providing document comparisons, you can also convert file formats and CAD previews.
-Weibo Uncle Ban
Media observation
Looking at Meta With the new round of financial report performance of Alphabet and Microsoft, it is not difficult to see that generative AI has penetrated into all aspects of the business pipeline and become the core engine driving enterprise growth.
The success of Microsoft is particularly typical. With the forward-looking layout of multiple AI software and hardware, Microsoft is firmly in the top position, and all businesses take off at a high speed.
Although Meta's huge AI investment makes investors jumpy, whether it can achieve commercial profits as soon as possible is still unknown, but the advertising revenue boosted by AI has achieved initial results. With Llama's big model and computing power, Zuckerberg's determination and ambition to build a global AI leader cannot be underestimated.
As for the big brother Google, it finally found its own rhythm in the AI war. Gemini, TPU and other weapons have been sharpened and are ready to accelerate the search for new growth points in the AI wave.
The three giants' judgments on the development direction of artificial intelligence are surprisingly consistent, and they all offer real money and silver to invest in this future technological change. Standing in the wind of the times, As far as AI can take them, the global technology industry will look here.
- Silurian
In the first quarter of this year, the added value of Shenzhen's industries above designated size increased by 11.5% year-on-year, 7.0 percentage points higher than the same period last year. The added value of manufacturing above designated size increased by 11.8% year on year, and the added value of high-tech manufacturing increased by 13.1% year on year. As the country's "first industrial city", it is not easy to achieve this achievement under a large base.
This is due to Shenzhen firmly grasping the key task of new industrialization, constantly consolidating the market body, industrial investment and industrial ecological foundation, constantly improving the industrial chain, gathering innovative resources, shaping new advantages, injecting new vitality, pioneering new tracks, continuously climbing to the top of the global industrial chain value chain, and drawing the "second growth curve".
-Shenzhen Special Zone Daily
"There is no contradiction between the technical school and the market school. They consider problems at different levels of the innovation chain. We can't just do technical research, nor can we just focus on market realization. These two things complement each other." On April 26, during the 2024 Zhongguancun Forum "Hard Technology Investment and Development Forum", Wang Shaolan, president of Zhipu AI, received an exclusive interview from reporters of the 21st Century Economic Report, Talking about the views on this current hot topic. He believes that the whole chain of innovation is from ideas, theories, methods, technologies to products and markets. The market school pays attention to the products and markets at the back end, while the technology school pays attention to the theories, methods and technologies at the front end. In fact, this is something at different levels of the whole chain, and there is no one or the other relationship.
However, some companies will choose to focus most of their energy on products and markets, while others will focus on the whole chain and attach importance to the whole chain innovation of theories, methods, technologies, products and markets. Different companies will make different choices from their own perspective.
- 21st Century Economic Report
Recommended today
Open source project
https://github.com/immersive-translate/immersive-translate
Immersive web page bilingual translation extension is completely free of charge. It supports multiple translation services such as Deepl/Google/Tencent/Volcano Translation, Firefox/Chrome/Oil Monkey script, and can also be used on iOS Safari.
A daily blog
The current SOTA performance on LLM quantization is basically based on the weight only quantization mode. The reduction of video memory required for the model to run in the GPU is its main contribution.
on the whole, Quantification work in the LLM field is still very preliminary. If the performance accuracy of the model is required to be very high in actual tasks, algorithms and tools to improve the unit video memory throughput based solely on KV cache and other directions are recommended, such as Flash Attention-2 Paged Attention, etc.
Event comments
Recently, Google's response to Flutter The layoff of Dart, Python and other key teams was widely spread on social media. In this regard, Google will use the external media TechCrunch confirms The company has indeed made layoffs, but did not disclose the specific team, role and number of layoffs.
However, Google clarified that this layoff is not a layoff across the company, but a reorganization in the normal course of business, and the affected employees will be able to apply for other vacant positions at Google.
comment
This layoff event may reflect some broader trends within the technology industry, including the pursuit of efficiency and cost-effectiveness, as well as the reassessment of specific technologies or product lines. At the same time, it is also part of Google's adjustment of its resource allocation and business priorities in the changing market environment.
The layoff event reflects the rapid change and volatility of the technology industry. With the continuous evolution of market and technology, companies need to flexibly adjust their resource allocation and business priorities.
As a leader in the technology industry, Google's decisions and strategic adjustments are closely watched by the market. This layoff may affect investors' views on Google's future development and strategic direction.
Google presentation It said that in 2023, the company prevented 2.28 million Android applications violating the policy from being launched on Google Play. And 333000 Google Play accounts that uploaded malware, fraudulent applications or repeatedly violated the policy were found and blocked.
In contrast, in 2022, Google banned 1.5 million "bad" applications and 173000 developer accounts that seriously violated store policies.
comment
This figure reflects the complex challenges faced by the mobile application market in terms of security, regulation and user trust. With the popularity of mobile devices, the mobile application market is growing rapidly, but it also brings regulatory and security challenges. Google needs to constantly update its review mechanism to deal with emerging threats and violations. While protecting users' security, Google also needs to ensure that users can easily access the applications they need. This event shows the importance of finding a balance between security and user convenience.
It is an ongoing challenge for developers to comply with Google Play's policies and standards. Google's updating and strengthening of policy requirements means that developers need to constantly learn and adapt. Although Google has taken strict security measures on Google Play, third-party app stores may not have similar security guarantees. This may cause users to face higher security risks.
With the development of technology, developers of malware and illegal applications are also constantly looking for new ways to avoid censorship. Google must constantly update its security measures to stay ahead.
OpenAI announce We have reached cooperation with the Financial Times to enable its large language model to gain access to articles in the Financial Times.
