Most of the non-technical people I see using LLMs to create software are not engaging in "cognitive surrender" as far as I can tell.
They didn't read the code before and they don't now. However, asking questions and improving understanding of the system and architecture is actually faster and easier for them now than it used to be.
They're evaluating what they can see, i.e, judging results. They build things for their own use. They test them, iterate on them to meet their needs and try to achieve their vision. This is roughly the same process they followed when they used to outsource to an agency except the loop is much tighter.
It seems to work well, mainly hurts when the problems are invisible, so security, backups, breaking migrations, perf, maintenance over time (language / runtime updates, dep updates, server updates etc). All the stuff that hurts if you have never been responsible for the uptime of something before. Not knowing that software is never "done". But junior devs trip on those exact same things.
I agree with the tail of the article; tooling and processes that make that stuff more visible would close a lot of the gaps.
On the same theme:
> the supreme goal of all theory is to make the irreducible basic elements as simple and as few as possible without having to surrender the adequate representation of a single datum of experience
- Einstein
> I would have written a shorter letter, but did not have the time
- Pascal
> Simplicity is a great virtue but requires hard work to achieve and education to appreciate, and, to make matters worse, complexity sells better.
- Dijkstra.
> La perfection est atteinte non pas lorsqu'il n'y a plus rien à ajouter, mais lorsqu'il n'y a plus rien à retirer.
> Perfection is attained not when there is nothing more to add, but when there is nothing more to remove.
Previously, pretty much everyone all agreed on YAGNI and KISS because code was so expensive to create and more expensive to maintain. Now the code generation is "free" - but at a design level, YAGNI and KISS do still matter. But they are enforced now only by sheer will power. You have to actively restrain the code from being created if you don't want it. Perfectly good code - I can't even fault it. I just know, if it doesn't need to exist yet, it's a mistake to create it earlier than needed.
This has become common now, and I've seen it ruin large projects until nobody understands the codebase full of verbose nonsensical code and comments. It becomes impossible to further develop the project without AI. Contributors not using AI for any reason are all leaving because no one enjoys working with mindless robots (and humans) as colleagues. It's demeaning and devaluing their labor and humanity. They're not getting "left behind", they are leaving the madness of crowds behind.
> ..the possibility of runaway dynamics: once a critical threshold is crossed, small increases in adoption can trigger rapid population-level shifts toward *persistent dependence, with abrupt losses in cognitive competence*.
We can all see it happening. As this technology permeates society, the challenge to maintain control, coherence, and quality will become greater and ever more serious - for example, AI in healthcare, military, management and governance.
Running all those lines, people don't even use it all that much.
Much more efficient to put everybody into camps, and simply have the people take turns using electrical appliances operating at 100% duty cycle.
Freight delivery speeds need to be brought down too, 35 MPH maximum.
Food? Give me break, get everyone on a Soylent Green diet.
1. people aren't using (GPU) RAM [1] for just running LLMs (in the local model sense), they are using it for other things... games, system displays, other models (Text-to-Speech, Speech Recognition, etc.);
2. being able to run models locally allows them to be used when the internet is down, use them on private data (e.g. medical records, proprietary company data, etc.), including training on that data;
3. the big AI companies are already storing unused GPUs on shelves waiting for their roll-out in data centres, as they can't build the capacity fast enough.
With point (3) that goes against Sam Altman's argument about utilization. The big companies have the hardware, but aren't utilizing it themselves.
[1] And if the argument extends to RAM in general you have all sorts of applications, especially on developer hardware -- editing and compiling software; running local servers; running CAD and other applications; etc.
I guess it all comes back to the fact that they don't have a clear division between data they have learned about the world, instructions and data under analysis.
You can ask them only to consider the 5 facts in front of them, but they won't do that, and cannot do that.
Exponential complexity is what you automatically get when your algorithm is "generate a possible solution, verify if it's correct, repeat if it's not". Polynomial complexity is what you get when the problem itself has some kind of structure that you can exploit to either directly construct a solution, or to generate a very small set (relative to the problem size) of possible solutions that have to be checked by hand.
Now, it is true that for all practical problem-solving purposes, "P = NP but only with algorithms that require n^10^80 or more time (with this as a proven lower bound for even the typical case, no hope of improving)" would be exactly equivalent to "P != NP". However, it would still be a mathematically extremely interesting result, as it would tell you that any problem where it is easy to verify if a solution is correct carries some hidden structure that makes it "easy" to generate the solutions. And it could still open the door to finding new sub-classes of actually easy instances of these problems that are not known today, even if the worse case remains just as intractable.
