<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: antirez</title><link>https://news.ycombinator.com/user?id=antirez</link><description>Hacker News RSS</description><docs>https://hnrss.org/</docs><generator>hnrss v2.1.1</generator><lastBuildDate>Thu, 19 Nov 2026 15:13:35 +0000</lastBuildDate><atom:link href="https://hnrss.org/user?id=antirez" rel="self" type="application/rss+xml"></atom:link><item><title><![CDATA[New comment by antirez in "Google's AI answers reducing search traffic by 40%"]]></title><description><![CDATA[
<p>Accept the reality that the AI answers are <i>for many searches</i> an improvement for the user. Would not be maybe in the web of 20 years ago. Now it was a shit party of useless sites, advertising of all the kind, and other garbage.</p>
]]></description><pubDate>Wed, 18 Nov 2026 10:53:32 +0000</pubDate><link>https://news.ycombinator.com/item?id=50544666</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=50544666</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50544666</guid></item><item><title><![CDATA[New comment by antirez in "Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU"]]></title><description><![CDATA[
<p>That's not the speed of the inference, but the fact that the default parameters are to max thinking. You can even disable thinking for many queries, or lower it to a smaller value.</p>
]]></description><pubDate>Mon, 09 Nov 2026 08:47:52 +0000</pubDate><link>https://news.ycombinator.com/item?id=50422865</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=50422865</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50422865</guid></item><item><title><![CDATA[New comment by antirez in "Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU"]]></title><description><![CDATA[
<p>To put this on perspective, consider that with 128GB of unified RAM it can run even <i>fast</i>: <a href="https://x.com/antirez/status/2119392658661749175" rel="nofollow">https://x.com/antirez/status/2119392658661749175</a>
Like, coding-agent acceptably fast. See later in the reply for the prefill at ~300 t/s but can go even faster for larger prefills.</p>
]]></description><pubDate>Sun, 08 Nov 2026 12:28:08 +0000</pubDate><link>https://news.ycombinator.com/item?id=50412688</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=50412688</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50412688</guid></item><item><title><![CDATA[New comment by antirez in "MaxxBench – "Nice DeepSWE score. But is it true?""]]></title><description><![CDATA[
<p>I bet we will find this is mostly that smaller models have issues to go against: 1. strong priors that are inverted here in the changes of the benchmarks. 2. underspecified specs for which hidden tests test things not clearly stated, but that very large models can guess knowing they are tested.</p>
]]></description><pubDate>Sat, 07 Nov 2026 22:50:42 +0000</pubDate><link>https://news.ycombinator.com/item?id=50408786</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=50408786</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50408786</guid></item><item><title><![CDATA[New comment by antirez in "The 'Astra can read and write weights to improve Qwen' post is misleading"]]></title><description><![CDATA[
<p>Why this post is here in the front page? The Qwen vector steering performed was a very selective operation that interested only specific layers and weights set, it is non trivial to identify and fix. This <i>is</i> remarkable. Also Experiment 3: no access to forward pass, so why here this post says the contrary of what the paper shows and everybody upvoted it? I don't understand.</p>
]]></description><pubDate>Tue, 03 Nov 2026 15:12:46 +0000</pubDate><link>https://news.ycombinator.com/item?id=50353001</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=50353001</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50353001</guid></item><item><title><![CDATA[New comment by antirez in "Restrooms along the Norwegian Scenic Routes"]]></title><description><![CDATA[
<p>As seen in Rick and Morty</p>
]]></description><pubDate>Sun, 01 Nov 2026 15:11:02 +0000</pubDate><link>https://news.ycombinator.com/item?id=50326394</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=50326394</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50326394</guid></item><item><title><![CDATA[New comment by antirez in "Jepsen: ParticleDB 1.1.4"]]></title><description><![CDATA[
<p>It is months that I find subtle bugs far more complicated than the most immediately findable issues in the report. Together with the rest of the world. So there must be something wrong with their AI usage.</p>
