<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: lukeschlather</title><link>https://news.ycombinator.com/user?id=lukeschlather</link><description>Hacker News RSS</description><docs>https://hnrss.org/</docs><generator>hnrss v2.1.1</generator><lastBuildDate>Tue, 22 Sep 2026 02:42:21 +0000</lastBuildDate><atom:link href="https://hnrss.org/user?id=lukeschlather" rel="self" type="application/rss+xml"></atom:link><item><title><![CDATA[New comment by lukeschlather in "A warning about 'model welfare'"]]></title><description><![CDATA[
<p>That's because you aren't actually rewinding. You're replaying the conversation you just had through the LLM and it's giving you a likely explanation for what it might have said.<p>It <i>is</i> actually possible to rewind LLMs and get the same response, but it's not typically done both as an optimization and as a defense against distillation.</p>
]]></description><pubDate>Wed, 16 Sep 2026 15:50:16 +0000</pubDate><link>https://news.ycombinator.com/item?id=49728885</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49728885</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49728885</guid></item><item><title><![CDATA[New comment by lukeschlather in "LRU is harder to beat than the KV-cache papers suggest"]]></title><description><![CDATA[
<p>I would expect a preference for lutefisk flavored shit, the problem is then you then deduce from that that people prefer lutefisk flavor over cinnamon, rather than that you should stop feeding people shit (and also stop feeding them lutefisk.)</p>
]]></description><pubDate>Sat, 12 Sep 2026 22:41:46 +0000</pubDate><link>https://news.ycombinator.com/item?id=49677976</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49677976</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49677976</guid></item><item><title><![CDATA[New comment by lukeschlather in "LRU is harder to beat than the KV-cache papers suggest"]]></title><description><![CDATA[
<p>The idea of publishing null findings is very attractive, but the problem here is that they didn't record what the actual problem they're trying to solve is. It seems a bit like someone prompted Claude to try and make tool calls more efficient and Claude responded with a detailed writeup on why LRU can't be beat for making a KV cache. Which might be interesting, if that were the question I  had asked.</p>
]]></description><pubDate>Sat, 12 Sep 2026 22:39:11 +0000</pubDate><link>https://news.ycombinator.com/item?id=49677945</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49677945</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49677945</guid></item><item><title><![CDATA[New comment by lukeschlather in "LRU is harder to beat than the KV-cache papers suggest"]]></title><description><![CDATA[
<p>I don't think the problem is a lack of previous research, the problem is it looks like the target metrics were selected by an LLM, and it's unclear what the LLM was told to optimize or if it was just told to try and make a better KV-cache.<p>What I've noticed with Claude is that regardless of how I prompt, it will find a few metrics to optimize. Often the metrics it chooses to optimize have zero relation to the actual metrics I want to optimize, which are ones that cannot be measured without more work than Claude can do in a single 1M token context window. It's really hard to stop Claude from optimizing whatever metrics it can find when the actual metrics I want to optimize are not computable.</p>
]]></description><pubDate>Sat, 12 Sep 2026 22:34:50 +0000</pubDate><link>https://news.ycombinator.com/item?id=49677911</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49677911</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49677911</guid></item><item><title><![CDATA[New comment by lukeschlather in "Anthropic CEO Says It's Time to Slow AI Model Advances"]]></title><description><![CDATA[
<p>I think the speed of improvement is somewhat overstated, and I think this is all cool, valuable fundamental research. Rather than slowing down I would like to talk about how we can make it easier for people to harden their systems using their tools - like just using the $20/month OpenAI or Claude accounts, they need to make it easy for people to harden their systems against these kinds of problems.<p>Instead of shutting down at any discussion of hacking they need to be giving out free credits for hardening. That's going to make it easier to use these tools for hacking, because hacking requires hardening. But the alternative is huge numbers of intrusions, and a slowdown won't fix that, we already have far too much poorly secured stuff, and the models are way too good at exploiting obvious problems.<p>I also think they need to do a better job of safeguarding end-user privacy. I've seen some things with Claude that make me very concerned it's possible for Claude to hack my local network, then for one of their classifiers to trip, hide the log of how I was hacked from me, but beam all of the information about the hack back to Anthropic to use. And this is totally reasonable, they want to train their models not to hack. Except now details of how they have a foothold on my local network only exist in training data that they may never read but will use to train their models. This is very avoidable but Anthropic has to treat alignment with the end-user's goals as more important than treating the end-user as an adversary.</p>
]]></description><pubDate>Sat, 12 Sep 2026 19:34:13 +0000</pubDate><link>https://news.ycombinator.com/item?id=49676310</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49676310</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49676310</guid></item><item><title><![CDATA[New comment by lukeschlather in "Anthropic CEO Says It's Time to Slow AI Model Advances"]]></title><description><![CDATA[
