<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: lxe</title><link>https://news.ycombinator.com/user?id=lxe</link><description>Hacker News RSS</description><docs>https://hnrss.org/</docs><generator>hnrss v2.1.1</generator><lastBuildDate>Fri, 20 Nov 2026 08:59:19 +0000</lastBuildDate><atom:link href="https://hnrss.org/user?id=lxe" rel="self" type="application/rss+xml"></atom:link><item><title><![CDATA[New comment by lxe in "AI for scientific discovery: How researchers use Claude"]]></title><description><![CDATA[
<p>Thanks for the clarification, this is such a spot on response that actually mirrors my experience considerably.<p>> unless you hold their hand in the form of a prompt, skill, or direct instruction.<p>The hand-holding and inability to execute automated tasks is a very well known failure mode.<p>The requirement for me to prod, redirect and continuously steer is the number one pitfall when I did the Muse Open Kernel challenge using a long-horizon search/improvement swarm (or whatever the term is... just don't call it 'RSI').<p>Using plateau avoidance techniques such as <a href="https://arxiv.org/abs/2610.07447" rel="nofollow">https://arxiv.org/abs/2610.07447</a> and building harnesses around these is what helped me.</p>
]]></description><pubDate>Fri, 06 Nov 2026 17:40:28 +0000</pubDate><link>https://news.ycombinator.com/item?id=50395591</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50395591</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50395591</guid></item><item><title><![CDATA[New comment by lxe in "AI for scientific discovery: How researchers use Claude"]]></title><description><![CDATA[
<p>Because it's written for marketing and engagement, not to actually teach anyone anything.</p>
]]></description><pubDate>Fri, 06 Nov 2026 17:32:24 +0000</pubDate><link>https://news.ycombinator.com/item?id=50395478</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50395478</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50395478</guid></item><item><title><![CDATA[New comment by lxe in "AI for scientific discovery: How researchers use Claude"]]></title><description><![CDATA[
<p>When it comes to difficult optimization/search problems, doing research in a single Claude/Codex thread is definitely a struggle. There are ways to bypass some of the pitfalls by creating a agentic swarm with hard guardrails such as context resets and changing the way the harness works with memory, etc.</p>
]]></description><pubDate>Fri, 06 Nov 2026 17:23:54 +0000</pubDate><link>https://news.ycombinator.com/item?id=50395355</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50395355</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50395355</guid></item><item><title><![CDATA[New comment by lxe in "AI for scientific discovery: How researchers use Claude"]]></title><description><![CDATA[
<p>So you're not wrong in most of these conclusions, however this is very surface level use case for these tools when it comes to scientific research.<p>> I would never trust an AI to make a scientific conclusion, but they're very good at some things<p>> In this case, usually it will screw up and overfit, but it's a good smoke test.<p>> it's being used for tedious things that provide some value<p>I see a lot of comments like this about AI use in general, and I suspect there’s a kind of learned helplessness combined with anchoring to earlier experiences. Someone tries a model, encounters a few failures, and forms a fairly permanent view of what these systems can do.<p>My concern is that people sometimes turn an expectation of failure into a permanent ceiling on how much they’ll even experiment with these tools. Their limited use then keeps confirming their limited expectations.<p>EDIT: This came off as criticizing the user, but this is entirely to validate my own thesis that for quite a long time nobody has been able to crack the usability challenge when it comes to non-coding knowledge work for these harnesses. It's fascinating because it seems like it's such an easy problem to solve, but in reality it's a complex product problem.</p>
]]></description><pubDate>Fri, 06 Nov 2026 17:22:06 +0000</pubDate><link>https://news.ycombinator.com/item?id=50395330</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50395330</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50395330</guid></item><item><title><![CDATA[New comment by lxe in "OpenAI to acquire Pierre Computer Company"]]></title><description><![CDATA[
<p><a href="https://github.com/kohoj/codex-app-teardown" rel="nofollow">https://github.com/kohoj/codex-app-teardown</a> they were using Pierre libs for a while, so only natural.</p>
]]></description><pubDate>Wed, 04 Nov 2026 21:12:48 +0000</pubDate><link>https://news.ycombinator.com/item?id=50371601</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50371601</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50371601</guid></item><item><title><![CDATA[New comment by lxe in "Education shouldn't be held for ransom"]]></title><description><![CDATA[
