You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
They list 7 cases of "recent progress", 3 of them are explicitly LLM-assisted or LLM-generated.
200
200
201
-
## metaphors for RSI
202
-
203
-
Drawing balls from an urn.
204
-
:
205
-
There's some distribution of values, then you get a very nice expression. The expected value of $N$ draws just depends on the extreme value distribution of $f(v)$. This is exactly Kortum (1997).
206
-
207
-
A tool factory makes better tools.
208
-
:
209
-
I think this is Jones' metaphor. *Distinct* from Kortum, because the returns to search now depends on the stock of ideas.
210
-
211
-
Machine tools and regular tools
212
-
:
213
-
1. The Dutch make machine tools.
214
-
2. The machine tools make regular tools.
215
-
3. The regular tools make products.
216
-
217
-
At some point the machine tools become good enough to make themselves, but it's a discrete jump.
218
-
219
-
Lego - combining ideas.
220
-
:
221
-
Weitzmann's recombinant search is like this: you combine ideas to make new ideas, now you have a larger stock of ideas to combine.
222
-
223
-
Recipes.
224
-
:
225
-
You try out recipes, which are combinations of prior recipes. @jones2023recipes -- something like a reconciliation of Weitzman & Kortum.
226
-
227
-
Blacksmith.
228
-
:
229
-
You have a hammer and you spend time making horseshoes, or working on a new hammer.
230
-
231
-
A harder hammer both (1) makes horseshoes faster, or (2) makes your hammer still harder.
232
201
233
202
234
203
@@ -289,3 +258,65 @@ Some optimization landscapes that were already conquered.
289
258
[4]: https://www.investors.com/news/management/leaders-and-success/herman-hollerith-pioneered-data-processing/?utm_source=chatgpt.com"Herman Hollerith Put A Punch Into Data Processing"
290
259
[5]: https://cs.stanford.edu/people/eroberts/courses/soco/projects/2008-09/colossus/history.html?utm_source=chatgpt.com"The History of the Lorenz Cipher and the Colossus Machine"
| computer game -- collecting loot | thinning out of loot | loot speeds you up |
267
+
| apple picking | thinning out of apples | apples make you walk faster |
268
+
| blacksmith |||
269
+
270
+
Platform game (GOOD).
271
+
:
272
+
- Two types of things you can collect: gems, which contribute to your score; powerups, which make you faster.
273
+
- You observe player spends 10% of their time getting powerups, 90% collecting gems.
274
+
- Two effects of collecting a powerup: (a) make you go faster; (b) thinning out. If only one effect then you'd either get explosive growth (if powerups instantly regenerated then you'd just sit there collecting them forever) or collapse (if ).
275
+
- You observe exponential growth in gems and in speedups.
276
+
- Note: here the spillovers are to *both* goods and ideas.
277
+
278
+
Model: each level is 1 minute, and you choose to spend $R$ share of your time on speedups.
279
+
280
+
$$\begin{aligned}
281
+
\Delta A &= (RA)^\gamma
282
+
&& \text{(speedups collected)}\\
283
+
\Delta G &= ((1-R)A)^\gamma
284
+
&& \text{(gems collected)}
285
+
\end{aligned}$$
286
+
287
+
1. If $R$ is constant, then $\frac{\Delta A}{A}$ and $\frac{\Delta G}{G}$ will decline.
288
+
2. If there's no fishing-out ($\gamma=1$), e.g. speedups instantly regenerate, then we get constant exponential growth (AK).
289
+
3. Here the shoulders-of-giants effect is exactly linear, I think meaning that economy-wide productivity improvements exactly transfer to AIR&D productivity.
290
+
291
+
Note: for AI R&D we can ignore gems, imagine you're only picking up speedups.
292
+
293
+
294
+
Drawing balls from an urn (Kortum)
295
+
:
296
+
There's some distribution of values, then you get a very nice expression. The expected value of $N$ draws just depends on the extreme value distribution of $f(v)$. This is exactly Kortum (1997).
297
+
298
+
A tool factory makes better tools (Jones)
299
+
:
300
+
I think this is Jones' metaphor. *Distinct* from Kortum, because the returns to search now depends on the stock of ideas.
301
+
302
+
Machine tools vs regular tools.
303
+
:
304
+
1. The Dutch make machine tools (high quality)
305
+
2. The machine tools make regular tools (low quality)
306
+
3. The regular tools make products.
307
+
308
+
At some point the machine tools become sufficiently good quality to make themselves, but it's a discrete jump.
309
+
310
+
Lego - combining ideas.
311
+
:
312
+
Weitzmann's recombinant search is like this: you combine ideas to make new ideas, now you have a larger stock of ideas to combine.
313
+
314
+
Recipes.
315
+
:
316
+
You try out recipes, which are combinations of prior recipes. @jones2023recipes -- something like a reconciliation of Weitzman & Kortum.
317
+
318
+
Blacksmith.
319
+
:
320
+
You have a hammer and you spend time making horseshoes, or working on a new hammer.
321
+
322
+
A harder hammer both (1) makes horseshoes faster, or (2) makes your hammer still harder.
0 commit comments