Fred Zhang (@FredZhang0) / X

Fred Zhang

322 posts

Fred Zhang

@FredZhang0

@Meta, prev research scientist @GoogleDeepMind, PhD @Berkeley_EECS, DM open

Berkeley, CA

Joined July 2019

Fred Zhang
@FredZhang0
Jul 21, 2025
This is the most scaling-pilled project I've ever been part of, and the team really cooked. TL;DR: With RL and inference scaling, Gemini perfectly solved 5 out of 6 problems, reaching a gold medal in IMO '25, all within the time constraints of 4.5hr.
Google DeepMind
@GoogleDeepMind
Jul 21, 2025
An advanced version of Gemini with Deep Think has officially achieved gold medal-level performance at the International Mathematical Olympiad. 🥇 It solved 5️⃣ out of 6️⃣ exceptionally difficult problems, involving algebra, combinatorics, geometry and number theory. Here’s how 🧵
57K
Fred Zhang
@FredZhang0
Jul 21, 2025
Replying to @simonw
No tool use, no internet access, and pure informal.
21K
Fred Zhang
@FredZhang0
Sep 29, 2023
Activation patching (AP) is a standard tool in LM interpretability, esp circuit analysis. It has multiple degrees of freedom, but the literature has little consensus on them. Our new work systematically studies these, and works towards best practices.
29K
Fred Zhang
@FredZhang0
Mar 1, 2024
Beating prediction markets with chatbots sounds cool. In a recent work arxiv.org/abs/2402.18563, we get somewhat close to that. As another perspective, forecasting is a great capability domain to benchmark LM reasoning, calibration, pre-training knowledge, and more. 🧵1/n
arxiv.org
Approaching Human-Level Forecasting with Language Models
Forecasting future events is important for policy and decision making. In this work, we study whether language models (LMs) can forecast at the level of competitive human forecasters. Towards this...
11K
Fred Zhang
@FredZhang0
Jul 21, 2025
Replying to @ankesh_anand
see here:
en.wikipedia.org
Millennium Prize Problems - Wikipedia
2.9K
Fred Zhang
@FredZhang0
Mar 5, 2025
Replying to @giffmana
> DeepMind, DeepMind, what are you doing? Huge loss for GDM. Hopefully a wake up call for the leadership.
8.3K
Fred Zhang
@FredZhang0
Aug 6, 2025
Replying to @oh_that_hat
> for those at Berkeley or Stanford It's partly proximity but also mindset and network
5.1K
Fred Zhang
@FredZhang0
Oct 29, 2024
alternative timeline: strong interp is information theoretically solvable, but never solved, due to computational complexity barriers same may apply to neuroscience and fundamental physics
James Campbell
@jam3scampbell
Oct 28, 2024
strong interpretability will be solved, it’s just a matter of when (before AGI/before ASI/after ASI) but when it is solved, it’ll mark a major shift from taming dragons to designing super-dragons
2.7K
Fred Zhang
@FredZhang0
Sep 29, 2024
Replying to @StephenLCasper and @GavinNewsom
gov.ca.gov/2024/09/29/gov… This also seems inconsistent with him bringing Li Feifei to the new initiative, who argued SB 1047 is too restrictive.
652
Fred Zhang
@FredZhang0
Jun 5, 2024
Replying to @willccbb
For distributed training, I find these 2 posts well-written: siboehm.com/articles/22/da… siboehm.com/articles/22/pi…
Data-Parallel Distributed Training of Deep Learning Models
From siboehm.com
1.1K
Fred Zhang
@FredZhang0
Mar 20, 2025
Every OOM improvement along this trendline can be qualitatively different and break the line itself. In particular, I expect an t-AGI, for t ~ 1 week, would automate a decent fraction of tasks in day-to-day AI R&D and accelerate the trend, potentially to superexponential rate.
METR
@METR_Evals
Mar 19, 2025
When will AI systems be able to carry out long projects independently? In new research, we find a kind of “Moore’s Law for AI agents”: the length of tasks that AIs can do is doubling about every 7 months.
2.3K
Fred Zhang
@FredZhang0
Feb 26, 2024
The top-down vs bottom-up distinction in mech interp is quite a misnomer (suggesting the former is superior). But in fact, activation steering and circuit analysis are at the same level of abstraction. The right terminology should be representation-space vs weight-space interp.
ML Safety Daily
@topofmlsafety
Feb 23, 2024
Opening the Black Box of Large Language Models: Two Views on Holistic Interpretability Survey paper on the applications, limitations, and challenges of representation engineering and mechanistic interpretability. arxiv.org/abs/2402.10688
1.2K
Fred Zhang
@FredZhang0
Feb 29, 2024
arxiv.org/abs/2402.18563 Check out our recent paper on automated forecasting with LM! Joint with Danny Halawi (@dannyhalawi15), Yueh-han Chen (@jcyhc_ai) and Jacob Steinhardt (@JacobSteinhardt).
Jacob Steinhardt
@JacobSteinhardt
Feb 29, 2024
Can we build an LLM system to forecast geo-political events at the level of human forecasters? Introducing our work Approaching Human-Level Forecasting with Language Models! Arxiv: arxiv.org/abs/2402.18563 Joint work with @dannyhalawi15, @FredZhang0, and @jcyhc_ai
arxiv.org
Approaching Human-Level Forecasting with Language Models
Forecasting future events is important for policy and decision making. In this work, we study whether language models (LMs) can forecast at the level of competitive human forecasters. Towards this...
1.4K
Fred Zhang
@FredZhang0
Oct 2, 2024
Glad to have played a small role in this new benchmark effort on evaluating LM for forecasting. TL;DR: it's a fully dynamic set that asks you to forecast the future and so is always contamination-free. We find frontier models still not as good as humans.
Forecasting Research Institute
@Research_FRI
Oct 1, 2024
Today, we're excited to announce ForecastBench: a new benchmark for evaluating AI and human forecasting capabilities. Our research indicates that AI remains worse at forecasting than expert forecasters. 🧵 Arxiv: arxiv.org/abs/2409.19839 Website: forecastbench.org
1.7K