@Jaidcel

0.30000000000000004× full-stuck slopware engineet

xitter.cf/search?q=list:2021282544344498672%20-filter%3Areplies&f=live
Joined April 2020
  • Tweets8.1K
  • Following292
  • Followers885
3.4KMedia
Pinned Tweet
Jaid
@JaidcelSep 7
You wouldn’t understand. You don’t know what I feel when my most load-bearing smoke test quietly hits a regression.
16
2,827
logicmogged
5
152
incorrect horse battery staple
3
140
They chose 4 to avoid the Chinese hackers.
91
Jaid
@JaidcelOct 22
2
6
317
Oh, it looks bad. 🥺 Twitter keeps jumping between low bitrates while watching. Can be fixed by manually pinning to something high.
76
1
1
383
49
WTF I love Grok Bot now
Nick Dobos
@NickADobosOct 22
Whelp grok @bot just added all 50+ of my bots to my DMs here and now I can’t read any notifications from humans lol
1
82
Jaid retweeted
Jaid
@JaidcelOct 21
1
3
79
Jaid
@JaidcelOct 21
Turned out to be a good idea to set this up. Even though the entries from Opus are usually superior in every way that matters and the competing PRs just end up getting closed by me, having this kind of comparisons in the first place is fun and I still learn a lot about models.
2
2
157
1
3
79
Replying to @Jaidcel
@Jaidcel
1
1
1
53
Turned out to be a good idea to set this up. Even though the entries from Opus are usually superior in every way that matters and the competing PRs just end up getting closed by me, having this kind of comparisons in the first place is fun and I still learn a lot about models.
2
1
3
408
1
1
124
$ 80 for a single task just to have Grok “correct” the punctuation in unrelated comments
1
2
82
watching Gemini burn $ 17 on my credit card
Jaid
@JaidcelOct 19
Replying to @Jaidcel
I noticed the exploration loop and still let it run for 3 more hours to learn about it, assuming this isn’t real money anyway. But it was. I usually layer endpoints so that it makes my subscriptions tank the cost and only fall back to OpenRouter when needed, but I did it wrong.
4
239
Gemini 4 fixed nothing. Its agentic behavior continues to be abhorrently bad. Unironically Llama-grade tool calling. I lost $ 17 on the first task I assigned to Argon. And that’s how much it made me pay before starting to do any kind of work.
1
2
189
I noticed the exploration loop and still let it run for 3 more hours to learn about it, assuming this isn’t real money anyway. But it was. I usually layer endpoints so that it makes my subscriptions tank the cost and only fall back to OpenRouter when needed, but I did it wrong.
1
1
2
355
So the money thing is my sole responsibility and a skill issue on my part, but the doomlooping that caused this in the first place is obviously a skill issue on DeepMind’s part. The others did fine on the same task in the same environment.
60
29 resolved Issues so far without any kind of inconvenience
Jaid
@JaidcelOct 16
This week I worked a lot on adding GitHub support to Mage, turned out well. It now automatically converts every Issue I post into a pull request.
1
1
163
The next step will be to pass every issue to several models who all separately propose their own solution. Then I (or – whenever I don’t want to spend too much time on curating – a model as a judge) pick my favorite implementation for merging and close the other pull requests.
1
1
1
44
The average qualitative gain of this is negligible right now. I want this for two other reasons: • I have a bunch of active subscriptions I don’t know what to burn their tokens on. • It passively yields high-quality eval data. I can benchmark whose PRs are picked most often.
1
1
40
Load more