压缩率是AGI的智商衡量标准[2][3][4]。

<aside> 💡 Proof of Work of Compression (PoWoC)

</aside>

压缩-例

定义

$t=$ 时间

$c(t)=$ 随时间$t$变化的foundation model的被推至的最好压缩率(space saving)函数

$\Delta t = const$ (推荐设置为 1天~1周)

$t_i=t_{i-1} + \Delta t$

$\Delta c_i = c(t_i) - c(t_{i-1})$

PoWoC 代币激励模型

每隔时间$\Delta t$ 固定产生1个coin,即在每个时间节点$t_i$固定产生1个coin,这1个coin的分配方案如下:

c[t_] := 1 - (1/Ceiling[(t + 0.5)*3] + 0.85);
d = 0.1;
e = 4.5
Show[
 Plot[c[t], {t, 1, e}, PlotRange -> {0, c[e]}, Filling -> None, 
  AxesLabel -> {t, c}, Ticks -> {Range[2, 10], Range[0, c[10], 0.02]},
   PlotLabel -> "PoWoC"],
 Plot[c[t], {t, 2, 3}, PlotRange -> {0, c[e]}, Filling -> 0],
 Plot[c[t], {t, 2, 3}, PlotRange -> {0, c[e]}, Filling -> c[2], 
  FillingStyle -> Directive[Yellow, Opacity[0.2]]],
 Plot[c[t], {t, 2.1 + 0/3, 2.1 + 0/3 + d}, PlotRange -> {0, c[e]}, 
  Filling -> c[2], FillingStyle -> Red],
 Plot[c[t], {t, 2.1 + 1/3, 2.1 + 1/3 + d}, PlotRange -> {0, c[e]}, 
  Filling -> c[2.5], FillingStyle -> Green],
 Plot[c[t], {t, 2.1 + 2/3, 2.1 + 2/3 + d}, PlotRange -> {0, c[e]}, 
  Filling -> c[2.8], FillingStyle -> Blue]
 ]

(有损)压缩率$c$

We denote $c$ as space saving [1]:

$$ c:=1-\frac{compressed}{original}= 1- \frac{p+(1-\alpha)d}{d}=\alpha-\frac{p}{d} $$

$p$为AGI模型的 自由参数量

$d$为训练数据的数据量

$\alpha\in [0,1]$ 为测试数据上在某种metric下的拟合率,$\alpha =1-loss$