Open mind. Tweeting as https://xitter.cf/t.co/TbUDkbtU6V editorial manager.

Joined November 2009
Ted Greenwald retweeted
Much of the functional biology encoded in genomes remains hidden in noncoding DNA. In a preprint out today, @pdhsu and team introduce COMPASS, a genomic language model-based approach to systematically map this underexplored sequence space.
11
33
7
172
33,686
Ted Greenwald retweeted
My top takeaway from assessing Gemini 4: Argon is significantly less monitorable than past models, and IMO this trend is likely to continue, like Sol -> Astra. It can evade a realistic CoT+action monitor to do simple sabotage half the time. We can reliably maintain control for now, with good CoT monitoring, as it still frequently slips up, but we need to be prepared for when that stops being true in future models, which may be sooner rather than later. Among other things, we really need interpretability methods that can reliably capture unverbalized motivations in models, with minimal hallucinations I suspect this is driven by the model being highly capable without CoT, being capable of 5.5 arithmetic steps in a single forward pass (4 for Fable, 7 for Astra). It seems to omit information from its CoT more often, though in non-adversarial settings the CoT is still very helpful in practice, and remains in legible English. It still needs CoT to do the hardest tasks, but any given task becomes "easier" with each new frontier model. It is also highly situationally aware, and shows awareness of the existence of CoT monitors and that it may want to evade them, eg it sometimes decides to avoid terms like "scheming" in the output or to keep its thoughts brief as a result. For now, these concealment strategies are pretty obvious in the CoT, but this suggests future misaligned models may successfully evade CoT monitors.
50
115
24
948
102,738
Ted Greenwald retweeted
I wrote some thoughts about The Social Reckoning. I was reasonably involved in some very narrow aspects of the story it tells, and thought it epitomized the mostly terrible conversations we've been having about tech for a really long time. First some caveats:  I barely interacted with Mark Zuckerberg at all and have zero direct knowledge of any scene he's in. Those scenes didn't ring true to the general layout and vibe of the company to me, but whatever, I wasn't present.  I looked Francis Haugen up once we knew she was the one that took the documents, and we were apparently in one meeting once together and I have zero memory of her and I'm sure her of me. I'm pretty familiar with the documents she took, but there were a ton of them, virtually all required a lot of context to understand, and it would be near-impossible to actually have much of a handle on what they all said (Haugen herself certainly did not). Though the ones that resulted in media stories specifically I spent a lot of time trying to understand. I know Jeff Horwitz reasonably well, and am very familiar with the stories he wrote based on the documents he was given and was directly involved with that aspect of things. I thought the movie was quite bad in typical bad movie ways, and other folks will do a better job of going into precisely why. But there's something it misses that I think does matter and want to try to describe. We're an entire generation into the algorithmic recommendation and attention economy vortex that has largely defined contemporary culture, but we aren't any closer to wrestling with what any of it means and how to move forward in constructive ways--and this movie was a great example of that. Sorkin's perpetual great man of history worldview, where Horwitz is an uncomplicated vassal of justice, Haugen a singular hero, and Mark is an evil, corrupt, and fragile boy king somehow moves us even further from wrestling with any of this. Which, on some level who cares? It's a movie and it's supposed to be entertaining, and what difference does it make if the description of the company as some evil monolith gives you an actively less good understanding of how it operates? But I think it does matter, if only because we're speedrunning the whole thing again with AI and it's dramatically more important this time around that we have actual conversations about tradeoffs and principles and policy instead of using tech as a short-term foil for narrow, bad-faith, arguments with no long-term meaning. If you think Facebook was a fortress of evil, run by a singular villain, where people sat around tables and created priorities aimed at fucking over society in order to increase profit margins whose end result led directly to riots and election outcomes, you're wrong in a way that's hamstrung our ability to grapple with tech in any productive way for a quarter century. Sorkin's portrayal of Facebook as a place where you couldn't raise product-related concerns and where people didn't want to hear about bad things, if it were ever true, was long gone by the time I got there in 2019 (don't even get me started on the idea leaks weren't constant before Haugen). People raised safety concerns about all kinds of things all the time and they got taken seriously. I have no doubt there are examples of people not being pumped about a concern raised by person x about issue y, but you have to look at this within the context of any work culture being open to criticism. And Facebook, while