Now, note the refutation of ETH with the new algorithm found by AI is not really that impressive, it doesn't seem to have that deep of an implication. Sub-exponential running times are not that fundamentally different from exponential ones - they basically just tell you that the problem has just enough structure to eliminate a certain class of possible solutions outright, instead of spending the time to actually verify them. If the algorithm found had been polynomial, even of a huge degree, this would have been a completely different discussion. As it is, we'll just replace the ETH with a "nearly-E"TH and nothing much changes.
At the same time, most of their junior devwork can be replaced completely by AI - so they really should start to learn how to work with it. I recommended everybody to get a private 20€/month claude subscription and use it for toy projects. I have younger devs (early 20s) in the team that told me they won't spend private money, or simply have no usecase in their private life for AI. We talk about software engineers here, and IMHO most of them are in denial of the fact that their skills have mostly become worthless if they don't start to pivot to something valueable.
https://jacquesmattheij.com/local-inference-with-large-langu...
If you have more questions feel free to mail or ask here in the thread and I'll do my best to answer.
The hardware related parts are near the end of the article.
If you have a bigger machine you can optimize some of this (V8), but these optimizations typically need memory and/or a JIT. (And Python struggles to optimize even on stronger machines).
By starting from scratch (with what we learned from V8 and Dart) you can design something that feels similar but is much cleaner and faster to run.
Except in this case they have not issued any fix, for several weeks, to an extremely critical issue that most other distros have already patched for some time now.
For many this now means their trust is broken.
sudo iptables -t raw -I PREROUTING -p tcp -m u32 ! --u32 "0>>24&0x0F=5" -j DROP
Most consumer routers run ancient stacks with all kinds of weird proprietary vendor patches that are tied to a specific kernel version. Looking at you Qualcomm!
However they generally also drop all traffic by default so from the WAN side they look like a black hole. I guess that's "safe" enough for most consumers.
For fun, take a look at the open PRs in the OpenWRT repos. Those people are doing some amazing work but it's also frustrating to see them struggle with things that _could_ be easy if the vendor made some slightly better choices.
I would strongly prefer something like OpenWRT though. Call me paranoid but after the last few weeks I'm increasingly less comfortable with connecting my devices to a blackbox.
Every consumer router-wirelessAP I've seen runs linux.
I never see security upgrades for the kernels they use. Not one. Not ever.
Is this something that should be a cause for concern? Even if you don't run a consumer-grade tplink, netgear, asus, whatever router yourself.
Update: The context which I should have said is that I think LLMs are giving script kids nation state level tools and expertise. This known bug is always less interesting than the next twenty.
It does not seem that complicated to patch, and kernel upgrades are a dime a dozen these days. So what is making them delay a fix for so long?
"Freenode" in 2026 is an entity actively hostile to open source software and while what they say here might be true, they are only saying it because it aligns with their biases. They are not a real news source.
My mitigation for this lack of patches has been using iptables.
iptables -A INPUT -p tcp -m u32 --u32 "0>>24&0xF=6:15" -j DROPI am not qualified to make a judgement on the research quality, but it strikes me as a little weird? And also weird to cite when trying to create an "AI is stalling" narrative, because the text of the paper doesn't really support that at all.
The claim seems to be based on a single figure that the authors present as additional context in an appendix, which shows the trends of four surveys, in which only one of the four shows a decline. The authors don't even seem to think the decline is significant enough to comment on except to say that "the plateau in late 2025 reflects declining interest and correlates with google search trends for ChatGPT and Gemini" (which they also show using a chart that does not look like a plateau or decline to me at all).
The methodology and survey questions are included, the numbers seem to be based on a straightforward "Do you use generative AI for your job"?
It's weirdly ChatGPT and Gemini centric, the word "Claude" only appears once, in a definition, and "Microsoft" and "Copilot" are nowhere to be seen. I don't know how much to read into that, it just strikes me as weird and noteworthy in a paper that by its own admission is detecting a trend that is the opposite of everyone else doing large scale surveys on the topic.
It would be VERY BAD if we gave politicians arbitrary veto power over who IPOs and who doesn't. The threat of corruption should be obvious.
Surprising. I couldn't find the study they're citing, and they didn't include a link. Anecdotally, my friends across various industries (tech, finance, medicine, scientific research) are all using AI more heavily as time goes on.
