Since the release of DeepSeek-V3 in late December, many friends have asked for my evaluation. We have seriously discussed training costs and the impact on computing power.
To be honest, while China's public equity fund industry increasingly resembles the "men's national football team" in the eyes of the public, the case of DeepSeek proves beyond doubt that this industry still concentrates the smartest and most diligent group of people. I believe practitioners in many of our industries are a combination of "talent + hard work," which includes many geniuses in domestic AI teams like DeepSeek, Kimi, Doubao, and Zhipu who are working incredibly hard.
However, I rarely publicly evaluate domestic models for two reasons: I dislike flashy demonstrations of incorrect usage, and I dislike the public opinion atmosphere that either praises them to high heaven or tramples them ruthlessly.
Yet, even the seemingly professional evaluations recently are, in my view, highly misleading. The biggest misconception is: AI is meant to be used. It's hard to imagine users paying for models styled like "acrostic poems" or "forum keyboard warriors," but they will pay for models that truly improve productivity and stably solve practical problems.
As it happens, I decided to take out my Leica MP, which has been in the dry box for a long time, to rediscover some of the photographic joy that has rapidly vanished over the past two years. Since I haven't used a film camera in a long while, I naturally needed to check the manual. The manual on the Leica website is 120 pages in German and English. After quickly browsing the English version, I had a sudden whim to ask the models to write a Cheat Sheet.
This scenario is actually very practical and represents one of the most important changes brought about by Large Language Models.
I tested DeepSeek, GPT, Gemini, and Claude simultaneously. The link to the camera documentation is:
Leica MP Manual, PDF format, over 10MB
https://leica-camera.com/sites/default/files/pm-73015-Leica-MP_Instructions_de-en.pdf
First, it should be noted that among the models tested, GPT-o1 does not support PDF input, so I used GPT-4o instead (in my view, the difference is minimal; I will explain later why so-called "thinking models" are undoubtedly an important research direction but have limited practical value at present).
I tested five models: DeepSeek-R1, GPT-4o, Gemini-1.5-Flash, Gemini-2.0-Flash, and Claude-3.5-Sonnet.
(If you want to skip directly to the conclusion, scroll to the end.)
First up is DeepSeek-R1: My prompt was simple: "Please produce a user-friendly cheat sheet on how to use the camera based on the document content."
I believe the "complete thinking process" shown in the image above has become a widely known feature of R1.
Leica MP Quick Operation Guide
📷 Master the classic experience of film photography
Basic Setup
- Battery Installation
- Use 2x 1.55V button cells (e.g., SR44) or 1x 3V lithium battery (DL 1/3N).
- Unscrew the battery compartment cover (2) counter-clockwise, insert batteries with the positive terminal facing up, and tighten clockwise.
- Tip: Remove batteries if not used for long periods to avoid leakage.
- Film Installation
- Opening: Lift the bottom latch (28) → Turn left → Remove bottom cover → Unfold back cover.
- Loading: Insert the film lead into the take-up spool (30), press lightly to ensure it's secure → Close back cover → Reinstall bottom cover and lock.
- Advancing: Toggle the rapid wind lever (19) → Release shutter → Repeat until counter shows "1".
- ISO Setting
- Turn the ISO dial (25) to align with the film's ISO value (range 6-6400).
Shooting Operations
- Lens Installation
- Mounting: Align the red dot on the lens with the red dot on the body → Rotate clockwise until it clicks.
- Dismounting: Press the lens release button (3) → Rotate counter-clockwise until dots align → Remove.
- Note: Avoid direct sunlight on the shutter curtain when changing lenses.
- Focusing and Rangefinding
- Viewfinder Display: The central bright frame is the rangefinding area. LEDs indicate exposure (▶◀ needs adjustment, ● correct exposure). The battery icon lights up when replacement is needed.
- Focusing Methods:
- Double Image (Portraits): Turn focus ring until double images coincide.
