Google for Developers
Region: US
Wednesday 24 June 2026 11:00:08 GMT
34700
1650
35
68
Music
Download
Comments
Dr Sean :
who is json?
2026-06-24 17:23:01
45
jagger :
we are so allergic to coding😒its so easy to pick out the json from a string of text.
2026-06-24 18:46:25
13
Míra :
why not just use json schema output?
2026-06-24 23:52:50
1
Vincent :
Or you hire a real programmer to do it properly….
2026-06-25 00:11:14
4
Michael Gray :
How do you practically do this though?
2026-06-24 21:31:17
1
Noor Research Collective :
This is because the labels are variables for an attractor space and are not intrinsic to it.
2026-06-24 18:12:25
2
unify_ai :
I just use decompositions.
2026-06-24 18:18:30
2
Calton Rodriquez :
what's the spec for landing an internship
2026-06-24 11:16:20
4
user320xt :
WHO is that man Json
2026-06-24 18:16:57
3
basil :
gemini literally has a deterministic option for structured output, for json etc
2026-06-24 21:24:41
1
ag2574373 :
But the second bit of JSON was valid
2026-06-24 16:58:55
0
Jawad Sadat✅ :
follow me
2026-06-24 11:03:30
1
……… :
I use trustcall
2026-06-24 21:42:56
0
ShooShoo :
my experience is different. I used a local AI. Task is so simple: check if a translation is smooth. The key is brief. If prompt is long then the AI isn't stable
2026-06-25 00:48:03
0
0xPytness :
llamacpp allows for forcing models to output a reduced set of tokens. so you can basically force it to produce json. its called grammar and you can provide a gbnf string that suits your needs. I created a tool a while ago that takes interfaces from typescript (with some added types) and transformed it into a gbnf file. Pretty useful to force a json schema back in the day
2026-06-25 04:30:05
0
just tall Greg :
brother just use deterministic programs not fuzzy programming
2026-06-25 00:33:46
0
المضحك :
I once fine tuned t5 with this very same technique and it kept hallucinating
2026-06-25 02:57:52
0
user4338058358228 :
never happened to me at all in the past year. if that's still a problem for Google you are doing something wrong.
2026-06-25 06:38:16
0
ecom_nik :
Yeah or just pre-fill the start of the response with a {. A lot easier than tuning a whole new model 😂
2026-06-24 19:50:15
0
Getmaybe50%ofreplieshonestly :
oh for gods sake you dont need more ai for every little thing just write a function for grabbing the json jfc this is absurd.
2026-06-25 06:19:09
0
Al :
Is it about teaching ur own LLM or building an agent? If it’s the latter is it using the agent creation in Gemini app
2026-06-24 20:39:34
0
Alex Kryazhev :
overengineering
2026-06-24 20:15:18
0
kvadrato :
how about to provide a JSON schema that is expected? and instruct the model to strictly follow it
2026-06-24 19:53:21
0
klietus :
This is true if you want to run fine tuning training every model upgrade. If there was only a way to get deterministic output from probabalistic models. Neuro symbolic AI unite! Or just use a much cheaper regex on the back end. Do you guys not understand tradeoffs any more?
2026-06-24 18:45:27
0
To see more videos from user @googlefordevs, please go to the Tikwm
homepage.