Additionally, some ETFs have built in protections to rule out businesses that are not yet profitable (like SpaceX and Anthropic), which I think is a great rule.
You had lots of choices, you made this choice, and you can always change your mind. Do we really need to have the government step in to prevent you from having to choose something else?
I built this to aggregate all the aggregators (like HackerTube, Hackerbooks etc.) that were popular a few days ago (they go back to 2011, at least).
A couple of months ago (so probably vibe-coded too):
This one even from 6 years ago:
https://www.wolfgangfaust.com/project/paper-hn/ (https://news.ycombinator.com/item?id=23669650, 627 points, 133 comments)
The challenge took me a few weeks to make because "calibrating" a disk (determining the geometry of the pregroove) is incredibly sensitive.
After the CTF I tried really hard to make a repeatable calibration process with the intent of building something releasable. I even ended up measuring the details on every CD-RW and CD-R I could buy on Amazon.
But despite trying very hard, I never did make calibration patterns that were reliable enough that I felt like other people would be willing to go through the process.
It seems someone else has succeeded where I failed. Great to see!
> In 2012, a LightScribe optical drive was successfully used by Maher El-Kady, a graduate of UCLA, to turn a specially prepared graphite oxide layer coated onto a DVD into graphene.[10] El-Kady and Richard Kaner, his lab professor, used an unmodified LightScribe drive. The disc was prepared by coating the disc with an aqueous solution of graphite oxide and allowing it to dry. Areas of the coating illuminated by the LightScribe laser were turned into graphene. Various shapes can be drawn, which allowed the scientist duo essentially to laser-print an ultracapacitor on graphene using consumer-grade technology.
In general my personal research thesis is that current deep learning is much less malleable and neuroplastic as it should be, I got into interpretability because of this. Many focus on safety, but I prioritize it as a toolbox that allows for greater shaping of the underlying intelligence system. If we can fully understand how learning happens, where information is stored and how it is processed, unlike in biological brains, we can apply edits and transformations in digital ones. And if I can use a technique to edit weights based on some underlying reasoning about a system (with a caveat that it leads to something worthwhile), so can an agent. Hence self-improvement. There are many interesting implication to weight space research, and I very much encourage people to read about it. But it is also very easy to overstate what is currently possible, in fact I'd say better to know what is not so one can make it possible.
Part of my research is in weight space learning. For starters, we have had the ability to edit weights, to generate new weights, to identify weights for a while. Here is an example for larger models as early as 2023: https://rome.baulab.info/
Weight space interpretability and learning is an active research community, has been for the past 3 years with many great papers. I suspect it is hitting non-interp folks harder because the authors hyped up their work, and people who are not in the field did not bother to read it before typing up X posts.
Some experiments here are done on very small models (20K params). We already know that small models are more interpretable. In fact, much of editing here you can do by hand. For Qwen, the model fixed a looping bug that may have been known apriori, and it is basically a steering vector, a very established technique. This is clear when they then attempt to "improve" math capability, and it doesn't work. For me, this paper rather shows that OpenAI has done training on mechinterp environment, in fact, maybe even on weight space learning datasets, which are open. But it does not show some insane "well past what we’d expect from humans or conventional algorithms" (from the blog post), more like past what I would expect from a large model with no specialized harness. I still think it is cool work, but do check out recent advances in autointerp, to ground your expectation.
On a personal note, I understand the need for promotion of your work, but I think healthy skepticism is fine too. Otherwise, your understanding of what problems need to be worked on and solved in science becomes warped.
that is why this is crazy,
give a human a legacy codebase and 5 years and he'll be able to understand very well.
if you give a human the weights of qwen 0.8B and 5 years it'll still be impossible to edit the weights in any legible manner.
Why is this not overhyped in my opinion? Nothing has worked this well, ever.
> We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights
This claim is backed up by the ability of Astra, given a text representation of the weights of tiny (20k) special-purpose models, to figure out what those models do and in some cases improve them beyond what RL could do just by reading the weights with no tool calls. They call this "weight-space programming".
> Likewise, asked to make Qwen3.5-0.8B better at math without gradients, Astra identifies an issue in its weights causing reasoning loops that Qwen’s own model card warns about, and fixes this by natively editing 0.016% of the weights
Here it did use tool calls, and seemingly used standard steering vector techniques to identify and fix reasoning loops in this small Qwen model. This doesn't seem that related to the previous results.
A model is _sort of_ designed to predict the next token in a string of text. But it's equally accurate to say it's designed to edit code until it passes a suite of unit tests.
