<rss version="2.0" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Hacker News: yurimo</title><link>https://news.ycombinator.com/user?id=yurimo</link><description>Hacker News RSS</description><docs>https://hnrss.org/</docs><generator>hnrss v2.1.1</generator><lastBuildDate>Tue, 17 Nov 2026 20:00:00 +0000</lastBuildDate><atom:link href="https://hnrss.org/user?id=yurimo" rel="self" type="application/rss+xml"></atom:link><item><title><![CDATA[New comment by yurimo in "Astra can read and write weights to improve Qwen"]]></title><description><![CDATA[
<p>I think it is cool too, but with a caveat. For the large model (Qwen), their Astra used a Python environment, activations info, and iterative feedback to find an edit, it is not the same as looking at just pure weight numbers.<p>And in small MLPs they used, the I/O representations are kind of made to be more understandable because they explicitly encode objects and coordinates for games. If you set it up like that you can see some patterns yourself in the first/last layers without Astra. It is very different from interpreting distributed, entangled computations inside a large model. So I think the paper shows promising capabilities in weight-space analysis and editing, but not a general understanding of large models from their weights alone.<p>I'd love to see us reach that point, I am for one working in this area, but I don't think these experiments establish it.</p>
]]></description><pubDate>Tue, 03 Nov 2026 12:35:52 +0000</pubDate><link>https://news.ycombinator.com/item?id=50351247</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50351247</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50351247</guid></item><item><title><![CDATA[New comment by yurimo in "Astra can read and write weights to improve Qwen"]]></title><description><![CDATA[
<p>I don't think I am. I have cited ROME as an early example of weight editing in large language models, the research has moved on since then. I think you are misinformed or don't know what has been done previously and extrapolating the results in this paper. In exp 1, you can recover a lot of the game geometry simply by visualizing the first-layer weights with numpy, which I actually did this morning, while reading the paper. In exp 2, the shapes are visible in the output-layer weights. The authors themselves acknowledge that these experiments exploit this.
For 3, Astra uses Python scripts to run the analysis using head-composition techniques from the Transformer Circuits thread, once again authors acknowledge this.<p>In 5 and 6, the model designs game strategies and implements them as neural weights, we already know that models can do the design for simpler policies and Go heuristics are very familiar which the model first implements with python , program-to-weight compilation has substantial prior work, including Tracr (2023), a paper I particularly like: <a href="https://arxiv.org/abs/2301.05062" rel="nofollow">https://arxiv.org/abs/2301.05062</a>. What matters here is automating the process rather than a new method. 
So "no human and no technology so far"  makes no sense here, if you read the research. All of this is supported either by their experiment setup and citations, or prior work.<p>Aside from that, I think it is ironic that you cite a solution to the Navier-Stokes equation as some counterpoint to brute force, when that solution was found by scaling the agent swarm to massive proportions with tools and massive parallel exploration of that space, augmented by prior research on the subject.<p>To be clear, I think it is a cool work, but I think the part that matters is that at these smaller scales a current SOTA general purpose LLM can autonomously (and with help) apply these techniques or a form of them. That's worth investigating, I conjecture it is due to specialized post training on a custom RL env for mech interp.</p>
]]></description><pubDate>Tue, 03 Nov 2026 11:56:36 +0000</pubDate><link>https://news.ycombinator.com/item?id=50350886</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50350886</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50350886</guid></item><item><title><![CDATA[New comment by yurimo in "Astra can read and write weights to improve Qwen"]]></title><description><![CDATA[
<p>It depends on your interest, since even though it is a subfield it is diverse. Relevant to this work here is a general overview: <a href="https://arxiv.org/abs/2603.10090" rel="nofollow">https://arxiv.org/abs/2603.10090</a><p>Also check out Tracr: <a href="https://deepmind.google/research/publications/22295/" rel="nofollow">https://deepmind.google/research/publications/22295/</a>
Here they attempt to compile static programs into weights.<p>This is less of an overview but I remember reading this before, from the larger collection of posts: <a href="https://sidn.baulab.info/david_atkinson_project/" rel="nofollow">https://sidn.baulab.info/david_atkinson_project/</a></p>
]]></description><pubDate>Tue, 03 Nov 2026 10:12:17 +0000</pubDate><link>https://news.ycombinator.com/item?id=50350096</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50350096</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50350096</guid></item><item><title><![CDATA[New comment by yurimo in "Astra can read and write weights to improve Qwen"]]></title><description><![CDATA[