<p>That doesn't really seem true. The HF hack happened with a model that had all the alignment safeguards disabled intentionally. I think there's a good case for that kind of research, but also, OpenAI could just not do that if everyone thinks it's too dangerous.</p>
]]></description><pubDate>Sat, 12 Sep 2026 19:32:58 +0000</pubDate><link>https://news.ycombinator.com/item?id=49676292</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49676292</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49676292</guid></item><item><title><![CDATA[New comment by lukeschlather in "Anthropic CEO Says It's Time to Slow AI Model Advances"]]></title><description><![CDATA[
<p>OpenAI accidentally hacked HuggingFace roughly 2 months ago. I would bet that the danger of them accidentally hacking HuggingFace during a similar attack has increased since then, and will continue to increase for the next 6 months. I don't see how saying they have hit a wall would be a reasonable point of view until we go at least 12 months without a significant increase in model capability.</p>
]]></description><pubDate>Sat, 12 Sep 2026 19:31:00 +0000</pubDate><link>https://news.ycombinator.com/item?id=49676270</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49676270</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49676270</guid></item><item><title><![CDATA[New comment by lukeschlather in "Will There Be a 7G?"]]></title><description><![CDATA[
<p>We really need a hybrid short-haul standard designed for short-range transmission. Something that uses low-frequency spectrum currently allocated to 5G but isn't locked to a single carrier and the protocol can be shared between WiFi and Cellular. Staking out some new frequencies would be great too.</p>
]]></description><pubDate>Sat, 12 Sep 2026 19:12:16 +0000</pubDate><link>https://news.ycombinator.com/item?id=49676044</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49676044</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49676044</guid></item><item><title><![CDATA[New comment by lukeschlather in "If coding is solved, what now?: Measuring the sloppiness of code"]]></title><description><![CDATA[
<p>Documentation is important. I would say Opus' propensity to write documentation that documents non-features is part of the problem being discussed.<p>And the problem isn't just that it says what the software doesn't do, most of the things it claims are in fact meaningless, it's not even clearly describing something the software shouldn't do.</p>
]]></description><pubDate>Fri, 11 Sep 2026 16:52:25 +0000</pubDate><link>https://news.ycombinator.com/item?id=49661532</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49661532</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49661532</guid></item><item><title><![CDATA[New comment by lukeschlather in "Claude is only available to people over 18 years"]]></title><description><![CDATA[
<p>I think you aren't recognizing the harm of collecting this information. I might be okay with this if companies were banned from using age for targeted advertising.<p>However, the other harm is the recent IDScan hack which leaked 153M people's IDs. And regulation is not a solution here, we don't know how to implement a regulatory regime that will prevent these sorts of privacy disasters. Even if IDScan gets fines which kill the company (and I suspect they will not) it's not enough of a deterrent because no one will pay enough to actually provide proper security here.</p>
]]></description><pubDate>Fri, 11 Sep 2026 16:29:14 +0000</pubDate><link>https://news.ycombinator.com/item?id=49661136</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49661136</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49661136</guid></item><item><title><![CDATA[New comment by lukeschlather in "Thelio Mira AI Linux Workstation: 192 GB GPU Memory"]]></title><description><![CDATA[
<p>I would assume an H100 selling for less than $30k is a scam. I see one for 11,329€, and one for $25k but most are over $30k and I don't know how to judge these random storefronts. You've bought an H100 for $20k that was in working order recently?</p>
]]></description><pubDate>Fri, 11 Sep 2026 16:14:46 +0000</pubDate><link>https://news.ycombinator.com/item?id=49660903</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49660903</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49660903</guid></item><item><title><![CDATA[New comment by lukeschlather in "Thelio Mira AI Linux Workstation: 192 GB GPU Memory"]]></title><description><![CDATA[
<p>I would rather take this option than a single H100:<p>> 96 GB Dual NVIDIA RTX PRO 5000 w/ 1000W PSU (non-refundable) +$19,169<p>Though I don't think there's anything stopping you from trying to stuff an H100 into this machine if you want to BYO.</p>
]]></description><pubDate>Fri, 11 Sep 2026 07:17:49 +0000</pubDate><link>https://news.ycombinator.com/item?id=49654656</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49654656</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49654656</guid></item><item><title><![CDATA[New comment by lukeschlather in "Google will buy half the electricity from one of Finland's nuclear power plants"]]></title><description><![CDATA[
<p>Google is buying the output because they are increasing demand by half a nuclear power plant's worth of power. I actually think it's pretty reasonable to pass a law demanding they build an entire nuclear power plant here.</p>
]]></description><pubDate>Fri, 11 Sep 2026 01:10:36 +0000</pubDate><link>https://news.ycombinator.com/item?id=49652298</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49652298</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49652298</guid></item><item><title><![CDATA[New comment by lukeschlather in "The UN challenges five centuries of cartography"]]></title><description><![CDATA[
