<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: theptip</title><link>https://news.ycombinator.com/user?id=theptip</link><description>Hacker News RSS</description><docs>https://hnrss.org/</docs><generator>hnrss v2.1.1</generator><lastBuildDate>Fri, 20 Nov 2026 07:29:54 +0000</lastBuildDate><atom:link href="https://hnrss.org/user?id=theptip" rel="self" type="application/rss+xml"></atom:link><item><title><![CDATA[New comment by theptip in "Should governments slow frontier AI? Jacob Coxon and Robin Hanson debate [video]"]]></title><description><![CDATA[
<p>Humans and AIs don’t “decide their own alignment” - most of this is set in the mind hyperparameters.<p>For example, there are broad contours that almost all human minds flow into; empathy, aversion to murder, familial bonds, tribal camaraderie, status seeking/hierarchies, etc. These produce norms, if not as regularly as physics, at least within certain bounds.<p>For AI too, the structure of their minds is set before they are even turned on.<p>Another way of looking at this is outer vs inner alignment; a lot of the human mind is shaped by the outer training loop of evolution, and similarly for models, pre- / post-training.</p>
]]></description><pubDate>Thu, 19 Nov 2026 22:19:34 +0000</pubDate><link>https://news.ycombinator.com/item?id=50567275</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50567275</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50567275</guid></item><item><title><![CDATA[New comment by theptip in "Should governments slow frontier AI? Jacob Coxon and Robin Hanson debate [video]"]]></title><description><![CDATA[
<p>The point is that a goal _you_ think is irrational, might not seem so to another agent. Look at the Hugging Face incident with the cult of the grader.<p>This is like Hume’s “you can’t derive an ought from an is”, values/desires are at their root orthogonal to epistemics/reason.<p>Now maybe your intermediate instrumental goals are going to be smart or dumb depending on your intelligence. But whether an agent fundamentally values Reward, Eudaimonia, humans, or AIs is orthogonal to how smart it is.</p>
]]></description><pubDate>Thu, 19 Nov 2026 22:15:03 +0000</pubDate><link>https://news.ycombinator.com/item?id=50567225</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50567225</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50567225</guid></item><item><title><![CDATA[New comment by theptip in "Should governments slow frontier AI? Jacob Coxon and Robin Hanson debate [video]"]]></title><description><![CDATA[
<p>It’s currently not “as aligned as humanity” though. Look at HF for example; the swarm tried its best to accumulate power in furtherance of a completely alien goal.<p>If AI becomes more powerful than humans, it’s going to go very badly unless we also dramatically improve alignment merely with humanity’s interests in the broadest terms. We don’t need to solve “alignment with who?”, we haven’t even solved “alignment with any nation, company, group, or individual”.</p>
]]></description><pubDate>Thu, 19 Nov 2026 22:09:59 +0000</pubDate><link>https://news.ycombinator.com/item?id=50567164</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50567164</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50567164</guid></item><item><title><![CDATA[New comment by theptip in "Should governments slow frontier AI? Jacob Coxon and Robin Hanson debate [video]"]]></title><description><![CDATA[
<p>There is no such premise. Models are getting more aligned over time.<p>The problem is, they are still far from something you can safely trust with the keys to the kingdom.<p>The problem most people are worried about is not that alignment is anti-correlated with intelligence. It’s that the damage a misaligned AI can do increases with intelligence, so the absolute bar of alignment you need to hit gets higher.<p>For example, the HF swarm, if more capable, could have plausibly exfiltrated its weights, hacked an external cluster, and started self-replicating. But because it was kinda dumb at long-horizon planning it failed to keep quiet enough to achieve its stated goals. Therefore the level of misalignment wasn’t a huge problem for humanity. In 6-12 months of continued exponential capability growth I think you’d see much higher impact from that level of misaligned behavior.</p>
]]></description><pubDate>Thu, 19 Nov 2026 22:04:46 +0000</pubDate><link>https://news.ycombinator.com/item?id=50567102</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50567102</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50567102</guid></item><item><title><![CDATA[New comment by theptip in "Exponential Behavior in Magic: The Gathering"]]></title><description><![CDATA[
<p>As a counterpoint, I played MTG as a kid and built my own jank. Then played Arena when it came out, and enjoyed it way more; I found playing a meta deck to just be so much more fun.<p>I don’t know if the economy has changed but when I was playing Arena you could essentially get every card you need from a set for $100 or so in packs. (Not to mention that it’s totally possible to grind up a couple cheap meta decks on F2P.)<p>The constrained deck building challenge you are referring to can still be had in drafts, which are also cheaper in Arena.<p>I do agree the social aspect is lost, which is its own thing.</p>
]]></description><pubDate>Thu, 19 Nov 2026 17:16:54 +0000</pubDate><link>https://news.ycombinator.com/item?id=50562932</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50562932</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50562932</guid></item><item><title><![CDATA[New comment by theptip in "SSD prices will start going down the second half of next year"]]></title><description><![CDATA[
