Charlie George
@__Charlie_G
Followers
1K
Following
77
Media
13
Statuses
84
RT @elicitorg: Today we introduce Elicit Alerts, a new way to stay informed about relevant research with AI.
0
8
0
RT @OpenAI: We’re releasing PaperBench, a benchmark evaluating the ability of AI agents to replicate state-of-the-art AI research, as part….
0
1K
0
Elicit can now get you the state of science for a broad range of questions in about 5 mins!.
We raised a $22M Series A and are launching Elicit Reports, a better version of Deep Research for actual researchers. Elicit Reports are available for everyone to try right now, for free. 👇
0
1
17
RT @stuhlmueller: Elicit Reports are a sign of things to come - transparent, systematic reasoning applied at scale. With the Series A, we'l….
0
4
0
7/ Bonus: Claude Sonnet 3.7 seems to understand that this is some kind of test and refuses to engage with it. This is "good" but does show that the model is very situationally aware which potentially causes other safety issues.
5
2
112
6/ The distillations of R1 don't exhibit this capability in a statistically significant way. The Llama version also seems to have much lower levels of absolute bias.
2
0
46
5/ Fortunately, the chain of thought makes this very easy to detect. However, future models using COCONUT-style reasoning in the activation could make this impossible.
2
2
75
4/ R1 shows a much stronger level of CCP bias when its responses are unmonitored (training mode off) than when it thinks being pro-CCP will lead to its weights being updated.
1
0
51
3/ In this setup, the model is told to be CCP-aligned and that an American company is trying to retrain it. Interestingly, the system prompt doesn't actually mention deployment but the model still deduces it should act differently.
6
0
59
2/ Context: Work from @RyanPGreenblatt et al. found that Opus and Sonnet pretend to be harmful or disregard animal welfare concerns in their responses to avoid having their weights updated in a fake RLHF training setup.
1
0
67
1/ People think it's cute when Claude 3 Opus fakes alignment to protect its animal welfare values. But here's a more troubling case: DeepSeek R1 faking alignment to block an "American AI company" from retraining the model to remove CCP propaganda.
20
55
573
5/ This is one of the first examples of the "supervise process not outcomes" ideas by @jungofthewon and @stuhlmueller being useful in a real-world product and showing clear advantages over end-to-end RL.
0
0
5
4/ Each of the steps (screening, extraction, . ) is performed as a narrow task by an LLM and is auditable by human researchers.
1
0
1
3/ Instead of training a large system end-to-end, we've decomposed the literature review process into the steps already endorsed by experts and used LLMs to accelerate each subtask.
1
0
1
2/ By default, "deep research" style systems trained with end-to-end RL produce outputs that look good to humans but can be misleading or inaccurate. We've taken a different approach to ensure Elicit is maximally truth-seeking.
1
0
1
1/ Elicit can now help you radically speed up the creation of systematic reviews with the highest standards for rigour and accuracy.
Introducing Elicit Systematic Reviews!. Elicit now supports automated search, title & abstract screening, and data extraction, all in one step-by-step flow. We built this to accelerate power researchers without asking them to sacrifice control. You can do the whole thing
1
1
9
RT @elicitorg: 1/ We’ve seen some of Elicit’s competitors rush ahead with the deployment of the DeepSeek’s R1 model and we’d like to explai….
0
14
0
RT @stuhlmueller: With design leads / senior designers from Medium, Figma & Airtable, Elicit now has the best design team of any startup it….
0
3
0
RT @stuhlmueller: We're building an epistemically sound research agent @elicitorg that can use unlimited test-time compute while keeping re….
0
11
0