]]></description><pubDate>Sat, 31 Oct 2026 20:38:02 +0000</pubDate><link>https://news.ycombinator.com/item?id=50319944</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=50319944</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50319944</guid></item><item><title><![CDATA[New comment by antirez in "Jepsen: ParticleDB 1.1.4"]]></title><description><![CDATA[
<p>Imagine asking for a review before checking with the brutally powerful agents we have today all the low hanging fruits. And, at the point we are, you can likely go at incredible levels of automated bugs seeking and do the whole test in house.</p>
]]></description><pubDate>Sat, 31 Oct 2026 18:41:52 +0000</pubDate><link>https://news.ycombinator.com/item?id=50318762</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=50318762</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50318762</guid></item><item><title><![CDATA[New comment by antirez in "Root shell on Samsung TV"]]></title><description><![CDATA[
<p>"Why you bother about local AI? LOL with the $ you spend for <insert hardware here> you can buy 5 years of 20x account". Yeah. 5 years where you are told what you can do.</p>
]]></description><pubDate>Sat, 24 Oct 2026 10:29:36 +0000</pubDate><link>https://news.ycombinator.com/item?id=50219893</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=50219893</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50219893</guid></item><item><title><![CDATA[The Commodore 64 hardware simulator: break chips and pins, see the results]]></title><description><![CDATA[
<p>Article URL: <a href="https://github.com/antirez/c64-hw-simulator">https://github.com/antirez/c64-hw-simulator</a></p>
<p>Comments URL: <a href="https://news.ycombinator.com/item?id=50100366">https://news.ycombinator.com/item?id=50100366</a></p>
<p>Points: 238</p>
<p># Comments: 33</p>
]]></description><pubDate>Thu, 15 Oct 2026 07:18:14 +0000</pubDate><link>https://github.com/antirez/c64-hw-simulator</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=50100366</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50100366</guid></item><item><title><![CDATA[New comment by antirez in "From the creator of Redis; run LLM locally with ds4"]]></title><description><![CDATA[
<p>Thanks, I believe the hate is also linked to my AI-positiveness... Funny enough GPT 6 Astra can easily produce a TLA+ model that can be verified under very real-world system model assumptions on the safety of RedLock, which is a paradox to the arguments I see in this thread and to the AI-rage.</p>
]]></description><pubDate>Mon, 05 Oct 2026 10:47:56 +0000</pubDate><link>https://news.ycombinator.com/item?id=49963178</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=49963178</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49963178</guid></item><item><title><![CDATA[New comment by antirez in "From the creator of Redis; run LLM locally with ds4"]]></title><description><![CDATA[
<p>Redis remained useful for 10+ years because of the design choices that were made, actually. Many other systems/companies failed since the idea, back then, was that a system needed to be linearizable or could not be useful at all. So what you point as a limit of my work was one of the better things I did in my career.</p>
]]></description><pubDate>Sun, 04 Oct 2026 16:44:29 +0000</pubDate><link>https://news.ycombinator.com/item?id=49955485</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=49955485</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49955485</guid></item><item><title><![CDATA[New comment by antirez in "CEO of Mistral: AI is software. It can be controlled"]]></title><description><![CDATA[
<p>I hope he didn't really say this: it is an incredibly basic take. More or less like atomic bomb is hardware, it can be controlled.</p>
]]></description><pubDate>Sat, 26 Sep 2026 13:50:52 +0000</pubDate><link>https://news.ycombinator.com/item?id=49856547</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=49856547</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49856547</guid></item><item><title><![CDATA[New comment by antirez in "Italian parliament votes for return to nuclear energy"]]></title><description><![CDATA[
<p>Fun fact: in the 60s Italy was after the US and UK the county with the most advanced nuclear energy plants. These days still North Italy produces key parts of nuclear reactors in other countries.</p>
]]></description><pubDate>Wed, 23 Sep 2026 19:34:41 +0000</pubDate><link>https://news.ycombinator.com/item?id=49821342</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=49821342</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49821342</guid></item><item><title><![CDATA[New comment by antirez in "Jev in 25 Lines of Python"]]></title><description><![CDATA[
