AI points of the week: Free AI "siege" ChatGPT, SD-XL public beta strikes
ChatGPT in 2023 is like the "vanguard" of the AI era. Within a "month", people will have the real feeling of "AI attack". . However, the world of AI will never be short of challengers. This week, Google has made new moves, and several free "ChatGPT-like" AIs have emerged one after another, either performing well, or open source and aggressive, making the technology circle every day. There are new topics on the stage.
Claude's free open test is known as ChatGPT's "strongest competitor"
Contrary to ChatGPT's cumbersome registration threshold, Claude can be said to have removed the "door handle". Starting last weekend, as long as you have a proper online posture and sign up for a free Slack account, you can complete the registration in a few minutes and use this Claude artifact out of the box.
Although OpenAI's ChatGPT and GPT-4 are recognized as powerful, Claude has extraordinary potential even if he is not evenly matched. Tested by netizens, it can play back and forth with ChatGPT whether it is arithmetic, design code, or Chinese writing, idiom Solitaire. Coupled with the current "free" bonus, compared to some restrictions of ChatGPT, it is indeed the "strongest competitor" of ChatGPT worth trying.
SD developers open source large language models
Can StableLM reproduce the SD trend?
Stable Diffusion is currently one of the most mainstream and important AI painting models. The bigwigs in the open source community have created many interesting plug-ins and models based on it. Before ChatGPT became popular, some friends had experienced the artifact of AI first.
And its developer, Stability AI, dropped another bomb on the open source community on April 29 (Wednesday). The large language model StableLM is still an open source and transparent model, free to use and modify the code . That's right, like Stable Diffusion, you can configure it freely, or even customize StableLM for your needs. Although compared with the 175 billion parameters of the GPT-3 model, the StableLM model with a maximum of 7 billion at present and a maximum of 65 billion in the future is indeed much smaller, but with more powerful extended data set training and strong support from the open source community, it may still be possible. Will repeat the success of Stable Diffusion again.
Open Source LLM Open Assistant Announced
The ultimate goal is to run on consumer-grade hardware
This week, in addition to the "extraordinary" StableLM coming to the open source community, there is also an Open Assistant called "the world's largest ChatGPT open source replacement" also met with us on April 16 (last Sunday).
This model is fine-tuned based on Pythia and LLaMA, and trained from a large database manually generated and annotated by more than 13,000 volunteers. It performs well in arithmetic, reasoning, and telling jokes, although it is not as powerful as Claude, but The timely and appropriate answers also let people see the possibility of it competing with ChatGPT in the future.
Like StableLM, large language models born in the open source community can have unlimited possibilities, not only in terms of model customization, but also in the possibility of localized operation in the future. Open source LLMs such as Open Assistant have shown some hope .
SD-XL ushers in public beta
Can draw hands, can write, simple prompt can also have good results
Last weekend, Stability AI announced that Stable Diffusion XL, which is under development, was open to the public for free trial. At the same time, the official confirmed that the model will be open source once the parameters are stable.
Compared with the previous SD v2.1 and v1.5, the pictures generated by SD-XL are more aesthetically pleasing, the structure of the human body and hands is more reasonable, the portraits are more realistic, and the text in the pictures is clearer and more readable. In addition, it can be used The ability to produce high-quality images with shorter prompts (or even unfilled negative prompts) is also an exciting new improvement. However, SD-XL also reported that it may not be compatible with the previous version of the plug-in. This is what netizens do not want to see. After all, the current open source community has accumulated many excellent plug-ins, LoRa models, and compatibility It may be a major factor in determining whether SD-XL will become popular in the future.
Google merges DeepMind and Google Brain
Committed to the next world-changing breakthrough
On April 21 (Friday), Google announced that it will merge two world-class AI teams, Google Brain and DeepMind, to become a new department of Google DeepMind, with DeepMind CEO Demis Hassabis as the CEO of the new department.
In this regard, Google CEO Sundar Pichai and DeepMind CEO Demis Hassabis sent letters to employees, hoping that this new team that brings together leading research talents in the field of AI can use Google's computing resources to "unify the strengths" and significantly accelerate our progress in the field of AI. progress and make more breakthroughs in the field of artificial intelligence.
WPS AI announced
Build Chinese GPT Office
On April 18 (Tuesday), Kingsoft Office demonstrated the first ChatGPT-like application on the domestic collaborative office track - WPS AI, which will be embedded in all Kingsoft Office products in the future.
Some media took the lead in testing the effect of WPS AI on WPS light documents, and found that compared with the earlier released Notion AI, WPS AI understands Chinese better in more fields, whether it is planning outlines, expanding abbreviations or other writing needs , or planning copywriting, which is a "worker" characteristic manuscript, performed well.
As Jinshan executives said, the biggest advantage of Kingsoft Office is that it has a very thorough understanding of Chinese and user scenarios, and WPS AI also shows this very well. Such a high degree of localized AI tools may also be a major direction for the future development of domestic AI.
Adobe's "Firefly" AI is "brilliant" again
No difficulty in video creation, materials and scripts can be played
A month ago, Adobe announced the launch of Firefly, which was initially only used to generate copyright-free images, but from April 18 (Tuesday), Firefly began to attack the video field. Compared with "creating videos", Firefly's current focus is on making it easy for anyone to edit videos. You can adjust the color division with just a few words. You can tell Firefly directly what music and sound effects you want. Can't make a title? It can be used to create animations for opening titles of dynamic fonts and graphic logos.
In the future, Firefly also promises to automatically convert scripts into storyboards and preview images, and even recommend clips to help users create better videos, which will greatly accelerate the workflow of pre- and post-production.
It is believed that with the gradual evolution of Firefly, the future professional design industry may usher in a new era.
This week's AI application recommendation
Magic Copy: One-click cutout in Chrome/Firefox
Claude: A strong competitor of ChatGPT
X Studio 3: Unlock your AI singer
MiniGPT-4: GPT-4-like multimodal LLM
Godmode: AutoGPT with U easy to use
AgentGPT: Another AutoGPT Application Case
StableLM: An open source language model released by Stability AI
From WPS light document with AI document editing function, to Adobe Firefly that can help you optimize video, some "job bowl" industries that we often use at work have begun to pour into the embrace of AI. The times are changing really fast. ChatGPT, which was worried about "developing too fast" a few days ago, now has many new opponents ready to go. In the rapidly developing and ever-changing AI era, there may be one thing that remains unchanged, and that is people's exploration of the infinite possibilities of the future.
Because it's there.