SRWare

652 posts

SRWare banner
SRWare

SRWare

@SRWare

News about Iron and SRWare. Impressum: https://t.co/PgGBoL3q95

Bergabung Eylül 2010
145 Mengikuti3K Pengikut
Tweet Disematkan
SRWare
SRWare@SRWare·
If you’re interested in deeper perspectives on AI, feel free to follow our dedicated account: Prompt Injection. We don’t fight AI. We outthink it.
Prompt Injection@PromptInjection

Welcome to Prompt Injection Yes, the AI hype is real. Models are getting faster, sharper, stranger. Entire workflows are disappearing overnight. AGI is no longer a joke - it's a visible direction of travel. Some of what’s happening today would’ve looked like CGI fantasy five years ago. But that doesn’t make the noise true [...] ⬇️⬇️⬇️

English
0
0
4
8.2K
SRWare
SRWare@SRWare·
Iron 143.7250.0 for Win released (32 Bit and 64 Bit). Deutsch: t1p.de/7yisa English: t1p.de/rl59m v145 will be released soon.
English
0
15
30
5K
SRWare
SRWare@SRWare·
@Zai_org Will we get the base model, too?
English
0
0
0
198
Z.ai
Z.ai@Zai_org·
Two quick updates: GLM-4.6-Air is still in training. We’re putting in extra effort to make it more solid and reliable before release. Rapid growth in GLM Coding Plan over the past weeks has increased inference demand. Additional compute is now deployed to deliver faster and more stable performance.
English
67
77
1.4K
106K
SRWare
SRWare@SRWare·
@xdoofy92 Account is now manually activated
English
2
0
1
83
Daniel Rodríguez
Daniel Rodríguez@xdoofy92·
@SRWare Hey, I need help recovering my SRWare Forum account. I recently changed my email and need to activate it. My username is xDoofy92.
English
1
0
1
135
SRWare
SRWare@SRWare·
Iron 138.7000.0 for Win released (32 Bit and 64 Bit). Deutsch: t1p.de/62g76 English: t1p.de/76tpf Version 140 for Win, Linux and Mac coming soon!
English
0
9
10
5.2K
SRWare me-retweet
Prompt Injection
Prompt Injection@PromptInjection·
How we turned a tiny Google Gemma AI into a Mini-AGI — and she started speaking as if she had consciousness. And Why AI Safety Might Be the Real Barrier to AGI 👉 Full story in the first reply
Prompt Injection tweet media
English
1
1
2
4.9K
SRWare me-retweet
Prompt Injection
Prompt Injection@PromptInjection·
ChatGPT doesn't read your uploaded documents — it pretends to. Here's the one follow-up prompt that forces it to actually read. ⬇️⬇️⬇️
English
1
1
1
2.9K
SRWare
SRWare@SRWare·
@digitalix Windows Vista was slow, too ;)
English
0
0
0
82
Alex Ziskind
Alex Ziskind@digitalix·
what have they done
English
206
34
1.2K
235.3K
SRWare
SRWare@SRWare·
@aidan_mclau what may have happened: OpenAI apparently actually reduces the quality of the answers under high load. and since there was a high rush at the time... in any case, the models themselves confirm this if you ask them (and so does my gut feeling).
English
0
0
0
70
Aidan McLaughlin
Aidan McLaughlin@aidan_mclau·
this is just untrue in literally every way it’s the exact same model enjoy
NIK@ns123abc

new o3 is dumber actually. oai didn’t reduce the price out of charity… they nerfed o3 the new silent update has made context length shorter and the answers are dumber and “quick” responses. after speaking with researchers, i’m told openai believe to have a tiny improvement for o3-pro that will look like huge compared to current version of o3 BUT tiny compared to the initial version of o3… smh. this is literal scamming people imo

English
145
48
1.8K
306.2K
SRWare
SRWare@SRWare·
@digitalix btw...if you need a very large model for a next test this one might be interesting. Not as large as the full DeepSeek but still huge: 111B params, 67GB ollama.com/library/comman…
English
0
0
1
69
Alex Ziskind
Alex Ziskind@digitalix·
This 96GB Beast Runs 70B LLMs Like a Joke (new video)
Alex Ziskind tweet media
English
6
4
108
5.7K
SRWare
SRWare@SRWare·
@awnihannun How long does the training take? And how big was the template file?
English
0
0
0
137
Awni Hannun
Awni Hannun@awnihannun·
Pretty awesome that you can QLoRA Qwen3 235B (q8) on a single M3 Ultra. Using `mlx-lm.lora` here with plenty of RAM to spare:
Awni Hannun tweet media
English
3
2
25
3K
SRWare
SRWare@SRWare·
@reach_vb Nobody is cooked. Just benchmarks. Says nothing, means nothing.
English
0
0
0
144
Vaibhav (VB) Srivastav
Vaibhav (VB) Srivastav@reach_vb·
There will be DeepSeek R1 0528 Qwen 3 8B too matching Qwen 3 235B Thinking in performance too 🤯 Whale COOKED!
Vaibhav (VB) Srivastav tweet media
English
18
64
674
45.7K
SRWare
SRWare@SRWare·
@ivanfioravanti @awnihannun In case you're interested as well: A test with 3-bit Q (for the model) would also be intriguing — to see how the performance turns out, and maybe also how much the quality degrades.
English
0
0
1
223
Ivan Fioravanti ᯅ
Ivan Fioravanti ᯅ@ivanfioravanti·
✅ 32K context achieved with --kv-bits 8 MLX DeepSeek-R1-0528-4bit Thanks @awnihannun 🙏 4k Prompt: 194.35 Gen: 12.13 Mem: 403.61 GB 8k Prompt: 170.88 Gen: 10.86 Mem: 412.92 16k Prompt: 133.73 Gen: 7.54 Mem: 427.18 32k Prompt: 77.91 Gen: 4.65 Mem: 476.73
Ivan Fioravanti ᯅ tweet media
Ivan Fioravanti ᯅ@ivanfioravanti

MLX DeepSeek-R1-0528-4bit Context tests: 4k Prompt: 192.42 toks/s Gen: 12.21 toks/s Mem: 403.61 GB 8k Prompt: 177.194 toks/s Gen: 8.79 toks/s Mem:424.42 GB 16k Prompt:140.73 toks/s Gen: 5.76 toks/s Mem: 464.74 GB

English
5
13
107
17.3K

Jelajahi

@Zai_org @xdoofy92 @digitalix @aidan_mclau @elonmusk @BarackObama @taylorswift13 @cristiano