Been Kim
@_beenkim
Followers
26K
Following
2K
Media
97
Statuses
840
Research Scientist at Google DeepMind, PhD from MIT. Make machines empower people.
Joined August 2011
1/8 Pareto Frontier 🤠for Human-centered AI 📈: We all want to build AI that is good for humans, but the path is often paralyzed by complexity. Either “oh my god, it’s too complicated😱” or delusional “I have a warm and fuzzy feeling of understanding 🥴”? "It’s hard because it
4
12
75
What a privilege it is to have time as your most valuable currency.
3
6
117
Safety-oriented interpretability researchers should be focused on AI systems, not individual model artifacts. A snippet from the NeurIPS CogInterp workshop panel on Sunday:
4
18
162
This post seems to describe substantially the same view that I offer here: https://t.co/LoNw7jFltD Why are people describing the GDM post as concluding that mech-interp is a failed project? Is it the renaming of the field and constant talk of "pivoting"?
The GDM mechanistic interpretability team has pivoted to a new approach: pragmatic interpretability Our post details how we now do research, why now is the time to pivot, why we expect this way to have more impact and why we think other interp researchers should follow suit
4
20
127
Tomorrow 9:30am #NeurIPS2025 Room 30A-E I'll talk about " 📈Towards Pareto frontier of interpretability: 15 years of interpretability research in 15 mins"🚅 @ mech interp workshop
4
10
80
Cyber Monday = Cyber DECEMBER! 🎅 Let the season of giving commence and last all month. Use Code CYBER25 for 25% Off Courses. Don't forget, we have Free & NameYourPrice options, too. Just Hacking Training (JHT) is a platform providing "Focused Technical Training for All Levels"
2
17
45
Our work out there in the wild 🥹
🔥 Proactive Co-Creator is officially LIVE in @GoogleAIStudio! Stop guessing prompts. Start collaborating. Use it now to remix ideas and generate images, stories, and video with an AI that proactively helps you create. 🔗 Try it here: https://t.co/YQ4JBZbYbJ 📍 At #NeurIPS2025?
0
3
27
🔥 Proactive Co-Creator is officially LIVE in @GoogleAIStudio! Stop guessing prompts. Start collaborating. Use it now to remix ideas and generate images, stories, and video with an AI that proactively helps you create. 🔗 Try it here: https://t.co/YQ4JBZbYbJ 📍 At #NeurIPS2025?
0
8
24
Awesome @NeurIPSConf keynote this morning by @YejinChoinka on The Art of (Artificial) Reasoning – and her broader thoughts and wishes on the future of Artificial Intelligence https://t.co/Zn5y7LWOV1
1
15
100
Highly skilled Operator vs Nauticus ToolKITT Extrapolation of the variance tells you the power of autonomy in improving operational performance - It is a must see. The transformation of autonomy to business value
13
35
84
Add: 9:30am on Sunday at Neurips, i'll touch upon this at the mech interp workshop keynote
0
1
4
8/8 Making AI benefit humans takes a village. 🌍 But a village needs a shared language. Let's stop guessing and start measuring the frontier.📷 a short write-up:
medium.com
We all want to build AI that is good for humans, but the path is often paralyzed by complexity. Not only is human evaluation a lot of work…
1
2
3
7/8 Example: We discovered Veo’s zero-shot capability all via prompting [ https://t.co/7IaYDyiwdD]. 🎥. Complexity: Low 📷 Efficiency: High - transfer is instant! Once you see this paper, you can start prompting Veo to do a task it wasn’t trained to do.
video-zero-shot.github.io
Video models like Veo 3 are on a path to become vision foundation models.
1
1
3
6/8 Example: We used AlphaZero to teach chess Grandmasters [ https://t.co/EgUnrMD5gw]. ♟️ Complexity: Superhuman (High ⬆️) Efficiency: Low (Took hours with the world's best). But it’s worth it - one student (Gukesh) became the youngest World Chess Champion.
pnas.org
AI systems have attained superhuman performance across various domains. If the hidden knowledge encoded in these highly capable systems can be leve...
1
1
3
5/8 By decomposing the problem, one can make a clear contribution by focusing on one axis (e.g., improving teaching efficiency of the same knowledge). It also makes the goal of human-centered AI clear: simply “discovering” the knowledge does not push the Pareto frontier without
1
0
2
4/8 The complexity can be measured by intrinsic concept complexity (e.g., MDL) and perceived complexity (e.g., NASA TLX).
1
0
2
3/8 The transfer efficiency can be measured by the differences in human understanding/performance/satisfaction on the target task divided by time spent. We can measure the "best" human to get at our upper limit (not always PhDs). Some tasks will have high complexity and low
1
0
3
2/8 LLMs use a Pareto curve (Performance vs. serving cost). Why don't we? Evaluating humans is messy, but we can project the chaos onto two axes: 📈 Y-Axis: Complexity of Knowledge or capability ⚡️ X-Axis: Transfer efficiency.
1
0
5
(Only for those who care to read this far) We have an internship position-if you are interested in discovering model capabilities. Email me with CV please:)
1
0
30
Navigating Privacy and Compliance in Blockchain Systems: a new report from Inco and @0xPredicate a review of the privacy + compliance landscape, exploring solutions from @AleoHQ, @AvaCloud, @fluidkey, @payy_link, @0xprivacypools, @RAILGUN_Project, @solana, @verulink, + @zksync↓
79
57
214
Coming to Neurips from 5-7th! Speaking at the mechanistic interpretability workshop https://t.co/MEoMMFgKRB on the 7th about unmechanistic interpretability (as requested by the organizers) 🙃🙂 While I’ll miss this, our work will be demoed at Google booth eg veo zeroshot
4
5
122
Students in the US (and many other countries) can get their hands on all the Gemini 3 Pro goodness for free!
With advanced reasoning and the ability to analyze hour-long videos, Gemini 3 is perfect for students. And now, college students in the US can get Gemini's Pro plan for free, with more access to Gemini 3 Pro (terms apply). See how our most accurate model can help ⬇️
83
130
3K