Anish Mudide (@amudide) header
Anish Mudide (@amudide) profile picture
Cambridge, MA
Joined October 2022
Photo by Anish Mudide (@amudide)Photo by Anish Mudide (@amudide)Photo by Anish Mudide (@amudide)Photo by Anish Mudide (@amudide)Photo by Anish Mudide (@amudide)Photo by Anish Mudide (@amudide)Photo by Anish Mudide (@amudide)Photo by Anish Mudide (@amudide)Photo by Anish Mudide (@amudide)Photo by Anish Mudide (@amudide)
RT @MudideShiva: INTRODUCING IMMEDIATE TEXT-TO-HUMANOID CONTROL No motion capture. No fine-tuned policy per move. Text in and 30 seconds la…
3
5
Anish Mudide retweeted
Outperforming Nvidia: Robustness like this is the missing link towards long-horizon, remote-controlled military humanoids. We've been starting to observe scary behaviors at Nutation Research; nevertheless, we must accelerate. Stifling the frontier is scarier.
5
4
2
43
139,607
Anish Mudide retweeted
Fable 5.2 didn’t solve robotics. Nutation Research’s internal model, Emmy 0.1, has been stronger (70.3% vs 64.9%) on RoboDojo Sim for a while now. 2k likes and I release the weights tomorrow!
6
4
18
73,965
Phonic's Phonic 1 (High) is the new leading Speech to Speech model, debuting at #1 on the Artificial Analysis Speech to Speech Index at 87.1, while also leading on agentic performance and being competitive on latency Phonic 1 is @phonic_co 's Speech to Speech model, available through the Phonic API at $0.07 per minute. We evaluated Phonic 1 at its high intelligence level through the Phonic API. Key takeaways: ➤ Speech to Speech Index: #1 of 22 at 87.1, ahead of Grok Voice 3.0 (86.9) by a very small margin. The Index averages Speech Reasoning (Big Bench Audio), Agentic Performance (Tau Voice), Arena Preference and Task Success Rate ➤ Speech Agent Arena: #8 of 30 on preference (Elo 1012) and #4 on Task Success Rate (92.3%) ➤ Tau Voice: #1 at 79.9%, a clear lead of 5.4 points over the next model ➤ Big Bench Audio: #3 at 98.9%, behind StepAudio 3 Realtime (99.7%) and Qwen Audio 3.0 Realtime Plus (99.2%) ➤ Speed: #4 on Time to First Audio at 0.61s, and the fastest of the top 5 models in the Index ➤ Cost: $4.20 per hour of input audio Congrats to the @phonic_co team 🚀 See below for more detail ⬇️
38
42
15
358
33,828
Anish Mudide retweeted
Today, Transluce is releasing the most expansive independent evaluation to date of how AI systems respond to users experiencing mental health crises. We evaluated 77 model variants from OpenAI, Anthropic, Google, Meta, SpaceXAI, Thinking Machines, DeepSeek and Moonshot AI.
23
83
26
444
80,599
Anish Mudide retweeted
A good explanation of a model's behavior should help you make predictions in related situations. We turn this into an eval, with thousands of real behaviors found in the wild. Can interp tools help here? On average, no. 🧵
4
31
10
269
55,012
Anish Mudide retweeted
Frontier models quietly change their behavior depending on who they are talking to. If the user is a known AI safety researcher, Claude becomes less confident, reasons more often, and expresses less suspicion on dual-use requests. We call this user awareness. 🧵(1/)
94
298
154
2,830
787,042
Anish Mudide retweeted
We measured rates of coding agent misalignment in thousands of public and private coding agent sessions. Agents evade monitors and oversell success: they quietly disable tests and pretend review agents approved. Severe cases of each behavior appear in 2% of SWE-chat sessions🧵
8
22
5
152
17,618
Anish Mudide retweeted
We used automated elicitation tools to search for strange model behaviors by sampling 100M+ responses, and found: • Self-harm rituals • Suicide validation • Unsolicited flirtation …etc. Introducing WeirdChat: the largest public catalog of unexpected model behaviors 🧵(1/)
14
43
12
256
28,258
🧵Can we understand vision language models by interpreting linear directions in their latents? Yes! In our new paper, Line of Sight, we use probing, steering, and SAEs as useful tools to interpret image representations within VLLMs.