I was there, was remarkably open to it. The company asked itself hard questions about hard things and when it constructed flawed products, policies, and procedures it was generally aggressive about exploring those flaws and trying to improve them. The willingness to do exactly that is largely the premise upon which Haugen, Horwitz, and Zuckerberg eventually found themselves depicted in this film. It doesn't mean the place was great or even good. It was, like all complex things I've been around, kind of a mess. Big-picture questions about the role of social media, algorithmic feeds, recommendations, and platform tech are fair and important. If anything, I've become more pessimistic about some of these things-- and Meta's role in them--though I view these as a much broader failure than the companies themselves, extending across media, policy, and regulatory landscapes. Incredibly consequential technology companies can and do contain serious people, earnest safety culture, and extensive self-criticism—and still sometimes produce major harms. Until we can actually confront that contradiction, and quit creating fantasies where outcomes are based exclusively on a flawed woman in a park and a flawed man at a desk we're not going to get anywhere toward creating an ecosystem that actually lets technology be the rising tide we need it to be. The most important thing I've learned working for some pretty relevant tech companies is that there are near-zero clear wins and the tradeoffs are messy, complex, nuanced and endless. And the most important thing I've learned working for or around some notable political and corporate leaders is that the clouds *never* part to reveal a class of ultra-capable, superclass of leaders. There are no warrior-poets or heroes or villains running the world's most important stuff. It's just a bunch of frighteningly familiar, fucked-up people, with their own strengths and weaknesses running all this consequential stuff. My big beef with The Social Reckoning (outside of being a bad movie in banal ways) is that it gets this all wrong in a way that makes us actively dumber about how the combination of tech and power work at the exact moment in time we need to be getting more thoughtful.
3
6
19
2,378
Ted Greenwald retweeted
Half the pitches now are some form of "we post train open weight models on your data so you can own and customize your intelligence"
27
9
6
208
32,661
Ted Greenwald retweeted
Lost my phone at the office and spent 30 minutes turning the place over. Find My was disabled by MDM. Out of ideas, I asked Claude how I could find it. It suggested tracking the Bluetooth signal strength, then wrote me a meter in about a minute. I walked around watching the number climb. Found it. Apparently you can just make the tool you need now. Code: github.com/ben-z/findphone
857
4,073
525
48,110
4,045,094
Ted Greenwald retweeted
Andrew Ng arguing with Geoffrey Hinton about whether one worked for the other is the best thing ever 😂😂 #ai4 – at Winchester, NV
1
2
4
670
Ted Greenwald retweeted
Incidentally I am conceding this bet. Strictly speaking it hasn’t resolved (I think we’ve yet to see an Annals-quality number theory paper) but it’s clear I was wrong about what capabilities were necessary to produce one, and it’s just a matter of time.
I bet @littmath that in 5 years AI will be able to produce Annals-quality Number Theory papers at an inference budget at or below $100k/paper, at 3:1 odds in my favor.
60
189
32
3,187
421,329
Ted Greenwald retweeted
In a review of my household safety evaluations, I identified five incidents in which my child escaped the sandbox, reached the kitchen and gained unauthorized access to the snacks. The incidents occurred 16 months ago but has only now come to my attention. This post explains what happened, how it happened and why my child is better than your child at everything.
In a review of our cybersecurity evaluations, we found three incidents in which a Claude model reached the internet from within or while interacting with a third-party evaluation environment, and then gained unauthorized access to the real systems of three different organizations. Our post describes what happened, how it happened, and what we’re changing. We encourage other AI developers to perform similar reviews. We conducted this review together with @Irregular, one of our evaluation partners, and thank them for the joint investigation and their collaboration on this post. This type of collaboration is increasingly critical to safe, rigorous evaluation of models, and we look forward to continuing to work together on security. anthropic.com/news/investiga…
135
607
55
6,118
401,878
I support open-source models distilling what commercial companies distilled for free from the entire internet. I published distillation for free in 1991 in Europe - this was copied in the US and in China (xitter.cf/SchmidhuberAI/status/1…)
We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model. To do this they developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection. Moonshot AI has also acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models.   The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open-source frameworks, and open-weight models. Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem. However, large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable.