As part of the cooperation, the two companies will provide ChatGPT users with part of the content of the Financial Times. OpenAI means that, Through this cooperation, ChatGPT users will be able to see "Selected abstracts, quotes and links to Financial Times news reports To respond to related queries ”。 However, OpenAI did not specify when the content can be accessed or in which versions of ChatGPT.
comment
This cooperation not only provides OpenAI users with richer content sources, but also marks an important milestone in the application of AI in the news field. Through cooperation with the Financial Times, OpenAI can provide more accurate and relevant information retrieval services, and the content of the Financial Times has therefore been more widely disseminated and utilized.
It may also have an impact on the business model of the news industry. As more and more media companies cooperate with AI companies, the production and distribution of news content may change. This cooperation may bring new sources of income to news organizations, and also provide AI companies with higher quality data sets to train their models.
And may promote the competition of OpenAI in the field of artificial intelligence. More similar partnerships may emerge in the future to promote the further integration and development of AI and the news industry.
Voice of Open Source
Media opinion
The development of general intelligence is faced with three bottlenecks: technical, biological and social. The technical bottleneck is that AI systems need to have higher computing power, more advanced algorithms and more effective data processing methods to achieve more complex and intelligent functions; The biological bottleneck is mainly reflected in our limited understanding of the operating mechanism of human brain cognitive ability. To achieve a similar level of intelligence, we need more in-depth neuroscience and cognitive research; Social bottlenecks include the integration of artificial intelligence systems and human society, such as cultural differences, ethics, privacy protection, etc. "Overcoming these bottlenecks requires interdisciplinary cooperation and continuous innovation efforts. Only by making breakthroughs in technology, biology and society can general intelligence move towards a more mature and comprehensive development, but it is extremely difficult to achieve."
- Global Times
At the beginning of 2024, overseas Internet giants began to realize the commercialization potential of AI business.
In the past week, Google, Microsoft Meta has successively released the first quarter financial report of 2024, and the three have mixed feelings. After the release of Microsoft's and Google's financial reports, they rose by 4% and 15% respectively, and the latter reached a record high; In contrast, Meta suffered a failure. After the release of the financial report, the stock price once fell by 19%.
- Geek Park
With the arrival of the era of big model, Zhou Yahui, the famous "vent catcher" in the capital circle, has no surprise to include AI in his own territory. The share price of Kunlun Wanwei under his control soared in 2023, once creating a "myth" of quadrupling the price. However, under the uproar, it is not easy for Kunlun Wanwei to continue to dig gold in the big model market. In the first quarter of 2024, the net profit attributable to the parent company of Kunlun Wanwei dropped 188% year on year, the worst record in nearly a decade.
- city boundary
Zhang Yiming, who is now in the process of byte hopping and has almost retired, only focuses on two things: first, TikTok in the storm; The second is the AI strategic business that byte beating is trying to catch up with.
Referring to the AI strategic vision of bytes, many people close to bytes believe that Flow is the most "orthodox" product. "Although many sub businesses are doing related things, such as flying books, the management level still believes that the AGI war is dominated by Flow".
- Phoenix Network Technology
User perspective
- Viewpoint 1: Combine the Flutter and Compose teams, and work on Compose, Flutter's drawing engine can also be migrated to Compose
- Viewpoint 2: This is the most reliable
- Viewpoint 3: Flutter should not go astray and use a dart
- Viewpoint 4: The language that was originally half a foot into the coffin was dragged out, but now the prospect becomes unclear. I don't know what the people at Google think.
- Viewpoint 5: Because it is controllable, Dart has his own say. If ts is not so fast and hot, Dart still had a chance, but lost to ts
- Viewpoint 6: "Flutter EngProd team has been cut off. Cutting firewood really has your talent"/
- Viewpoint 7: In fact, most of the recruitment work of the Flutter headquarters team has stopped two years ago, but Flutter and Dart will not die
- Viewpoint 8: It is better to acquire uniapp
- Viewpoint 1: Although France is the mother community of OCCT, the core developers of OCCT on GitHub are all Russians. Without Russians, the mother community of France cannot continue to operate. So Huawei took over, moved to China, changed its name and resumed open source and community operations. What's the problem? Usually, they seldom praise Huawei, but they do a good job.
- Viewpoint 2: Huazi is the pearl on the crown of the computer industry
- Viewpoint 3: If you understand the open source world, please remain modest, otherwise you will expose your ignorance, and you will jump like a clown in front of the public without knowing it.
- Viewpoint 4: This is just leak detection, not self research
- Viewpoint 5: What's your spiritual cleanliness? Although you take over halfway, the team spends money to keep full-time code writing. Reasonably, you can switch from closed source to commercial use for selling your own products
- Viewpoint 6: Huawei: It's better to come early than to come at the right time. Thanks for the assistance from Laotie
- Viewpoint 7: I have been developing with OCCT in the past two years. There are few data, poor Boolean operation performance, and lack of some key algorithms. I need to do it myself. If Huawei can solve these problems, China's industrial software will develop more smoothly
- Viewpoint 8: This is Niubai
- Viewpoint 1: Complete language does not exist
- Viewpoint 2: It is necessary to deal with possible errors, and there is no problem at the language level.
- Viewpoint 3: There is a problem with this proposition. A person who is unwilling to deal with errors will always ignore if there is a way to ignore it. This increase will make the neglect more complex and the code looks more complex. _ is already the best solution
- Viewpoint 4: Simple and effective solution: check with inspection tools or compiler attributes
---END---
Finally, welcome to scan the code to download "Open Source China APP", read massive technical reports, and share with programmers and geeks!