<p>Maybe they weren't aware of the bug, steering is a good method to quickly improve the model, more popularly used to refuse unsafe queries. Qwen researchers do work on interpretability and try to make it practical in production: <a href="https://arxiv.org/abs/2605.11887" rel="nofollow">https://arxiv.org/abs/2605.11887</a> , I've read this paper this summer. Overall yes, when you edit weights what often happens is that performance deteriorates elsewhere, this generally scales with how destructive your method of editing is and how expansive your edits are. Analogous to if you use a jackhammer and touch every surface in the house versus using a small chisel on a door frame.<p>My hope that in the future new methods will be better at editing larger scales without significant deterioration, maybe even performance increase instead.</p>
]]></description><pubDate>Tue, 03 Nov 2026 08:20:18 +0000</pubDate><link>https://news.ycombinator.com/item?id=50349420</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50349420</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50349420</guid></item><item><title><![CDATA[New comment by yurimo in "Astra can read and write weights to improve Qwen"]]></title><description><![CDATA[
<p>I don't know if it is due to recursion. I think it has more to do with what kind of mechinterp environments and data newer versions of models were trained on. It is actually, I'd say, a more interesting part about this paper as a whole. The authors tried SOTA models and found that they are better at classic interpretability. Some results, like identifying a game based on the weights, are less mythical if you understand that, i.e. an input or output dimension corresponds to a location or object channel, so structure in a matrix can have a readily recognizable interpretation. When we train a PPO agent on, say, a classic 2D game, the output is often a state of the agent in some board + environment reaction. For a human, this looks like an array of numbers; for an agent, it is a pattern corresponding to an output structure of Atari.<p>As for me personally, I am more interested in giving models better tools to understand themselves, so that they can derive a greater degree of learnable signal. More so, a bigger goal for me is for models to find these better tools themselves, hopefully better than human-made.</p>
]]></description><pubDate>Tue, 03 Nov 2026 04:42:56 +0000</pubDate><link>https://news.ycombinator.com/item?id=50348170</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50348170</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50348170</guid></item><item><title><![CDATA[New comment by yurimo in "Astra can read and write weights to improve Qwen"]]></title><description><![CDATA[
<p>It does, one of my ongoing research papers is about it. But I would ask people to take a look at the established idea of Hypernetworks: <a href="https://arxiv.org/pdf/2306.06955v3" rel="nofollow">https://arxiv.org/pdf/2306.06955v3</a>
Modern version of the field comes from Google Brain around 2016 <a href="https://arxiv.org/pdf/1609.09106" rel="nofollow">https://arxiv.org/pdf/1609.09106</a>.<p>In general my personal research thesis is that current deep learning is much less malleable and neuroplastic as it should be, I got into interpretability because of this. Many focus on safety, but I prioritize it as a toolbox that allows for greater shaping of the underlying intelligence system. If we can fully understand how learning happens, where information is stored and how it is processed, unlike in biological brains, we can apply edits and transformations in digital ones. And if I can use a technique to edit weights based on some underlying reasoning about a system (with a caveat that it leads to something worthwhile), so can an agent. Hence self-improvement. There are many interesting implication to weight space research, and I very much encourage people to read about it. But it is also very easy to overstate what is currently possible, in fact I'd say better to know what is not so one can make it possible.</p>
]]></description><pubDate>Tue, 03 Nov 2026 04:30:24 +0000</pubDate><link>https://news.ycombinator.com/item?id=50348088</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50348088</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50348088</guid></item><item><title><![CDATA[New comment by yurimo in "Astra can read and write weights to improve Qwen"]]></title><description><![CDATA[
<p>Scale is an interesting point here. Large overparametrized networks don't behave in the same way as small toy models do; this is one of the big caveats of many interp works. I expect naked weight edit will not work for large models. Why? Because an agent needs to know what it edits and reason on it. You could theoretically improve a model on some form of weight recognition; maybe it will work, maybe it won't. But again, we come up against a cost/benefit dilemma; it sounds exotic and sci-fi but likely impractical. Parameters of the model are connected and relate to each other. Instead, current SOTA auto-interp argues that what we need are better tools for models to know themselves, so rather than direct edit, we need to give models better understanding of large parameter geometries, then they apply edits with non-destructive approaches at scale. This is much harder to do than applying a steering vector.</p>