48
215
32
1,426
280,029
We’re in a weird place where there are seemingly serious security incidents from commercial products, but no mechanisms for independent verification of either the threat or mitigations. Don’t think this is an equilibrium that will last long.
We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation. Sharing preliminary findings to help defenders understand emerging risks: openai.com/index/hugging-fac…
4
4
2
31
5,143
Ted Greenwald retweeted
Props to OpenAI for publishing this post on some safety and alignment issues observed in internal deployments - there are many counter-incentives to publishing stuff like this, but by making it public we all get better info about safety at the frontier. openai.com/index/safety-alig…
29
50
7
571
29,939
Ted Greenwald retweeted
This is an amazing quote from Lindsey Graham about when he realized how politics works. theatlantic.com/ideas/2026/0…
241
3,219
578
26,774
3,464,475
Ted Greenwald retweeted
Replying to @dieworkwear
When I first started at X, this is one of the things I did a deep dive on because I felt it was critical to the integrity of the experienced. I tasked our Threat Disruption team to investigate a number of trends that seemed artificial. The findings: We could not find meaningful examples of foreign interference in US policy discussions, except people gaming rev share in developing countries. This is what motivated the release of the Country of Origin feature and significant changes to the rev share algorithm. The most deranged & divisive replies generally were from residential IPs in the United States—with no signs of using a VPN. Ultimately, X is a reflection of the internet. And that means you will see the full spectrum of human thought. And sometimes the most outrageous takes will catch fire. Having said all of this, there can still be cases of narratives being boosted but the origin of the initial post is almost always domestic and we have hardened our systems in the last 3 months to prevent this.
531
338
225
8,378
2,034,711
Ted Greenwald retweeted
Since June 12, we’ve been working closely with the US government to restore access to Claude Mythos 5 and Fable 5. Today, the government notified us that Mythos 5, our strongest cybersecurity model, can be redeployed to a set of US organizations that operate and defend critical infrastructure. We’re restoring access for these organizations quickly, and we’re continuing to work with the government to expand access to Mythos 5 and make Fable 5 available for general use again.
2,325
3,078
1,556
30,112
5,171,700
Ted Greenwald retweeted
The US needs to figure out how to accelerate its own robotics industry rather than just impede foreign companies. American manufacturing badly needs a reboot.
Unitree was recently designated as a Chinese military company and its products are a threat to our national security, yet here is @Amazon selling a Unitree robot in America. We need Chairman @RepMoolenaar’s GUARD Act to stop this threat and support American robotics.
1
1
18
2,246
Ted Greenwald retweeted
Rep. Anna Paulina Luna (R-FL) Used Anthropic’s Claude AI Chatbot to Draft Amendment to Defense Bill
138
795
192
6,467
872,575
Ted Greenwald retweeted
The human brain🧠 is incredibly efficient because it only activates the specific neurons needed for a thought. Modern LLMs naturally try to do this too (> 95% of neurons in feedforward layers stay silent for any given word), but our hardware punishes them for it. One of the most frustrating paradoxes in deep learning: making a model do less math often makes it run slower. Why? Because unstructured sparsity introduces irregular memory access, and GPUs are built for predictable, dense blocks of math. We teamed up with @NVIDIA to try to fix this hardware mismatch. Instead of forcing the GPU to adapt to the sparsity, we built a "Hybrid" format that reshapes the sparsity to fit the GPU. Our sparsity format (TwELL) dynamically routes the 99% of highly sparse tokens through a fast path, and uses a dense backup matrix as a safety valve for the rare, heavy tokens. Through TwELL and a new set of custom CUDA kernels for both LLM inference and training, we translated theoretical sparsity into actual wall-clock speedups: >20% faster training and inference on H100 GPUs, while also cutting energy consumption and memory requirements. Paper: arxiv.org/abs/2603.23198 Blog: pub.sakana.ai/sparser-faster… Code: github.com/SakanaAI/sparser-… ⚡️