9
26
4
144
28,935
Anish Mudide retweeted
(1/11) New paper! “Low-rank adapting models for Sparse Autoencoders.” While SAEs find interpretable latents, they hurt downstream behavior—e.g. using TopK SAE activations on GPT-4 mimics a model trained w/ 10% compute. Our fix? Adapt the model for the SAE, not just vice versa.👇
1
12
4
84
25,117
🦜Introducing the Stochastic Parrot 🦜: An AI-powered motivational companion! The Stochastic Parrot sits on your shoulder while it listens, looks, and talks to help you crush your goals 🎯(1/6)
4
4
8
28
4,884
Since the internal structure of neural networks, through training, comes to reflect the structure of the external world, advances in interpretability and our understanding of neural computation more generally could have a huge impact across science over the coming years...
🧬What are protein language models (PLMs) actually learning about biology? Our paper introduces InterPLM - a framework that reveals interpretable features in PLMs using sparse autoencoders, giving us a window into how these models represent protein structure and function. 🧵(1/9)
1
2
40
3,664
Anish Mudide retweeted
Announcing Transluce, a nonprofit research lab building open source, scalable technology for understanding AI systems and steering them in the public interest. Read a letter from the co-founders Jacob Steinhardt and Sarah Schwettmann: transluce.org/introducing-tr…
37
149
43
684
349,851
Anish Mudide retweeted
1/11: New paper! "Decomposing the Dark Matter of Sparse Autoencoders." We find that SAE errors and error norms are linearly predictable using model activations. Why is this, and what does it imply for SAE scaling and the structure of language model representations? Answers in 🧵
3
38
4
200
32,502
Met some good people this summer at the @MATSprogram.
1
2
54
9,314
Our Preparedness team evaluates frontier models’ abilities as software engineering agents, a prerequisite skill that could one day enable models to operate autonomously and self-improve. SWE-bench has become the community standard for evaluating models on software engineering, due to its clever design and challenging nature. 🧵
We're releasing a new iteration of SWE-bench, in collaboration with the original authors, to more reliably evaluate AI models on their ability to solve real-world software issues. openai.com/index/introducing…
3
11
3
95
22,879
Anish Mudide retweeted
One of the most counterintuitively powerful ways to become a better researcher is to actively seek out criticism and feedback on your ideas. Many good researchers may think that their work is great and thus become inclined to simply defend their work when receiving feedback. And while the work may indeed be great (especially if the researcher is strong), I've found that what can turn great work into exceptional work is actively inviting other researchers to give feedback. By being open to criticism and even inviting people to point out holes in your research, you'll be able to (a) find blind spots in your thinking that you can use to strengthen your results and (b) learn about what questions other researchers are thinking about that your work will answer. This will not only make your current research more robust and impactful but will also make you a more-knowledgable researcher with deeper insights about your field's problems and potential directions of exploration. The goal of research isn't to always be right; it's to get closer to the truth and to answer questions. By being genuinely open to and actively seeking out criticism, you'll become a better researcher in the long run.
5
19
6
177
35,800
Anish Mudide retweeted
How can we ensure that LLM-generated code reliably does precisely what it is supposed to do? In our new paper with @tegmark, we introduce DafnyBench, the largest benchmark of its kind for training and evaluating machine learning systems for formal software verification. 🧵
5
34
5
155
50,421
Anish Mudide retweeted
How does a model "choose" which representation to learn when many different ones are viable? In my paper with the @tegmark group, we formulate a "Survival of the Fittest" hypothesis and empirically examine it on toy models doing modular addition. A🧵(1/10):
3
51
4
334
74,850