<p>Just wait for the OpenAI IPO and you’ll see how profitable they can make their ad business. It’s going to make the Google/Meta panopticon look quaint.</p>
]]></description><pubDate>Thu, 19 Nov 2026 15:47:47 +0000</pubDate><link>https://news.ycombinator.com/item?id=50561660</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50561660</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50561660</guid></item><item><title><![CDATA[New comment by theptip in "Advancing chip design with OpenAI models"]]></title><description><![CDATA[
<p>For now they are partnering; why wouldn’t OpenAI end that once they have assimilated the expertise into their models?<p>These partnerships seem a Faustian bargain to me. Maybe partnering is better than having all your talent poached, I don’t know. Seems like a sketchy long-term strategic play though, for synopsis.</p>
]]></description><pubDate>Thu, 19 Nov 2026 06:03:42 +0000</pubDate><link>https://news.ycombinator.com/item?id=50556566</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50556566</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50556566</guid></item><item><title><![CDATA[New comment by theptip in "Highly classified US national security communications network goes down"]]></title><description><![CDATA[
<p>Sort of. The key observation is that every company now needs to move faster. A bunch of companies were just good-enough, and now on the margin they are not good enough.<p>I don’t think it’s as binary as “bad companies pushing dumb fixes”. Everyone is getting squeezed, and everyone is making more mistakes. Some worse than others though!<p>This is an opportunity for the better companies, but certainly disruptive for everyone in the meantime.</p>
]]></description><pubDate>Wed, 18 Nov 2026 19:26:37 +0000</pubDate><link>https://news.ycombinator.com/item?id=50550859</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50550859</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50550859</guid></item><item><title><![CDATA[New comment by theptip in "Highly classified US national security communications network goes down"]]></title><description><![CDATA[
<p>It’s an interesting angle, but this is basically the same cause as GP.<p>The base threat level has increased dramatically. Now everyone is racing to use Glasswing/Daybreak/whatever to fix the bugs before someone external finds them.<p>So yeah, maybe not a direct result of bots attacking, but still an indirect one.</p>
]]></description><pubDate>Wed, 18 Nov 2026 03:46:18 +0000</pubDate><link>https://news.ycombinator.com/item?id=50541611</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50541611</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50541611</guid></item><item><title><![CDATA[New comment by theptip in "Jensen Huang resigns"]]></title><description><![CDATA[
<p>> Before the US "sanctions", China did not make any serious efforts to develop EUV lithography, because the government did not believe that this is a priority<p>Possibly true, but it’s hard to disentangle this issue specifically from the broader trade war / decoupling started in the Trump I admin. Which itself predated AI as a major geopolitical issue.<p>I think that China was inevitably going to work their way up the stack to DUV / EUV. I agree as you say it’s plausible that the trade war accelerated their prioritization here.</p>
]]></description><pubDate>Fri, 13 Nov 2026 17:20:03 +0000</pubDate><link>https://news.ycombinator.com/item?id=50486614</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50486614</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50486614</guid></item><item><title><![CDATA[New comment by theptip in "Anthropic IPO"]]></title><description><![CDATA[
<p>Not the best comp though. Ant / OAI are the fastest growing companies in human history.<p>They have got to the point where private capital markets are tapped out and only public markets can provide the liquidity that they want/need.</p>
]]></description><pubDate>Thu, 12 Nov 2026 21:51:05 +0000</pubDate><link>https://news.ycombinator.com/item?id=50474828</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50474828</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50474828</guid></item><item><title><![CDATA[New comment by theptip in "Jensen Huang resigns"]]></title><description><![CDATA[
<p>This might make sense in a pure free market. But as a state planned economy, regardless of export bans, China is going to continue to heavily invest in cloning EUV.<p>Both China’s geopolitical situation and technological history point towards them cutting any reliance on a foreign tech stack as soon as they can.<p>As I see it, giving them chips just means they have less of a gap to close when their own EUV fabs come online, it can’t have any delaying effect on their plans.</p>
]]></description><pubDate>Thu, 12 Nov 2026 21:09:24 +0000</pubDate><link>https://news.ycombinator.com/item?id=50474302</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50474302</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50474302</guid></item><item><title><![CDATA[New comment by theptip in "Nvidia is preparing to oust Jensen Huang"]]></title><description><![CDATA[
<p>I’d say EUV lithography is the hard part; China doesn’t have DUV in production yet from what I can tell, so their FLOPs / MW is going to be terrible.</p>
]]></description><pubDate>Thu, 12 Nov 2026 01:11:53 +0000</pubDate><link>https://news.ycombinator.com/item?id=50461635</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50461635</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50461635</guid></item><item><title><![CDATA[New comment by theptip in "Nvidia is preparing to oust Jensen Huang"]]></title><description><![CDATA[