Yes, they generate sequentially, so "next token predictor" isn't technically incorrect (which is why I chose the word "meaningless"). But it's pretty clear there's a mechanism inside them that allows them to "plan" what they generate. I might as well call you a next word predictor.
Parameter Prediction for Unseen Deep Architectures [1]
NeuroGen: Neural Network Parameter Generation via Large Language Models [2]
Diffusion-Based Neural Network Weights Generation [3]
The critical thing to me is that Astra is capable of doing it. I don't know if it was trained to do so.
All this does not make it any less crazy in my books. And I still don't know how I can sleep easy tonight.
[1] https://arxiv.org/abs/2110.13100
As others have said, while still absolutely insane, I don't find it completely out-of-the-blue surprising. One of the recurring abilities of (current-generation) LLMs seem to be to rapidly grasp very complex systems, where humans would need months or years to reach the same level of expertise.
You can see this in the way a coding agent can "understand" and explain sprawling legacy codebases in seconds.
You can also see it in the projects that are using them to reverse-engineer proprietary IoT protocols or computer games.
But those were all systems where we had detailed knowledge about what structures are there to be reverse-enginered.
If the models can transfer this ability to some other systems where we don't know the underlying structure then I think we will be in for a lot more craziness in the future. (Protein interaction networks? DNA sequences? Ecosystem interactions? Animal languages?)
As for weights, if the models really can analyze the semantics behind specific weights or directions, I have my hopes that they could also give a textual explanation or help designing tools to make "handcrafting" weights more practical in general.
IYKYK
The work done in this experiment could have also been done by a sufficiently motivated human with enough time to explore, though obviously we would need more tools to visualize and track things. There have been a lot of interesting LLM introspection tools that allow you to peek into the model and see what's being activated and run experiments to trigger it.
Are you familiar with abliterated or "uncensored" models where weights have been modified to avoid triggering refusal patterns? You take an LLM, create test cases that trigger the refusal pattern, observe weights involved in that response, then tweak those weights until the response is reduced. This is the same type of work, applied to a specific type of error that was impeding math within Qwen. That work can also be done with an LLM in a harness running the experiments. The amazing part here is that Astra was large enough to walk through the problem without other tooling doing it experimentally.
So I don't think your analogy of a human modifying a neuron is quite applicable. LLM weights are not neurons and they've been inspectable, by humans or LLMs in a harness, for a long time.
Something similar happen in early GPT models which had a lot of chess game in standard notation in their training data which made them very good at chess as long as your prompt had the right format.
We'll see if this attempt at RSI worked.
They also tried the same with Gemma 3 270M, which doesn't loop and is just bad at math, and Astra couldn't improve it. Still impressive, but it's more in the realm of what I'd expect, given labs are probably training their models on ML research and interpretability work.
So it's not able to insert new knowledge, but it's able to adjust the routing/looping/weighing of existing one.
As I understand, it can replace post-training like LoRA/GRPO, but if it'd cheaper the authors don't go into
So it's not able to insert new knowledge, but it's able to adjust the routing/looping/weighing of existing one.
As I understand, it can replace post-training like LoRA/GRPO, but if it'd cheaper the authors don't go into
It’s no secret that these big labs see “recursive self improvement” as a primary goal. It seems natural that the models would be explicitly trained to excel at all sorts of ML related tasks as a result. Is this one of those tasks, or is this emergent?
I wouldn’t mind if AI progress would stop right here and now, but I fear it will continue, and increase inequality in society which is contrary to what the CEOs say (as usual).
HA! Now that's cool. Not necessarily the improvements, those are ancient and probably overtrained benchmarks, but it's so cool that the model found a singular direction for this "loop bug". I imagine it could possibly do much the same thing in finding other "single directions" on a variety of topics (if allowed by guardrails). This is really really cool, especially for the mechinterp folks.
Astra can, without tooling, determine what games small networks are trained on, can edit the weights of networks directly to improve outcomes, and can hand edit a small qwen model to improve its math skills by like 5x.
I'm shook. If you'd asked me this morning whether LLMs have native understanding of their weights that is less than human <-> neuron understanding, the same as human or more than human, I would have said "same to less" depending on definitions.
No human can go edit a brain to become better at math. This is just very, very surprising to me and extremely interesting. I would guess this upsets quite a number of interp theories and philosophical takes on what's going on with these networks.
Where dw works, this game teaches vlllld instead, which is a bad habit to teach.
