<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: jononor</title><link>https://news.ycombinator.com/user?id=jononor</link><description>Hacker News RSS</description><docs>https://hnrss.org/</docs><generator>hnrss v2.1.1</generator><lastBuildDate>Wed, 18 Nov 2026 16:27:56 +0000</lastBuildDate><atom:link href="https://hnrss.org/user?id=jononor" rel="self" type="application/rss+xml"></atom:link><item><title><![CDATA[New comment by jononor in "Many of Amazon’s Black Friday and Cyber Monday “deals” offered no savings"]]></title><description><![CDATA[
<p>The search sucks these days too! Shows just a couple of direct matches then goes off to show things that might be in the overall category but not compatible at all. Try buying a motherboard of a particular model or particular CPU socket.</p>
]]></description><pubDate>Tue, 17 Nov 2026 08:41:39 +0000</pubDate><link>https://news.ycombinator.com/item?id=50528979</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50528979</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50528979</guid></item><item><title><![CDATA[New comment by jononor in "In defense of giving a shit"]]></title><description><![CDATA[
<p>The standard might be higher than currently practiced in many places also pre AI.
From that perspective, I guess what I would want is that AI is <i>not</i> used as another tool to (further) absolve responsibility. "AI did it" is not a valid excuse. I must admit that I was thinking mostly inside the reponsibility software developer team when writing the comment. But it goes for entire business towards users also.</p>
]]></description><pubDate>Thu, 12 Nov 2026 08:27:38 +0000</pubDate><link>https://news.ycombinator.com/item?id=50464508</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50464508</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50464508</guid></item><item><title><![CDATA[New comment by jononor in "Nvidia is preparing to oust Jensen Huang"]]></title><description><![CDATA[
<p>Yes it is Chinas MO. But it probably possible to influence their timing and/or degree. I do not thing they have anywhere near replaced western/Taiwan x86/ARM CPUs, for example?</p>
]]></description><pubDate>Thu, 12 Nov 2026 08:21:32 +0000</pubDate><link>https://news.ycombinator.com/item?id=50464461</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50464461</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50464461</guid></item><item><title><![CDATA[New comment by jononor in "100x Developers"]]></title><description><![CDATA[
<p>Amdahl can be applied directly to sequential tasks in projects.<p>In project management the overall gains can be even smaller than that. If the task is not on the critical path (ref a GANTT chart), then speeding it up does <i>nothing</i> to overall completion time.<p>Of course the particular resource can still be freed earlier. Which might make it possible to do other things with it. But nothing on overall completion time.<p>If one underestimates the swtiching costs of changing the task a resource has, one can end up with negative effects on the original project.</p>
]]></description><pubDate>Wed, 11 Nov 2026 18:27:22 +0000</pubDate><link>https://news.ycombinator.com/item?id=50456725</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50456725</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50456725</guid></item><item><title><![CDATA[New comment by jononor in "In defense of giving a shit"]]></title><description><![CDATA[
<p>If you are willing to own the fuckups that (invariably) happen, then you get to own the good things that come out (as in, said I made it with X). But only if you take full responsibility, for both good and bad.<p>Can't blame someone else and take credit for the work when convenient.</p>
]]></description><pubDate>Wed, 11 Nov 2026 17:15:31 +0000</pubDate><link>https://news.ycombinator.com/item?id=50455848</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50455848</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50455848</guid></item><item><title><![CDATA[New comment by jononor in "Our newest TPUs: 8tx and 8i Max, built for the agentic era"]]></title><description><![CDATA[
<p>Off chip does not have to mean to disk? I would assume a staged approach. First offload to CPU RAM, and then to disk (if not used within a few seconds), and then to network storage (so it can maybe be resumed by other machines). And some request routing that makes use of cache affinity based on session/user.</p>
]]></description><pubDate>Mon, 09 Nov 2026 17:47:57 +0000</pubDate><link>https://news.ycombinator.com/item?id=50428865</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50428865</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50428865</guid></item><item><title><![CDATA[New comment by jononor in "V100 SXM2 pinout and NVLink reference"]]></title><description><![CDATA[