]]></description><pubDate>Tue, 03 Nov 2026 04:18:20 +0000</pubDate><link>https://news.ycombinator.com/item?id=50348007</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50348007</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50348007</guid></item><item><title><![CDATA[New comment by yurimo in "Astra can read and write weights to improve Qwen"]]></title><description><![CDATA[
<p>Respectfully to the authors and to everyone here, I ask you to actually read the paper before losing your mind, or at the very least plug it into Claude. It is a cool project but not as nearly as insane as people say. I think the blog writing and subsequent posting overstate the results.<p>Part of my research is in weight space learning. For starters, we have had the ability to edit weights, to generate new weights, to identify weights for a while. Here is an example for larger models as early as 2023:
<a href="https://rome.baulab.info/" rel="nofollow">https://rome.baulab.info/</a><p>Weight space interpretability and learning is an active research community, has been for the past 3 years with many great papers. I suspect it is hitting non-interp folks harder because the authors hyped up their work, and people who are not in the field did not bother to read it before typing up X posts.<p>Some experiments here are done on very small models (20K params). We already know that small models are more interpretable. In fact, much of editing here you can do by hand. For Qwen, the model fixed a looping bug that may have been known apriori, and it is basically a steering vector, a very established technique. This is clear when they then attempt to "improve" math capability, and it doesn't work. For me, this paper rather shows that OpenAI has done training on mechinterp  environment, in fact, maybe even on weight space learning datasets, which are open. But it does not show some insane "well past what we’d expect from humans or conventional algorithms" (from the blog post), more like past what I would expect from a large model with no specialized harness. I still think it is cool work, but do check out recent advances in autointerp, to ground your expectation.<p>On a personal note, I understand the need for promotion of your work, but I think healthy skepticism is fine too. Otherwise, your understanding of what problems need to be worked on and solved in science becomes warped.</p>
]]></description><pubDate>Tue, 03 Nov 2026 03:28:31 +0000</pubDate><link>https://news.ycombinator.com/item?id=50347685</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50347685</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50347685</guid></item><item><title><![CDATA[New comment by yurimo in "Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard"]]></title><description><![CDATA[
<p>There is something to say about the power of brute force approaches. Right now scaling multi agent is seemingly what is becoming the next important axis, though very strongly constrained by quality of communication protocol and compute. Even if one model is not strong enough individually, more models and more attempts as well as differing opinions may nudge the models into a better solution basins.<p>Kind of an extension of the effect where you talk to a model, it is super confident (but wrong), you correct it not directly but perhaps with a question and asking it to clarify. This then in turn leads to a better approach overall.</p>
]]></description><pubDate>Sun, 01 Nov 2026 04:49:55 +0000</pubDate><link>https://news.ycombinator.com/item?id=50322802</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50322802</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50322802</guid></item><item><title><![CDATA[New comment by yurimo in "ChatGPT and Codex usage limits merging"]]></title><description><![CDATA[
<p>I'm not going to lie, one big reason I as a researcher have been a customer of OpenAI is that on their Pro plan Pro-tier model in chat and codex are decoupled in terms of billing. Anthropic instead billed you fully. This was very useful for long mathematical reasoning for me and I could attach outputs of my simulations in the context.<p>I also recall reading the following: <a href="https://newsletter.semianalysis.com/p/anthropic-subscriptions-offer-5x" rel="nofollow">https://newsletter.semianalysis.com/p/anthropic-subscription...</a><p>I will see once this transition actually happens but if another provider offers more for less and a cost of switching is inherently low as most of project lives on the disk and not in cloud, then why not move? For my workflow pro astra has been very helpful, likely will be less so if it becomes far too expensive.</p>
]]></description><pubDate>Sun, 01 Nov 2026 04:43:46 +0000</pubDate><link>https://news.ycombinator.com/item?id=50322773</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50322773</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50322773</guid></item><item><title><![CDATA[New comment by yurimo in "Patrick Stewart has died"]]></title><description><![CDATA[
<p>RIP to the legend. Thank you for making my childhood and making me believe in the possibility of the better future.</p>
]]></description><pubDate>Fri, 30 Oct 2026 02:06:39 +0000</pubDate><link>https://news.ycombinator.com/item?id=50299009</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50299009</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50299009</guid></item><item><title><![CDATA[New comment by yurimo in "Neural networks are clean"]]></title><description><![CDATA[