Ctrl+A increments the first number after the cursor, which is easier than, say, f6r7 in cases where they do the same thing. Similarly, Ctrl+X decrements.
t:t: is weird when fY is much easier.
The Vim simulator has a minor difference compared to real Vim:
0
0
Start on line one and do CTRL-V then j then CTRL-A. Real Vim would increment both. The simulator only increments the one under the cursor.[1] https://app.opencve.io/cve/CVE-2026-98432 [2] https://aws.amazon.com/security/security-bulletins/2026-160-...
Red Hat disagrees[0][1]. They might be parroting Anthropic, and Anthropic might be exagerating the impact (likely!), but that's the current consensus about the vulnerability. If Ubuntu disagrees, they should have said something and not pasted "Cvss 3 Severity Score 9.8 · Critical" at the top of the page.
> A motivated attacker could probably do this on your network
I find that's still catastrophic given how centralized cloud environments are.
[0] https://access.redhat.com/security/cve/cve-2026-98432
[1] https://access.redhat.com/security/vulnerabilities/RHSB-2026...
From the article the bits you need to know:
1. CVE-2026-98432 is a critical (CVSS 9.8) bug.
2. Anthropic says it built a working exploit that gets root on a server.
I took the lazy Mint route with my most recent machine since I wanted to use it for light gaming with minimal fuss. I'm going to redo it with Nix sometime soon because the value prop simply isn't there and I'd like to try a declaratively-configured distro.
And this is how Ubuntu describes the vulnerability to sysadmins:
In the Linux kernel, the following vulnerability has been resolved:
ipv4: free inet_opt and ireq_opt after an RCU grace period
tcp_v4_syn_recv_sock() transfers ownership of ireq->ireq_opt to the
child socket (newinet->inet_opt) without copying it.
Another cpu can concurrently process a retransmitted SYN for the same
request socket, and send a SYNACK from tcp_check_req().
tcp_v4_send_synack() and inet_csk_route_req() read ireq->ireq_opt
under rcu_read_lock() only, and ip_build_and_send_pkt() and
ip_options_build() then read opt->optlen twice.
Note that the SYNACK timer itself is not an issue: it holds a
reference on its request socket, and inet_csk_reqsk_queue_drop()
calls timer_delete_sync() before the child can be freed.
Since commit 079096f103fa (“tcp/dccp: install syn_recv requests
into ehash table”), request sockets are processed without holding
the listener lock, so nothing prevents the child socket from being
freed while the SYNACK is still being built. TCP child sockets do
not have SOCK_RCU_FREE, and inet_sock_destruct() frees inet_opt
with a plain kfree(), leading to a use-after-free in
ip_options_build().
A similar issue exists with request socket migration
(net.ipv4.tcp_migrate_req=1, or a BPF_SK_REUSEPORT_SELECT_OR_MIGRATE
program). reqsk_timer_handler() clones the request socket with
inet_reqsk_clone(), so that the old request socket and its clone
share the same ireq_opt, then reqsk_migrate_reset() clears the
pointer in the old one. Another cpu holding a reference on the old
request socket can still be using these options (sending a SYNACK,
or creating a child in tcp_v4_syn_recv_sock()) when the clone is
freed, and tcp_v4_reqsk_destructor() also uses a plain kfree().
Readers already use RCU, and other paths replacing inet_opt
(do_ip_setsockopt(), cipso_v4_sock_setattr()...) already use
kfree_rcu(). Use kfree_rcu() in inet_sock_destruct() and
tcp_v4_reqsk_destructor() as well.
IPv6 is not affected by the first issue, because tcp_v6_syn_recv_sock()
duplicates the options. tcp_v6_reqsk_destructor() has the same
migration issue with ipv6_opt, which is only set by CALIPSO. This
will be addressed in a separate patch.
How is this in any way, shape, or form acceptable two weeks after the CVE was published?Some of my clients may be surprised by this change, as I once advocated for Ubuntu as a good choice for a development and deployment platform. While that has not been true for quite a while, they have now jumped the shark into eel-infested waters; I cannot take them seriously any longer.
Yes it does? And Ubuntu is selling security subscriptions and doing a poor job of it, These aren't open source volunteers, this is a paid product.
* https://launchpad.net/ubuntu/+source/linux
You can pull them down with a "-t" added to your apt-get (but will not appear by default).