<p>Because of masked attention in LLMs, if you put the options <i>before</i> the body (the email to analyze), the transformer already knows what it needs to look for, and can use more tokens to create state to address that specific task (BERT has no mask in the attention, so tokens attend also to next tokens). You could also do a few examples in the system prompt to improve calibration.<p>Another trick that works is to repeat the question two times: "I'm repeating the task and labels for clarity: ..."</p>
]]></description><pubDate>Wed, 23 Sep 2026 09:09:40 +0000</pubDate><link>https://news.ycombinator.com/item?id=49813417</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=49813417</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49813417</guid></item><item><title><![CDATA[New comment by antirez in "Claude is only available to people over 18 years"]]></title><description><![CDATA[
<p>Anthropic is incredibly good at avoiding all the useless AI risks, while not doing anything serious about the real risks (that is: uncontrolled growth). I'm as pro-AI as I think that eventually it will remove suffering from humans, and will allow us to prosper more and help us with tons of problems we created. However the problem is not job loss or minors using AI (minors are fucked because of cell phones and social networks), the problem is avoiding extinction. For this position I was accused of AI psychosis multiple times, but I'm in good company (Hinton, for instance), so I bet some comment that trivializes this issue will surely reply to this one, but at this point it is important to see what everyone really believes. Anthropic only touches the surface of AI security, consistently.</p>
]]></description><pubDate>Fri, 11 Sep 2026 11:24:59 +0000</pubDate><link>https://news.ycombinator.com/item?id=49656574</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=49656574</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49656574</guid></item><item><title><![CDATA[New comment by antirez in "Debian votes to allow "responsible use of generative AI""]]></title><description><![CDATA[
<p>That's perfectly wrong. Since strong coding AI, people venture into huge rewrites and other big changes that automatically make sense but otherwise would not.</p>
]]></description><pubDate>Sat, 29 Aug 2026 15:56:36 +0000</pubDate><link>https://news.ycombinator.com/item?id=49490950</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=49490950</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49490950</guid></item><item><title><![CDATA[New comment by antirez in "Anthropic's best AI model struggles to attract users as cheaper tools thrive"]]></title><description><![CDATA[
<p>Sol for low level programming is consistently better, can work alone for more time, and is faster. If you think Fable is so superior, you need to work with Sol ways more.</p>
]]></description><pubDate>Mon, 24 Aug 2026 11:28:18 +0000</pubDate><link>https://news.ycombinator.com/item?id=49418230</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=49418230</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49418230</guid></item><item><title><![CDATA[New comment by antirez in "To become a better writer, read as much as you can"]]></title><description><![CDATA[
<p>Yes. The article says that you need to read a lot to write well. You actually need to write a lot, and to read carefully, high quality, and potentially low quantity too (to use more time to write).</p>
]]></description><pubDate>Sun, 23 Aug 2026 19:40:10 +0000</pubDate><link>https://news.ycombinator.com/item?id=49411814</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=49411814</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49411814</guid></item><item><title><![CDATA[New comment by antirez in "To become a better writer, read as much as you can"]]></title><description><![CDATA[
<p>This is simply not true, even if it is commonly repeated among the folks that usually don't really write. Like how reading every source code that comes handy will not turn yourself into a great programmer, to be a good writer you need to: 1. Read <i>selected</i> books, and re-read good books more often then reading new stuff, to understand <i>why</i> they are good. 2. Read, from time to time, some bad book, and understand why it is bad. 3. And obviously you need to write a lot to become good at writing. And writing, in order to improve, is writing remembering, at the same time, the vibrations of the good authors you loved, and especially making the act of choosing of every word you put in the blank page, one after the other.</p>
]]></description><pubDate>Sun, 23 Aug 2026 12:09:48 +0000</pubDate><link>https://news.ycombinator.com/item?id=49408167</link><dc:creator>antirez</dc:creator><comments>https://news.ycombinator.com/item?id=49408167</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49408167</guid></item></channel></rss>