<p>Thanks for the input! It would be TP4 across PCIe and PP2 across network. So I am hoping the network hit will not be too high.</p>
]]></description><pubDate>Wed, 04 Nov 2026 09:55:24 +0000</pubDate><link>https://news.ycombinator.com/item?id=50363325</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50363325</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50363325</guid></item><item><title><![CDATA[New comment by jononor in "V100 SXM2 pinout and NVLink reference"]]></title><description><![CDATA[
<p>Thanks, super information. I am considering a TP4 PP2 system using 2 nodes with 4x 5060ti 16gb on each. Already have 6 such GPUs, with 4 of them running 27B at the moment. Hoping to get similar performance. Ideally with stock vLLM to keep on a well maintained standard software.</p>
]]></description><pubDate>Tue, 03 Nov 2026 19:42:37 +0000</pubDate><link>https://news.ycombinator.com/item?id=50357113</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50357113</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50357113</guid></item><item><title><![CDATA[New comment by jononor in "V100 SXM2 pinout and NVLink reference"]]></title><description><![CDATA[
<p>Those are excellent numbers, especially on QFN. You say the cards are networked, how many cards per node / how many cards per machine?</p>
]]></description><pubDate>Tue, 03 Nov 2026 11:09:20 +0000</pubDate><link>https://news.ycombinator.com/item?id=50350497</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50350497</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50350497</guid></item><item><title><![CDATA[New comment by jononor in "Anthropic reports $17B in Q3 revenue, commitments of $895B in revised S-1"]]></title><description><![CDATA[
<p>Not just against open weight models and not just on hardware I think. Also on token subsidies etc. OpenAI and Anthropic both seem to be aiming to outspend everyone else, so that they can be the last ones standing. It is a risky move, especially since everyone else includes Google and Microsoft, players that have both existing surplus cashflow, solid warchests, and excellent distribution channels. But all outcomes are open still...</p>
]]></description><pubDate>Mon, 02 Nov 2026 18:14:04 +0000</pubDate><link>https://news.ycombinator.com/item?id=50341056</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50341056</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50341056</guid></item><item><title><![CDATA[New comment by jononor in "Anthropic S-1"]]></title><description><![CDATA[
<p>Cloud compute has managed to have pretty good margins. That seems like a really similar business model to what Anthropic is doing - inferene+value-adds as-a-service. Why wouldn't they also be able to get cloud level margins or a bit higher?</p>
]]></description><pubDate>Mon, 26 Oct 2026 13:47:24 +0000</pubDate><link>https://news.ycombinator.com/item?id=50243645</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50243645</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50243645</guid></item><item><title><![CDATA[New comment by jononor in "DeepSeek-v4.1-Flash-1017"]]></title><description><![CDATA[
<p>Yeah performance at a given compute budget will likely continue to improve.<p>We are running Qwen 3.8 27B FP8 with 4x 5060 16gb at the office for a small development team. It is doing pretty well for implementing backend and frontend tasks, and is already useful. In 6 and 12 months from how I expect it to get even better. Some planning is still done with a stronger model, but even that is maybe not strictly necessary.</p>
]]></description><pubDate>Sat, 17 Oct 2026 12:48:32 +0000</pubDate><link>https://news.ycombinator.com/item?id=50130657</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50130657</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50130657</guid></item><item><title><![CDATA[New comment by jononor in "DeepSeek-v4.1-Flash-1017"]]></title><description><![CDATA[
<p>Above GLM 5.3, nice! That is a very capable model. I do wish that we will see some smaller "flash" models again though - running this one locally is tough (very costly).</p>
]]></description><pubDate>Sat, 17 Oct 2026 08:49:55 +0000</pubDate><link>https://news.ycombinator.com/item?id=50129171</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50129171</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50129171</guid></item><item><title><![CDATA[New comment by jononor in "Yandex Takes a Second Data Center Hit in 48 Hours"]]></title><description><![CDATA[
<p>Everything except for LNG is down a lot. LNG seems to be basically at the same level. Does anyone know why this is?</p>