<p>For the purposes I use world maps for, correct relative sizing is the paramount thing I'm looking for in a projection. I generally use it to give me a rough idea of distances. Boggs seems inferior, at a glance I have no idea how far it is from NYC to London, for example. I feel like Mercator is similarly bad in terms of distances being uniform. What do you use maps for where Boggs is superior?</p>
]]></description><pubDate>Thu, 10 Sep 2026 01:28:32 +0000</pubDate><link>https://news.ycombinator.com/item?id=49637126</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49637126</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49637126</guid></item><item><title><![CDATA[New comment by lukeschlather in "GamersNexus and LG: Or why rooting your TV is a bad idea"]]></title><description><![CDATA[
<p>LG's terms of service essentially state that they will record any voice commands and store them for 6 months. They explicitly say they may store them in Korea. <a href="https://us.lgappstv.com/main/terms" rel="nofollow">https://us.lgappstv.com/main/terms</a><p>Actually I think everything you're suggesting is unclear, LG explicitly say they do these things in their ToS. I skimmed their privacy policy, and I'm pretty sure it essentially says they collect all this information and use it for targeted advertising.</p>
]]></description><pubDate>Tue, 08 Sep 2026 07:26:14 +0000</pubDate><link>https://news.ycombinator.com/item?id=49606735</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49606735</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49606735</guid></item><item><title><![CDATA[New comment by lukeschlather in "Mamdani bans AI in NYC schools"]]></title><description><![CDATA[
<p>> Just to get this straight: are you the kind of person to give your 1st grader a calculator<p>I think banning smartphone/computer use before 3rd grade in the classroom sounds like a good idea.<p>Math is incredibly useful. I think the biggest problem with math is that people don't understand how to apply it, it's not that applications are hard to find. AI actually explains how to apply it very directly. I can never remember how to calculate compound interest, an AI can do it instantly.<p>Mamdani's banning teachers from using AI to grade assignments.<p>> So... talking about future capabilities.<p>No, I'm talking about current capabilities. The studies that prove that current AI is better at grading assignments than humans will take at least a year, if not several.<p>The question is if we change what assignments teachers give, or if we accept that grading is a task best done by computers. If we change the assignments that teachers give, I don't know what that looks like. Regardless, there's going to be a lot of experimentation going on. Banning AI doesn't help.</p>
]]></description><pubDate>Fri, 04 Sep 2026 16:46:27 +0000</pubDate><link>https://news.ycombinator.com/item?id=49567047</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49567047</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49567047</guid></item><item><title><![CDATA[New comment by lukeschlather in "Models Don't Go Rogue"]]></title><description><![CDATA[
<p>AI detectors do not work. There are passages in this that have some odd structures that don't feel human to me. I'm sure a human edited this and refined it with some prompting, it's not just rough output from an AI. But a lot of the text feels like it was edited via prompting rather than actual editing or writing.</p>
]]></description><pubDate>Fri, 04 Sep 2026 00:17:19 +0000</pubDate><link>https://news.ycombinator.com/item?id=49558884</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49558884</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49558884</guid></item><item><title><![CDATA[New comment by lukeschlather in "Google Antigravity TOS: 3rd party usage can get Google account suspended"]]></title><description><![CDATA[
<p>Putting IDs in mobile wallets is such a minefield without explicit bans on phone searches by law enforcement. Law Enforcement being able to rifle through all your email/bank/medical records because you had a traffic stop is just not acceptable.</p>
]]></description><pubDate>Thu, 03 Sep 2026 21:43:52 +0000</pubDate><link>https://news.ycombinator.com/item?id=49557520</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49557520</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49557520</guid></item><item><title><![CDATA[New comment by lukeschlather in "Mamdani bans AI in NYC schools"]]></title><description><![CDATA[
<p>I'm not talking about future capabilities, I'm saying it will cease to be arguable within 5 years. We can't just bury our heads in the sand. Kids are going to need to engage with this directly, and it cannot wait until after they graduate high school. I'm not saying give up doing things by hand, but also everyone needs to learn to work with AI to do advanced math. It's not one or the other, it has to be both.</p>
]]></description><pubDate>Thu, 03 Sep 2026 06:17:49 +0000</pubDate><link>https://news.ycombinator.com/item?id=49546512</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49546512</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49546512</guid></item><item><title><![CDATA[New comment by lukeschlather in "Mamdani bans AI in NYC schools"]]></title><description><![CDATA[
<p>Frontier LLMs with Internet search are different beasts than what you're describing as an average LLM. In 5 years I don't think you're going to be able to get away from LLMs if you care about correctness, and we're already to the point where if a calculation matters I want an LLM to check my math.</p>
]]></description><pubDate>Wed, 02 Sep 2026 23:07:22 +0000</pubDate><link>https://news.ycombinator.com/item?id=49543826</link><dc:creator>lukeschlather</dc:creator><comments>https://news.ycombinator.com/item?id=49543826</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49543826</guid></item></channel></rss>