- Split Image (Architecture): Align split lines until they are continuous.
- Exposure Settings
- Shutter Speed (16): 1s-1/1000s, B-bulb for long exposures (metering off).
- Aperture Ring (13): Rotate to match shutter speed (LEDs indicate direction).
- Metering: Lightly press shutter to activation point; turns off automatically after 14 seconds.
Advanced Features
- Flash Usage
- Sync Speed: 1/50s (⚡ symbol).
- Connection: Hot shoe (20) or X-contact (22); supports multi-flash sync.
- Viewfinder Expansion
- Frame Selector (9): Toggle to select 28+90mm, 35+135mm, or 50+75mm frame lines.
- Magnifier (1.25x): Increases focusing accuracy (optional accessory).
- Depth of Field Preview
- Aperture affects depth of field; refer to the DoF scale (11b) for adjustments.
Maintenance and Tips
- Rewinding Film
- If advancing jams → Toggle rewind button (5) → Pull out rewind lever (10a/b) and rotate clockwise until film is fully rewound → Open cover to remove film.
- Cleaning and Storage
- Lens: Use a soft brush or dedicated cloth; avoid chemical reagents.
- Storage: Dry environment; fire the shutter periodically to prevent lubricant solidification.
- Important Warnings
- Do not let sunlight shine directly into the lens (may burn the shutter curtain).
- Use with caution in high temperature/humidity; desiccant recommended.
Appendix: Quick Memo Table
| Function | Operation Location | Key Tip |
|---|---|---|
| Shutter Speed | Top Dial (16) | B-bulb turns off metering |
| Aperture Adjustment | Lens Ring (13) | Lower value = more light |
| ISO Setting | Rear Dial (25) | Align with white dot |
| Rewind Film | Bottom Rewind Lever (10a/b) | Toggle R release button (5) first |
🎞️ Enjoy the unique charm of every frame; Leica MP records time with you.
The text above looks a bit messy because it uses Markdown syntax, which doesn't always pair well with WeChat accounts. However, on DeepSeek's own page, it looks like this:
Evaluation: This result is quite good and error-free. However, the "thinking process" is a bit confusing to me; it looks like a prompt generation process on the surface, but behind it likely lies an RL-based Agent independent of the base model. "Overthinking" is a real phenomenon where the base model's (V3?) understanding of the prompt seems limited compared to the complexity of the thinking process.
The second model is GPT-4o. It switched directly to Canvas mode, which offers a better user experience. The output was also in Markdown.
Leica MP Quick User Guide (Cheat Sheet)
Evaluation: Frankly, there were more details.
The third set includes Gemini 1.5 Flash and 2.0 Flash. Since the release of Gemini 1.5, AI Studio has been my most important production platform. Gemini is the most reliable productivity tool.
Evaluation: If you look closely, Gemini's output is better than both R1 and GPT-4o: 1. It removed references to part numbers like "battery cover (2)." While subjective—numbers are helpful with a diagram—in this text-only context, Gemini made a discerning choice. 2. Gemini adjusted the order to "Loading - Shooting - Rewinding," which follows the standard photography workflow, making it more user-friendly.
The last one is Claude 3.5.
Evaluation: Objectively, the output is informative and accurate, yet it feels "unfriendly." Claude 3.5's greatest strength has always been code generation, but in other fields, it seems a gap is opening up compared to other models.
Summary:
DeepSeek-R1's output is genuinely good. Its significance lies in being a model comparable to the best while having much lower API costs and being "open source." For all users, it's an additional "free" model without access restrictions. For "pioneer users," local deployment and calling leading APIs go hand-in-hand.
Finally, this simple experiment, especially with R1's clearly displayed "thinking process," makes me doubt the true value of so-called "thinking models." I believe reinforcement learning is a path with more certainty, but in the absence of "memory" and "self-correction," how significant is pure thinking?
Is the path of thinking models truly correct?
Daoming, Public Account: Daoming Digital Nomad Life