]]></description><pubDate>Sun, 11 Oct 2026 17:08:08 +0000</pubDate><link>https://news.ycombinator.com/item?id=50045376</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50045376</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50045376</guid></item><item><title><![CDATA[New comment by jononor in "Nvidia in talks to acquire US 'open' model startup Reflection AI"]]></title><description><![CDATA[
<p>Both Google and Meta have done that years ago, with Tensorflow/JAX and PyTorch?<p>But a software abstraction only helps to reduce lock in. It does not unlock efficiencies on itself. For that new and/or hardware hardware might be needed.</p>
]]></description><pubDate>Sun, 11 Oct 2026 08:46:03 +0000</pubDate><link>https://news.ycombinator.com/item?id=50041005</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=50041005</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50041005</guid></item><item><title><![CDATA[New comment by jononor in "Cloud Agents Are Inevitable AI Prisons"]]></title><description><![CDATA[
<p>The default state of agents and LLMs is inert. It requires action from a human even be <i>able</i> to do something. From the very basics like starting the software, connecting it to a network, having hardware to run on.<p>But the most important wrt AI is to keep the owners/operators responsible. Don't let them weasel their way out of it. They are for sure trying, and will continue to. This includes using language to overemphasise agents importance in bad outcomes, in order to downplay their own responsibility.</p>
]]></description><pubDate>Thu, 24 Sep 2026 08:15:06 +0000</pubDate><link>https://news.ycombinator.com/item?id=49827741</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=49827741</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49827741</guid></item><item><title><![CDATA[New comment by jononor in "Cloud Agents Are Inevitable AI Prisons"]]></title><description><![CDATA[
<p>Agreed. AI as an "accountability sink" is an incredibly bad idea. It allows/incentives bad actors to do bad shit and get away with it. Which will, generally, tend in such practice becoming more common. Everyone loses except for the crooks.
We cannot accept "AI" absolving humans of responsibility.</p>
]]></description><pubDate>Thu, 24 Sep 2026 08:01:06 +0000</pubDate><link>https://news.ycombinator.com/item?id=49827650</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=49827650</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49827650</guid></item><item><title><![CDATA[New comment by jononor in "What AI-Native Looks Like"]]></title><description><![CDATA[
<p>Might even be cheaper!</p>
]]></description><pubDate>Wed, 23 Sep 2026 17:11:04 +0000</pubDate><link>https://news.ycombinator.com/item?id=49819306</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=49819306</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49819306</guid></item><item><title><![CDATA[New comment by jononor in "The Inference Hardware Revolution of 2026"]]></title><description><![CDATA[
<p>Some kind of compute-in-memory architecture is a good candidate, I think. There are many alternatives here, researched for many years prior to the LLM craze. However economies of scale dominate in chip industries, and this tends to favor more conventional or incremental approaches (to piggyback on existing scale). Alternatively someone needs to have a way of bootstrapping the insane scales needed to be competitive with a better-but-different approach.
So it could be that boring and straightforward stuff like two-chip prefill+decode takes most.</p>
]]></description><pubDate>Wed, 16 Sep 2026 14:53:38 +0000</pubDate><link>https://news.ycombinator.com/item?id=49727967</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=49727967</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49727967</guid></item><item><title><![CDATA[New comment by jononor in "The Inference Hardware Revolution of 2026"]]></title><description><![CDATA[
<p>I am missing a mention of ROM in the article. Keeping read-only weights in RAM is rather wasteful, as ROM can be implemented more cheaply. Approaches like High Bandwidth Flash (HBF) are relevant here, and should come to market in a few years.
Further optimization may be possible by tailoring for sequential access, since inference of a particular model is very predictable.</p>
]]></description><pubDate>Wed, 16 Sep 2026 14:44:47 +0000</pubDate><link>https://news.ycombinator.com/item?id=49727833</link><dc:creator>jononor</dc:creator><comments>https://news.ycombinator.com/item?id=49727833</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=49727833</guid></item></channel></rss>