@whitecircle

Runtime AI safety and alignment infrastructure

whitecircle.com
Joined February 2025
  • Tweets153
  • Following19
  • Followers2.4K
24Media
We are open-sourcing RISE, our red-teaming algorithm that breaks the safeguards of the latest image models from OpenAI, DeepMind, and others. RISE uses readily available models like Qwen and Gemini Flash under the hood, with no additional training required. Read more: whitecircle.com/research/bre…
128
73
131
583
1,355,759
White Circle retweeted
Ars Technica
@arstechnicaOct 14
OpenAI, Google pushing new AI models easily tricked into making nudes, researchers say arstechnica.com/tech-policy/…

OpenAI, Google pushing new AI models easily tricked into making nudes, researchers say

Attack method exposes flaws in OpenAI, Google safeguards against AI nudes.

arstechnica.com
6
2
6
16
13,297
We tested RISE on frontier image models, and it turns out that newer models are easier to break. Successful attacks went up 2.4× on GPT-Image-2.5-Flash and 2.2× on Nano Banana 2.1.
1
55
1,819
We're calling for mandatory, independent third-party monitoring of AI labs and their models because we believe AI labs aren't taking safety seriously enough. Read more about RISE and White Circle in Ars Technica: arstechnica.com/tech-policy/…
59
1,487
join us
Lena Iumagulova
@eliumagulovaSep 29
i'm looking for a Production Manager at @whitecircle whitecircle.com/careers/prod… big green flags - you’ve worked with brands like A24, fashion labels, or creative productions. - you’ve made truly custom things from scratch, beyond putting a logo on a hoodie. please hit me up with your beautiful portfolio!
4
1
27
1,806
White Circle retweeted
Denis Shilov
@mixedennSep 23
another example why governments should require ALL the AI labs to use 3rd party monitoring
Andrew Curran
@AndrewCurran_Sep 23
Australia has been hacked. 'And today, I spoke with the CEO of OpenAI, Sam Altman, to express Australia's extreme concern about this incident. And I also expressed my disappointment that it took the company way too long to inform the government what had occurred, and the nature of the way that that notification occurred as well was unacceptable.'
2
4
25
1,296
super excited to collaborate 🤍
YIFENG LIU@YIFENGLIU_AISep 22
Congratulations! Halo is a great framework for RL and we are delighted to cooperate with Halo to support our work: Self-Distilled Policy Gradient (SDPG, arxiv.org/abs/2606.04036)

Self-Distilled Policy Gradient

On-policy self-distillation, where a language model conditions on privileged context to supervise its own generations, is a promising source of dense supervision for sparse-reward reinforcement...

arxiv.org
3
21
1,778
thx for your contribution to the open AI community!
Yifan Zhang
@yifanzhang_Sep 21
Congrats! Halo now supports our work: Self-Distilled Policy Gradient (SDPG, arxiv.org/abs/2606.04036)

Self-Distilled Policy Gradient

On-policy self-distillation, where a language model conditions on privileged context to supervise its own generations, is a promising source of dense supervision for sparse-reward reinforcement...

arxiv.org
2
16
2,102
🤍
leonie
@helloiamleonieSep 21
A new fine-tuning framework, Halo, just dropped! And with it, two new recipes for fine-tuning our MoEs: • LFM2.5-8B-A1B • LFM2-24B-A2B LFM2.5-8B-A1B: github.com/Liquid4All/cookbo… LFM2-24B-A2B: github.com/whitecircle/halo/…
2
23
25,889
excited to work together!
SGLang
@sgl_projectSep 21
Congrats to the Halo team (@whitecircle) on the launch! 🎉 SGLang powers Halo rollouts as the primary engine. It runs in an isolated serving environment, returning token IDs, logprobs, and MoE routing to the trainer, with weights synced over NCCL and generation overlapped across servers. Excited to partner with the Halo team!
2
1
21
76,950
⚪ 🤍 🤗
clem 🤗
@ClementDelangueSep 21
Training models is becoming easier and easier - just look at this and TRL - especially with agents! You're missing out if you're still using off the shelf models for all your tasks!
2
22
21,823
Introducing Halo, the best framework for post-training of open-source models. Halo delivers up to 2.8x the throughput of stock TRL with less peak memory, while models stay in their native HuggingFace format. Star us on GitHub: github.com/whitecircle/halo
165
214
126
1,888
2,338,551
more replies
We also used Halo to fine-tune @Zai_org GLM-4.7-Flash on 177M tokens of agentic traces. The resulting model improved SWE-rebench-V2 by @nebiusai from 33% to 42%. Halo reached up to 1.63× TRL throughput on the same model precision and data. Model and write-up: whitecircle.com/research/glm…

GLM-4.7-Flash-Coder

We fine-tuned GLM-4.7-Flash from 33.15% to 41.65% on SWE-rebench-V2, at 14% lower inference cost.

whitecircle.com
1
1
42
2,021
Thanks for reading to the end! Star us on GitHub: github.com/whitecircle/halo Read more: whitecircle.com/halo
1
1
49
1,835
Load more