@theo 90% of my energy past few weeks has been spent trying to figure out why agents keep overcomplicating releases and wanting to do perform 100 irrelevant checks.
Finally got into place where it works now - but incredibly frustrating.
@AndreasSteno@btdk Ballon d'or har altid vaeret en blanding af hold og individuel. Mbappe slaar rekorder for ikke at arbejde defensivt og hans tidligere hold er blevet bedre efter han er skiftet.
Hvis det var pure "hvem er verdens bedste bedste spiller" ville jeg stadig pege paa Messi.
@craptor When you filter outcome-based, i.e. best attacking/defending team over the period you will always get bias.
The way to do unbiased filtering is to only use information that's available prior to the match. I.e. your own ratings or the market lines.
@theo I think Opus 5.5 lasts longer than 5.6. Constantly ran ran out of tokens on the $200 subscription (whereas 5.5 was better for that). Would say price 6/10.
@shrewdstone@johncrickett Unit tests don't guarantee that.
Unit test can ensure the public function produces correct input and output, but there is no guarantee that it fits correctly within the larger scope of the codebase.
You need e2e test for it.
@alperortac@mattpocockuk I tried the intent approach. It can't stand alone.
The only way to prevent it from wandering off and doing bad stuff is to have rigorous patterns setup that make it impractical to break (along with documentation as to why they exist)
@thorstenball Ask the agent on how to split your 2K PR into smaller separated components and it will be pretty good at that.
I don't see how this is particularly valueable skill.
@blueemi99 Why do you need it to be a big upgrade? For anyone with limited tokens, this is a big win. And if you have unlimited, use Astra.
The Opus 5.5 is confusing to me because it competes with Fable.
@michael_timbs If your agent wants to write unit-test to debug, let it do that. The point is that humans should focus on setting up structures in which the end functionality works correctly and is tested thoroughly.
@nateberkopec The time saved on writing code - spend that time figuring out how to effectively test e2e. Once you have that setup, you can actually trust what the agent produces and release much faster.
@thdxr I think this is one of the things most ppl get wrong about AI and code reviews. If the final output in terms of the functionality and desired metrics work - code quality doesn't matter much. At least not in the short term - and the cost of rewriting/cleaning later is much lower.
@Gingfacekillah Depends when you are betting. Are you only betting against closing line - then sure. Otherwise you can't use the closing line directly in your modelling process.
@craptor How do you measure CLV when you backtest? It requires that your model only has information that's available at the time of the bet, i.e. injury or other lineup news that may first be available at kickoff.
@PlusEVAnalytics@d_feustel If you can sgp it its most likely done through a sim. The variance around the players skill might not be taking into account though.
@JackJGaming In theory this could be solved by a a live-win probability model. Even something like tempo could be measured through positional coordinates (although complex).
The bigger problem is that it's harder to attribute (skill) contribution to an action.