I connected the fruit fly's brain to Othello.
The kids at home have recently become obsessed with Othello. Just last week, I saw some news about fruit flies (maybe one of the hottest topics lately besides thin-skinned issues?...) and found that the fruit flies were performing quite well playing games, so I had a sudden idea over the weekend to let the fruit flies try out Othello, a game that requires a certain level of memory and decision-making.
The results from the experiment were a mix of joy and worry: they were obviously better than random moves, but could not match the control group that didn't use the fruit fly’s brain.
Chess strength evaluation:
- Random moves: 0.275
- Disrupted fruit fly neural connections: 0.4472
- Real fruit fly neural connections: 0.4578
- Completely without fruit fly: 0.5122
(The values represent the consistency with the decisions made by the decision tree engine.)
The difference between real and disrupted was only 0.011, p = 0.347, statistically indistinguishable; while the completely non-fruit-fly-controlled group was significantly stronger instead.
Is it because I didn't train the fruit fly?
No. Firstly, the neural connection structure of the fruit fly has already been determined, so I initially did not make any modifications/disruptions to avoid creating "neural chaos" (the disrupted neural connections mentioned above were primarily for the control experiment).
The principle of "training" mainly involves shooting arrows first and then drawing the target.
First, I took inputs from the fruit fly's visual relay layer to observe subsequent neuron activation patterns, to see how these activations could correspond to placing pieces, aiming for a better game state (matching to a depth 3 search engine).
After all, it is not hard to understand that it is unlikely to expect a fruit fly to directly manipulate the pieces with its front legs or click on the board; it can only rely on "translating" neuron activities to correspond to the board.
An unexpected finding from the training: the fruit fly's eyes are completely useless.
The initial design was to render the chessboard onto the fruit fly's compound eyes and drive the photoreceptors, which was the most biologically faithful approach. But this ended up completely non-functional. Not a single downstream neuron discharged, even when the driving intensity was increased fourfold.
The reason is that the fruit fly's photoreceptors and early visual neurons use "graded potentials," which do not fire action potentials at all, whereas the whole brain model assumes all neurons do. The signal dies in the medulla.
This actually aligns with biology: a fruit fly's brain does not need a pixel map; it only needs a rough signal (for example, something approaching from the left front). Feeding it detailed chessboard data is actually information that this evolved biological architecture is designed to discard.
However, demos like Flappy Bird that only require "avoiding obstacles" might be different. Such neuronal pathways do exist in fruit flies as shortcuts: approach-detection neurons connect directly to jump-command neurons, bypassing the central brain, completing the process in about 30 milliseconds. So when you try to swat a fly, it can evade you.
Fruit fly trading might actually be feasible, for instance, issuing an order immediately upon seeing a news item or an outstanding interest event.
---
However, none of this prevents the kids from having a great time playing.
Another thought: this truly is a great era; as long as you have an idea, you can quickly create a weekend project/demo using AI without the hesitation of the past: Is it really worth spending my weekend free time on this kind of pointless thing?
Now, there’s no need to consider the energy cost; as long as you’re willing to spend tokens, it suffices.
Code: https://github.com/hulatown/flyothello
Application: https://flyothello.vercel.app/
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