26LTS had the new kernel go to the updates/security repos on October 28.
server: nginx/1.14.0 (Ubuntu)
The post gives no details about what this vulnerability is, how critical it is, how it is impacting them personally, how vulnerable their servers are, how they are working around it.
Just a title designed to generate the maximum outrage followed by a few meaningless lines.
> Why is this CVE high priority?
> See ubuntu cvss score
That's it, we don't even get a period on this TBD of a sentence. It's followed by a generic link that explains what "Ubuntu Priority" is, but that page contradicts their choice. Note that it doesn't offer any mitigations either, and even the (horrendously unclear) description is copied from the NIST page, so there's no actual message from Ubuntu.
I already had a bad feeling the first time I saw "Get more security updates through Ubuntu Pro" on `apt upgrade`, this is now the kick I needed to move to a different distro.
[0] https://ubuntu.com/security/CVE-2026-98432#priority-reason
Surprised how many answers here are on "social" reasons, and ignoring that Ubuntu does provide genuine additions to Debian.
Ubuntu includes additional proprietary drivers to Debian. So there are legitimate hardware cases were Ubuntu will work out the box and Debian won't. I can't speak for how many people need those drivers, and how much of choice is down to defaults/inertia etc.
From hearing interviews with PopOS folk, I think this is also a big reason people choose Ubuntu as a base for a lot of Distros instead of Debian.
And don't get me started with Debian testing. I want a LTS distro not a testing one.
A gradually increasing number of popular userland applications have been publishing their own apt repositories however so this is less of a problem now than it used to be.
Over the years though I've found that I much prefer Debian's relative stability and more thoughtful, if slightly plodding, release cadence.
https://ubuntu.com/security/CVE-2026-98432
It’s still, over a week after disclosure and patches, unpatched in Ubuntu.
As for me, a part of me wants to move over to Rocky Linux 10.
Canonical absolutely is marketing enterprise security updates (at least for those who pay). I don't see this in the paid ESM updates channel yet either?
I may be biased as a Debian user, but if Ubuntu now reacts slower than Debian this may push some people to reconsider their choice.
Maybe it's time for StealMyBitcoinV2…
I don't remember why exactly but it was definitely just 'linux VPS = Ubuntu' for a long time. I think they got reputation as being the 'just works' option because they were happier to include some non-free graphics drivers etc where Debian was the more pure option but most end users didn't care
But since they added snap and I started running into issues with trying to install Docker etc a few years ago I default to Debian now and most of the time I don't even know which it is.
Which would support your argument, but I don’t think this is particularly new.
- Cursor: https://cursor.com/blog/git-at-any-scale
- GitHub: https://github.blog/engineering/architecture-optimization/bu...
- Anthropic: https://www.anthropic.com/research/redesigning-source-contro...
- GitHub: https://github.blog/engineering/architecture-optimization/th...
Maybe also tangentially East River Source Control's too: https://ersc.io/blog/what-comes-after-git.
I'm wondering how much cross-pollination of ideas and subsequent implementation details there are between people working at these companies. It does seem like all of this could potentially converge.
As a user of Astra, it writes code that appears to work. It does not necessarily write sensible code that is architected in a way that makes sense based on what it does. I just watched it do a major rework of a couple thousand lines of code by keeping the original organization and kludging up an adapter to make old code have new semantics.
I would expect that this approach, coupled with lots of fun asking for changes and none of the grunt work of actually thinking about the result, could result in exponentially increasing amounts of mess.
I haven’t used a recent Anthropic model - maybe it’s better at this.
I think it’s simply a matter of dogfooding and new feature development, they do have much higher velocity than companies of equivalent size and an unlimited token budget helps as well.
Why would the volume that any of this be increasing exponentially? And what are all the engineers doing when (besides the model quality) their products don't seem to be changing very much?
It would be very useful if I could just boot a kernel and get an instant report of what kernel config, firmware and modules are actually used/needed. You can kind of do this but
yet its title is:
> Small Business Cybersecurity: Non-Employer Firms (NIST Cybersecurity White Paper 50)
Small businesses don't identify as 'Non-Employer Firms', have no idea what 'NIST' is, most wouldn't know what a White Paper is, nor number 50.
'A Crash Course in Cybersecurity for Small Businesses' would be an improvement.
It’s a darn fancy algorithm and a cool tool. Putting it on a pedestal instead of grounding it to reality is not going to help anyone. My condolences to the authors.
IMO this misses the mark. Yes, it’s a decent attempt at simplifying cybersecurity, but if this is aimed at entry-level, non-technical people, it’s not gonna work.
