- youtube ad blocking
- Xinhua Research Institute releases report on domestic large-scale models
- Ministry of Commerce Network
- Multimodal Visual Language Model: NeVA
youtube ad blocking#
YouTube recently tested stricter anti-ad blocking measures. In June, when YouTube detects that a user is blocking ads, it will display a prompt reminding the user to enable ad blocking. In this case, the user can only watch up to three videos. In other words, either watch ads or keep seeing pop-ups.
Creators provide content, platforms provide infrastructure, and users provide traffic and consumption. This business model is relatively healthy, and YouTube's advertising strategy is already considered very user-friendly. However, some people still find ads troublesome and come up with various ways to block them, turning it into a game between users blocking ads and platforms countering ad blocking.
I personally recommend the model of carpooling in the low-price area, which allows users to enjoy the benefits of being a legitimate user at a low price.
Xinhua Research Institute releases report on domestic large-scale models#
As can be seen, Xinhua Research Institute has conducted a systematic examination and scoring of several mainstream large-scale models in China. From the announced examination and evaluation methods, combined with my own usage experience, it is relatively consistent with their actual capabilities.
However, we must face several issues:
- There is a duplication of efforts in developing large-scale models. The mainstream algorithms of large-scale models have already been open-sourced to a considerable extent. Personally, I believe that the core differences are not significant. The main differences lie in some algorithms, the input data, and the amount of training. Under the background of the US-China technological blockade, GPU computing power is actually a scarce resource. ChatGPT consumes $700,000 daily. It may not be a good phenomenon for domestic technology companies to rush into this field.
- There are limitations in application scenarios. In fact, for many application scenarios, there is no need for such a large parameter size as ChatGPT. For example, as mentioned in the previous issue, Xiaomi's Xiao Ai Assistant may only need a parameter size in the range of tens of billions to hundreds of billions, which is sufficient for mobile phone applications. My personal opinion is that in the case of limited data, it is more appropriate to explore as many application scenarios as possible. Only by actually applying it can we discover and train for specific problems. This business model can also contribute to the investment in training to some extent.
Ministry of Commerce Network#
The State Council has issued opinions on further optimizing the foreign investment environment and increasing efforts to attract foreign investment:
- Support foreign investment in establishing research and development centers in China to undertake major scientific research projects.
- Encourage foreign-invested enterprises to conduct clinical trials of cell and gene therapy drugs that have been listed overseas in accordance with the law.
- Prudently increase the opening up of value-added telecommunications services such as domestic Internet virtual private network (VPN) business (with foreign shareholding not exceeding 50%) in pilot areas.
This is a small section about the network in the entire "Opinions," which indicates further opening up of legal access to foreign networks by enterprises. We can see that the economy is under great pressure, even if we don't understand it. From the relaxation of the network, we can see that efforts are being made to stabilize the economy.
For users who need to bypass the Great Firewall, the biggest benefit may be the availability of more dedicated line nodes.
Multimodal Visual Language Model: NeVA#
NVIDIA has released the NVIDIA AI Playground, where you can directly interact with NeVA in your browser by uploading an image and asking questions related to the image.
The recognition rate is quite good as an online general model. It is already very good to have this recognition rate. Personally, I think local deployment and training for specific scenarios are still promising. This is also a direction for the development of large-scale image models. After all, not everyone wants to do mid-journey. Combining with specific industries is a path with good development prospects.