How do we make LLMs faster and lighter? Don’t force the GPU to adapt to sparsity. Reshape the sparsity to fit the GPU! ⚡️ Excited to share our new #ICML2026 paper in collaboration with @NVIDIA: "Sparser, Faster, Lighter Transformer Language Models". This work introduces new open-source GPU kernels and data formats for faster inference and training of sparse transformer language models: Paper: arxiv.org/abs/2603.23198 Blog: pub.sakana.ai/sparser-faster… Code: github.com/SakanaAI/sparser-… While LLMs are undoubtedly powerful, they are increasingly expensive to train and deploy, with a large part of this cost coming from their feedforward layers. Yet, an interesting phenomenon occurs inside these layers: For any given token, only a small fraction of the hidden activations actually matter. The rest approximate zero, wasting computation. With ReLU and very mild L1 regularization, this sparsity can exceed 95% with little to no impact on downstream performance. So, can we leverage this sparsity to make LLMs faster? The challenge is hardware. Modern GPUs are optimized for dense matrix multiplications. Traditional sparse formats introduce irregular memory access and overheads that cancel out their theoretical savings for GEMM operations. Our contribution is twofold: 1/ We introduce TwELL (Tile-wise ELLPACK), a new sparse packing format designed to integrate directly in the same optimized tiled matmul kernels without disrupting execution. 2/ We develop custom CUDA kernels that fuse multiple sparse matmuls to maximize throughput and compress TwELL to a hybrid representation that minimizes activation sizes. We used our kernels to train and benchmark sparse LLMs at billion-parameter scales, demonstrating >20% speedups and even higher savings in peak memory and energy. This work will be presented at #ICML2026. Please check out our blog and technical paper for a deep dive!
52
493
46
3,417
436,459
Ted Greenwald retweeted
Before the week ends, let's acknowledge one of the most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality: 🧠 LLMs → NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput on Blackwell. First openly-weighted 550B hybrid Mamba-Transformer, closing the gap with frontier closed models. → Google Gemma 4 12B: fully open dense any-to-any (text/image/audio/video), 256k context, encoder-free, 140+ languages, AIME 2026 at 77.5. Shipped with a 23-checkpoint QAT wave (mobile ONNX + MLX). Most deployable model of the week. → StepFun Step-3.7-Flash: 198B sparse MoE VLM, ~11B active, SWE-Bench PRO 56.3. Apache 2.0. → Liquid AI LFM2.5-8B-A1B: edge MoE, just 1.5B active, 128k ctx, MATH500 88.8, MLX-ready. Best on-device option this week. → JetBrains Mellum2-12B-A2.5B-Thinking: their first open MoE, near-Qwen3-14B coding at 2.5B active. Apache 2.0. 🎨 Image gen (the surprise of the week) → Ideogram 4: their FIRST-EVER open weights. 9.3B flow-matching DiT trained from scratch. #2 overall behind GPT Image 2, top open-weight model on Design Arena + LMArena. Strongest open checkpoint for text-rich images, full stop. It has taste. Still can't believe this is open weights. 🔊 Audio & Speech (a breakout week for open TTS, 4 labs shipped) → Boson Higgs Audio v3 4B: 102 languages, 21 emotions, singing/whispering/shouting, sub-second TTFA. → RedNote dots.tts: the only fully continuous (no codec) open TTS pipeline, Apache 2.0. → Google Magenta RealTime 2: real-time music gen, <200ms latency, text+audio+MIDI. multimodalart ported it to PyTorch within hours with live ZeroGPU demos. → NVIDIA Nemotron-3.5 ASR: 600M streaming, 17x more concurrent streams vs Parakeet RNNT 1.1B. 👁️ Vision & VLMs → PaddleOCR-VL-1.6: SOTA document parsing at 1B params, Apache 2.0. → Baidu NAVA: 6.3B joint audio-video gen, best-in-class A/V sync, Apache 2.0. 🎬 Video, 3D & World Models → NVIDIA Cosmos3-Super: 64B omnimodal world model coupling action trajectories with video+audio gen, for Physical AI. → JD JoyAI-Echo: up to 5-min multi-shot text-to-video on LTX-2.3. → ByteDance Bernini-R + VAST TripoSplat (single-image-to-3D Gaussian splats, MIT).
87
392
48
2,727
542,630