@abacaj

Software

Joined August 2009
  • Tweets13.4K
  • Following613
  • Followers48.5K
2.2KMedia
Almost all of the examples of Jev that I have seen talk about the speed and not the quality. Coincidence? Or does no one care about output anymore
53
5
1
129
20,207
GPT-6 is a lucky seed that’s all it is, they got a lucky run
This quoted post is unavailable.
Try viewing anyway · View on X
7
1
18
9,195
Ok tell them it’s AGI Now tell them it’s like AGI but not really Ok now tell them the next model will really be AGI
32
88
4
1,494
36,645
Hate to say it but total Dario victory once again. Was able to use their model when it was announced. Also Fable looks far ahead in code still
17
3
215
24,179
“AGI” scores the same as previous non AGI model?
24
58
12,239
It’s over (for OpenAI)
6
11
2,771
Alright I'm tired of being linked this example and saying repeatedly it's not my experience with GLM 5.3 Flash. So I ran it myself. @skalskip92 what precision are you running GLM 5.3 Flash here? Are you using an API? Can I just download all your examples unlabled somewhere?
SkalskiP
@skalskip92Aug 26
as expected, GLM 5.3 Flash is pretty bad at vision interesting how different models end up specializing in completely different things some are great at coding, others at computer vision no AGI yet, I guess 1/n aerial and satellite images, crops link: playground.roboflow.com/shar…
19
7
7
416
80,538
Was using GPT 5.6 Luna to process documents, replaced it all with GLM 5.3 flash past two days for zero quality drop and it’s still cheaper than their 80% discount… You can just switch models
121
67
22
2,406
303,414
The model is also really good as a browser agent, drop in replacement didn’t change any of the prompts/tools
7
1
147
19,661
they need to make models that think less, seems to be the only lever they have is how many tokens it generates while thinking
2
32
4,846
and this is fable 5, supposedly the best model
anton
@abacajAug 27
I really hope the next models from anthropic don't talk like this
8
2
1
131
10,508
I really hope the next models from anthropic don't talk like this
9
3
1
57
14,842
GLM 5.3 flash on high effort is perfect, not too verbose and still correct. I think the best model possible (if you can run it locally). Very comparable to opus 4.8
16
6
1
356
24,519
Yea this one is gonna be interesting. The form factor, unified memory and efficiency of macs > big heavy GPUs
3
1
12
4,167
byte-for-byte, byte-identical, byte-verified. bet you can tell which model this is
38
4
4
895
50,030
K3 makes a lot of mistakes but funny enough it can fix them if you keep asking it to. Fable makes much less mistakes, but K3 is a surprisingly capable model just wish it was cheaper
10
1
66
5,101
I used it for about a week while my Fable usage was capped out. The code was fine and not much worse than Fable, though differences become more noticeable as context grows
2
7
1,552
I tried using Fable to train a model (LFM 2.6B) because I didn’t want to spend time on the data. Turns out neither did Fable and it ended up making the model worse on every attempt until I decided to look at the data. It was using the wrong chat template on like 1/3 of the data and had started to import HF datasets that didn’t align with the task at all. Sometimes I wonder if it was intentional sabotage or if it was just lazy
vie ⟢
@viemccoyAug 17
if you're training a model and you aren't inspecting the data, you actually aren't training a model - the model is training you
57
18
7
545
52,634
Anthropic sucks man, this is getting ridiculous
68
19
5
848
42,690
Harness will be less important as models become more capable. Can your harness bring gpt-3 performance up? Or is your harness impressive because the model is impressive?
37
1
7
101
11,685
Example xitter.cf/jcrwhittington/status/…
James Whittington
@jcrwhittingtonAug 12
Replying to @jcrwhittington
We find that 1) Opus 5 & Fable marked a real increase in ability on these games, but they still struggle with the hardest tiers of games (Tiers 6 and 7) 2) Agentic harnesses like Prime don’t offer any benefit over a basic (non coding!) harness 3) There’s a large gap between closed-source (Opus 5) and open-source models (Kimi K3) and another big gap to smaller open-source models (Qwen 3.6 27B) Finally, rule discovery really is the bottleneck in these games: when we give the ground-truth rules in plain language, even the previously weakest models score near perfect. (4/5)
1
7
2,264
Load more