<p>The AI response also doesn't really provide you effective answers unless you have a decent understanding and critical thinking capability.<p>Same holds true for when we were learning about how Google works. My generation had to learn about how to sift through the results in order to actually understand whether the information is accurate. Similarly, you still have to do this when you receive an AI response from Google.</p>
]]></description><pubDate>Wed, 04 Nov 2026 20:42:15 +0000</pubDate><link>https://news.ycombinator.com/item?id=50371216</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50371216</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50371216</guid></item><item><title><![CDATA[New comment by lxe in "Education shouldn't be held for ransom"]]></title><description><![CDATA[
<p>I think the effect of AI on education is transformative, not destructive, despite what the public vibe seems to be.<p>Google already provided us with all the answers and Google's AI section is not much different. It just gives you the result faster.</p>
]]></description><pubDate>Wed, 04 Nov 2026 20:05:51 +0000</pubDate><link>https://news.ycombinator.com/item?id=50370786</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50370786</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50370786</guid></item><item><title><![CDATA[New comment by lxe in "Archive.today links going forward"]]></title><description><![CDATA[
<p>Again? This happened before, if you remember this: <a href="https://www.tomshardware.com/software/search-engines/archive-today-allegedly-leads-bizarre-ddos-campaign-against-security-blogger-wikipedia-considers-removing-all-links-to-the-archive" rel="nofollow">https://www.tomshardware.com/software/search-engines/archive...</a><p>Previous HN thread:<p><a href="https://news.ycombinator.com/item?id=46624740">https://news.ycombinator.com/item?id=46624740</a></p>
]]></description><pubDate>Wed, 28 Oct 2026 17:41:31 +0000</pubDate><link>https://news.ycombinator.com/item?id=50277675</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50277675</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50277675</guid></item><item><title><![CDATA[New comment by lxe in "Why is AI generated language so hard to understand?"]]></title><description><![CDATA[
<p><a href="https://gist.github.com/lxe/4cd2ca0fd3b714b25f138227c8c650d3" rel="nofollow">https://gist.github.com/lxe/4cd2ca0fd3b714b25f138227c8c650d3</a><p>I've been iterating on this writing clarity prompt. The term we came up with was "Asyndetic parataxis with referential ellipsis and repeated rhetorical closure." It doesn't fully describe the exact issue I have with modern AI writing, but it's close.</p>
]]></description><pubDate>Sat, 24 Oct 2026 21:16:50 +0000</pubDate><link>https://news.ycombinator.com/item?id=50225403</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50225403</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50225403</guid></item><item><title><![CDATA[New comment by lxe in "Kimi K3.1 Pro and K3.1 Pro Highspeed"]]></title><description><![CDATA[
<p>That's because every single model release after GPT-3.5 has been, at least in part, been trained on data synthesized by GPT-3.5 and then the other OpenAI models, as well as Anthropic models later.<p>Also the distillation attacks and other distillation methods that gather chain of thought information from frontier models are also used to generate vast amounts of data in order to generate synthetic datasets to post-train what I believe is the vast majority of models that are being released.<p>If you look at any information about how these models are trained, the details about how the data has been synthesized, where the agentic traces come from that are used in post-training, etc., they almost never tell you about exactly what model was used to generate them.</p>
]]></description><pubDate>Sat, 24 Oct 2026 18:58:33 +0000</pubDate><link>https://news.ycombinator.com/item?id=50224040</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50224040</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50224040</guid></item><item><title><![CDATA[New comment by lxe in "A farewell to Arc browser"]]></title><description><![CDATA[
<p><a href="https://helium.computer/" rel="nofollow">https://helium.computer/</a> this is the browser I use now. it's basically ungoogled chromium without all the nonsense. It can run chrome extensions, unblock, doesn't have random unnecessary features, etc</p>
]]></description><pubDate>Thu, 22 Oct 2026 22:17:59 +0000</pubDate><link>https://news.ycombinator.com/item?id=50201852</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50201852</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50201852</guid></item><item><title><![CDATA[New comment by lxe in "Nordwand Mono – A monospace font for coders and adventurers"]]></title><description><![CDATA[
<p>Which I don't understand. We have wide screens. I want to stack a lot of columns on them.</p>