One of my earliest jobs was to redesign training materials for one of the worlds largest companies. I stripped the training to its core, the true basics, and started there. The results were remarkable. I feel the urge to do the same with this document…
Good gosh. Goldman probably has 50-100 people, inc. maybe 20 Engineers just on that alone, full time.
And yes, Intel can fail 100%. They can absolutely go into receivership and/or be broken up and bought for pennies by other entities.
It would be a strategic dismantling, obviously the government won't allow certain things to happen, but they could 100% allow investors to be wiped out.
Boeing is a bit different as there are no substitutes, but it could go into receivership. Some parts of Boeing will 'always exist' but the corporate entity can fail.
It is not in the national interest to make sure that the Intel common stock remains valued above 0. Plenty of companies that are of national interest were saved directly from bankruptcy, by nationalizing them in a way that wipes out the common stock.
That's kinda the point; something as important as this shouldn't be a casino.
Advanced chip manufacturing is so incredibly hard. The finance bozos are clueless, again.
They are somewhat rare, have some cultural overlap, and definitely get free upvotes.
But the truest measure of their potential to default, their credit default swaps, are pricing in a 20-22% chance to default before 2031.
- Oracle defaulting on its debt commitments.
- Anti trust action into Skydance forcing a breakup.
- Larry forced to cover the difference by selling family silver.
- Lanai going back into the hands of the people of Hawaii.
Just for the laughs and for forcing me to bear with some of the worst applications of all time.
> an Oracle facility designed, financed, and contracted around an older GPU
> generation can be overtaken before it is fully online or fully monetized.
Every user of Nvidia's gear has thought of this, and Nvidia provides a solution: Newer components are (usually) compatible with older racks, and Nvidia will buy back parts you replace with newer hardware.
For example, let's say you have an NV72 with Blackwell-ultra. Rubin comes out, and you decide you want to upgrade. There is a SKU of Rubin designed specifically to work in the racks shipped with Blackwell. Use it, and you only need to replace the compute board (not the power system, cooling, networking, etc.)
Nvidia will buy back the Blackwell boards you remove (and sell them to someone else, with a warrantee).
If you want to do an upgrade that increases power use, you may have to get a new rack. All of the parts are modular, and all parts can be sold back to Nvidia.
Nvidia does this because they have to. Without it, many customers would consider other vendors when upgrading, or delay purchases in case something better comes along later.
AI hardware is advancing rapidly. If NVIDIA's next generations deliver radically more tokens per watt and per dollar, an Oracle facility designed, financed, and contracted around an older GPU generation can be overtaken before it is fully online or fully monetized.
That's a big risk.
And to finance this they are making deep cuts to the ERP business where Oracle does have proprietary IP.
https://arstechnica.com/gadgets/2026/10/from-impulse-purchas...
https://github.com/aroum/fm1-custom-fw
see also:
> Reverse Engineering of the M-VAVE FM-1 Pocket Synthesizer Firmware
https://news.ycombinator.com/item?id=49994065 (75 points - 50 comments)
"Out-of-Warranty Repairs Available: In the unlikely event that something ever goes wrong with your hardware while using custom firmware, it is not covered under standard warranty. However, Polyend is always happy to help get your instrument back up and running through its official out-of-warranty repair service (standard service, parts, and shipping fees apply."Your iPhone has more sophisticated security measures inside of it along with an entire computer capable of playing relatively recent AAA games and it costs less than some of these keys.
The actual answer is that dealerships make their money on service and demand that the manufacturers throw them bones like this.
We had finally heard about the Microsoft store on campus, and had heard about the super discounted software. Turned out you had to be an employee. We were crushed because retail copies of Win2k were expensive for high schoolers. So we sat out in front of the store with signs "will work for cheap Microsoft software." Finally this girl came by and asked what we were doing, she was an intern or a new hire or something like that not much older than us, still in college. She went into to store and came out with boxed copies of Win2k Workstation for $20 a pop for us. We were stoked and tried to give he some cash, but she was just happy to help us out.
Those were some good times. I wonder if stuff like that still happens today.
Hey, I played with that when I was approximately 8 years old! Also: wireless printing via infrared, instead of those huge parallel printer cables.
WorkOS's production systems are secure. No customer platform data was accessed. This incident was isolated to internal GTM systems.
The attack involved ARTEX, an autonomous AI agent framework used to carry out offensive security operations. It's a striking example of how quickly the capabilities of attackers are evolving.