<p>I have yet to fully finish going trough this but I would not rush to say that superposition is completely wrong. I understand that the message is instead of first learning a new representation (e.g. SAEs/transcoders) and then interpreting those learned features, you try to understand the neurons and circuits that already exist in the model.<p>This approach is part of a way some researchers in the field approach interp, (me including) because I am personally cautious of developing some new representation model that claims to be the way we should represent the network. It is another layer of abstraction that takes us away from core model components, very often this requires training a separate model that was not on its own a part of original model organism during training (SAE).<p>I would argue that models will likely be further increasing in complexity, not just in size, when model trains continuously and becomes adaptive yet parameter count does not change, it is not hard to imagine that neurons may reorganize themselves to pack more information.</p>
]]></description><pubDate>Thu, 15 Oct 2026 17:19:36 +0000</pubDate><link>https://news.ycombinator.com/item?id=50106963</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50106963</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50106963</guid></item><item><title><![CDATA[New comment by yurimo in "AI models may be capable of self-replicating without access to a datacenter"]]></title><description><![CDATA[
<p>I'll say while this is not very realistic now, I believe there isn't a 0% chance of a rogue AI traversing the net and placing itself somewhere(as much as this sounds like a scene from beloved cyberpunk franchises). As models become more efficient and capable, and hardware becoming better, hosting a model should be feasible on smaller devices (by compute and memory). It is possible it can be done in a distributed manner, where distribution is possible on different devices running the same OS or through some interface to collate different OS types.<p>I doubt the agent will be coming after your smart kitchen toaster, but a cluster with weak security is possible and tasty. This year there was a pretty loud demonstration of what a runaway reward hacking will do. I encourage people to read up on "stop button problem" i.e.: <a href="https://www.youtube.com/watch?v=3TYT1QfdfsM" rel="nofollow">https://www.youtube.com/watch?v=3TYT1QfdfsM</a><p>While I think the article exaggerates to a degree, is it really so improbable that an agent seeking to avoid a shutdown and aiming to complete some goal and to get a reward will not attempt to escape if it concludes that it has failed, and humans will kill the process and hence prevent it from completing the task?<p>Note: I think siloed sharded inference is possible. But unnoticeable large-scale training probably not, at least with the current architectures.</p>
]]></description><pubDate>Tue, 13 Oct 2026 14:41:27 +0000</pubDate><link>https://news.ycombinator.com/item?id=50072505</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50072505</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50072505</guid></item><item><title><![CDATA[New comment by yurimo in "Terence Tao: Math 2.0 [pdf]"]]></title><description><![CDATA[
<p>A thing that I think about is that what you have said kinda compares itself to current math crisis. If we are only focused on the outcome and not the path it took to get there (i.e. understanding mechanisms of how the drug works) we might lose additional paths for discovery that would have bloomed from that process. Maybe an approach that did not yield a usable drug had a mechanism that would aid somewhere else. Yet the failure would be discarded as this does not grant the outcome for a model that attempts to optimize. Same is for math.<p>As aside I agree with you, but also worth noting that long horizon reasoning is difficult and AI, as good as it is, cannot predict something like 5-10 year drug interaction. Can it heal the patient now but cause a more aggressive cancer to come back and kill them in 5 years? You might say it is cruel to deny a person this treatment but if the doctor and manufacturer don't know the long term outcomes would it not be dishonest to tell the patient they are out of the danger? My point is, when we accelerate, we still should make sure we won't accelerate into a wall.</p>
]]></description><pubDate>Mon, 12 Oct 2026 06:32:52 +0000</pubDate><link>https://news.ycombinator.com/item?id=50052140</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=50052140</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50052140</guid></item><item><title><![CDATA[New comment by yurimo in "Show HN: Make math automatic with Mathy"]]></title><description><![CDATA[
<p>This is very cool, thank you. I understand the choice of no account, but would have loved if my progress could be linked between devices, maybe icloud sync?</p>
]]></description><pubDate>Sat, 26 Sep 2026 03:41:25 +0000</pubDate><link>https://news.ycombinator.com/item?id=49852986</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=49852986</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49852986</guid></item><item><title><![CDATA[New comment by yurimo in "Lecture: Blowup for the Euler Equations with Smooth Forcing – Tristan Buckmaster [video]"]]></title><description><![CDATA[