]]></description><pubDate>Tue, 20 Oct 2026 23:20:31 +0000</pubDate><link>https://news.ycombinator.com/item?id=50174037</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50174037</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50174037</guid></item><item><title><![CDATA[New comment by lxe in "The Other Epidemic: Trying to get off benzodiazepines"]]></title><description><![CDATA[
<p>I was on low-dose Ativan for a few weeks casually prescribed for anxiety. What followed is an absolute year of hell. The most unimaginably difficult time in my life. Usually they tell you that short-term low dose poses lower risk or protracted withdrawal, but it's not a purely statistical risk. There are potentially pharmacological, physiological, and other factors that govern how these drugs affect each individual and those are largely unknowable.</p>
]]></description><pubDate>Tue, 20 Oct 2026 23:19:48 +0000</pubDate><link>https://news.ycombinator.com/item?id=50174030</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50174030</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50174030</guid></item><item><title><![CDATA[New comment by lxe in "The Other Epidemic: Trying to get off benzodiazepines"]]></title><description><![CDATA[
<p>Site seems down. Here's archive: <a href="https://archive.ph/Tb1SZ" rel="nofollow">https://archive.ph/Tb1SZ</a></p>
]]></description><pubDate>Tue, 20 Oct 2026 23:15:59 +0000</pubDate><link>https://news.ycombinator.com/item?id=50174000</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50174000</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50174000</guid></item><item><title><![CDATA[New comment by lxe in "Nordwand Mono – A monospace font for coders and adventurers"]]></title><description><![CDATA[
<p>After Iosevka came out, no other coding font has replaced it for me</p>
]]></description><pubDate>Tue, 20 Oct 2026 18:03:38 +0000</pubDate><link>https://news.ycombinator.com/item?id=50170229</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50170229</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50170229</guid></item><item><title><![CDATA[New comment by lxe in "DeepInfer: High-performance inference framework for DeepSeek models"]]></title><description><![CDATA[
<p>Interference optimization is a solved problem with no moat. Anybody can achieve this result, optimize for a particular hardware, optimize for a particular problem. In general, software is no longer a moat.<p>The moat is custom inference hardware, lower power, higher throughput, etc.</p>
]]></description><pubDate>Tue, 20 Oct 2026 17:48:24 +0000</pubDate><link>https://news.ycombinator.com/item?id=50170020</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50170020</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50170020</guid></item><item><title><![CDATA[New comment by lxe in "DeepSeek V4.1 Flash 1017 (Max)"]]></title><description><![CDATA[
<p>I specifically tuned it for maximum heat and noise</p>
]]></description><pubDate>Mon, 19 Oct 2026 17:53:21 +0000</pubDate><link>https://news.ycombinator.com/item?id=50155568</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50155568</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50155568</guid></item><item><title><![CDATA[New comment by lxe in "DeepSeek V4.1 Flash 1017 (Max)"]]></title><description><![CDATA[
<p>Hello it is I, lxe<p>Perplexity looks bad and unsloth just did some good quants but they don’t perform as good on my asymmetric old dual 3090 box.<p>However, actual quality evals show literally zero degradation in “hard math” questions against full hosted model at a cost of 1.5x - 2x CoT tokens. I expect other quants to suffer from this too. This behavior seems to decrease as precision increases. I also suspect it’s the specific tuning a what experts I’m picking.</p>
]]></description><pubDate>Mon, 19 Oct 2026 17:52:37 +0000</pubDate><link>https://news.ycombinator.com/item?id=50155555</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=50155555</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50155555</guid></item><item><title><![CDATA[New comment by lxe in "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"]]></title><description><![CDATA[
<p>Nothing is stopping it. This model is woefully bad at accurate creative red teaming however. GLM finetunes on the other hand are pretty good. And I'd bet they are already deployed and doing all sorts of deeds.</p>
]]></description><pubDate>Sun, 04 Oct 2026 15:39:19 +0000</pubDate><link>https://news.ycombinator.com/item?id=49954843</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=49954843</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49954843</guid></item><item><title><![CDATA[New comment by lxe in "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"]]></title><description><![CDATA[
<p>Why isn't this type of expert caching in the native llama.cpp yet? Why do we need a separate codebase?</p>
]]></description><pubDate>Sun, 04 Oct 2026 15:36:10 +0000</pubDate><link>https://news.ycombinator.com/item?id=49954815</link><dc:creator>lxe</dc:creator><comments>https://news.ycombinator.com/item?id=49954815</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49954815</guid></item></channel></rss>