We've contained the incident, brought in independent experts to investigate, and are working through the details.
Security is something we take extremely seriously, and I believe in being transparent when things go wrong. We'll continue to share updates as we learn more.
Here’s the full post: https://workos.com/security/bulletins/october-30th-2026-data...
If you are a customer and would like to have a call with our security team, please send a note to security@workos.com
And what does “hardware and performance is there” mean? The latest Apple TV 4K has an A19 chip, same as the base iPhone 17. What console market are they eating alive with that exactly?
The AI data center is analogous to the mainframe.
If I can run Qwen 3.8 Flash Next on my DDR4 closet server with a $500 16GB GPU at 50 tok/s (which I can) then I just about guarantee you that in 2-3 years a pretty acceptable AI model is going to be running off of iPhones and M7/M8 Macs.
The other option that reduces Apple’s long-term risk is that they could just buy an AI company.
1. What you like doing
2. What you are good at doing
3. What people will pay you to do
For math, if your goal is to prove something that no one else has proven, then all three are under attack.I think the "egoless" path is taking a 50,000-foot view of mathematics that ignores the above lens. Sure, mathematics will survive and even prosper with AI, but what about the mathematicians?
In time, the profession of mathematics will evolve so a new task is at the center of the diagram--maybe it will be explaining math results to humans or teaching math to future mathematicians. I don't know. But it's easy to see that not everyone is going to be happy with the transition, and many will be left behind.
For software engineering, it depends on what your center is. If you like "writing code", then #3 is under attack. But for those whose goal is, "make something work", all three are still viable and even better than before (or at least faster).
That's why we see so much disagreement on HN about whether AI is good or bad for software engineering.
I'm sure lots of other people have made this observation before.
But it's pretty clear you can chuck "human pursuits" in the garbage, in favor of what really matters: intensification of the market system and wealth accumulation for its own sake.
Your technology leader thinks: Humanity? People? Who the fuck cares about that BS? Fuck 'em (or "I'll fuck 'em and make them like it"). I've got stuff to build and money to make.
> In the pursuit of knowledge, what ultimately drives people is what they themselves can understand, and the meaning these understanding provides, both based on their own experience and people whose opinions they value.
It's pretty clear the message we're getting is "what [people] themselves can understand, and the meaning these understanding provides" doesn't matter, and if you care about that you're an sentimental piece of garbage.
I think the first implying the second might be a bit of a stretch. At some point, companies are going to start looking at their AI spending and start negotiating deals. OpenAI is right behind them and Anthropic is around 8x more expensive (Sol 6.1 vs Opus 5.5 based on Artificial Intelligence's cost per task; cost per token is a somewhat useless metric when different models use vastly different numbers of tokens for the same thing). Chinese models like Deepseek 4.1 Flash are cheap, can be self-hosted, and while certainly behind Claude Opus 5.5, are still good.
Lots of companies buy "inferior" products like Microsoft Teams. Over the next 5 years, it seems likely that many models will get really good to the point that people often won't be able to tell any difference. Even if something isn't "definitively" better, companies might say "our engineers can't tell the difference and we're cutting our costs by 85%."
It's not absurd to think that a company would use its own harness or an open source harness and/or put a proxy in to route requests variably. For two months, they have engineers using different models unbeknownst to them and they measure things like how often the engineers have to correct or redirect the model. At the end of two months, they might find little difference in performance and switch to a much lower cost model.
It's not like it'd be crazy hard to analyze workers frustration as they did their jobs and figure out if one model was actually better. They could log all the conversations and it's not like AI can't easily determine things like "no, that's completely wrong" as a phrase showing that the agent was really off base.
I think a big part of the next 5-10 years is going to be about cost/efficiency.
I want AI and these companies to succeed. And I especially want new companies to succeed. But I want them to be valued far more realistically. Because if they have to be worth trillions, then there’s a real motivation to do what it takes to realize that valuation. And that’s how you birth a Zuckerberg.
I’m not rooting for them to fail. I’m rooting for the correction to happen sooner than later.
As to the author's core complaint (CAT5/6 degrading to 100mbps), I've found that the issue is almost always the termination at the jacks. Reseating the wires with a punchdown tool, or failing that, cutting a bit more off the ends and ensuring that the pair twists run all the way to the punchdowns has been enough to restore 1000mbps negotiation the few times I've had this issue.
Also, the items in the game are very artfully arranged.
This version (the part I played) just dumps everything on the floor and the music is lifeless.