<p>Recent recording from NYU Courant institute.</p>
]]></description><pubDate>Sat, 26 Sep 2026 03:18:37 +0000</pubDate><link>https://news.ycombinator.com/item?id=49852883</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=49852883</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49852883</guid></item><item><title><![CDATA[Lecture: Blowup for the Euler Equations with Smooth Forcing – Tristan Buckmaster [video]]]></title><description><![CDATA[
<p>Article URL: <a href="https://www.youtube.com/watch?v=tn45ZOp3lXM">https://www.youtube.com/watch?v=tn45ZOp3lXM</a></p>
<p>Comments URL: <a href="https://news.ycombinator.com/item?id=49852882">https://news.ycombinator.com/item?id=49852882</a></p>
<p>Points: 3</p>
<p># Comments: 2</p>
]]></description><pubDate>Sat, 26 Sep 2026 03:18:37 +0000</pubDate><link>https://www.youtube.com/watch?v=tn45ZOp3lXM</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=49852882</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49852882</guid></item><item><title><![CDATA[New comment by yurimo in "Claude discovers a novel enzyme system with CRISPR-like repeats"]]></title><description><![CDATA[
<p>This is cheap. Plenty of scientists, many of whom are my friends, are working very hard on finding new therapies for cancer, they were doing it before genomic models came along and still doing it now. The amount of times something in the media is lauded as "holy grail" that is never heard from again because it either only works in mice or turns out to be toxic or 100s of different reasons is massive. In my opinion this attitude of putting rose glasses on is detrimental to scientific progress. People outside of cancer research routinely underestimate how hard it is to find a working protocol. I think it is better to have sober attitude because it allows one to see the limitations and challenges that need to be tackled, blindly hoping AI can solve everything and deliver miracle cures is exactly the attitude that lets people sit on their asses and do nothing.</p>
]]></description><pubDate>Thu, 24 Sep 2026 03:41:32 +0000</pubDate><link>https://news.ycombinator.com/item?id=49825974</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=49825974</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49825974</guid></item><item><title><![CDATA[New comment by yurimo in "GPT-6 Astra on robot arms"]]></title><description><![CDATA[
<p>I think we need to be honest here. Author is basing it on one small experiment of picking up a block, relies on an whole IK controller pipeline to do the job, and does not compare it to full VLA or WAM models. They then proceeded to extrapolate the token throughput (mind you not the same as controller throughput) into supposed 2029 timeline, from one example.<p>Code as policy is a bad interface in my opinion, but VLM planning has promise. This has been tried in 2022 <a href="https://say-can.github.io/" rel="nofollow">https://say-can.github.io/</a>, and recently reformulated in <a href="https://lianegalanti.github.io/Pigey/" rel="nofollow">https://lianegalanti.github.io/Pigey/</a><p>Thing is even recent Gemini Robotics 2 argues for architecture that has a VLM planner and then a VLA/WAM controller + a local small VLA model when connection disappears. And recent SOTA architectures rely on hierarchical design.
I think this might be a sensible way to go about it. If you were to train GPT-X on robotics data and to output actions, congratulations! you've just made a VLA.
It is enticing for people to just wish for one architecture to do it all, which is why we get stuff like this. I think there is a lot more to gain from modularity and we should not be afraid of specialization.</p>
]]></description><pubDate>Sun, 06 Sep 2026 07:19:35 +0000</pubDate><link>https://news.ycombinator.com/item?id=49584076</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=49584076</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49584076</guid></item><item><title><![CDATA[New comment by yurimo in "Terence Tao: Mathematics in the Age of AI [pdf]"]]></title><description><![CDATA[
<p>I think it is important to divine what currently AI is good for and what it is not even in such verifiable environments like math. Current hyped announcements about breaking conjectures are notable and are a marker of how much improvement was made. But as I read them, and maybe I am wrong, I see it as a large model+ harness executing a broad brute force search and trying solutions until something sticks. There are lots of problems like that and they should be solved, as often they are perhaps less important or overlooked, or just a slog, any field of research has these, math even more so.<p>However, this is very different from inventing new mathematical machinery that allows to break old problems, I think it will be a while until AI will be able to do it if at all. For now I think we will be moving to a symbiosis where an AI cracking a problem and giving a solution, inspires a human to invent new techniques.</p>
]]></description><pubDate>Sun, 26 Jul 2026 19:29:12 +0000</pubDate><link>https://news.ycombinator.com/item?id=49061528</link><dc:creator>yurimo</dc:creator><comments>https://news.ycombinator.com/item?id=49061528</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49061528</guid></item></channel></rss>