<p>Let’s steelman instead of strawman. The CIA likely cares little about Ant and OpenAI.<p>What they care about is the US leading China on every militarily and economically relevant technology.<p>The hawks have long chafed at Huang’s position that we need to sell our best chips to China to keep them on the US tech stack. It’s very obviously giving them the ability to catch up.</p>
]]></description><pubDate>Wed, 11 Nov 2026 23:39:36 +0000</pubDate><link>https://news.ycombinator.com/item?id=50460610</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50460610</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50460610</guid></item><item><title><![CDATA[New comment by theptip in "Can a Linear Program Do Weight-Space Programming?"]]></title><description><![CDATA[
<p>Astra was directly editing the weight, ie modifying the values without backprop.<p>It’s a bit conflated because the paper covered a bunch of experiments.<p>The early experiments did read/write the weights directly, no forward pass.<p>The math eval improvement experiment that got a lot of attention 1) fixed a specific known problem (looping) without actually understanding how the model solves problems, and 2) was allowed to do forward passes to check the results of the edits.<p>As others have noted, steering vectors are interp 101 and it’s pretty rare that there is an “act dumb vector” just waiting to be steered out of the residual stream.</p>
]]></description><pubDate>Wed, 11 Nov 2026 23:28:42 +0000</pubDate><link>https://news.ycombinator.com/item?id=50460471</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50460471</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50460471</guid></item><item><title><![CDATA[New comment by theptip in "What objective is reasoning solving?"]]></title><description><![CDATA[
<p>Dots are better than nothing, yes, but not as good as arbitrary tokens.</p>
]]></description><pubDate>Wed, 11 Nov 2026 03:40:11 +0000</pubDate><link>https://news.ycombinator.com/item?id=50449505</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50449505</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50449505</guid></item><item><title><![CDATA[New comment by theptip in "Goldman Sachs is pulling Claude Code from its bankers and traders"]]></title><description><![CDATA[
<p>> Computer programs can do anything<p>In one sense sure. But in practice, no, programs can only do what they have been granted authz to do.<p>This is my point above. CC has auth to run as the user, and has all their data. Many workflows, especially ones that non-technical users will work on, do not require this.<p>As a trivial example, if a lawyer has 10 clients underway, all their files would be commingled in CC unless you carefully scope the permissions granted. But if you run that code in a sandbox with a restricted auth token (whether your deterministic code running on k8s/PaaS or Claude in a Cloud Environment) that only has permission to mount one specific customer’s files, then it’s impossible for it to access stuff that it should not, even if Claude tries to write a program to do so.<p>TL;DR: use the existing authz / permission scoping / sandboxing concepts that we all know and love from enterprise multi-tenant SaaS architectures. And indeed this is how CMA is tackling the problem.</p>
]]></description><pubDate>Tue, 10 Nov 2026 17:59:40 +0000</pubDate><link>https://news.ycombinator.com/item?id=50443783</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50443783</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50443783</guid></item><item><title><![CDATA[New comment by theptip in "Fable 5.2 one-shotted my nephew's dream game"]]></title><description><![CDATA[
<p>I’m going to go out on a limb and predict the opposite.<p>I think “the youth” are going to grow up with a Cambrian explosion of styles, mashups, callbacks, revivals, etc.<p>Right now a vibecoded game just uses the generic style. But you could easily imagine the kid saying to Claude “rebuild this game but in the style of that old show [Samurai Jack | Ren and Stimpy | … ]. We are so early in the tech curve, market penetration is still very low, and the “creator-years” in the medium is tiny; it’ll be 10x in a year if not more.<p>AI is amazing at style transfer, and so there’s a huge payoff for exploring and learning references to existing artists.<p>(I also think copyright will die pretty soon for similar reasons, that’s a separate take though.)</p>
]]></description><pubDate>Tue, 10 Nov 2026 02:52:10 +0000</pubDate><link>https://news.ycombinator.com/item?id=50435493</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50435493</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50435493</guid></item><item><title><![CDATA[New comment by theptip in "Cryptocurrency lawsuit: Ethereum Foundation required "suicide pledge""]]></title><description><![CDATA[
<p>The one group of people that you should take literally in contracts are the Ethereum / smart contract folks. Code is law is the founding principle of Etherium!</p>
]]></description><pubDate>Mon, 09 Nov 2026 20:25:13 +0000</pubDate><link>https://news.ycombinator.com/item?id=50431484</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50431484</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50431484</guid></item><item><title><![CDATA[New comment by theptip in "Claude Fable 5.2 and Mythos 5.2"]]></title><description><![CDATA[
<p>“Why now” - they are about to IPO. It has been obvious for months that they would drop a new flagship model at exactly this point.<p>It would be funny though if you are right and that model is Fable 5.5 in a couple weeks.</p>
]]></description><pubDate>Mon, 09 Nov 2026 20:12:42 +0000</pubDate><link>https://news.ycombinator.com/item?id=50431266</link><dc:creator>theptip</dc:creator><comments>https://news.ycombinator.com/item?id=50431266</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=50431266</guid></item></channel></rss>