Claude-3.5 has released a version update, which not only significantly improves coding capabilities: this morning, a bug caused by OpenAI's o1 model in Cursor was fixed by the new Claude-3.5 model in one go.
Most eye-catching is the new experimental feature Claude introduced with this release: Computer Use. As the name suggests, it uses the model to operate a computer.
Of course, for user experience and security reasons, the new model does not directly operate our actual desktop; instead, it creates a new virtual machine and completes all operations within it. Some colleagues tested it immediately and released a video.
[One-Take] Claude 3.5 Computer Use Feature Test Official Account: NB Lab
Live test: Claude's 'computer use' controlling a computer to download research papers
To be honest, this feature is indeed something many "programmers," including myself, have long awaited. From the implementation path, it is very similar to Rabbit R1's LAM feature. However, because the Rabbit R1 is a physical piece of hardware you hold in your hand, it feels a bit cooler.
Here is a brief evaluation of Claude's new feature:
- It is long-awaited, and in my view, models should ideally implement such functions;
- Processing speed is still very slow because it requires constantly sending screenshots to the cloud model for processing—with so many steps, slowness is natural;
- Looking at the process, the model's logic and reasoning abilities are fully present, but accuracy needs improvement;
- To improve accuracy, training the model is an important foundation, but increasing image resolution and using higher-precision models on the inference side might be even more critical;
- So, where is the problem? It lies in model "inference": it needs higher precision, faster speeds, and lower costs.
How to achieve this? For leading model developers like Anthropic and OpenAI, the answer lies in NVIDIA's latest GPUs. I have explained this before: How is the 30x increase in GB200 inference performance achieved?
Through the NVL-72 configuration, inference performance increases by at least 30 times. Combined with software optimization for inference, the cost for the same model can easily drop to one percent of what it was before.
Today, as model usage rates rise and new features are introduced, the demand for such inference is plainly visible. Therefore, it is logical that demand for NVIDIA's next-generation GPU (Blackwell) continues to soar. Jensen Huang's statement that more and more computing power is being used for inference aligns with the current situation.
Thus, NVIDIA GPUs remain in high demand, but the construction methods are starting to change. This part involves shifts in investment opportunities that are not suitable for discussion on this public account.
On the other hand, whether it's the earlier Rabbit R1, mobile AI from Google and Apple, Microsoft's AIPC, or now Claude-3.5's computer use—and even the humanoid robots confirmed for mass production in the coming years—the trend toward "Edge" AI will be stronger than the Cloud.
With higher precision, faster speed, and lower cost, the real competition for edge inference chips has just begun. Naturally, the process will be much more interesting than the cloud market where NVIDIA dominates alone.
In terms of edge chips, Apple is the undisputed overall number one. I have written more than one article about this, such as my piece during the release of Apple's M3 generation, so I won't repeat it here.
The latest news comes from the lawsuit between Qualcomm and ARM.
This litigation has actually been ongoing for several years, originating from Qualcomm's acquisition of the chip company Nuvia (which has deep ties to Apple's M-series chips). Qualcomm's newly released X-Elite chips for AIPC are essentially derived from that company's designs. These chips are based on the ARM architecture, similar to Qualcomm's Snapdragon mobile chips. However, Qualcomm "cleverly saved" a large amount in ARM patent licensing fees.
Now, ARM is expanding this case. Their real intention is to target the AIPC chip market, which has massive potential. There have been rumors that ARM will release its own PC-based SoC. Although the market performance of Qualcomm's X-Elite has been slightly below expectations, it has at least taken the lead on the Windows system. This is very unfavorable for ARM's long-term chip strategy. Therefore, in my view, this matter may not be settled quickly.
This is just one side reflecting the intensity of the competition. However, the competition as AI enters the "edge" is only just beginning. The best is yet to come, and plenty of opportunities lie ahead.
Compared to the increasingly boring "Cloud," I am full of expectations for the "Edge."